Add Batch 1bdc3cf8-13fb-4089-a5dc-88458813af7d data
Browse filesThis view is limited to 50 files because it contains too many changes. See raw diff
- .gitattributes +16 -0
- 2024/Momentum Auxiliary Network for Supervised Local Learning/eb53acf2-10ec-44f2-923d-fc539ed31362_content_list.json +1733 -0
- 2024/Momentum Auxiliary Network for Supervised Local Learning/eb53acf2-10ec-44f2-923d-fc539ed31362_model.json +2280 -0
- 2024/Momentum Auxiliary Network for Supervised Local Learning/eb53acf2-10ec-44f2-923d-fc539ed31362_origin.pdf +3 -0
- 2024/Momentum Auxiliary Network for Supervised Local Learning/full.md +295 -0
- 2024/Momentum Auxiliary Network for Supervised Local Learning/images.zip +3 -0
- 2024/Momentum Auxiliary Network for Supervised Local Learning/layout.json +0 -0
- 2024/Mono-ViFI_ A Unified Learning Framework for Self-supervised Single- and Multi-frame Monocular Depth Estimation/47dcf543-78f7-47b3-955c-778904eede62_content_list.json +1893 -0
- 2024/Mono-ViFI_ A Unified Learning Framework for Self-supervised Single- and Multi-frame Monocular Depth Estimation/47dcf543-78f7-47b3-955c-778904eede62_model.json +0 -0
- 2024/Mono-ViFI_ A Unified Learning Framework for Self-supervised Single- and Multi-frame Monocular Depth Estimation/47dcf543-78f7-47b3-955c-778904eede62_origin.pdf +3 -0
- 2024/Mono-ViFI_ A Unified Learning Framework for Self-supervised Single- and Multi-frame Monocular Depth Estimation/full.md +369 -0
- 2024/Mono-ViFI_ A Unified Learning Framework for Self-supervised Single- and Multi-frame Monocular Depth Estimation/images.zip +3 -0
- 2024/Mono-ViFI_ A Unified Learning Framework for Self-supervised Single- and Multi-frame Monocular Depth Estimation/layout.json +0 -0
- 2024/MonoTTA_ Fully Test-Time Adaptation for Monocular 3D Object Detection/cfa6dde5-bde3-4186-b95b-8bbf6c77c339_content_list.json +1590 -0
- 2024/MonoTTA_ Fully Test-Time Adaptation for Monocular 3D Object Detection/cfa6dde5-bde3-4186-b95b-8bbf6c77c339_model.json +0 -0
- 2024/MonoTTA_ Fully Test-Time Adaptation for Monocular 3D Object Detection/cfa6dde5-bde3-4186-b95b-8bbf6c77c339_origin.pdf +3 -0
- 2024/MonoTTA_ Fully Test-Time Adaptation for Monocular 3D Object Detection/full.md +271 -0
- 2024/MonoTTA_ Fully Test-Time Adaptation for Monocular 3D Object Detection/images.zip +3 -0
- 2024/MonoTTA_ Fully Test-Time Adaptation for Monocular 3D Object Detection/layout.json +0 -0
- 2024/MonoWAD_ Weather-Adaptive Diffusion Model for Robust Monocular 3D Object Detection/4d73977a-6d19-47be-84e6-b3aec99a372e_content_list.json +1866 -0
- 2024/MonoWAD_ Weather-Adaptive Diffusion Model for Robust Monocular 3D Object Detection/4d73977a-6d19-47be-84e6-b3aec99a372e_model.json +0 -0
- 2024/MonoWAD_ Weather-Adaptive Diffusion Model for Robust Monocular 3D Object Detection/4d73977a-6d19-47be-84e6-b3aec99a372e_origin.pdf +3 -0
- 2024/MonoWAD_ Weather-Adaptive Diffusion Model for Robust Monocular 3D Object Detection/full.md +337 -0
- 2024/MonoWAD_ Weather-Adaptive Diffusion Model for Robust Monocular 3D Object Detection/images.zip +3 -0
- 2024/MonoWAD_ Weather-Adaptive Diffusion Model for Robust Monocular 3D Object Detection/layout.json +0 -0
- 2024/Monocular Occupancy Prediction for Scalable Indoor Scenes/325d434b-3564-4d20-a82e-4ec4e345eb58_content_list.json +1824 -0
- 2024/Monocular Occupancy Prediction for Scalable Indoor Scenes/325d434b-3564-4d20-a82e-4ec4e345eb58_model.json +2456 -0
- 2024/Monocular Occupancy Prediction for Scalable Indoor Scenes/325d434b-3564-4d20-a82e-4ec4e345eb58_origin.pdf +3 -0
- 2024/Monocular Occupancy Prediction for Scalable Indoor Scenes/full.md +323 -0
- 2024/Monocular Occupancy Prediction for Scalable Indoor Scenes/images.zip +3 -0
- 2024/Monocular Occupancy Prediction for Scalable Indoor Scenes/layout.json +0 -0
- 2024/Motion Aware Event Representation-driven Image Deblurring/a9105a02-b537-456a-9778-80b745c423ab_content_list.json +2161 -0
- 2024/Motion Aware Event Representation-driven Image Deblurring/a9105a02-b537-456a-9778-80b745c423ab_model.json +0 -0
- 2024/Motion Aware Event Representation-driven Image Deblurring/a9105a02-b537-456a-9778-80b745c423ab_origin.pdf +3 -0
- 2024/Motion Aware Event Representation-driven Image Deblurring/full.md +389 -0
- 2024/Motion Aware Event Representation-driven Image Deblurring/images.zip +3 -0
- 2024/Motion Aware Event Representation-driven Image Deblurring/layout.json +0 -0
- 2024/Motion Keyframe Interpolation for Any Human Skeleton using Point Cloud-based Human Motion Data Homogenisation/391ac41a-5023-4e3a-b209-fdaecfa62b6a_content_list.json +1685 -0
- 2024/Motion Keyframe Interpolation for Any Human Skeleton using Point Cloud-based Human Motion Data Homogenisation/391ac41a-5023-4e3a-b209-fdaecfa62b6a_model.json +0 -0
- 2024/Motion Keyframe Interpolation for Any Human Skeleton using Point Cloud-based Human Motion Data Homogenisation/391ac41a-5023-4e3a-b209-fdaecfa62b6a_origin.pdf +3 -0
- 2024/Motion Keyframe Interpolation for Any Human Skeleton using Point Cloud-based Human Motion Data Homogenisation/full.md +303 -0
- 2024/Motion Keyframe Interpolation for Any Human Skeleton using Point Cloud-based Human Motion Data Homogenisation/images.zip +3 -0
- 2024/Motion Keyframe Interpolation for Any Human Skeleton using Point Cloud-based Human Motion Data Homogenisation/layout.json +0 -0
- 2024/Motion Mamba_ Efficient and Long Sequence Motion Generation/1b586349-9221-4bad-a56d-e274ec79c852_content_list.json +1595 -0
- 2024/Motion Mamba_ Efficient and Long Sequence Motion Generation/1b586349-9221-4bad-a56d-e274ec79c852_model.json +0 -0
- 2024/Motion Mamba_ Efficient and Long Sequence Motion Generation/1b586349-9221-4bad-a56d-e274ec79c852_origin.pdf +3 -0
- 2024/Motion Mamba_ Efficient and Long Sequence Motion Generation/full.md +305 -0
- 2024/Motion Mamba_ Efficient and Long Sequence Motion Generation/images.zip +3 -0
- 2024/Motion Mamba_ Efficient and Long Sequence Motion Generation/layout.json +0 -0
- 2024/Motion and Structure from Event-based Normal Flow/2dddd958-d905-447a-8afe-98428ea7b185_content_list.json +1909 -0
.gitattributes
CHANGED
|
@@ -1428,3 +1428,19 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
|
|
| 1428 |
2024/Modeling[[:space:]]Label[[:space:]]Correlations[[:space:]]with[[:space:]]Latent[[:space:]]Context[[:space:]]for[[:space:]]Multi-Label[[:space:]]Recognition/db360f7c-5171-4a2a-8147-499c0fbf9014_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 1429 |
2024/Modeling[[:space:]]and[[:space:]]Driving[[:space:]]Human[[:space:]]Body[[:space:]]Soundfields[[:space:]]through[[:space:]]Acoustic[[:space:]]Primitives/3e3fdfc7-9a44-402d-ae28-13ceca875c26_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 1430 |
2024/Modelling[[:space:]]Competitive[[:space:]]Behaviors[[:space:]]in[[:space:]]Autonomous[[:space:]]Driving[[:space:]]Under[[:space:]]Generative[[:space:]]World[[:space:]]Model/ed2fa938-68fa-4d43-85a5-df3d24fff9a3_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1428 |
2024/Modeling[[:space:]]Label[[:space:]]Correlations[[:space:]]with[[:space:]]Latent[[:space:]]Context[[:space:]]for[[:space:]]Multi-Label[[:space:]]Recognition/db360f7c-5171-4a2a-8147-499c0fbf9014_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 1429 |
2024/Modeling[[:space:]]and[[:space:]]Driving[[:space:]]Human[[:space:]]Body[[:space:]]Soundfields[[:space:]]through[[:space:]]Acoustic[[:space:]]Primitives/3e3fdfc7-9a44-402d-ae28-13ceca875c26_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 1430 |
2024/Modelling[[:space:]]Competitive[[:space:]]Behaviors[[:space:]]in[[:space:]]Autonomous[[:space:]]Driving[[:space:]]Under[[:space:]]Generative[[:space:]]World[[:space:]]Model/ed2fa938-68fa-4d43-85a5-df3d24fff9a3_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 1431 |
+
2024/Momentum[[:space:]]Auxiliary[[:space:]]Network[[:space:]]for[[:space:]]Supervised[[:space:]]Local[[:space:]]Learning/eb53acf2-10ec-44f2-923d-fc539ed31362_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 1432 |
+
2024/Mono-ViFI_[[:space:]]A[[:space:]]Unified[[:space:]]Learning[[:space:]]Framework[[:space:]]for[[:space:]]Self-supervised[[:space:]]Single-[[:space:]]and[[:space:]]Multi-frame[[:space:]]Monocular[[:space:]]Depth[[:space:]]Estimation/47dcf543-78f7-47b3-955c-778904eede62_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 1433 |
+
2024/MonoTTA_[[:space:]]Fully[[:space:]]Test-Time[[:space:]]Adaptation[[:space:]]for[[:space:]]Monocular[[:space:]]3D[[:space:]]Object[[:space:]]Detection/cfa6dde5-bde3-4186-b95b-8bbf6c77c339_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 1434 |
+
2024/MonoWAD_[[:space:]]Weather-Adaptive[[:space:]]Diffusion[[:space:]]Model[[:space:]]for[[:space:]]Robust[[:space:]]Monocular[[:space:]]3D[[:space:]]Object[[:space:]]Detection/4d73977a-6d19-47be-84e6-b3aec99a372e_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 1435 |
+
2024/Monocular[[:space:]]Occupancy[[:space:]]Prediction[[:space:]]for[[:space:]]Scalable[[:space:]]Indoor[[:space:]]Scenes/325d434b-3564-4d20-a82e-4ec4e345eb58_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 1436 |
+
2024/Motion[[:space:]]Aware[[:space:]]Event[[:space:]]Representation-driven[[:space:]]Image[[:space:]]Deblurring/a9105a02-b537-456a-9778-80b745c423ab_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 1437 |
+
2024/Motion[[:space:]]Keyframe[[:space:]]Interpolation[[:space:]]for[[:space:]]Any[[:space:]]Human[[:space:]]Skeleton[[:space:]]using[[:space:]]Point[[:space:]]Cloud-based[[:space:]]Human[[:space:]]Motion[[:space:]]Data[[:space:]]Homogenisation/391ac41a-5023-4e3a-b209-fdaecfa62b6a_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 1438 |
+
2024/Motion[[:space:]]Mamba_[[:space:]]Efficient[[:space:]]and[[:space:]]Long[[:space:]]Sequence[[:space:]]Motion[[:space:]]Generation/1b586349-9221-4bad-a56d-e274ec79c852_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 1439 |
+
2024/Motion[[:space:]]and[[:space:]]Structure[[:space:]]from[[:space:]]Event-based[[:space:]]Normal[[:space:]]Flow/2dddd958-d905-447a-8afe-98428ea7b185_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 1440 |
+
2024/Motion-Guided[[:space:]]Latent[[:space:]]Diffusion[[:space:]]for[[:space:]]Temporally[[:space:]]Consistent[[:space:]]Real-world[[:space:]]Video[[:space:]]Super-resolution/dbea7f31-cd1b-433d-b163-70053b8d34db_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 1441 |
+
2024/Motion-Oriented[[:space:]]Compositional[[:space:]]Neural[[:space:]]Radiance[[:space:]]Fields[[:space:]]for[[:space:]]Monocular[[:space:]]Dynamic[[:space:]]Human[[:space:]]Modeling/6e1b3c80-b2f1-4690-8b23-807237bf60da_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 1442 |
+
2024/Motion-prior[[:space:]]Contrast[[:space:]]Maximization[[:space:]]for[[:space:]]Dense[[:space:]]Continuous-Time[[:space:]]Motion[[:space:]]Estimation/34d0d1c3-d92d-4348-bbb4-52308320fede_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 1443 |
+
2024/MotionChain_[[:space:]]Conversational[[:space:]]Motion[[:space:]]Controllers[[:space:]]via[[:space:]]Multimodal[[:space:]]Prompts/24d4fe30-81b0-462c-a698-38fed56b9fc4_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 1444 |
+
2024/MotionDirector_[[:space:]]Motion[[:space:]]Customization[[:space:]]of[[:space:]]Text-to-Video[[:space:]]Diffusion[[:space:]]Models/8325ffaa-f0c0-4f28-929f-63aa2f3d7c76_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 1445 |
+
2024/MotionLCM_[[:space:]]Real-time[[:space:]]Controllable[[:space:]]Motion[[:space:]]Generation[[:space:]]via[[:space:]]Latent[[:space:]]Consistency[[:space:]]Model/69320bc9-c9fe-4fc1-8e78-3d61fd4c6087_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 1446 |
+
2024/Multi-Granularity[[:space:]]Sparse[[:space:]]Relationship[[:space:]]Matrix[[:space:]]Prediction[[:space:]]Network[[:space:]]for[[:space:]]End-to-End[[:space:]]Scene[[:space:]]Graph[[:space:]]Generation/6b85535a-3ad0-4a54-882a-f4d4244fbf81_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
2024/Momentum Auxiliary Network for Supervised Local Learning/eb53acf2-10ec-44f2-923d-fc539ed31362_content_list.json
ADDED
|
@@ -0,0 +1,1733 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
{
|
| 3 |
+
"type": "text",
|
| 4 |
+
"text": "Momentum Auxiliary Network for Supervised Local Learning",
|
| 5 |
+
"text_level": 1,
|
| 6 |
+
"bbox": [
|
| 7 |
+
238,
|
| 8 |
+
140,
|
| 9 |
+
763,
|
| 10 |
+
186
|
| 11 |
+
],
|
| 12 |
+
"page_idx": 0
|
| 13 |
+
},
|
| 14 |
+
{
|
| 15 |
+
"type": "text",
|
| 16 |
+
"text": "Junhao Su $^{1*}$ , Changpeng Cai $^{1*}$ , Feiyu Zhu $^{2*}$ , Chenghao He $^{4}$",
|
| 17 |
+
"bbox": [
|
| 18 |
+
282,
|
| 19 |
+
215,
|
| 20 |
+
720,
|
| 21 |
+
229
|
| 22 |
+
],
|
| 23 |
+
"page_idx": 0
|
| 24 |
+
},
|
| 25 |
+
{
|
| 26 |
+
"type": "text",
|
| 27 |
+
"text": "Xiaojie $\\mathrm{Xu}^5$ , Dongzhi Guan $^{1\\boxtimes}$ , and Chenyang Si $^{3\\boxtimes}$",
|
| 28 |
+
"bbox": [
|
| 29 |
+
316,
|
| 30 |
+
229,
|
| 31 |
+
684,
|
| 32 |
+
247
|
| 33 |
+
],
|
| 34 |
+
"page_idx": 0
|
| 35 |
+
},
|
| 36 |
+
{
|
| 37 |
+
"type": "text",
|
| 38 |
+
"text": "$^{1}$ Southeast University",
|
| 39 |
+
"bbox": [
|
| 40 |
+
423,
|
| 41 |
+
257,
|
| 42 |
+
580,
|
| 43 |
+
272
|
| 44 |
+
],
|
| 45 |
+
"page_idx": 0
|
| 46 |
+
},
|
| 47 |
+
{
|
| 48 |
+
"type": "text",
|
| 49 |
+
"text": "2 University of Shanghai for Science and Technology",
|
| 50 |
+
"bbox": [
|
| 51 |
+
323,
|
| 52 |
+
272,
|
| 53 |
+
676,
|
| 54 |
+
287
|
| 55 |
+
],
|
| 56 |
+
"page_idx": 0
|
| 57 |
+
},
|
| 58 |
+
{
|
| 59 |
+
"type": "text",
|
| 60 |
+
"text": "<sup>3</sup> Nanyang Technological University",
|
| 61 |
+
"bbox": [
|
| 62 |
+
377,
|
| 63 |
+
286,
|
| 64 |
+
622,
|
| 65 |
+
301
|
| 66 |
+
],
|
| 67 |
+
"page_idx": 0
|
| 68 |
+
},
|
| 69 |
+
{
|
| 70 |
+
"type": "text",
|
| 71 |
+
"text": "$^{4}$ East China University of Science and Technology",
|
| 72 |
+
"bbox": [
|
| 73 |
+
328,
|
| 74 |
+
300,
|
| 75 |
+
671,
|
| 76 |
+
314
|
| 77 |
+
],
|
| 78 |
+
"page_idx": 0
|
| 79 |
+
},
|
| 80 |
+
{
|
| 81 |
+
"type": "text",
|
| 82 |
+
"text": "<sup>5</sup> The Hong Kong University of Science and Technology",
|
| 83 |
+
"bbox": [
|
| 84 |
+
313,
|
| 85 |
+
314,
|
| 86 |
+
689,
|
| 87 |
+
328
|
| 88 |
+
],
|
| 89 |
+
"page_idx": 0
|
| 90 |
+
},
|
| 91 |
+
{
|
| 92 |
+
"type": "text",
|
| 93 |
+
"text": "Abstract. Deep neural networks conventionally employ end-to-end backpropagation for their training process, which lacks biological credibility and triggers a locking dilemma during network parameter updates, leading to significant GPU memory use. Supervised local learning, which segments the network into multiple local blocks updated by independent auxiliary networks. However, these methods cannot replace end-to-end training due to lower accuracy, as gradients only propagate within their local block, creating a lack of information exchange between blocks. To address this issue and establish information transfer across blocks, we propose a Momentum Auxiliary Network (MAN) that establishes a dynamic interaction mechanism. The MAN leverages an exponential moving average (EMA) of the parameters from adjacent local blocks to enhance information flow. This auxiliary network, updated through EMA, helps bridge the informational gap between blocks. Nevertheless, we observe that directly applying EMA parameters has certain limitations due to feature discrepancies among local blocks. To overcome this, we introduce learnable biases, further boosting performance. We have validated our method on four image classification datasets (CIFAR-10, STL-10, SVHN, ImageNet), attaining superior performance and substantial memory savings. Notably, our method can reduce GPU memory usage by more than $45\\%$ on the ImageNet dataset compared to end-to-end training, while achieving higher performance. The Momentum Auxiliary Network thus offers a new perspective for supervised local learning. Our code is available at: https://github.com/JunhaoSuO/MAN.",
|
| 94 |
+
"bbox": [
|
| 95 |
+
261,
|
| 96 |
+
397,
|
| 97 |
+
743,
|
| 98 |
+
731
|
| 99 |
+
],
|
| 100 |
+
"page_idx": 0
|
| 101 |
+
},
|
| 102 |
+
{
|
| 103 |
+
"type": "text",
|
| 104 |
+
"text": "Keywords: Local Learning $\\cdot$ Image Classification $\\cdot$ Momentum Auxiliary Network",
|
| 105 |
+
"bbox": [
|
| 106 |
+
259,
|
| 107 |
+
746,
|
| 108 |
+
740,
|
| 109 |
+
776
|
| 110 |
+
],
|
| 111 |
+
"page_idx": 0
|
| 112 |
+
},
|
| 113 |
+
{
|
| 114 |
+
"type": "page_footnote",
|
| 115 |
+
"text": "*Equal Contribution.",
|
| 116 |
+
"bbox": [
|
| 117 |
+
230,
|
| 118 |
+
797,
|
| 119 |
+
375,
|
| 120 |
+
811
|
| 121 |
+
],
|
| 122 |
+
"page_idx": 0
|
| 123 |
+
},
|
| 124 |
+
{
|
| 125 |
+
"type": "page_footnote",
|
| 126 |
+
"text": "Corresponding Authors: Chengyang Si (chenyang.si.mail@gmail.com) and Dongzhi",
|
| 127 |
+
"bbox": [
|
| 128 |
+
230,
|
| 129 |
+
811,
|
| 130 |
+
785,
|
| 131 |
+
825
|
| 132 |
+
],
|
| 133 |
+
"page_idx": 0
|
| 134 |
+
},
|
| 135 |
+
{
|
| 136 |
+
"type": "page_footnote",
|
| 137 |
+
"text": "Guan (guandongzhi@seu.edu.cn).",
|
| 138 |
+
"bbox": [
|
| 139 |
+
230,
|
| 140 |
+
825,
|
| 141 |
+
455,
|
| 142 |
+
839
|
| 143 |
+
],
|
| 144 |
+
"page_idx": 0
|
| 145 |
+
},
|
| 146 |
+
{
|
| 147 |
+
"type": "image",
|
| 148 |
+
"img_path": "images/59658765abdb05b2d54ec752161d340c1c89205054a19306ed2110965081b8ad.jpg",
|
| 149 |
+
"image_caption": [
|
| 150 |
+
"Fig. 1: Comparison between different methods with MAN and the original methods in terms of accuracy. Results are obtained using ResNet-110 (K=55) on the various datasets."
|
| 151 |
+
],
|
| 152 |
+
"image_footnote": [],
|
| 153 |
+
"bbox": [
|
| 154 |
+
217,
|
| 155 |
+
145,
|
| 156 |
+
787,
|
| 157 |
+
328
|
| 158 |
+
],
|
| 159 |
+
"page_idx": 1
|
| 160 |
+
},
|
| 161 |
+
{
|
| 162 |
+
"type": "text",
|
| 163 |
+
"text": "1 Introduction",
|
| 164 |
+
"text_level": 1,
|
| 165 |
+
"bbox": [
|
| 166 |
+
217,
|
| 167 |
+
411,
|
| 168 |
+
375,
|
| 169 |
+
428
|
| 170 |
+
],
|
| 171 |
+
"page_idx": 1
|
| 172 |
+
},
|
| 173 |
+
{
|
| 174 |
+
"type": "text",
|
| 175 |
+
"text": "In deep learning, the prevalent use of end-to-end backpropagation [35] is essential for training complex neural networks, which requires heavy computational work for loss evaluation and successive gradient descent through network layers to refine parameters [17, 24, 29]. This method starkly contrasts with the local signal processing of biological synapses [6, 7, 40] and imposes a 'locking' constraint [19] that delays parameter updates until the entire forward and backward pass is complete. Such constraints exacerbate challenges like reduced parallelism and increased GPU memory usage [28, 38], compromising training efficiency and scalability. An emerging alternative to traditional end-to-end (E2E) training is Local Learning [2, 3, 5, 18, 19, 29, 32, 37, 41], which promises to mitigate the drawbacks of traditional methods. These methods divide the network into discrete, gradient-isolated blocks, each of which is updated by its dedicated auxiliary network according to distinct local objectives [3, 32, 37]. This training strategy enables immediate parameter updates upon local error reception, circumventing the sequential update bottleneck of E2E training, thereby significantly enhancing the efficiency of parallel training [13, 22]. Moreover, local learning reduces the demand on GPU memory by only retaining gradients for the local and auxiliary networks, thus eliminating the need to store extensive global gradient information and saving computational resources [28, 38]. However, even with these advantages, there is still a significant performance gap between local learning and traditional E2E methods, especially as the network is divided into more local blocks. Current local learning techniques mainly focus on improving the design of auxiliary network structures [3] and making local loss functions better to close this performance gap [32, 37]. Yet, these improvements do not completely address the inherent short-sightedness of local learning: the separation into blocks can make each part of the network only focus on its local objectives, possibly ig-",
|
| 176 |
+
"bbox": [
|
| 177 |
+
212,
|
| 178 |
+
446,
|
| 179 |
+
787,
|
| 180 |
+
843
|
| 181 |
+
],
|
| 182 |
+
"page_idx": 1
|
| 183 |
+
},
|
| 184 |
+
{
|
| 185 |
+
"type": "page_number",
|
| 186 |
+
"text": "2",
|
| 187 |
+
"bbox": [
|
| 188 |
+
217,
|
| 189 |
+
114,
|
| 190 |
+
228,
|
| 191 |
+
126
|
| 192 |
+
],
|
| 193 |
+
"page_idx": 1
|
| 194 |
+
},
|
| 195 |
+
{
|
| 196 |
+
"type": "header",
|
| 197 |
+
"text": "Junhao Su et al.",
|
| 198 |
+
"bbox": [
|
| 199 |
+
271,
|
| 200 |
+
114,
|
| 201 |
+
382,
|
| 202 |
+
126
|
| 203 |
+
],
|
| 204 |
+
"page_idx": 1
|
| 205 |
+
},
|
| 206 |
+
{
|
| 207 |
+
"type": "text",
|
| 208 |
+
"text": "noring the overall objectives of the network. This can lead to the discarding of globally beneficial information due to the lack of inter-block communication.",
|
| 209 |
+
"bbox": [
|
| 210 |
+
212,
|
| 211 |
+
146,
|
| 212 |
+
782,
|
| 213 |
+
175
|
| 214 |
+
],
|
| 215 |
+
"page_idx": 2
|
| 216 |
+
},
|
| 217 |
+
{
|
| 218 |
+
"type": "text",
|
| 219 |
+
"text": "In this paper, we introduce the Momentum Auxiliary Network (MAN), a novel network architecture designed to mitigate the inherent limitations of supervised local learning by enhancing inter-block communication. Specifically, MAN not only accepts the current local block as input but also absorbs the parameters of its next block. Upon completion of a forward pass, parameter updates are performed using local gradients refined through Exponential Moving Average (EMA) techniques [14]. This novel approach enables each local block to integrate information from the following block, thereby extending the operational perspective of each block beyond its immediate objectives and further aligning with the global objective of the network. This addresses the critical limitations in local learning frameworks, ensuring a more cohesive alignment with the overarching network objectives. Furthermore, we identify that directly applying EMA parameters is constrained by the feature incongruities across gradient-isolated blocks. To mitigate this, we implement additional learnable bias terms to each auxiliary block, enhancing their ability to share information effectively. The proposed MAN requires only a minimal increase in memory use but significantly improves performance by enhancing the information sharing ability of local blocks. The efficacy of the MAN approach is validated on a suite of benchmark image classification datasets, including CIFAR-10 [21], STL-10 [9], SVHN [30], and ImageNet [12]. The results demonstrate the effectiveness of our method in surpassing the limitations of traditional supervised local learning.",
|
| 220 |
+
"bbox": [
|
| 221 |
+
212,
|
| 222 |
+
176,
|
| 223 |
+
787,
|
| 224 |
+
493
|
| 225 |
+
],
|
| 226 |
+
"page_idx": 2
|
| 227 |
+
},
|
| 228 |
+
{
|
| 229 |
+
"type": "text",
|
| 230 |
+
"text": "The contributions of this paper could be summarized as follows:",
|
| 231 |
+
"bbox": [
|
| 232 |
+
238,
|
| 233 |
+
494,
|
| 234 |
+
700,
|
| 235 |
+
508
|
| 236 |
+
],
|
| 237 |
+
"page_idx": 2
|
| 238 |
+
},
|
| 239 |
+
{
|
| 240 |
+
"type": "list",
|
| 241 |
+
"sub_type": "text",
|
| 242 |
+
"list_items": [
|
| 243 |
+
"- We propose Momentum Auxiliary Network (MAN), designed to facilitate inter-block communication with Exponential Moving Average (EMA), mitigating the short-sightedness issue in conventional supervised local learning techniques and culminating in an elevated overall performance of the network.",
|
| 244 |
+
"- MAN is a versatile, plug-and-play approach that can seamlessly integrate with any supervised local learning technique, markedly amplifying their efficacy while requiring only a negligible increase in memory usage.",
|
| 245 |
+
"- MAN has demonstrated its effectiveness through experiments on benchmark image classification datasets, achieving state-of-the-art performance. Notably, on the ImageNet [12] dataset, it outperforms E2E training while using significantly less memory."
|
| 246 |
+
],
|
| 247 |
+
"bbox": [
|
| 248 |
+
225,
|
| 249 |
+
518,
|
| 250 |
+
782,
|
| 251 |
+
700
|
| 252 |
+
],
|
| 253 |
+
"page_idx": 2
|
| 254 |
+
},
|
| 255 |
+
{
|
| 256 |
+
"type": "text",
|
| 257 |
+
"text": "2 Related Work",
|
| 258 |
+
"text_level": 1,
|
| 259 |
+
"bbox": [
|
| 260 |
+
215,
|
| 261 |
+
736,
|
| 262 |
+
387,
|
| 263 |
+
752
|
| 264 |
+
],
|
| 265 |
+
"page_idx": 2
|
| 266 |
+
},
|
| 267 |
+
{
|
| 268 |
+
"type": "text",
|
| 269 |
+
"text": "2.1 Local Learning",
|
| 270 |
+
"text_level": 1,
|
| 271 |
+
"bbox": [
|
| 272 |
+
215,
|
| 273 |
+
768,
|
| 274 |
+
385,
|
| 275 |
+
784
|
| 276 |
+
],
|
| 277 |
+
"page_idx": 2
|
| 278 |
+
},
|
| 279 |
+
{
|
| 280 |
+
"type": "text",
|
| 281 |
+
"text": "Local learning is first proposed as an innovative deep learning algorithm with the intention of utilizing memory more efficiently and adhering more closely to principles of biological plausibility [10]. This approach emerges as a response to",
|
| 282 |
+
"bbox": [
|
| 283 |
+
212,
|
| 284 |
+
794,
|
| 285 |
+
784,
|
| 286 |
+
840
|
| 287 |
+
],
|
| 288 |
+
"page_idx": 2
|
| 289 |
+
},
|
| 290 |
+
{
|
| 291 |
+
"type": "header",
|
| 292 |
+
"text": "Momentum Auxiliary Network for Supervised Local Learning",
|
| 293 |
+
"bbox": [
|
| 294 |
+
321,
|
| 295 |
+
114,
|
| 296 |
+
732,
|
| 297 |
+
130
|
| 298 |
+
],
|
| 299 |
+
"page_idx": 2
|
| 300 |
+
},
|
| 301 |
+
{
|
| 302 |
+
"type": "page_number",
|
| 303 |
+
"text": "3",
|
| 304 |
+
"bbox": [
|
| 305 |
+
774,
|
| 306 |
+
116,
|
| 307 |
+
784,
|
| 308 |
+
126
|
| 309 |
+
],
|
| 310 |
+
"page_idx": 2
|
| 311 |
+
},
|
| 312 |
+
{
|
| 313 |
+
"type": "image",
|
| 314 |
+
"img_path": "images/21efcd67c2d31a85b36be6d1f25ab9a282f7e5aedfdad7e1deda6dcee912e8b2.jpg",
|
| 315 |
+
"image_caption": [
|
| 316 |
+
"Fig. 2: Comparison of (a) end-to-end backpropagation, (b) other supervised local learning methods, and (c) our proposed method. Unlike E2E, supervised local learning separates the network into K gradient-isolated local blocks."
|
| 317 |
+
],
|
| 318 |
+
"image_footnote": [],
|
| 319 |
+
"bbox": [
|
| 320 |
+
218,
|
| 321 |
+
142,
|
| 322 |
+
789,
|
| 323 |
+
292
|
| 324 |
+
],
|
| 325 |
+
"page_idx": 3
|
| 326 |
+
},
|
| 327 |
+
{
|
| 328 |
+
"type": "text",
|
| 329 |
+
"text": "the limitations presented by global E2E training [16], prompting the development of alternative supervised local learning methodologies. Examples of these methodologies include a differentiable search algorithm, which decouples network blocks for block-level learning and selects a manually assisted network for each local block [33], as well as a self-supervised contrasting loss function that leverages its local learning rules [18, 41]. In supervised local learning, training primarily relies on the design of reasonable supervised local loss functions or the manual construction of auxiliary networks.",
|
| 330 |
+
"bbox": [
|
| 331 |
+
212,
|
| 332 |
+
388,
|
| 333 |
+
787,
|
| 334 |
+
511
|
| 335 |
+
],
|
| 336 |
+
"page_idx": 3
|
| 337 |
+
},
|
| 338 |
+
{
|
| 339 |
+
"type": "text",
|
| 340 |
+
"text": "Supervised Local Loss: InfoPro [37] presents a method that involves the reconstruction loss for each hidden layer, coupled with cross-entropy loss. This approach forces the hidden layers to preserve vital input information, thereby preventing the early layers from discarding information relevant to the task. PredSim [32] employs layer-wise loss functions for network training. It makes use of two distinct supervised loss functions to generate local error signals for hidden layers. The optimization of these losses takes place during supervised local learning, thus eliminating the necessity to propagate global errors back to the hidden layers. These methods have achieved advanced performance in the field of supervised local learning, but there is still a significant gap compared to the performance of E2E training.",
|
| 341 |
+
"bbox": [
|
| 342 |
+
212,
|
| 343 |
+
516,
|
| 344 |
+
787,
|
| 345 |
+
683
|
| 346 |
+
],
|
| 347 |
+
"page_idx": 3
|
| 348 |
+
},
|
| 349 |
+
{
|
| 350 |
+
"type": "text",
|
| 351 |
+
"text": "Auxiliary Networks: InfoPro [37] incorporates two auxiliary networks into its approach. One network consists of a single convolutional layer followed by two fully-connected layers, which are employed for the cross-entropy loss. The other network utilizes two convolutional layers with upsampling for the reconstruction loss. Alternatively, DGL [3] designs a structure that uses three consecutive convolutional layers connected to a global pooling layer, followed by three consecutive fully connected layers. Moreover, DGL restricts the parameter of the auxiliary network to only $5\\%$ of the main network capacity, offering advantages in terms of speed and memory savings. However, these method still perform poorly when the network is divided into a large number of local blocks.",
|
| 352 |
+
"bbox": [
|
| 353 |
+
212,
|
| 354 |
+
689,
|
| 355 |
+
787,
|
| 356 |
+
840
|
| 357 |
+
],
|
| 358 |
+
"page_idx": 3
|
| 359 |
+
},
|
| 360 |
+
{
|
| 361 |
+
"type": "page_number",
|
| 362 |
+
"text": "4",
|
| 363 |
+
"bbox": [
|
| 364 |
+
217,
|
| 365 |
+
114,
|
| 366 |
+
228,
|
| 367 |
+
126
|
| 368 |
+
],
|
| 369 |
+
"page_idx": 3
|
| 370 |
+
},
|
| 371 |
+
{
|
| 372 |
+
"type": "header",
|
| 373 |
+
"text": "Junhao Su et al.",
|
| 374 |
+
"bbox": [
|
| 375 |
+
271,
|
| 376 |
+
114,
|
| 377 |
+
382,
|
| 378 |
+
126
|
| 379 |
+
],
|
| 380 |
+
"page_idx": 3
|
| 381 |
+
},
|
| 382 |
+
{
|
| 383 |
+
"type": "text",
|
| 384 |
+
"text": "2.2 Alternative Learning Rules to E2E Training",
|
| 385 |
+
"text_level": 1,
|
| 386 |
+
"bbox": [
|
| 387 |
+
214,
|
| 388 |
+
146,
|
| 389 |
+
625,
|
| 390 |
+
162
|
| 391 |
+
],
|
| 392 |
+
"page_idx": 4
|
| 393 |
+
},
|
| 394 |
+
{
|
| 395 |
+
"type": "text",
|
| 396 |
+
"text": "Due to certain limitations inherent in E2E training, the pursuit of alternative methods to E2E training has gradually garnered attention in recent years [27]. Several efforts have been dedicated to addressing the biologically implausible aspects of E2E training, such as the training methodology of target propagation [4, 23, 25] attempts to train a specialized backward connection by utilizing local reconstruction targets. Additionally, some recent studies have attempted to completely avoid backpropagation in neural networks through forward gradient learning [11, 34]. Meanwhile, the weight transport problem [10]. This has been tackled either by employing distinct feedback connections [1, 26] or by directly disseminating global errors to each hidden unit [8, 31]. While these methods have somewhat alleviated the biological implausibility of E2E training, they still rely on global objectives, which fundamentally differ from biological neural networks that rely on local synapses for information transmission. Furthermore, these methods currently struggle to be effective on large datasets [12].",
|
| 397 |
+
"bbox": [
|
| 398 |
+
212,
|
| 399 |
+
171,
|
| 400 |
+
787,
|
| 401 |
+
383
|
| 402 |
+
],
|
| 403 |
+
"page_idx": 4
|
| 404 |
+
},
|
| 405 |
+
{
|
| 406 |
+
"type": "text",
|
| 407 |
+
"text": "3 Method",
|
| 408 |
+
"text_level": 1,
|
| 409 |
+
"bbox": [
|
| 410 |
+
215,
|
| 411 |
+
406,
|
| 412 |
+
330,
|
| 413 |
+
422
|
| 414 |
+
],
|
| 415 |
+
"page_idx": 4
|
| 416 |
+
},
|
| 417 |
+
{
|
| 418 |
+
"type": "text",
|
| 419 |
+
"text": "3.1 Preliminaries",
|
| 420 |
+
"text_level": 1,
|
| 421 |
+
"bbox": [
|
| 422 |
+
214,
|
| 423 |
+
436,
|
| 424 |
+
372,
|
| 425 |
+
452
|
| 426 |
+
],
|
| 427 |
+
"page_idx": 4
|
| 428 |
+
},
|
| 429 |
+
{
|
| 430 |
+
"type": "text",
|
| 431 |
+
"text": "To set the stage, we first provide a brief overview of traditional end-to-end supervised learning and backpropagation mechanisms. We denote a data sample as $x$ and its corresponding ground-truth label as $y$ . The entire deep network is fragmented into several local blocks. During the forward propagation process, the output from the j-th block serves as the input for the (j+1)-th block, expressed as $x_{j+1} = f_{\\theta_j}(x_j)$ . Here, $\\theta_j$ symbolizes the parameters of the j-th local block and $f(\\cdot)$ represents the forward calculation of the block. We evaluate the loss function $\\mathcal{L}(\\hat{y}, y)$ between the output of the last block and the ground truth label, and propagate it back iteratively to the preceding blocks.",
|
| 432 |
+
"bbox": [
|
| 433 |
+
212,
|
| 434 |
+
462,
|
| 435 |
+
787,
|
| 436 |
+
598
|
| 437 |
+
],
|
| 438 |
+
"page_idx": 4
|
| 439 |
+
},
|
| 440 |
+
{
|
| 441 |
+
"type": "text",
|
| 442 |
+
"text": "Supervised local learning strategies [3, 32, 37] integrate auxiliary networks for local supervision. For each local block, an auxiliary network is affixed. The output from a local block is fed to its corresponding auxiliary network, generating the local supervisory signal as $\\hat{y}_j = g_{\\gamma_j}(x_{j + 1})$ . Here, $\\gamma_{j}$ denotes the parameters of the j-th auxiliary network.",
|
| 443 |
+
"bbox": [
|
| 444 |
+
212,
|
| 445 |
+
598,
|
| 446 |
+
787,
|
| 447 |
+
672
|
| 448 |
+
],
|
| 449 |
+
"page_idx": 4
|
| 450 |
+
},
|
| 451 |
+
{
|
| 452 |
+
"type": "text",
|
| 453 |
+
"text": "In this setup, we update the parameters of the j-th auxiliary network and local block, $\\gamma_{j},\\theta_{j}$ , as follows:",
|
| 454 |
+
"bbox": [
|
| 455 |
+
212,
|
| 456 |
+
674,
|
| 457 |
+
787,
|
| 458 |
+
705
|
| 459 |
+
],
|
| 460 |
+
"page_idx": 4
|
| 461 |
+
},
|
| 462 |
+
{
|
| 463 |
+
"type": "equation",
|
| 464 |
+
"text": "\n$$\n\\gamma_ {j} \\leftarrow \\gamma_ {j} - \\eta_ {a} \\times \\nabla_ {\\gamma_ {j}} \\mathcal {L} (\\hat {y} _ {j}, y) \\tag {1}\n$$\n",
|
| 465 |
+
"text_format": "latex",
|
| 466 |
+
"bbox": [
|
| 467 |
+
403,
|
| 468 |
+
720,
|
| 469 |
+
784,
|
| 470 |
+
736
|
| 471 |
+
],
|
| 472 |
+
"page_idx": 4
|
| 473 |
+
},
|
| 474 |
+
{
|
| 475 |
+
"type": "equation",
|
| 476 |
+
"text": "\n$$\n\\theta_ {j} \\leftarrow \\theta_ {j} - \\eta_ {l} \\times \\nabla_ {\\theta_ {j}} \\mathcal {L} \\left(y _ {\\hat {j}}, y\\right) \\tag {2}\n$$\n",
|
| 477 |
+
"text_format": "latex",
|
| 478 |
+
"bbox": [
|
| 479 |
+
406,
|
| 480 |
+
757,
|
| 481 |
+
784,
|
| 482 |
+
773
|
| 483 |
+
],
|
| 484 |
+
"page_idx": 4
|
| 485 |
+
},
|
| 486 |
+
{
|
| 487 |
+
"type": "text",
|
| 488 |
+
"text": "where $\\eta_{a},\\eta_{l}$ are the learning rates of the auxiliary networks and local blocks, respectively. By attaching auxiliary networks, each local block becomes gradient-isolated and can be updated with local supervision rather than global backpropagation.",
|
| 489 |
+
"bbox": [
|
| 490 |
+
212,
|
| 491 |
+
779,
|
| 492 |
+
787,
|
| 493 |
+
840
|
| 494 |
+
],
|
| 495 |
+
"page_idx": 4
|
| 496 |
+
},
|
| 497 |
+
{
|
| 498 |
+
"type": "header",
|
| 499 |
+
"text": "Momentum Auxiliary Network for Supervised Local Learning",
|
| 500 |
+
"bbox": [
|
| 501 |
+
321,
|
| 502 |
+
114,
|
| 503 |
+
732,
|
| 504 |
+
130
|
| 505 |
+
],
|
| 506 |
+
"page_idx": 4
|
| 507 |
+
},
|
| 508 |
+
{
|
| 509 |
+
"type": "page_number",
|
| 510 |
+
"text": "5",
|
| 511 |
+
"bbox": [
|
| 512 |
+
774,
|
| 513 |
+
116,
|
| 514 |
+
784,
|
| 515 |
+
126
|
| 516 |
+
],
|
| 517 |
+
"page_idx": 4
|
| 518 |
+
},
|
| 519 |
+
{
|
| 520 |
+
"type": "text",
|
| 521 |
+
"text": "3.2 Momentum Auxiliary Network",
|
| 522 |
+
"text_level": 1,
|
| 523 |
+
"bbox": [
|
| 524 |
+
215,
|
| 525 |
+
146,
|
| 526 |
+
517,
|
| 527 |
+
162
|
| 528 |
+
],
|
| 529 |
+
"page_idx": 5
|
| 530 |
+
},
|
| 531 |
+
{
|
| 532 |
+
"type": "text",
|
| 533 |
+
"text": "Existing techniques incorporate supervision signals into individual local blocks, which allows for parallel parameter updates and reduces memory overhead. However, this approach can lead to a short-sightedness problem, where each local block neglects the information from subsequent blocks, ultimately resulting in suboptimal final accuracy.",
|
| 534 |
+
"bbox": [
|
| 535 |
+
212,
|
| 536 |
+
172,
|
| 537 |
+
787,
|
| 538 |
+
247
|
| 539 |
+
],
|
| 540 |
+
"page_idx": 5
|
| 541 |
+
},
|
| 542 |
+
{
|
| 543 |
+
"type": "text",
|
| 544 |
+
"text": "To address this issue, we propose a comprehensive information exchange module—the Momentum Auxiliary Network (MAN). The MAN employs the Exponential Moving Average (EMA) mechanism [14] as a conduit for transferring information from subsequent blocks to the current block. In the context of MAN, we update the parameters of the j-th auxiliary network and local block as follows:",
|
| 545 |
+
"bbox": [
|
| 546 |
+
212,
|
| 547 |
+
248,
|
| 548 |
+
787,
|
| 549 |
+
324
|
| 550 |
+
],
|
| 551 |
+
"page_idx": 5
|
| 552 |
+
},
|
| 553 |
+
{
|
| 554 |
+
"type": "equation",
|
| 555 |
+
"text": "\n$$\n\\gamma_ {j} \\leftarrow \\gamma_ {j} - \\eta_ {a} \\times \\nabla_ {\\gamma_ {j}} \\mathcal {L} \\left(\\hat {y} _ {j}, y\\right) \\tag {3}\n$$\n",
|
| 556 |
+
"text_format": "latex",
|
| 557 |
+
"bbox": [
|
| 558 |
+
403,
|
| 559 |
+
338,
|
| 560 |
+
785,
|
| 561 |
+
356
|
| 562 |
+
],
|
| 563 |
+
"page_idx": 5
|
| 564 |
+
},
|
| 565 |
+
{
|
| 566 |
+
"type": "equation",
|
| 567 |
+
"text": "\n$$\n\\gamma_ {j} \\leftarrow E M A \\left(\\gamma_ {j}, \\theta_ {j + 1}\\right) \\tag {4}\n$$\n",
|
| 568 |
+
"text_format": "latex",
|
| 569 |
+
"bbox": [
|
| 570 |
+
424,
|
| 571 |
+
377,
|
| 572 |
+
785,
|
| 573 |
+
393
|
| 574 |
+
],
|
| 575 |
+
"page_idx": 5
|
| 576 |
+
},
|
| 577 |
+
{
|
| 578 |
+
"type": "equation",
|
| 579 |
+
"text": "\n$$\n\\theta_ {j} \\leftarrow \\theta_ {j} - \\eta_ {l} \\times \\nabla_ {\\theta_ {j}} \\mathcal {L} (\\hat {y} _ {j}, y) \\tag {5}\n$$\n",
|
| 580 |
+
"text_format": "latex",
|
| 581 |
+
"bbox": [
|
| 582 |
+
403,
|
| 583 |
+
415,
|
| 584 |
+
785,
|
| 585 |
+
431
|
| 586 |
+
],
|
| 587 |
+
"page_idx": 5
|
| 588 |
+
},
|
| 589 |
+
{
|
| 590 |
+
"type": "text",
|
| 591 |
+
"text": "where $\\gamma_{j}$ represents the parameters of the j-th auxiliary network and $\\theta_{j}$ represents those of the j-th local block. After updating with local gradients, $\\gamma_{j}$ undergoes further refinement by incorporating the parameters of the subsequent local block via the EMA, which is a weighted sum operation.",
|
| 592 |
+
"bbox": [
|
| 593 |
+
212,
|
| 594 |
+
438,
|
| 595 |
+
787,
|
| 596 |
+
500
|
| 597 |
+
],
|
| 598 |
+
"page_idx": 5
|
| 599 |
+
},
|
| 600 |
+
{
|
| 601 |
+
"type": "text",
|
| 602 |
+
"text": "However, through experimental results, it becomes apparent that directly applying EMA parameters to update the auxiliary network has its limitations, and it provides limited improvements to each local block. Upon analyzing, we find that the features learned between each local block vary to a certain degree, which hampers the effectiveness of the EMA update method. As a consequence, a learnable bias is introduced to augment the learning capability of the hidden layers within the local block, while also compensating for the deficiencies of the EMA update method.",
|
| 603 |
+
"bbox": [
|
| 604 |
+
212,
|
| 605 |
+
500,
|
| 606 |
+
486,
|
| 607 |
+
739
|
| 608 |
+
],
|
| 609 |
+
"page_idx": 5
|
| 610 |
+
},
|
| 611 |
+
{
|
| 612 |
+
"type": "image",
|
| 613 |
+
"img_path": "images/3b716f4cdbd8f877ca15468385def71a2b8632344813761894cbd88dfa245ebb.jpg",
|
| 614 |
+
"image_caption": [
|
| 615 |
+
"Fig. 3: Details of the Momentum Auxiliary Network. Local $(\\mathrm{i} + 1)$ represents the $(\\mathrm{i} + 1)$ -th gradient-isolated local block, which contains layers from layer m to layer $(\\mathrm{m} + \\mathrm{n})$ , totaling n+1 layers $(\\mathrm{n} \\geqslant 0)$ . We only use the parameters of the first layer to ensure a balance in GPU memory usage."
|
| 616 |
+
],
|
| 617 |
+
"image_footnote": [],
|
| 618 |
+
"bbox": [
|
| 619 |
+
498,
|
| 620 |
+
527,
|
| 621 |
+
784,
|
| 622 |
+
632
|
| 623 |
+
],
|
| 624 |
+
"page_idx": 5
|
| 625 |
+
},
|
| 626 |
+
{
|
| 627 |
+
"type": "text",
|
| 628 |
+
"text": "Therefore, the parameter update method for the j-th auxiliary network can be written as:",
|
| 629 |
+
"bbox": [
|
| 630 |
+
212,
|
| 631 |
+
739,
|
| 632 |
+
787,
|
| 633 |
+
770
|
| 634 |
+
],
|
| 635 |
+
"page_idx": 5
|
| 636 |
+
},
|
| 637 |
+
{
|
| 638 |
+
"type": "equation",
|
| 639 |
+
"text": "\n$$\n\\left(\\gamma_ {j}, b _ {j}\\right) \\leftarrow \\left(\\gamma_ {j}, b _ {j}\\right) - \\eta_ {a} \\times \\nabla_ {\\left(\\gamma_ {j}, b _ {j}\\right)} \\mathcal {L} \\left(\\hat {y} _ {j}, y\\right) \\tag {6}\n$$\n",
|
| 640 |
+
"text_format": "latex",
|
| 641 |
+
"bbox": [
|
| 642 |
+
354,
|
| 643 |
+
785,
|
| 644 |
+
785,
|
| 645 |
+
805
|
| 646 |
+
],
|
| 647 |
+
"page_idx": 5
|
| 648 |
+
},
|
| 649 |
+
{
|
| 650 |
+
"type": "equation",
|
| 651 |
+
"text": "\n$$\n\\gamma_ {j} \\leftarrow E M A \\left(\\gamma_ {j}, \\theta_ {j + 1}\\right) \\tag {7}\n$$\n",
|
| 652 |
+
"text_format": "latex",
|
| 653 |
+
"bbox": [
|
| 654 |
+
423,
|
| 655 |
+
824,
|
| 656 |
+
785,
|
| 657 |
+
842
|
| 658 |
+
],
|
| 659 |
+
"page_idx": 5
|
| 660 |
+
},
|
| 661 |
+
{
|
| 662 |
+
"type": "page_number",
|
| 663 |
+
"text": "6",
|
| 664 |
+
"bbox": [
|
| 665 |
+
217,
|
| 666 |
+
114,
|
| 667 |
+
228,
|
| 668 |
+
126
|
| 669 |
+
],
|
| 670 |
+
"page_idx": 5
|
| 671 |
+
},
|
| 672 |
+
{
|
| 673 |
+
"type": "header",
|
| 674 |
+
"text": "Junhao Su et al.",
|
| 675 |
+
"bbox": [
|
| 676 |
+
271,
|
| 677 |
+
114,
|
| 678 |
+
383,
|
| 679 |
+
127
|
| 680 |
+
],
|
| 681 |
+
"page_idx": 5
|
| 682 |
+
},
|
| 683 |
+
{
|
| 684 |
+
"type": "text",
|
| 685 |
+
"text": "where $(\\gamma_j, b_j)$ represent the parameters of the j-th auxiliary network. During this process, we update the parameters $\\gamma_j$ of the j-th auxiliary network jointly through the learnable bias $b_j$ and EMA.",
|
| 686 |
+
"bbox": [
|
| 687 |
+
212,
|
| 688 |
+
146,
|
| 689 |
+
782,
|
| 690 |
+
191
|
| 691 |
+
],
|
| 692 |
+
"page_idx": 6
|
| 693 |
+
},
|
| 694 |
+
{
|
| 695 |
+
"type": "text",
|
| 696 |
+
"text": "Essentially, the proposed MAN facilitates communication between subsequent blocks through the EMA mechanism [14], effectively resolving the short-sightedness problem inherent in traditional supervised local learning methods. Concurrently, it introduces an independent learnable bias to mitigate information discrepancies caused by feature variations among different local blocks, thereby aligning the output of each block closer to the global target.",
|
| 697 |
+
"bbox": [
|
| 698 |
+
212,
|
| 699 |
+
191,
|
| 700 |
+
782,
|
| 701 |
+
281
|
| 702 |
+
],
|
| 703 |
+
"page_idx": 6
|
| 704 |
+
},
|
| 705 |
+
{
|
| 706 |
+
"type": "text",
|
| 707 |
+
"text": "In addition to the two innovative methods mentioned above, MAN demonstrates strong versatility. It can be seamlessly integrated into existing supervised local learning methods, and it excels across various datasets, reflecting its flexibility and generalizability.",
|
| 708 |
+
"bbox": [
|
| 709 |
+
212,
|
| 710 |
+
282,
|
| 711 |
+
782,
|
| 712 |
+
343
|
| 713 |
+
],
|
| 714 |
+
"page_idx": 6
|
| 715 |
+
},
|
| 716 |
+
{
|
| 717 |
+
"type": "text",
|
| 718 |
+
"text": "4 Experiments",
|
| 719 |
+
"text_level": 1,
|
| 720 |
+
"bbox": [
|
| 721 |
+
214,
|
| 722 |
+
364,
|
| 723 |
+
375,
|
| 724 |
+
382
|
| 725 |
+
],
|
| 726 |
+
"page_idx": 6
|
| 727 |
+
},
|
| 728 |
+
{
|
| 729 |
+
"type": "text",
|
| 730 |
+
"text": "4.1 Experimental Setup",
|
| 731 |
+
"text_level": 1,
|
| 732 |
+
"bbox": [
|
| 733 |
+
214,
|
| 734 |
+
393,
|
| 735 |
+
428,
|
| 736 |
+
410
|
| 737 |
+
],
|
| 738 |
+
"page_idx": 6
|
| 739 |
+
},
|
| 740 |
+
{
|
| 741 |
+
"type": "text",
|
| 742 |
+
"text": "We conduct experiments using four widely adopted datasets: CIFAR-10 [21], SVHN [30], STL-10 [9], and ImageNet [12], with ResNets [15] of varying depths serving as the network architectures.",
|
| 743 |
+
"bbox": [
|
| 744 |
+
212,
|
| 745 |
+
417,
|
| 746 |
+
782,
|
| 747 |
+
462
|
| 748 |
+
],
|
| 749 |
+
"page_idx": 6
|
| 750 |
+
},
|
| 751 |
+
{
|
| 752 |
+
"type": "text",
|
| 753 |
+
"text": "Three state-of-the-art supervised local learning methods are selected for comparison: PredSim [32], DGL [3], and InfoPro [37]. We then partition the network into K local blocks, each containing an approximately equal number of layers. Our proposed Momentum Auxiliary Network is incorporated only into the first K-1 local blocks. The K-th local block does not employ an auxiliary network and is directly connected to the output classifier. We compare these configurations against traditional E2E and original supervised local learning methods to maintain consistent training settings and eliminate confounding variables.",
|
| 754 |
+
"bbox": [
|
| 755 |
+
212,
|
| 756 |
+
463,
|
| 757 |
+
784,
|
| 758 |
+
584
|
| 759 |
+
],
|
| 760 |
+
"page_idx": 6
|
| 761 |
+
},
|
| 762 |
+
{
|
| 763 |
+
"type": "text",
|
| 764 |
+
"text": "4.2 Implement Details",
|
| 765 |
+
"text_level": 1,
|
| 766 |
+
"bbox": [
|
| 767 |
+
214,
|
| 768 |
+
604,
|
| 769 |
+
415,
|
| 770 |
+
619
|
| 771 |
+
],
|
| 772 |
+
"page_idx": 6
|
| 773 |
+
},
|
| 774 |
+
{
|
| 775 |
+
"type": "text",
|
| 776 |
+
"text": "In our experiments on CIFAR-10 [21], SVHN [30], and STL-10 [9] datasets with ResNet-32 [15] and ResNet-110 [15], we utilize the SGD optimizer with Nesterov momentum set at 0.9 and an L2 weight decay factor of 1e-4. We employ batch sizes of 1024 for CIFAR-10 and SVHN and 128 for STL-10. The training duration spans 400 epochs, starting with initial learning rates of 0.8 for CIFAR-10 / SVHN and 0.1 for STL-10, following a cosine annealing scheduler [9].",
|
| 777 |
+
"bbox": [
|
| 778 |
+
212,
|
| 779 |
+
628,
|
| 780 |
+
782,
|
| 781 |
+
719
|
| 782 |
+
],
|
| 783 |
+
"page_idx": 6
|
| 784 |
+
},
|
| 785 |
+
{
|
| 786 |
+
"type": "text",
|
| 787 |
+
"text": "In our experiments with ImageNet [12], we adopt different training settings for various architectures. We train VGG13 [36] for 90 epochs with an initial learning rate of 0.025. For ResNet-101 [15], ResNet-152 [15], and ResNeXt-101,32 $\\times$ 8d [39], we train them for 90 epochs as well, with initial learning rates of 0.05, 0.05, and 0.025, respectively. We set the batch sizes to 64 for VGG13, 128 for ResNet-101 and ResNet-152, and 64 for ResNeXt-101,32 $\\times$ 8d. We maintain consistency with other training configurations as previously described for CIFAR-10 [21].",
|
| 788 |
+
"bbox": [
|
| 789 |
+
212,
|
| 790 |
+
719,
|
| 791 |
+
784,
|
| 792 |
+
839
|
| 793 |
+
],
|
| 794 |
+
"page_idx": 6
|
| 795 |
+
},
|
| 796 |
+
{
|
| 797 |
+
"type": "header",
|
| 798 |
+
"text": "Momentum Auxiliary Network for Supervised Local Learning",
|
| 799 |
+
"bbox": [
|
| 800 |
+
321,
|
| 801 |
+
114,
|
| 802 |
+
732,
|
| 803 |
+
130
|
| 804 |
+
],
|
| 805 |
+
"page_idx": 6
|
| 806 |
+
},
|
| 807 |
+
{
|
| 808 |
+
"type": "page_number",
|
| 809 |
+
"text": "7",
|
| 810 |
+
"bbox": [
|
| 811 |
+
774,
|
| 812 |
+
116,
|
| 813 |
+
784,
|
| 814 |
+
126
|
| 815 |
+
],
|
| 816 |
+
"page_idx": 6
|
| 817 |
+
},
|
| 818 |
+
{
|
| 819 |
+
"type": "text",
|
| 820 |
+
"text": "Based on our multiple experimental results, we select the hyperparameter momentum value as 0.995, due to its consistently stable and superior performance. We will provide the experimental results of other momentum hyperparameters in the supplementary materials.",
|
| 821 |
+
"bbox": [
|
| 822 |
+
212,
|
| 823 |
+
146,
|
| 824 |
+
787,
|
| 825 |
+
207
|
| 826 |
+
],
|
| 827 |
+
"page_idx": 7
|
| 828 |
+
},
|
| 829 |
+
{
|
| 830 |
+
"type": "text",
|
| 831 |
+
"text": "4.3 Results on Image Classification Datasets",
|
| 832 |
+
"text_level": 1,
|
| 833 |
+
"bbox": [
|
| 834 |
+
215,
|
| 835 |
+
229,
|
| 836 |
+
598,
|
| 837 |
+
244
|
| 838 |
+
],
|
| 839 |
+
"page_idx": 7
|
| 840 |
+
},
|
| 841 |
+
{
|
| 842 |
+
"type": "text",
|
| 843 |
+
"text": "Results on Image Classification Benchmarks: We start by assessing the accuracy performance of our approach using the CIFAR-10 [21], SVHN [30], and STL-10 [9] datasets. We employ ResNet-32 [15], partitioned into 8 and 16 local blocks, and ResNet-110 [15], divided into 32 and 55 local blocks. As illustrated in Table 1, our MAN significantly bolsters the accuracy of all methods.",
|
| 844 |
+
"bbox": [
|
| 845 |
+
212,
|
| 846 |
+
253,
|
| 847 |
+
787,
|
| 848 |
+
330
|
| 849 |
+
],
|
| 850 |
+
"page_idx": 7
|
| 851 |
+
},
|
| 852 |
+
{
|
| 853 |
+
"type": "table",
|
| 854 |
+
"img_path": "images/4f3d82f689a3e522b8bb8528ab4cf6588240a70062adc67c2439434322184a52.jpg",
|
| 855 |
+
"table_caption": [
|
| 856 |
+
"Table 1: Performance of different networks with varying numbers of local blocks. The average test error is obtained by 5 experiments. The * means addition of our MAN."
|
| 857 |
+
],
|
| 858 |
+
"table_footnote": [],
|
| 859 |
+
"table_body": "<table><tr><td rowspan=\"2\">Dataset</td><td rowspan=\"2\">Method</td><td colspan=\"2\">ResNet-32</td><td colspan=\"2\">ResNet-110</td></tr><tr><td>K = 8 (Test Error)</td><td>K = 16 (Test Error)</td><td>K = 32 (Test Error)</td><td>K = 55 (Test Error)</td></tr><tr><td rowspan=\"6\">CIFAR-10(E2E(ResNet-32)=6.37,E2E(ResNet-110)=5.42)</td><td>PredSim [32]</td><td>20.62</td><td>22.71</td><td>22.08</td><td>24.74</td></tr><tr><td>PredSim*</td><td>14.29(↓6.33)</td><td>15.58(↓7.13)</td><td>17.05(↓5.03)</td><td>18.02(↓6.72)</td></tr><tr><td>DGL [3]</td><td>11.63</td><td>14.08</td><td>12.51</td><td>14.45</td></tr><tr><td>DGL*</td><td>8.42(↓3.21)</td><td>9.11(↓4.97)</td><td>9.65(↓2.86)</td><td>9.73(↓4.72)</td></tr><tr><td>InfoPro [37]</td><td>11.51</td><td>12.93</td><td>12.26</td><td>13.22</td></tr><tr><td>InfoPro*</td><td>9.32(↓2.19)</td><td>9.65(↓3.28)</td><td>9.06(↓3.20)</td><td>9.77(↓3.45)</td></tr><tr><td rowspan=\"6\">STL-10(E2E(ResNet-32)=19.35,E2E(ResNet-110)=19.67)</td><td>PredSim [32]</td><td>31.97</td><td>32.90</td><td>32.05</td><td>33.27</td></tr><tr><td>PredSim*</td><td>29.97(↓2.00)</td><td>29.99(↓2.91)</td><td>30.48(↓1.57)</td><td>31.55(↓1.72)</td></tr><tr><td>DGL [3]</td><td>25.05</td><td>27.14</td><td>25.67</td><td>28.16</td></tr><tr><td>DGL*</td><td>20.74(↓4.31)</td><td>21.37(↓5.77)</td><td>22.54(↓3.13)</td><td>22.69(↓5.47)</td></tr><tr><td>InfoPro [37]</td><td>27.32</td><td>29.28</td><td>28.58</td><td>29.20</td></tr><tr><td>InfoPro*</td><td>23.17(↓4.15)</td><td>23.54(↓5.74)</td><td>24.08(↓4.50)</td><td>24.74(↓4.46)</td></tr><tr><td rowspan=\"6\">SVHN(E2E(ResNet-32)=2.99,E2E(ResNet-110)=2.92)</td><td>PredSim [32]</td><td>6.91</td><td>8.08</td><td>9.12</td><td>10.47</td></tr><tr><td>PredSim*</td><td>5.54(↓1.37)</td><td>6.39(↓1.69)</td><td>7.27(↓1.85)</td><td>8.24(↓2.23)</td></tr><tr><td>DGL [3]</td><td>4.83</td><td>5.05</td><td>5.12</td><td>5.36</td></tr><tr><td>DGL*</td><td>3.80(↓1.03)</td><td>4.04(↓1.01)</td><td>4.08(↓1.04)</td><td>4.52(↓0.84)</td></tr><tr><td>InfoPro [37]</td><td>5.61</td><td>5.97</td><td>5.89</td><td>6.11</td></tr><tr><td>InfoPro*</td><td>4.49(↓1.12)</td><td>5.19(↓0.78)</td><td>4.85(↓1.04)</td><td>4.99(↓1.12)</td></tr></table>",
|
| 860 |
+
"bbox": [
|
| 861 |
+
220,
|
| 862 |
+
398,
|
| 863 |
+
787,
|
| 864 |
+
597
|
| 865 |
+
],
|
| 866 |
+
"page_idx": 7
|
| 867 |
+
},
|
| 868 |
+
{
|
| 869 |
+
"type": "text",
|
| 870 |
+
"text": "On the CIFAR-10 dataset [21], our method exhibits considerable improvements in diminishing test errors across various methods. In the relatively shallower network of ResNet-32 (K=16), where individual layers function as gradient-isolated local blocks, we record a reduction in test errors for PredSim [32], DGL [3], and InfoPro [37], from 22.71, 14.08, and 12.93 to 15.58, 9.11, and 9.65 respectively. This translates to a performance enhancement exceeding $25\\%$ for all methods. Even though the performance across all methods in the comparatively deeper network, ResNet-110 (K=55), is somewhat inferior due to the inherent need for more global information in such networks, our method still delivers exceptional performance. It achieves approximately a $20\\%$ improvement across all methods, underscoring the robust effectiveness of MAN in deeper networks.",
|
| 871 |
+
"bbox": [
|
| 872 |
+
212,
|
| 873 |
+
628,
|
| 874 |
+
787,
|
| 875 |
+
794
|
| 876 |
+
],
|
| 877 |
+
"page_idx": 7
|
| 878 |
+
},
|
| 879 |
+
{
|
| 880 |
+
"type": "text",
|
| 881 |
+
"text": "When applied to other datasets, MAN can also reduce the test error of Pred-Sim [32], DGL [3], and InfoPro [37] by at least $5\\%$ , $12\\%$ , and $16\\%$ on the STL-10 [9] dataset. On the SVHN [30] dataset, our improvements over the three",
|
| 882 |
+
"bbox": [
|
| 883 |
+
212,
|
| 884 |
+
795,
|
| 885 |
+
787,
|
| 886 |
+
840
|
| 887 |
+
],
|
| 888 |
+
"page_idx": 7
|
| 889 |
+
},
|
| 890 |
+
{
|
| 891 |
+
"type": "page_number",
|
| 892 |
+
"text": "8",
|
| 893 |
+
"bbox": [
|
| 894 |
+
217,
|
| 895 |
+
114,
|
| 896 |
+
228,
|
| 897 |
+
126
|
| 898 |
+
],
|
| 899 |
+
"page_idx": 7
|
| 900 |
+
},
|
| 901 |
+
{
|
| 902 |
+
"type": "header",
|
| 903 |
+
"text": "Junhao Su et al.",
|
| 904 |
+
"bbox": [
|
| 905 |
+
271,
|
| 906 |
+
114,
|
| 907 |
+
382,
|
| 908 |
+
127
|
| 909 |
+
],
|
| 910 |
+
"page_idx": 7
|
| 911 |
+
},
|
| 912 |
+
{
|
| 913 |
+
"type": "text",
|
| 914 |
+
"text": "methods also surpass $20\\%$ , $21\\%$ , and $13\\%$ . As can be seen, the improvement our MAN introduces to all methods is quite remarkable—comparable even to the accuracy of E2E training—and it significantly mitigates the overwhelming performance issue that has continually plagued supervised local learning.",
|
| 915 |
+
"bbox": [
|
| 916 |
+
212,
|
| 917 |
+
146,
|
| 918 |
+
782,
|
| 919 |
+
207
|
| 920 |
+
],
|
| 921 |
+
"page_idx": 8
|
| 922 |
+
},
|
| 923 |
+
{
|
| 924 |
+
"type": "text",
|
| 925 |
+
"text": "Results on ImageNet: We further validate the effectiveness of our approach on ImageNet [12] using four networks of varying depths (ResNets [15] and VGG13 [36]). As depicted in Table 2, when we employ VGG13 as the backbone and divide the network into 10 blocks, DGL [3] achieves merely a Top1-Error of 35.60 and a Top5-Error of 14.2, representing a substantial gap when compared to the E2E method. However, with the introduction of our MAN, the Top1-Error reduces by 3.61 points, and the Top5-Error decreases by 3.36 points for DGL. This significant enhancement brings the performance closer to the E2E method.",
|
| 926 |
+
"bbox": [
|
| 927 |
+
212,
|
| 928 |
+
209,
|
| 929 |
+
784,
|
| 930 |
+
330
|
| 931 |
+
],
|
| 932 |
+
"page_idx": 8
|
| 933 |
+
},
|
| 934 |
+
{
|
| 935 |
+
"type": "text",
|
| 936 |
+
"text": "As illustrated in Table 2, when we use ResNet-101 [15], ResNet-152 [15], ResNeXt-101, $32 \\times 8\\mathrm{d}$ [39] as backbones and divide the network into four blocks, the performance of InfoPro [37] is already below that of E2E. After incorporating our MAN, the Top-1 Error of these three backbone networks can be reduced by approximately $6\\%$ compared to the original, surpassing the performance of E2E training. These results underscore the effectiveness of our MAN on the large-scale ImageNet [12] dataset, even when using deeper networks.",
|
| 937 |
+
"bbox": [
|
| 938 |
+
212,
|
| 939 |
+
332,
|
| 940 |
+
785,
|
| 941 |
+
439
|
| 942 |
+
],
|
| 943 |
+
"page_idx": 8
|
| 944 |
+
},
|
| 945 |
+
{
|
| 946 |
+
"type": "table",
|
| 947 |
+
"img_path": "images/580e507a3e07a6eb64b7d026a5fde080f06d3b55364e2ca96f02c81de23ec32c.jpg",
|
| 948 |
+
"table_caption": [
|
| 949 |
+
"Table 2: Results on the validation set of ImageNet"
|
| 950 |
+
],
|
| 951 |
+
"table_footnote": [],
|
| 952 |
+
"table_body": "<table><tr><td>Network</td><td>Method</td><td>Top1-Error</td><td>Top5-Error</td></tr><tr><td rowspan=\"5\">ResNet-101</td><td>E2E</td><td>22.03</td><td>5.93</td></tr><tr><td>InfoPro(K=2) [37]</td><td>21.85</td><td>5.89</td></tr><tr><td>InfoPro*(K=2)</td><td>21.65(↓0.20)</td><td>5.49(↓0.40)</td></tr><tr><td>InfoPro(K=4) [37]</td><td>22.81</td><td>6.54</td></tr><tr><td>InfoPro*(K=4)</td><td>21.73(↓1.08)</td><td>5.81(↓0.73)</td></tr><tr><td rowspan=\"5\">ResNet-152</td><td>E2E</td><td>21.60</td><td>5.92</td></tr><tr><td>InfoPro(K=2) [37]</td><td>21.45</td><td>5.84</td></tr><tr><td>InfoPro*(K=2)</td><td>21.23(↓0.22)</td><td>5.53(↓0.31)</td></tr><tr><td>InfoPro(K=4) [37]</td><td>22.93</td><td>6.71</td></tr><tr><td>InfoPro*(K=4)</td><td>21.59(↓1.34)</td><td>5.89(↓0.82)</td></tr><tr><td rowspan=\"5\">ResNeXt-101, 32 × 8d</td><td>E2E</td><td>20.64</td><td>5.40</td></tr><tr><td>InfoPro(K=2) [37]</td><td>20.35</td><td>5.28</td></tr><tr><td>InfoPro*(K=2)</td><td>20.11(↓0.24)</td><td>5.18(↓0.10)</td></tr><tr><td>InfoPro(K=4) [37]</td><td>21.69</td><td>6.11</td></tr><tr><td>InfoPro*(K=4)</td><td>20.37(↓1.32)</td><td>5.34(↓0.77)</td></tr><tr><td rowspan=\"3\">VGG13</td><td>E2E</td><td>28.41</td><td>9.63</td></tr><tr><td>DGL [3]</td><td>35.60</td><td>14.20</td></tr><tr><td>DGL*(K=10)</td><td>31.99(↓3.61)</td><td>10.84(↓3.36)</td></tr></table>",
|
| 953 |
+
"bbox": [
|
| 954 |
+
233,
|
| 955 |
+
496,
|
| 956 |
+
767,
|
| 957 |
+
696
|
| 958 |
+
],
|
| 959 |
+
"page_idx": 8
|
| 960 |
+
},
|
| 961 |
+
{
|
| 962 |
+
"type": "text",
|
| 963 |
+
"text": "Training-Accuracy Curve Analysis: As depicted in the accuracy-epoch curve in Fig. 4, our Momentum Auxiliary Network consistently outperforms the original method in terms of accuracy throughout the entire training process. This underscores its reliability and stability during the training process of classification tasks. Moreover, our MAN achieves a higher accuracy earlier in the later stages of training and attains stability sooner, indicating a faster convergence rate—a critical attribute in large-scale and complex tasks.",
|
| 964 |
+
"bbox": [
|
| 965 |
+
212,
|
| 966 |
+
734,
|
| 967 |
+
787,
|
| 968 |
+
840
|
| 969 |
+
],
|
| 970 |
+
"page_idx": 8
|
| 971 |
+
},
|
| 972 |
+
{
|
| 973 |
+
"type": "header",
|
| 974 |
+
"text": "Momentum Auxiliary Network for Supervised Local Learning",
|
| 975 |
+
"bbox": [
|
| 976 |
+
321,
|
| 977 |
+
114,
|
| 978 |
+
732,
|
| 979 |
+
128
|
| 980 |
+
],
|
| 981 |
+
"page_idx": 8
|
| 982 |
+
},
|
| 983 |
+
{
|
| 984 |
+
"type": "page_number",
|
| 985 |
+
"text": "9",
|
| 986 |
+
"bbox": [
|
| 987 |
+
774,
|
| 988 |
+
116,
|
| 989 |
+
784,
|
| 990 |
+
126
|
| 991 |
+
],
|
| 992 |
+
"page_idx": 8
|
| 993 |
+
},
|
| 994 |
+
{
|
| 995 |
+
"type": "image",
|
| 996 |
+
"img_path": "images/b07c01938291e0a448ebb3858e0be9c8c8bb17e68a298b6cdb7c67ccbacb0085.jpg",
|
| 997 |
+
"image_caption": [
|
| 998 |
+
"Fig. 4: Training-Accuracy curves, the left uses ResNet-32 (K=16) as the backbone, while the right uses ResNet-110 (K=55). Both are utilizing the CIFAR-10 dataset."
|
| 999 |
+
],
|
| 1000 |
+
"image_footnote": [],
|
| 1001 |
+
"bbox": [
|
| 1002 |
+
230,
|
| 1003 |
+
161,
|
| 1004 |
+
480,
|
| 1005 |
+
308
|
| 1006 |
+
],
|
| 1007 |
+
"page_idx": 9
|
| 1008 |
+
},
|
| 1009 |
+
{
|
| 1010 |
+
"type": "image",
|
| 1011 |
+
"img_path": "images/df3d2ab998231fd5d3874f9afe2e9b6e29d6c8f6cb317212026fe1b08f1dd61a.jpg",
|
| 1012 |
+
"image_caption": [],
|
| 1013 |
+
"image_footnote": [],
|
| 1014 |
+
"bbox": [
|
| 1015 |
+
511,
|
| 1016 |
+
161,
|
| 1017 |
+
759,
|
| 1018 |
+
306
|
| 1019 |
+
],
|
| 1020 |
+
"page_idx": 9
|
| 1021 |
+
},
|
| 1022 |
+
{
|
| 1023 |
+
"type": "text",
|
| 1024 |
+
"text": "In summary, our Momentum Auxiliary Network significantly enhances the accuracy of traditional supervised local learning methods by promoting information exchange between local blocks. The beneficial information exchange facilitated by Momentum Auxiliary Network significantly mitigates the pervasive issue of shortsightedness within the supervised local learning domain, thereby offering advantages in terms of training stability and convergence speed. This is crucial for training models more effectively and efficiently.",
|
| 1025 |
+
"bbox": [
|
| 1026 |
+
212,
|
| 1027 |
+
378,
|
| 1028 |
+
787,
|
| 1029 |
+
486
|
| 1030 |
+
],
|
| 1031 |
+
"page_idx": 9
|
| 1032 |
+
},
|
| 1033 |
+
{
|
| 1034 |
+
"type": "table",
|
| 1035 |
+
"img_path": "images/ec04e57e122a5123e26bcd488997ca24aaf363e36720f62d19a5682b3eb8f0ab.jpg",
|
| 1036 |
+
"table_caption": [
|
| 1037 |
+
"Table 3: Comparison of GPU memory usage using InfoPro as baseline with different backbones on the ImageNet dataset."
|
| 1038 |
+
],
|
| 1039 |
+
"table_footnote": [],
|
| 1040 |
+
"table_body": "<table><tr><td>Method</td><td>ResNet-101\nGPU Memory(GB)</td><td>ResNet-152\nGPU Memory(GB)</td><td>ResNeXt-101, 32×8d\nGPU Memory(GB)</td></tr><tr><td>E2E</td><td>19.71</td><td>26.29</td><td>19.22</td></tr><tr><td>InfoPro(K=2) [37]</td><td>12.06(↓ 38.8%)</td><td>15.53(↓ 40.9%)</td><td>11.55(↓ 39.9%)</td></tr><tr><td>InfoPro*(K=2)</td><td>12.32(↓ 37.5%)</td><td>15.93(↓ 39.4%)</td><td>11.73(↓ 38.9%)</td></tr><tr><td>InfoPro(K=4) [37]</td><td>10.37(↓ 47.3%)</td><td>13.48(↓ 48.7%)</td><td>10.24(↓ 46.7%)</td></tr><tr><td>InfoPro*(K=4)</td><td>10.69(↓ 45.8%)</td><td>13.91(↓ 47.1%)</td><td>10.49(↓ 45.5%)</td></tr></table>",
|
| 1041 |
+
"bbox": [
|
| 1042 |
+
236,
|
| 1043 |
+
551,
|
| 1044 |
+
761,
|
| 1045 |
+
672
|
| 1046 |
+
],
|
| 1047 |
+
"page_idx": 9
|
| 1048 |
+
},
|
| 1049 |
+
{
|
| 1050 |
+
"type": "text",
|
| 1051 |
+
"text": "Results on GPU memory requirement: Supervised local learning markedly conserves GPU memory by limiting gradient propagation to within local blocks, thereby ensuring that backpropagation transpires exclusively within each respective local block and its corresponding auxiliary network. This strategy significantly diminishes the storage demands for activation parameters and gradient information, which would typically disseminate across local blocks. Consequently, this facilitates substantial savings in GPU memory utilization. Our analysis, featuring a comprehensive comparison of GPU memory usage on the ImageNet dataset, illustrates that the implementation of a Momentum Auxiliary Network",
|
| 1052 |
+
"bbox": [
|
| 1053 |
+
212,
|
| 1054 |
+
703,
|
| 1055 |
+
787,
|
| 1056 |
+
840
|
| 1057 |
+
],
|
| 1058 |
+
"page_idx": 9
|
| 1059 |
+
},
|
| 1060 |
+
{
|
| 1061 |
+
"type": "page_number",
|
| 1062 |
+
"text": "10",
|
| 1063 |
+
"bbox": [
|
| 1064 |
+
217,
|
| 1065 |
+
114,
|
| 1066 |
+
235,
|
| 1067 |
+
126
|
| 1068 |
+
],
|
| 1069 |
+
"page_idx": 9
|
| 1070 |
+
},
|
| 1071 |
+
{
|
| 1072 |
+
"type": "header",
|
| 1073 |
+
"text": "Junhao Su et al.",
|
| 1074 |
+
"bbox": [
|
| 1075 |
+
271,
|
| 1076 |
+
114,
|
| 1077 |
+
382,
|
| 1078 |
+
126
|
| 1079 |
+
],
|
| 1080 |
+
"page_idx": 9
|
| 1081 |
+
},
|
| 1082 |
+
{
|
| 1083 |
+
"type": "table",
|
| 1084 |
+
"img_path": "images/e2eb95355a30cf53f58dc887b70013841e734ecb8adc3080590c32728f2664ff.jpg",
|
| 1085 |
+
"table_caption": [
|
| 1086 |
+
"Table 4: Abalation study of MAN. (a) Using DGL as baseline and ResNet-32 $(\\mathrm{K} = 16)$ as backbone on the CIFAR-10 dataset. (b) Using InfoPro as baseline and ResNet-101 $(\\mathrm{K} = 4)$ as backbone on the ImageNet dataset. LB stands for Learnable Bias."
|
| 1087 |
+
],
|
| 1088 |
+
"table_footnote": [],
|
| 1089 |
+
"table_body": "<table><tr><td>EMA</td><td>LB</td><td>Test Error</td></tr><tr><td>×</td><td>×</td><td>14.08</td></tr><tr><td>✓</td><td>×</td><td>11.07(↓3.01)</td></tr><tr><td>✓</td><td>✓</td><td>9.11(↓4.97)</td></tr><tr><td colspan=\"3\">(a)</td></tr></table>",
|
| 1090 |
+
"bbox": [
|
| 1091 |
+
217,
|
| 1092 |
+
196,
|
| 1093 |
+
390,
|
| 1094 |
+
272
|
| 1095 |
+
],
|
| 1096 |
+
"page_idx": 10
|
| 1097 |
+
},
|
| 1098 |
+
{
|
| 1099 |
+
"type": "table",
|
| 1100 |
+
"img_path": "images/cd15f42c2505970c794118f68718ff70719e972a1612ae1cbdaaf7ae424e2159.jpg",
|
| 1101 |
+
"table_caption": [],
|
| 1102 |
+
"table_footnote": [],
|
| 1103 |
+
"table_body": "<table><tr><td>EMA</td><td>LB</td><td>Top1-Error</td><td>Top5-Error</td></tr><tr><td>×</td><td>×</td><td>22.81</td><td>6.54</td></tr><tr><td>✓</td><td>×</td><td>22.09(↓ 0.72)</td><td>6.07(↓ 0.47)</td></tr><tr><td>✓</td><td>✓</td><td>21.73(↓ 1.08)</td><td>5.81(↓ 0.73)</td></tr><tr><td colspan=\"4\">(b)</td></tr></table>",
|
| 1104 |
+
"bbox": [
|
| 1105 |
+
500,
|
| 1106 |
+
198,
|
| 1107 |
+
764,
|
| 1108 |
+
272
|
| 1109 |
+
],
|
| 1110 |
+
"page_idx": 10
|
| 1111 |
+
},
|
| 1112 |
+
{
|
| 1113 |
+
"type": "text",
|
| 1114 |
+
"text": "can elicit notable performance enhancements while incurring only a minimal uptick in GPU memory consumption.",
|
| 1115 |
+
"bbox": [
|
| 1116 |
+
212,
|
| 1117 |
+
301,
|
| 1118 |
+
782,
|
| 1119 |
+
332
|
| 1120 |
+
],
|
| 1121 |
+
"page_idx": 10
|
| 1122 |
+
},
|
| 1123 |
+
{
|
| 1124 |
+
"type": "text",
|
| 1125 |
+
"text": "As shown in Table 3, it indicates the substantial GPU memory reduction on the ImageNet [12]. When applying our method to InfoPro [37] on ResNet-101 [15], ResNet-152 [15], and ResNeXt-101, $32 \\times 8d$ [39], dividing the backbone into two local blocks results in a GPU memory saving of $37.5\\%$ to $39.4\\%$ . When the network is divided into four local blocks, the degree of GPU memory savings is further enhanced, exceeding $45\\%$ . Combined with Table 2, it can be seen that we achieve better performance with almost half the GPU memory required for end-to-end training. Further analysis of the memory usage comparison with the original method shows that our approach only increases the GPU memory by about $1\\%$ over the original method, while achieving a performance improvement of over $5\\%$ . These results highlight the excellent balance that the MAN method achieves in terms of GPU memory usage and performance enhancement.",
|
| 1126 |
+
"bbox": [
|
| 1127 |
+
212,
|
| 1128 |
+
333,
|
| 1129 |
+
784,
|
| 1130 |
+
513
|
| 1131 |
+
],
|
| 1132 |
+
"page_idx": 10
|
| 1133 |
+
},
|
| 1134 |
+
{
|
| 1135 |
+
"type": "text",
|
| 1136 |
+
"text": "4.4 Ablation Studies",
|
| 1137 |
+
"text_level": 1,
|
| 1138 |
+
"bbox": [
|
| 1139 |
+
215,
|
| 1140 |
+
539,
|
| 1141 |
+
400,
|
| 1142 |
+
553
|
| 1143 |
+
],
|
| 1144 |
+
"page_idx": 10
|
| 1145 |
+
},
|
| 1146 |
+
{
|
| 1147 |
+
"type": "text",
|
| 1148 |
+
"text": "We conduct an ablation study on the CIFAR-10 [21] dataset and ImageNet dataset to assess the impact of the EMA method [14] and learnable bias in the MAN on performance. For this analysis, we use ResNet-32 (K=16) [15] as the backbone and the original DGL [3] method as a comparison baseline.",
|
| 1149 |
+
"bbox": [
|
| 1150 |
+
212,
|
| 1151 |
+
566,
|
| 1152 |
+
784,
|
| 1153 |
+
626
|
| 1154 |
+
],
|
| 1155 |
+
"page_idx": 10
|
| 1156 |
+
},
|
| 1157 |
+
{
|
| 1158 |
+
"type": "text",
|
| 1159 |
+
"text": "As shown in Table 4(a), when we only use the EMA [14] in MAN to promote information exchange with subsequent blocks, without adding a learnable bias, the test error decreases from 14.08 to 11.07. When we further add a learnable bias, the test error decreases from 11.07 to 9.11, it is evident that both the EMA method and learnable bias contribute to performance improvement. We hypothesize that the EMA and learnable bias are somewhat complementary in terms of the features they learn. To validate this, we conduct feature visualization.",
|
| 1160 |
+
"bbox": [
|
| 1161 |
+
212,
|
| 1162 |
+
628,
|
| 1163 |
+
784,
|
| 1164 |
+
733
|
| 1165 |
+
],
|
| 1166 |
+
"page_idx": 10
|
| 1167 |
+
},
|
| 1168 |
+
{
|
| 1169 |
+
"type": "text",
|
| 1170 |
+
"text": "As depicted in Fig. 5(a), we can observe that the original method without the addition of MAN learns very limited and chaotic features. After adding the EMA method alone, the features it learns are clearly more characteristic and defined, indicating that it indeed receives some global features from the information in the subsequent block. However, there are still many blurry features, which may be due to the EMA parameter update method [14], causing information imbalance due to feature differences between different local blocks. When we",
|
| 1171 |
+
"bbox": [
|
| 1172 |
+
212,
|
| 1173 |
+
734,
|
| 1174 |
+
785,
|
| 1175 |
+
839
|
| 1176 |
+
],
|
| 1177 |
+
"page_idx": 10
|
| 1178 |
+
},
|
| 1179 |
+
{
|
| 1180 |
+
"type": "header",
|
| 1181 |
+
"text": "Momentum Auxiliary Network for Supervised Local Learning",
|
| 1182 |
+
"bbox": [
|
| 1183 |
+
321,
|
| 1184 |
+
114,
|
| 1185 |
+
732,
|
| 1186 |
+
128
|
| 1187 |
+
],
|
| 1188 |
+
"page_idx": 10
|
| 1189 |
+
},
|
| 1190 |
+
{
|
| 1191 |
+
"type": "page_number",
|
| 1192 |
+
"text": "11",
|
| 1193 |
+
"bbox": [
|
| 1194 |
+
767,
|
| 1195 |
+
116,
|
| 1196 |
+
782,
|
| 1197 |
+
126
|
| 1198 |
+
],
|
| 1199 |
+
"page_idx": 10
|
| 1200 |
+
},
|
| 1201 |
+
{
|
| 1202 |
+
"type": "text",
|
| 1203 |
+
"text": "add learnable bias to the original method alone, the features it learns are more concentrated and clear, indicating that the learnable bias significantly enhances the learning ability of the current local block's hidden layer, but it still lacks some details because it does not receive more global information from the subsequent block. When we add the EMA method and learnable bias simultaneously, the learned features are not only clear and specific, but also more comprehensive in detail, proving that their learning abilities are complementary, and greatly improving the performance of the original method. The feature visualization in Fig. 5 verifies our previous thoughts.",
|
| 1204 |
+
"bbox": [
|
| 1205 |
+
212,
|
| 1206 |
+
146,
|
| 1207 |
+
787,
|
| 1208 |
+
282
|
| 1209 |
+
],
|
| 1210 |
+
"page_idx": 11
|
| 1211 |
+
},
|
| 1212 |
+
{
|
| 1213 |
+
"type": "image",
|
| 1214 |
+
"img_path": "images/6a5e04a5245e930abbb02927da9156858b05b05c4cd1fdc37fa7c71ca7236623.jpg",
|
| 1215 |
+
"image_caption": [
|
| 1216 |
+
"(a)"
|
| 1217 |
+
],
|
| 1218 |
+
"image_footnote": [],
|
| 1219 |
+
"bbox": [
|
| 1220 |
+
243,
|
| 1221 |
+
316,
|
| 1222 |
+
336,
|
| 1223 |
+
386
|
| 1224 |
+
],
|
| 1225 |
+
"page_idx": 11
|
| 1226 |
+
},
|
| 1227 |
+
{
|
| 1228 |
+
"type": "image",
|
| 1229 |
+
"img_path": "images/e86b2ebc2f6adfda5e94bb461ad8384510581d9974c641ed47a94b62988cc62c.jpg",
|
| 1230 |
+
"image_caption": [
|
| 1231 |
+
"(b)",
|
| 1232 |
+
"Fig.5: Feature maps comparison of (a) Original Method without MAN, (b) MAN with only EMA, (c) MAN with only learnable bias and (d) MAN with both EMA and learnable bias. The feature map are obtained using ResNet-32 $(\\mathrm{K} = 16)$ on CIFAR-10."
|
| 1233 |
+
],
|
| 1234 |
+
"image_footnote": [],
|
| 1235 |
+
"bbox": [
|
| 1236 |
+
383,
|
| 1237 |
+
316,
|
| 1238 |
+
475,
|
| 1239 |
+
386
|
| 1240 |
+
],
|
| 1241 |
+
"page_idx": 11
|
| 1242 |
+
},
|
| 1243 |
+
{
|
| 1244 |
+
"type": "image",
|
| 1245 |
+
"img_path": "images/cfd441b347f678abb1f0f8ef3af1fa200eff7a0e5621a1276fcbfe90fcc6b4e5.jpg",
|
| 1246 |
+
"image_caption": [
|
| 1247 |
+
"(c)"
|
| 1248 |
+
],
|
| 1249 |
+
"image_footnote": [],
|
| 1250 |
+
"bbox": [
|
| 1251 |
+
524,
|
| 1252 |
+
316,
|
| 1253 |
+
616,
|
| 1254 |
+
386
|
| 1255 |
+
],
|
| 1256 |
+
"page_idx": 11
|
| 1257 |
+
},
|
| 1258 |
+
{
|
| 1259 |
+
"type": "image",
|
| 1260 |
+
"img_path": "images/f99499f3083090cf91f95005d7233b93e4bcdb12d86301673c6056e6733178ae.jpg",
|
| 1261 |
+
"image_caption": [
|
| 1262 |
+
"(d)"
|
| 1263 |
+
],
|
| 1264 |
+
"image_footnote": [],
|
| 1265 |
+
"bbox": [
|
| 1266 |
+
666,
|
| 1267 |
+
316,
|
| 1268 |
+
756,
|
| 1269 |
+
386
|
| 1270 |
+
],
|
| 1271 |
+
"page_idx": 11
|
| 1272 |
+
},
|
| 1273 |
+
{
|
| 1274 |
+
"type": "text",
|
| 1275 |
+
"text": "4.5 The Effectiveness of EMA",
|
| 1276 |
+
"text_level": 1,
|
| 1277 |
+
"bbox": [
|
| 1278 |
+
214,
|
| 1279 |
+
492,
|
| 1280 |
+
477,
|
| 1281 |
+
507
|
| 1282 |
+
],
|
| 1283 |
+
"page_idx": 11
|
| 1284 |
+
},
|
| 1285 |
+
{
|
| 1286 |
+
"type": "text",
|
| 1287 |
+
"text": "To verify the effectiveness of the EMA update method in MAN, we conduct further ablation experiments. As shown in Table 5, using only the parameters of the next layer in MAN results in very limited improvement. In contrast, employing the EMA method to utilize the parameters of the next layer leads to more significant enhancements. This is because the subsequent local block and the current local block learn different features and have different learning objectives, leading to minimal improvement. However, using the EMA method can gradually promote information exchange between local blocks and reduce parameter fluctuations during training, leading to smoother parameter updates. This beneficial interaction brings significant improvements to MAN.",
|
| 1288 |
+
"bbox": [
|
| 1289 |
+
212,
|
| 1290 |
+
517,
|
| 1291 |
+
787,
|
| 1292 |
+
670
|
| 1293 |
+
],
|
| 1294 |
+
"page_idx": 11
|
| 1295 |
+
},
|
| 1296 |
+
{
|
| 1297 |
+
"type": "table",
|
| 1298 |
+
"img_path": "images/46b62afa0abb81d15ce5a5cd531fdbc3b17ede73bf9922d85d3ace06fdbb81dc.jpg",
|
| 1299 |
+
"table_caption": [
|
| 1300 |
+
"Table 5: Abalation study of EMA. (a) Using DGL as baseline and ResNet-32 $(\\mathrm{K} = 16)$ as backbone on the CIFAR-10 dataset. (b) Using InfoPro as baseline and ResNet-101 $(\\mathrm{K} = 4)$ as backbone on the ImageNet dataset. Parameter signifies whether each local block utilizes the parameters of the adjacent next layer connected to it."
|
| 1301 |
+
],
|
| 1302 |
+
"table_footnote": [],
|
| 1303 |
+
"table_body": "<table><tr><td>Parameter</td><td>EMA</td><td>Test</td><td>Error</td></tr><tr><td>×</td><td>×</td><td>14.08</td><td></td></tr><tr><td>✓</td><td>×</td><td>12.93</td><td></td></tr><tr><td>✓</td><td>✓</td><td>11.07</td><td></td></tr></table>",
|
| 1304 |
+
"bbox": [
|
| 1305 |
+
217,
|
| 1306 |
+
762,
|
| 1307 |
+
410,
|
| 1308 |
+
823
|
| 1309 |
+
],
|
| 1310 |
+
"page_idx": 11
|
| 1311 |
+
},
|
| 1312 |
+
{
|
| 1313 |
+
"type": "table",
|
| 1314 |
+
"img_path": "images/da282c873bf314ae3e071496196c78c221ed3675ba80279aa6d185ca4a1fad4d.jpg",
|
| 1315 |
+
"table_caption": [
|
| 1316 |
+
"(a)"
|
| 1317 |
+
],
|
| 1318 |
+
"table_footnote": [],
|
| 1319 |
+
"table_body": "<table><tr><td>Parameter</td><td>EMA</td><td>Top1-Error</td><td>Top5-Error</td></tr><tr><td>×</td><td>×</td><td>22.81</td><td>6.54</td></tr><tr><td>✓</td><td>×</td><td>22.64</td><td>6.41</td></tr><tr><td>✓</td><td>✓</td><td>22.09</td><td>6.07</td></tr></table>",
|
| 1320 |
+
"bbox": [
|
| 1321 |
+
500,
|
| 1322 |
+
762,
|
| 1323 |
+
772,
|
| 1324 |
+
823
|
| 1325 |
+
],
|
| 1326 |
+
"page_idx": 11
|
| 1327 |
+
},
|
| 1328 |
+
{
|
| 1329 |
+
"type": "text",
|
| 1330 |
+
"text": "(b)",
|
| 1331 |
+
"bbox": [
|
| 1332 |
+
625,
|
| 1333 |
+
823,
|
| 1334 |
+
645,
|
| 1335 |
+
835
|
| 1336 |
+
],
|
| 1337 |
+
"page_idx": 11
|
| 1338 |
+
},
|
| 1339 |
+
{
|
| 1340 |
+
"type": "page_number",
|
| 1341 |
+
"text": "12",
|
| 1342 |
+
"bbox": [
|
| 1343 |
+
217,
|
| 1344 |
+
114,
|
| 1345 |
+
235,
|
| 1346 |
+
126
|
| 1347 |
+
],
|
| 1348 |
+
"page_idx": 11
|
| 1349 |
+
},
|
| 1350 |
+
{
|
| 1351 |
+
"type": "header",
|
| 1352 |
+
"text": "Junhao Su et al.",
|
| 1353 |
+
"bbox": [
|
| 1354 |
+
271,
|
| 1355 |
+
114,
|
| 1356 |
+
380,
|
| 1357 |
+
126
|
| 1358 |
+
],
|
| 1359 |
+
"page_idx": 11
|
| 1360 |
+
},
|
| 1361 |
+
{
|
| 1362 |
+
"type": "text",
|
| 1363 |
+
"text": "4.6 Linear Separability Analysis",
|
| 1364 |
+
"text_level": 1,
|
| 1365 |
+
"bbox": [
|
| 1366 |
+
215,
|
| 1367 |
+
146,
|
| 1368 |
+
493,
|
| 1369 |
+
162
|
| 1370 |
+
],
|
| 1371 |
+
"page_idx": 12
|
| 1372 |
+
},
|
| 1373 |
+
{
|
| 1374 |
+
"type": "text",
|
| 1375 |
+
"text": "To demonstrate the effectiveness of our Momentum Auxiliary Network, we freeze the parameters of the main network and train a classifier for each local block to obtain the classification accuracy of each local block. As depicted in Fig. 6, we use DGL [3] as a baseline, and after integrating our method, the accuracy of the earlier layers decreases, while the accuracy of the middle and later layers significantly improve. The decrease in accuracy of the earlier layers suggests that they have learned more generalized features. Although these features are not beneficial for optimizing local objectives, they are beneficial from a global perspective. The middle and later layers, receiving these globally beneficial features, have seen a significant increase in accuracy. These results illustrate that our proposed method has facilitated information interaction between gradient-isolated local blocks. This addresses the shortsightedness issue existing in the current supervised local learning field and significantly enhances their performance.",
|
| 1376 |
+
"bbox": [
|
| 1377 |
+
212,
|
| 1378 |
+
171,
|
| 1379 |
+
787,
|
| 1380 |
+
368
|
| 1381 |
+
],
|
| 1382 |
+
"page_idx": 12
|
| 1383 |
+
},
|
| 1384 |
+
{
|
| 1385 |
+
"type": "image",
|
| 1386 |
+
"img_path": "images/6db996f1514a558097af01f9586d67f5704f0d5d7d835416605ab9f89d57a5e7.jpg",
|
| 1387 |
+
"image_caption": [
|
| 1388 |
+
"Fig. 6: Comparison of layer-wise linear separability across different learning rules on ResNet-32 and ResNet-110."
|
| 1389 |
+
],
|
| 1390 |
+
"image_footnote": [],
|
| 1391 |
+
"bbox": [
|
| 1392 |
+
246,
|
| 1393 |
+
395,
|
| 1394 |
+
498,
|
| 1395 |
+
560
|
| 1396 |
+
],
|
| 1397 |
+
"page_idx": 12
|
| 1398 |
+
},
|
| 1399 |
+
{
|
| 1400 |
+
"type": "image",
|
| 1401 |
+
"img_path": "images/759438b2911d841f2adcd073e32c9fb54746344980fa8fe40b9372997a11ed2c.jpg",
|
| 1402 |
+
"image_caption": [],
|
| 1403 |
+
"image_footnote": [],
|
| 1404 |
+
"bbox": [
|
| 1405 |
+
504,
|
| 1406 |
+
396,
|
| 1407 |
+
754,
|
| 1408 |
+
560
|
| 1409 |
+
],
|
| 1410 |
+
"page_idx": 12
|
| 1411 |
+
},
|
| 1412 |
+
{
|
| 1413 |
+
"type": "text",
|
| 1414 |
+
"text": "4.7 Representation Similarity Analysis",
|
| 1415 |
+
"text_level": 1,
|
| 1416 |
+
"bbox": [
|
| 1417 |
+
215,
|
| 1418 |
+
633,
|
| 1419 |
+
549,
|
| 1420 |
+
648
|
| 1421 |
+
],
|
| 1422 |
+
"page_idx": 12
|
| 1423 |
+
},
|
| 1424 |
+
{
|
| 1425 |
+
"type": "text",
|
| 1426 |
+
"text": "To further validate the efficacy of MAN, we conduct a Centered Kernel Alignment (CKA) experiment [20]. CKA serves as a metric to assess the similarity between feature representations. If a method's CKA score is closer to 1 in relation to the E2E training method, it indicates that the method's feature learning process is more aligned with that of E2E training. As depicted in Fig. 7, we use DGL [3] as a baseline and incorporate our method. It can be observed that whether adding the EMA method or learnable bias alone, the CKA scores significantly improve compared to the original method. When both the EMA and learnable bias are used in MAN, the CKA score further improves and performs more stably. Notably, the most significant increase in the CKA score occurs in the early and late layers. In conjunction with our previous analysis of the linear separability experiment, this is because if the early layers' learning method is closer",
|
| 1427 |
+
"bbox": [
|
| 1428 |
+
212,
|
| 1429 |
+
657,
|
| 1430 |
+
787,
|
| 1431 |
+
840
|
| 1432 |
+
],
|
| 1433 |
+
"page_idx": 12
|
| 1434 |
+
},
|
| 1435 |
+
{
|
| 1436 |
+
"type": "header",
|
| 1437 |
+
"text": "Momentum Auxiliary Network for Supervised Local Learning",
|
| 1438 |
+
"bbox": [
|
| 1439 |
+
321,
|
| 1440 |
+
114,
|
| 1441 |
+
730,
|
| 1442 |
+
128
|
| 1443 |
+
],
|
| 1444 |
+
"page_idx": 12
|
| 1445 |
+
},
|
| 1446 |
+
{
|
| 1447 |
+
"type": "page_number",
|
| 1448 |
+
"text": "13",
|
| 1449 |
+
"bbox": [
|
| 1450 |
+
767,
|
| 1451 |
+
116,
|
| 1452 |
+
784,
|
| 1453 |
+
126
|
| 1454 |
+
],
|
| 1455 |
+
"page_idx": 12
|
| 1456 |
+
},
|
| 1457 |
+
{
|
| 1458 |
+
"type": "text",
|
| 1459 |
+
"text": "to E2E training, they will focus more on learning general features to optimize the global objective, rather than focusing narrowly on local objectives. While this may result in poorer classification capabilities in the early layers, it greatly contributes to the overall performance improvement of the network. Through the analysis of images and experimental results, it can be demonstrated that MAN enables information interaction between gradient-isolated local blocks, solving the myopia problem present in current supervised local learning methods.",
|
| 1460 |
+
"bbox": [
|
| 1461 |
+
212,
|
| 1462 |
+
146,
|
| 1463 |
+
787,
|
| 1464 |
+
252
|
| 1465 |
+
],
|
| 1466 |
+
"page_idx": 13
|
| 1467 |
+
},
|
| 1468 |
+
{
|
| 1469 |
+
"type": "image",
|
| 1470 |
+
"img_path": "images/41e853fbfca688fd107444a9f7f8ce88d84afe36a94bfae8e47b9a49985d920f.jpg",
|
| 1471 |
+
"image_caption": [
|
| 1472 |
+
"Fig. 7: Assessment of Similarity in Layer-wise Representations. We use Centered Kernel Alignment (CKA) [20] to quantify the degree of similarity in the layer-wise representations between the E2E backpropagation (BP) and our proposed MAN."
|
| 1473 |
+
],
|
| 1474 |
+
"image_footnote": [],
|
| 1475 |
+
"bbox": [
|
| 1476 |
+
246,
|
| 1477 |
+
280,
|
| 1478 |
+
500,
|
| 1479 |
+
445
|
| 1480 |
+
],
|
| 1481 |
+
"page_idx": 13
|
| 1482 |
+
},
|
| 1483 |
+
{
|
| 1484 |
+
"type": "image",
|
| 1485 |
+
"img_path": "images/8773d9182a59cd09b3ff8c23529e403f422635454fb0f3e22848b615db1e5763.jpg",
|
| 1486 |
+
"image_caption": [],
|
| 1487 |
+
"image_footnote": [],
|
| 1488 |
+
"bbox": [
|
| 1489 |
+
501,
|
| 1490 |
+
280,
|
| 1491 |
+
756,
|
| 1492 |
+
445
|
| 1493 |
+
],
|
| 1494 |
+
"page_idx": 13
|
| 1495 |
+
},
|
| 1496 |
+
{
|
| 1497 |
+
"type": "text",
|
| 1498 |
+
"text": "5 Conclusion",
|
| 1499 |
+
"text_level": 1,
|
| 1500 |
+
"bbox": [
|
| 1501 |
+
215,
|
| 1502 |
+
534,
|
| 1503 |
+
359,
|
| 1504 |
+
549
|
| 1505 |
+
],
|
| 1506 |
+
"page_idx": 13
|
| 1507 |
+
},
|
| 1508 |
+
{
|
| 1509 |
+
"type": "text",
|
| 1510 |
+
"text": "This study addresses the performance disparities between traditional supervised local learning and end-to-end (E2E) methods in deep learning. We introduce a versatile Momentum Auxiliary Network to tackle the short-sighted problem in the early optimization work related to supervised local learning, facilitating information exchange between gradient-isolated local blocks. We integrate our Momentum Auxiliary Network into three advanced supervised local learning approaches and evaluate their performance across network architectures with varying depths on four widely adopted datasets. The results demonstrate our method's ability to significantly enhance the ultimate output performance of original supervised local learning methods. Particularly when combined with InfoPro [37], our method significantly reduces GPU memory usage while consistently maintaining performance levels closely aligned with E2E approaches.",
|
| 1511 |
+
"bbox": [
|
| 1512 |
+
212,
|
| 1513 |
+
566,
|
| 1514 |
+
787,
|
| 1515 |
+
748
|
| 1516 |
+
],
|
| 1517 |
+
"page_idx": 13
|
| 1518 |
+
},
|
| 1519 |
+
{
|
| 1520 |
+
"type": "text",
|
| 1521 |
+
"text": "Limitations and future works: Despite the superior performance on large-scale problems like ImageNet [12], our method still performs less accurately than E2E on some conventional image classification datasets. This may be due to the MAN using too few information interaction layers when the network is divided into a larger number of local blocks. In future work, we could explore deepening these information interaction layers to achieve better precision performance.",
|
| 1522 |
+
"bbox": [
|
| 1523 |
+
212,
|
| 1524 |
+
750,
|
| 1525 |
+
787,
|
| 1526 |
+
839
|
| 1527 |
+
],
|
| 1528 |
+
"page_idx": 13
|
| 1529 |
+
},
|
| 1530 |
+
{
|
| 1531 |
+
"type": "page_number",
|
| 1532 |
+
"text": "14",
|
| 1533 |
+
"bbox": [
|
| 1534 |
+
217,
|
| 1535 |
+
114,
|
| 1536 |
+
235,
|
| 1537 |
+
126
|
| 1538 |
+
],
|
| 1539 |
+
"page_idx": 13
|
| 1540 |
+
},
|
| 1541 |
+
{
|
| 1542 |
+
"type": "header",
|
| 1543 |
+
"text": "Junhao Su et al.",
|
| 1544 |
+
"bbox": [
|
| 1545 |
+
271,
|
| 1546 |
+
114,
|
| 1547 |
+
380,
|
| 1548 |
+
126
|
| 1549 |
+
],
|
| 1550 |
+
"page_idx": 13
|
| 1551 |
+
},
|
| 1552 |
+
{
|
| 1553 |
+
"type": "text",
|
| 1554 |
+
"text": "Acknowledgement",
|
| 1555 |
+
"text_level": 1,
|
| 1556 |
+
"bbox": [
|
| 1557 |
+
217,
|
| 1558 |
+
143,
|
| 1559 |
+
392,
|
| 1560 |
+
162
|
| 1561 |
+
],
|
| 1562 |
+
"page_idx": 14
|
| 1563 |
+
},
|
| 1564 |
+
{
|
| 1565 |
+
"type": "text",
|
| 1566 |
+
"text": "This work was supported by the National Natural Science Foundation of China (52278154), the Natural Science Foundation of Jiangsu (BK20231429), the Fundamental Research Funds for the Central Universities (2242024RCB0008), and assist from the program of Zhishan Young Scholar of Southeast University.",
|
| 1567 |
+
"bbox": [
|
| 1568 |
+
215,
|
| 1569 |
+
175,
|
| 1570 |
+
784,
|
| 1571 |
+
234
|
| 1572 |
+
],
|
| 1573 |
+
"page_idx": 14
|
| 1574 |
+
},
|
| 1575 |
+
{
|
| 1576 |
+
"type": "text",
|
| 1577 |
+
"text": "References",
|
| 1578 |
+
"text_level": 1,
|
| 1579 |
+
"bbox": [
|
| 1580 |
+
217,
|
| 1581 |
+
258,
|
| 1582 |
+
321,
|
| 1583 |
+
273
|
| 1584 |
+
],
|
| 1585 |
+
"page_idx": 14
|
| 1586 |
+
},
|
| 1587 |
+
{
|
| 1588 |
+
"type": "list",
|
| 1589 |
+
"sub_type": "ref_text",
|
| 1590 |
+
"list_items": [
|
| 1591 |
+
"1. Akrout, M., Wilson, C., Humphreys, P., Lillicrap, T., Tweed, D.B.: Deep learning without weight transport. Advances in neural information processing systems 32 (2019)",
|
| 1592 |
+
"2. Belilovsky, E., Eickenberg, M., Oyallon, E.: Greedy layerwise learning can scale to imagenet. In: International conference on machine learning. pp. 583-593. PMLR (2019)",
|
| 1593 |
+
"3. Belilovsky, E., Eickenberg, M., Oyallon, E.: Decoupled greedy learning of cnns. In: International Conference on Machine Learning. pp. 736-745. PMLR (2020)",
|
| 1594 |
+
"4. Bengio, Y.: How auto-encoders could provide credit assignment in deep networks via target propagation. arXiv preprint arXiv:1407.7906 (2014)",
|
| 1595 |
+
"5. Bengio, Y., Lamblin, P., Popovici, D., Larochelle, H.: Greedy layer-wise training of deep networks. Advances in neural information processing systems 19 (2006)",
|
| 1596 |
+
"6. Bengio, Y., Lee, D.H., Bornschein, J., Mesnard, T., Lin, Z.: Towards biologically plausible deep learning. arXiv preprint arXiv:1502.04156 (2015)",
|
| 1597 |
+
"7. Caporale, N., Dan, Y.: Spike timing-dependent plasticity: a hebbian learning rule. Annu. Rev. Neurosci. 31, 25-46 (2008)",
|
| 1598 |
+
"8. Clark, D., Abbott, L., Chung, S.: Credit assignment through broadcasting a global error vector. Advances in Neural Information Processing Systems 34, 10053-10066 (2021)",
|
| 1599 |
+
"9. Coates, A., Ng, A., Lee, H.: An analysis of single-layer networks in unsupervised feature learning. In: Proceedings of the fourteenth international conference on artificial intelligence and statistics. pp. 215-223. JMLR Workshop and Conference Proceedings (2011)",
|
| 1600 |
+
"0. Crick, F.: The recent excitement about neural networks. Nature 337(6203), 129-132 (1989)",
|
| 1601 |
+
"1. Dellafererra, G., Kreiman, G.: Error-driven input modulation: solving the credit assignment problem without a backward pass. In: International Conference on Machine Learning. pp. 4937-4955. PMLR (2022)",
|
| 1602 |
+
"2. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: 2009 IEEE conference on computer vision and pattern recognition. pp. 248-255. IEEE (2009)",
|
| 1603 |
+
"3. Gomez, A.N., Key, O., Perlin, K., Gou, S., Frosst, N., Dean, J., Gal, Y.: Interlocking backpropagation: Improving depthwise model-parallelism. The Journal of Machine Learning Research 23(1), 7714-7741 (2022)",
|
| 1604 |
+
"4. He, K., Fan, H., Wu, Y., Xie, S., Girshick, R.: Momentum contrast for unsupervised visual representation learning. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 9729-9738 (2020)",
|
| 1605 |
+
"5. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016)"
|
| 1606 |
+
],
|
| 1607 |
+
"bbox": [
|
| 1608 |
+
225,
|
| 1609 |
+
287,
|
| 1610 |
+
785,
|
| 1611 |
+
839
|
| 1612 |
+
],
|
| 1613 |
+
"page_idx": 14
|
| 1614 |
+
},
|
| 1615 |
+
{
|
| 1616 |
+
"type": "header",
|
| 1617 |
+
"text": "Momentum Auxiliary Network for Supervised Local Learning",
|
| 1618 |
+
"bbox": [
|
| 1619 |
+
321,
|
| 1620 |
+
114,
|
| 1621 |
+
730,
|
| 1622 |
+
128
|
| 1623 |
+
],
|
| 1624 |
+
"page_idx": 14
|
| 1625 |
+
},
|
| 1626 |
+
{
|
| 1627 |
+
"type": "page_number",
|
| 1628 |
+
"text": "15",
|
| 1629 |
+
"bbox": [
|
| 1630 |
+
767,
|
| 1631 |
+
116,
|
| 1632 |
+
784,
|
| 1633 |
+
126
|
| 1634 |
+
],
|
| 1635 |
+
"page_idx": 14
|
| 1636 |
+
},
|
| 1637 |
+
{
|
| 1638 |
+
"type": "list",
|
| 1639 |
+
"sub_type": "ref_text",
|
| 1640 |
+
"list_items": [
|
| 1641 |
+
"16. Hinton, G.E., Osindero, S., Teh, Y.W.: A fast learning algorithm for deep belief nets. Neural computation 18(7), 1527-1554 (2006)",
|
| 1642 |
+
"17. Huang, G., Liu, Z., Van Der Maaten, L., Weinberger, K.Q.: Densely connected convolutional networks. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4700-4708 (2017)",
|
| 1643 |
+
"18. Illing, B., Ventura, J., Bellec, G., Gerstner, W.: Local plasticity rules can learn deep representations using self-supervised contrastive predictions. Advances in Neural Information Processing Systems 34, 30365-30379 (2021)",
|
| 1644 |
+
"19. Jaderberg, M., Czarnecki, W.M., Osindero, S., Vinyals, O., Graves, A., Silver, D., Kavukcuoglu, K.: Decoupled neural interfaces using synthetic gradients. In: International conference on machine learning. pp. 1627-1635. PMLR (2017)",
|
| 1645 |
+
"20. Kornblith, S., Norouzi, M., Lee, H., Hinton, G.: Similarity of neural network representations revisited. In: International conference on machine learning. pp. 3519-3529. PMLR (2019)",
|
| 1646 |
+
"21. Krizhevsky, A., Hinton, G., et al.: Learning multiple layers of features from tiny images (2009)",
|
| 1647 |
+
"22. Laskin, M., Metz, L., Nabarro, S., Saroufim, M., Noune, B., Luschi, C., Sohl-Dickstein, J., Abbeel, P.: Parallel training of deep networks with local updates. arXiv preprint arXiv:2012.03837 (2020)",
|
| 1648 |
+
"23. Le Cun, Y.: Learning process in an asymmetric threshold network. In: Disordered systems and biological organization, pp. 233-240. Springer (1986)",
|
| 1649 |
+
"24. LeCun, Y., Bengio, Y., Hinton, G.: Deep learning. nature 521(7553), 436-444 (2015)",
|
| 1650 |
+
"25. Lee, D.H., Zhang, S., Fischer, A., Bengio, Y.: Difference target propagation. In: Machine Learning and Knowledge Discovery in Databases: European Conference, ECML PKDD 2015, Porto, Portugal, September 7-11, 2015, Proceedings, Part I 15. pp. 498-515. Springer (2015)",
|
| 1651 |
+
"26. Lillicrap, T.P., Cownden, D., Tweed, D.B., Akerman, C.J.: Random synaptic feedback weights support error backpropagation for deep learning. Nature communications 7(1), 13276 (2016)",
|
| 1652 |
+
"27. Lillicrap, T.P., Santoro, A., Marris, L., Akerman, C.J., Hinton, G.: Backpropagation and the brain. Nature Reviews Neuroscience 21(6), 335-346 (2020)",
|
| 1653 |
+
"28. Löwe, S., O'Connor, P., Veeling, B.: Putting an end to end-to-end: Gradient-isolated learning of representations. Advances in neural information processing systems 32 (2019)",
|
| 1654 |
+
"29. Mostafa, H., Ramesh, V., Cauwenberghs, G.: Deep supervised learning using local errors. Frontiers in neuroscience 12, 608 (2018)",
|
| 1655 |
+
"30. Netzer, Y., Wang, T., Coates, A., Bissacco, A., Wu, B., Ng, A.Y.: Reading digits in natural images with unsupervised feature learning (2011)",
|
| 1656 |
+
"31. Nokland, A.: Direct feedback alignment provides learning in deep neural networks. Advances in neural information processing systems 29 (2016)",
|
| 1657 |
+
"32. Nokland, A., Eidnes, L.H.: Training neural networks with local error signals. In: International conference on machine learning. pp. 4839-4850. PMLR (2019)",
|
| 1658 |
+
"33. Pyeon, M., Moon, J., Hahn, T., Kim, G.: Sedona: Search for decoupled neural networks toward greedy block-wise learning. In: International Conference on Learning Representations (2020)",
|
| 1659 |
+
"34. Ren, M., Kornblith, S., Liao, R., Hinton, G.: Scaling forward gradient with local losses. arXiv preprint arXiv:2210.03310 (2022)",
|
| 1660 |
+
"35. Rumelhart, D.E., Hinton, G.E., Williams, R.J.: Learning representations by backpropagating errors. nature 323(6088), 533-536 (1986)"
|
| 1661 |
+
],
|
| 1662 |
+
"bbox": [
|
| 1663 |
+
215,
|
| 1664 |
+
146,
|
| 1665 |
+
785,
|
| 1666 |
+
840
|
| 1667 |
+
],
|
| 1668 |
+
"page_idx": 15
|
| 1669 |
+
},
|
| 1670 |
+
{
|
| 1671 |
+
"type": "page_number",
|
| 1672 |
+
"text": "16",
|
| 1673 |
+
"bbox": [
|
| 1674 |
+
217,
|
| 1675 |
+
114,
|
| 1676 |
+
235,
|
| 1677 |
+
126
|
| 1678 |
+
],
|
| 1679 |
+
"page_idx": 15
|
| 1680 |
+
},
|
| 1681 |
+
{
|
| 1682 |
+
"type": "header",
|
| 1683 |
+
"text": "Junhao Su et al.",
|
| 1684 |
+
"bbox": [
|
| 1685 |
+
271,
|
| 1686 |
+
114,
|
| 1687 |
+
380,
|
| 1688 |
+
126
|
| 1689 |
+
],
|
| 1690 |
+
"page_idx": 15
|
| 1691 |
+
},
|
| 1692 |
+
{
|
| 1693 |
+
"type": "list",
|
| 1694 |
+
"sub_type": "ref_text",
|
| 1695 |
+
"list_items": [
|
| 1696 |
+
"36. Simonyan, K., Zisserman, A.: Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556 (2014)",
|
| 1697 |
+
"37. Wang, Y., Ni, Z., Song, S., Yang, L., Huang, G.: Revisiting locally supervised learning: an alternative to end-to-end training. arXiv preprint arXiv:2101.10832 (2021)",
|
| 1698 |
+
"38. Wu, B., Nair, S., Martin-Martin, R., Fei-Fei, L., Finn, C.: Greedy hierarchical variational autoencoders for large-scale video prediction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2318-2328 (2021)",
|
| 1699 |
+
"39. Xie, S., Girshick, R., Dollar, P., Tu, Z., He, K.: Aggregated residual transformations for deep neural networks. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1492-1500 (2017)",
|
| 1700 |
+
"40. Xie, X., Seung, H.S.: Equivalence of backpropagation and contrastive hebbian learning in a layered network. Neural computation 15(2), 441-454 (2003)",
|
| 1701 |
+
"41. Xiong, Y., Ren, M., Urtasun, R.: Loco: Local contrastive representation learning. Advances in neural information processing systems 33, 11142-11153 (2020)"
|
| 1702 |
+
],
|
| 1703 |
+
"bbox": [
|
| 1704 |
+
212,
|
| 1705 |
+
146,
|
| 1706 |
+
787,
|
| 1707 |
+
369
|
| 1708 |
+
],
|
| 1709 |
+
"page_idx": 16
|
| 1710 |
+
},
|
| 1711 |
+
{
|
| 1712 |
+
"type": "header",
|
| 1713 |
+
"text": "Momentum Auxiliary Network for Supervised Local Learning",
|
| 1714 |
+
"bbox": [
|
| 1715 |
+
321,
|
| 1716 |
+
114,
|
| 1717 |
+
730,
|
| 1718 |
+
128
|
| 1719 |
+
],
|
| 1720 |
+
"page_idx": 16
|
| 1721 |
+
},
|
| 1722 |
+
{
|
| 1723 |
+
"type": "page_number",
|
| 1724 |
+
"text": "17",
|
| 1725 |
+
"bbox": [
|
| 1726 |
+
767,
|
| 1727 |
+
116,
|
| 1728 |
+
784,
|
| 1729 |
+
126
|
| 1730 |
+
],
|
| 1731 |
+
"page_idx": 16
|
| 1732 |
+
}
|
| 1733 |
+
]
|
2024/Momentum Auxiliary Network for Supervised Local Learning/eb53acf2-10ec-44f2-923d-fc539ed31362_model.json
ADDED
|
@@ -0,0 +1,2280 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
[
|
| 3 |
+
{
|
| 4 |
+
"type": "title",
|
| 5 |
+
"bbox": [
|
| 6 |
+
0.24,
|
| 7 |
+
0.141,
|
| 8 |
+
0.764,
|
| 9 |
+
0.187
|
| 10 |
+
],
|
| 11 |
+
"angle": 0,
|
| 12 |
+
"content": "Momentum Auxiliary Network for Supervised Local Learning"
|
| 13 |
+
},
|
| 14 |
+
{
|
| 15 |
+
"type": "text",
|
| 16 |
+
"bbox": [
|
| 17 |
+
0.283,
|
| 18 |
+
0.216,
|
| 19 |
+
0.721,
|
| 20 |
+
0.231
|
| 21 |
+
],
|
| 22 |
+
"angle": 0,
|
| 23 |
+
"content": "Junhao Su\\(^{1*}\\), Changpeng Cai\\(^{1*}\\), Feiyu Zhu\\(^{2*}\\), Chenghao He\\(^{4}\\)"
|
| 24 |
+
},
|
| 25 |
+
{
|
| 26 |
+
"type": "text",
|
| 27 |
+
"bbox": [
|
| 28 |
+
0.318,
|
| 29 |
+
0.231,
|
| 30 |
+
0.686,
|
| 31 |
+
0.248
|
| 32 |
+
],
|
| 33 |
+
"angle": 0,
|
| 34 |
+
"content": "Xiaojie \\(\\mathrm{Xu}^5\\), Dongzhi Guan\\(^{1\\boxtimes}\\), and Chenyang Si\\(^{3\\boxtimes}\\)"
|
| 35 |
+
},
|
| 36 |
+
{
|
| 37 |
+
"type": "text",
|
| 38 |
+
"bbox": [
|
| 39 |
+
0.424,
|
| 40 |
+
0.258,
|
| 41 |
+
0.581,
|
| 42 |
+
0.273
|
| 43 |
+
],
|
| 44 |
+
"angle": 0,
|
| 45 |
+
"content": "\\(^{1}\\) Southeast University"
|
| 46 |
+
},
|
| 47 |
+
{
|
| 48 |
+
"type": "text",
|
| 49 |
+
"bbox": [
|
| 50 |
+
0.325,
|
| 51 |
+
0.273,
|
| 52 |
+
0.678,
|
| 53 |
+
0.288
|
| 54 |
+
],
|
| 55 |
+
"angle": 0,
|
| 56 |
+
"content": "2 University of Shanghai for Science and Technology"
|
| 57 |
+
},
|
| 58 |
+
{
|
| 59 |
+
"type": "text",
|
| 60 |
+
"bbox": [
|
| 61 |
+
0.379,
|
| 62 |
+
0.287,
|
| 63 |
+
0.623,
|
| 64 |
+
0.302
|
| 65 |
+
],
|
| 66 |
+
"angle": 0,
|
| 67 |
+
"content": "<sup>3</sup> Nanyang Technological University"
|
| 68 |
+
},
|
| 69 |
+
{
|
| 70 |
+
"type": "text",
|
| 71 |
+
"bbox": [
|
| 72 |
+
0.329,
|
| 73 |
+
0.301,
|
| 74 |
+
0.673,
|
| 75 |
+
0.315
|
| 76 |
+
],
|
| 77 |
+
"angle": 0,
|
| 78 |
+
"content": "\\(^{4}\\) East China University of Science and Technology"
|
| 79 |
+
},
|
| 80 |
+
{
|
| 81 |
+
"type": "text",
|
| 82 |
+
"bbox": [
|
| 83 |
+
0.315,
|
| 84 |
+
0.315,
|
| 85 |
+
0.69,
|
| 86 |
+
0.329
|
| 87 |
+
],
|
| 88 |
+
"angle": 0,
|
| 89 |
+
"content": "<sup>5</sup> The Hong Kong University of Science and Technology"
|
| 90 |
+
},
|
| 91 |
+
{
|
| 92 |
+
"type": "text",
|
| 93 |
+
"bbox": [
|
| 94 |
+
0.262,
|
| 95 |
+
0.398,
|
| 96 |
+
0.744,
|
| 97 |
+
0.732
|
| 98 |
+
],
|
| 99 |
+
"angle": 0,
|
| 100 |
+
"content": "Abstract. Deep neural networks conventionally employ end-to-end backpropagation for their training process, which lacks biological credibility and triggers a locking dilemma during network parameter updates, leading to significant GPU memory use. Supervised local learning, which segments the network into multiple local blocks updated by independent auxiliary networks. However, these methods cannot replace end-to-end training due to lower accuracy, as gradients only propagate within their local block, creating a lack of information exchange between blocks. To address this issue and establish information transfer across blocks, we propose a Momentum Auxiliary Network (MAN) that establishes a dynamic interaction mechanism. The MAN leverages an exponential moving average (EMA) of the parameters from adjacent local blocks to enhance information flow. This auxiliary network, updated through EMA, helps bridge the informational gap between blocks. Nevertheless, we observe that directly applying EMA parameters has certain limitations due to feature discrepancies among local blocks. To overcome this, we introduce learnable biases, further boosting performance. We have validated our method on four image classification datasets (CIFAR-10, STL-10, SVHN, ImageNet), attaining superior performance and substantial memory savings. Notably, our method can reduce GPU memory usage by more than \\(45\\%\\) on the ImageNet dataset compared to end-to-end training, while achieving higher performance. The Momentum Auxiliary Network thus offers a new perspective for supervised local learning. Our code is available at: https://github.com/JunhaoSuO/MAN."
|
| 101 |
+
},
|
| 102 |
+
{
|
| 103 |
+
"type": "text",
|
| 104 |
+
"bbox": [
|
| 105 |
+
0.261,
|
| 106 |
+
0.747,
|
| 107 |
+
0.741,
|
| 108 |
+
0.777
|
| 109 |
+
],
|
| 110 |
+
"angle": 0,
|
| 111 |
+
"content": "Keywords: Local Learning \\(\\cdot\\) Image Classification \\(\\cdot\\) Momentum Auxiliary Network"
|
| 112 |
+
},
|
| 113 |
+
{
|
| 114 |
+
"type": "page_footnote",
|
| 115 |
+
"bbox": [
|
| 116 |
+
0.232,
|
| 117 |
+
0.798,
|
| 118 |
+
0.376,
|
| 119 |
+
0.812
|
| 120 |
+
],
|
| 121 |
+
"angle": 0,
|
| 122 |
+
"content": "*Equal Contribution."
|
| 123 |
+
},
|
| 124 |
+
{
|
| 125 |
+
"type": "page_footnote",
|
| 126 |
+
"bbox": [
|
| 127 |
+
0.232,
|
| 128 |
+
0.812,
|
| 129 |
+
0.787,
|
| 130 |
+
0.827
|
| 131 |
+
],
|
| 132 |
+
"angle": 0,
|
| 133 |
+
"content": "Corresponding Authors: Chengyang Si (chenyang.si.mail@gmail.com) and Dongzhi"
|
| 134 |
+
},
|
| 135 |
+
{
|
| 136 |
+
"type": "page_footnote",
|
| 137 |
+
"bbox": [
|
| 138 |
+
0.232,
|
| 139 |
+
0.827,
|
| 140 |
+
0.456,
|
| 141 |
+
0.84
|
| 142 |
+
],
|
| 143 |
+
"angle": 0,
|
| 144 |
+
"content": "Guan (guandongzhi@seu.edu.cn)."
|
| 145 |
+
}
|
| 146 |
+
],
|
| 147 |
+
[
|
| 148 |
+
{
|
| 149 |
+
"type": "page_number",
|
| 150 |
+
"bbox": [
|
| 151 |
+
0.218,
|
| 152 |
+
0.116,
|
| 153 |
+
0.23,
|
| 154 |
+
0.127
|
| 155 |
+
],
|
| 156 |
+
"angle": 0,
|
| 157 |
+
"content": "2"
|
| 158 |
+
},
|
| 159 |
+
{
|
| 160 |
+
"type": "header",
|
| 161 |
+
"bbox": [
|
| 162 |
+
0.272,
|
| 163 |
+
0.115,
|
| 164 |
+
0.383,
|
| 165 |
+
0.127
|
| 166 |
+
],
|
| 167 |
+
"angle": 0,
|
| 168 |
+
"content": "Junhao Su et al."
|
| 169 |
+
},
|
| 170 |
+
{
|
| 171 |
+
"type": "image",
|
| 172 |
+
"bbox": [
|
| 173 |
+
0.218,
|
| 174 |
+
0.146,
|
| 175 |
+
0.788,
|
| 176 |
+
0.329
|
| 177 |
+
],
|
| 178 |
+
"angle": 0,
|
| 179 |
+
"content": null
|
| 180 |
+
},
|
| 181 |
+
{
|
| 182 |
+
"type": "image_caption",
|
| 183 |
+
"bbox": [
|
| 184 |
+
0.214,
|
| 185 |
+
0.339,
|
| 186 |
+
0.788,
|
| 187 |
+
0.382
|
| 188 |
+
],
|
| 189 |
+
"angle": 0,
|
| 190 |
+
"content": "Fig. 1: Comparison between different methods with MAN and the original methods in terms of accuracy. Results are obtained using ResNet-110 (K=55) on the various datasets."
|
| 191 |
+
},
|
| 192 |
+
{
|
| 193 |
+
"type": "title",
|
| 194 |
+
"bbox": [
|
| 195 |
+
0.218,
|
| 196 |
+
0.412,
|
| 197 |
+
0.377,
|
| 198 |
+
0.429
|
| 199 |
+
],
|
| 200 |
+
"angle": 0,
|
| 201 |
+
"content": "1 Introduction"
|
| 202 |
+
},
|
| 203 |
+
{
|
| 204 |
+
"type": "text",
|
| 205 |
+
"bbox": [
|
| 206 |
+
0.214,
|
| 207 |
+
0.448,
|
| 208 |
+
0.789,
|
| 209 |
+
0.844
|
| 210 |
+
],
|
| 211 |
+
"angle": 0,
|
| 212 |
+
"content": "In deep learning, the prevalent use of end-to-end backpropagation [35] is essential for training complex neural networks, which requires heavy computational work for loss evaluation and successive gradient descent through network layers to refine parameters [17, 24, 29]. This method starkly contrasts with the local signal processing of biological synapses [6, 7, 40] and imposes a 'locking' constraint [19] that delays parameter updates until the entire forward and backward pass is complete. Such constraints exacerbate challenges like reduced parallelism and increased GPU memory usage [28, 38], compromising training efficiency and scalability. An emerging alternative to traditional end-to-end (E2E) training is Local Learning [2, 3, 5, 18, 19, 29, 32, 37, 41], which promises to mitigate the drawbacks of traditional methods. These methods divide the network into discrete, gradient-isolated blocks, each of which is updated by its dedicated auxiliary network according to distinct local objectives [3, 32, 37]. This training strategy enables immediate parameter updates upon local error reception, circumventing the sequential update bottleneck of E2E training, thereby significantly enhancing the efficiency of parallel training [13, 22]. Moreover, local learning reduces the demand on GPU memory by only retaining gradients for the local and auxiliary networks, thus eliminating the need to store extensive global gradient information and saving computational resources [28, 38]. However, even with these advantages, there is still a significant performance gap between local learning and traditional E2E methods, especially as the network is divided into more local blocks. Current local learning techniques mainly focus on improving the design of auxiliary network structures [3] and making local loss functions better to close this performance gap [32, 37]. Yet, these improvements do not completely address the inherent short-sightedness of local learning: the separation into blocks can make each part of the network only focus on its local objectives, possibly ig-"
|
| 213 |
+
}
|
| 214 |
+
],
|
| 215 |
+
[
|
| 216 |
+
{
|
| 217 |
+
"type": "header",
|
| 218 |
+
"bbox": [
|
| 219 |
+
0.323,
|
| 220 |
+
0.115,
|
| 221 |
+
0.733,
|
| 222 |
+
0.131
|
| 223 |
+
],
|
| 224 |
+
"angle": 0,
|
| 225 |
+
"content": "Momentum Auxiliary Network for Supervised Local Learning"
|
| 226 |
+
},
|
| 227 |
+
{
|
| 228 |
+
"type": "page_number",
|
| 229 |
+
"bbox": [
|
| 230 |
+
0.775,
|
| 231 |
+
0.117,
|
| 232 |
+
0.785,
|
| 233 |
+
0.127
|
| 234 |
+
],
|
| 235 |
+
"angle": 0,
|
| 236 |
+
"content": "3"
|
| 237 |
+
},
|
| 238 |
+
{
|
| 239 |
+
"type": "text",
|
| 240 |
+
"bbox": [
|
| 241 |
+
0.214,
|
| 242 |
+
0.147,
|
| 243 |
+
0.784,
|
| 244 |
+
0.176
|
| 245 |
+
],
|
| 246 |
+
"angle": 0,
|
| 247 |
+
"content": "noring the overall objectives of the network. This can lead to the discarding of globally beneficial information due to the lack of inter-block communication."
|
| 248 |
+
},
|
| 249 |
+
{
|
| 250 |
+
"type": "text",
|
| 251 |
+
"bbox": [
|
| 252 |
+
0.214,
|
| 253 |
+
0.178,
|
| 254 |
+
0.788,
|
| 255 |
+
0.494
|
| 256 |
+
],
|
| 257 |
+
"angle": 0,
|
| 258 |
+
"content": "In this paper, we introduce the Momentum Auxiliary Network (MAN), a novel network architecture designed to mitigate the inherent limitations of supervised local learning by enhancing inter-block communication. Specifically, MAN not only accepts the current local block as input but also absorbs the parameters of its next block. Upon completion of a forward pass, parameter updates are performed using local gradients refined through Exponential Moving Average (EMA) techniques [14]. This novel approach enables each local block to integrate information from the following block, thereby extending the operational perspective of each block beyond its immediate objectives and further aligning with the global objective of the network. This addresses the critical limitations in local learning frameworks, ensuring a more cohesive alignment with the overarching network objectives. Furthermore, we identify that directly applying EMA parameters is constrained by the feature incongruities across gradient-isolated blocks. To mitigate this, we implement additional learnable bias terms to each auxiliary block, enhancing their ability to share information effectively. The proposed MAN requires only a minimal increase in memory use but significantly improves performance by enhancing the information sharing ability of local blocks. The efficacy of the MAN approach is validated on a suite of benchmark image classification datasets, including CIFAR-10 [21], STL-10 [9], SVHN [30], and ImageNet [12]. The results demonstrate the effectiveness of our method in surpassing the limitations of traditional supervised local learning."
|
| 259 |
+
},
|
| 260 |
+
{
|
| 261 |
+
"type": "text",
|
| 262 |
+
"bbox": [
|
| 263 |
+
0.24,
|
| 264 |
+
0.496,
|
| 265 |
+
0.702,
|
| 266 |
+
0.51
|
| 267 |
+
],
|
| 268 |
+
"angle": 0,
|
| 269 |
+
"content": "The contributions of this paper could be summarized as follows:"
|
| 270 |
+
},
|
| 271 |
+
{
|
| 272 |
+
"type": "text",
|
| 273 |
+
"bbox": [
|
| 274 |
+
0.226,
|
| 275 |
+
0.52,
|
| 276 |
+
0.784,
|
| 277 |
+
0.594
|
| 278 |
+
],
|
| 279 |
+
"angle": 0,
|
| 280 |
+
"content": "- We propose Momentum Auxiliary Network (MAN), designed to facilitate inter-block communication with Exponential Moving Average (EMA), mitigating the short-sightedness issue in conventional supervised local learning techniques and culminating in an elevated overall performance of the network."
|
| 281 |
+
},
|
| 282 |
+
{
|
| 283 |
+
"type": "text",
|
| 284 |
+
"bbox": [
|
| 285 |
+
0.226,
|
| 286 |
+
0.596,
|
| 287 |
+
0.784,
|
| 288 |
+
0.64
|
| 289 |
+
],
|
| 290 |
+
"angle": 0,
|
| 291 |
+
"content": "- MAN is a versatile, plug-and-play approach that can seamlessly integrate with any supervised local learning technique, markedly amplifying their efficacy while requiring only a negligible increase in memory usage."
|
| 292 |
+
},
|
| 293 |
+
{
|
| 294 |
+
"type": "text",
|
| 295 |
+
"bbox": [
|
| 296 |
+
0.226,
|
| 297 |
+
0.642,
|
| 298 |
+
0.784,
|
| 299 |
+
0.702
|
| 300 |
+
],
|
| 301 |
+
"angle": 0,
|
| 302 |
+
"content": "- MAN has demonstrated its effectiveness through experiments on benchmark image classification datasets, achieving state-of-the-art performance. Notably, on the ImageNet [12] dataset, it outperforms E2E training while using significantly less memory."
|
| 303 |
+
},
|
| 304 |
+
{
|
| 305 |
+
"type": "list",
|
| 306 |
+
"bbox": [
|
| 307 |
+
0.226,
|
| 308 |
+
0.52,
|
| 309 |
+
0.784,
|
| 310 |
+
0.702
|
| 311 |
+
],
|
| 312 |
+
"angle": 0,
|
| 313 |
+
"content": null
|
| 314 |
+
},
|
| 315 |
+
{
|
| 316 |
+
"type": "title",
|
| 317 |
+
"bbox": [
|
| 318 |
+
0.216,
|
| 319 |
+
0.737,
|
| 320 |
+
0.388,
|
| 321 |
+
0.753
|
| 322 |
+
],
|
| 323 |
+
"angle": 0,
|
| 324 |
+
"content": "2 Related Work"
|
| 325 |
+
},
|
| 326 |
+
{
|
| 327 |
+
"type": "title",
|
| 328 |
+
"bbox": [
|
| 329 |
+
0.216,
|
| 330 |
+
0.769,
|
| 331 |
+
0.386,
|
| 332 |
+
0.785
|
| 333 |
+
],
|
| 334 |
+
"angle": 0,
|
| 335 |
+
"content": "2.1 Local Learning"
|
| 336 |
+
},
|
| 337 |
+
{
|
| 338 |
+
"type": "text",
|
| 339 |
+
"bbox": [
|
| 340 |
+
0.214,
|
| 341 |
+
0.795,
|
| 342 |
+
0.785,
|
| 343 |
+
0.842
|
| 344 |
+
],
|
| 345 |
+
"angle": 0,
|
| 346 |
+
"content": "Local learning is first proposed as an innovative deep learning algorithm with the intention of utilizing memory more efficiently and adhering more closely to principles of biological plausibility [10]. This approach emerges as a response to"
|
| 347 |
+
}
|
| 348 |
+
],
|
| 349 |
+
[
|
| 350 |
+
{
|
| 351 |
+
"type": "page_number",
|
| 352 |
+
"bbox": [
|
| 353 |
+
0.218,
|
| 354 |
+
0.116,
|
| 355 |
+
0.23,
|
| 356 |
+
0.127
|
| 357 |
+
],
|
| 358 |
+
"angle": 0,
|
| 359 |
+
"content": "4"
|
| 360 |
+
},
|
| 361 |
+
{
|
| 362 |
+
"type": "header",
|
| 363 |
+
"bbox": [
|
| 364 |
+
0.272,
|
| 365 |
+
0.115,
|
| 366 |
+
0.383,
|
| 367 |
+
0.127
|
| 368 |
+
],
|
| 369 |
+
"angle": 0,
|
| 370 |
+
"content": "Junhao Su et al."
|
| 371 |
+
},
|
| 372 |
+
{
|
| 373 |
+
"type": "image",
|
| 374 |
+
"bbox": [
|
| 375 |
+
0.22,
|
| 376 |
+
0.143,
|
| 377 |
+
0.79,
|
| 378 |
+
0.294
|
| 379 |
+
],
|
| 380 |
+
"angle": 0,
|
| 381 |
+
"content": null
|
| 382 |
+
},
|
| 383 |
+
{
|
| 384 |
+
"type": "image_caption",
|
| 385 |
+
"bbox": [
|
| 386 |
+
0.214,
|
| 387 |
+
0.306,
|
| 388 |
+
0.788,
|
| 389 |
+
0.348
|
| 390 |
+
],
|
| 391 |
+
"angle": 0,
|
| 392 |
+
"content": "Fig. 2: Comparison of (a) end-to-end backpropagation, (b) other supervised local learning methods, and (c) our proposed method. Unlike E2E, supervised local learning separates the network into K gradient-isolated local blocks."
|
| 393 |
+
},
|
| 394 |
+
{
|
| 395 |
+
"type": "text",
|
| 396 |
+
"bbox": [
|
| 397 |
+
0.214,
|
| 398 |
+
0.39,
|
| 399 |
+
0.788,
|
| 400 |
+
0.512
|
| 401 |
+
],
|
| 402 |
+
"angle": 0,
|
| 403 |
+
"content": "the limitations presented by global E2E training [16], prompting the development of alternative supervised local learning methodologies. Examples of these methodologies include a differentiable search algorithm, which decouples network blocks for block-level learning and selects a manually assisted network for each local block [33], as well as a self-supervised contrasting loss function that leverages its local learning rules [18, 41]. In supervised local learning, training primarily relies on the design of reasonable supervised local loss functions or the manual construction of auxiliary networks."
|
| 404 |
+
},
|
| 405 |
+
{
|
| 406 |
+
"type": "text",
|
| 407 |
+
"bbox": [
|
| 408 |
+
0.214,
|
| 409 |
+
0.517,
|
| 410 |
+
0.788,
|
| 411 |
+
0.684
|
| 412 |
+
],
|
| 413 |
+
"angle": 0,
|
| 414 |
+
"content": "Supervised Local Loss: InfoPro [37] presents a method that involves the reconstruction loss for each hidden layer, coupled with cross-entropy loss. This approach forces the hidden layers to preserve vital input information, thereby preventing the early layers from discarding information relevant to the task. PredSim [32] employs layer-wise loss functions for network training. It makes use of two distinct supervised loss functions to generate local error signals for hidden layers. The optimization of these losses takes place during supervised local learning, thus eliminating the necessity to propagate global errors back to the hidden layers. These methods have achieved advanced performance in the field of supervised local learning, but there is still a significant gap compared to the performance of E2E training."
|
| 415 |
+
},
|
| 416 |
+
{
|
| 417 |
+
"type": "text",
|
| 418 |
+
"bbox": [
|
| 419 |
+
0.214,
|
| 420 |
+
0.69,
|
| 421 |
+
0.788,
|
| 422 |
+
0.841
|
| 423 |
+
],
|
| 424 |
+
"angle": 0,
|
| 425 |
+
"content": "Auxiliary Networks: InfoPro [37] incorporates two auxiliary networks into its approach. One network consists of a single convolutional layer followed by two fully-connected layers, which are employed for the cross-entropy loss. The other network utilizes two convolutional layers with upsampling for the reconstruction loss. Alternatively, DGL [3] designs a structure that uses three consecutive convolutional layers connected to a global pooling layer, followed by three consecutive fully connected layers. Moreover, DGL restricts the parameter of the auxiliary network to only \\(5\\%\\) of the main network capacity, offering advantages in terms of speed and memory savings. However, these method still perform poorly when the network is divided into a large number of local blocks."
|
| 426 |
+
}
|
| 427 |
+
],
|
| 428 |
+
[
|
| 429 |
+
{
|
| 430 |
+
"type": "header",
|
| 431 |
+
"bbox": [
|
| 432 |
+
0.322,
|
| 433 |
+
0.115,
|
| 434 |
+
0.733,
|
| 435 |
+
0.131
|
| 436 |
+
],
|
| 437 |
+
"angle": 0,
|
| 438 |
+
"content": "Momentum Auxiliary Network for Supervised Local Learning"
|
| 439 |
+
},
|
| 440 |
+
{
|
| 441 |
+
"type": "page_number",
|
| 442 |
+
"bbox": [
|
| 443 |
+
0.775,
|
| 444 |
+
0.117,
|
| 445 |
+
0.785,
|
| 446 |
+
0.127
|
| 447 |
+
],
|
| 448 |
+
"angle": 0,
|
| 449 |
+
"content": "5"
|
| 450 |
+
},
|
| 451 |
+
{
|
| 452 |
+
"type": "title",
|
| 453 |
+
"bbox": [
|
| 454 |
+
0.215,
|
| 455 |
+
0.147,
|
| 456 |
+
0.627,
|
| 457 |
+
0.163
|
| 458 |
+
],
|
| 459 |
+
"angle": 0,
|
| 460 |
+
"content": "2.2 Alternative Learning Rules to E2E Training"
|
| 461 |
+
},
|
| 462 |
+
{
|
| 463 |
+
"type": "text",
|
| 464 |
+
"bbox": [
|
| 465 |
+
0.214,
|
| 466 |
+
0.172,
|
| 467 |
+
0.788,
|
| 468 |
+
0.385
|
| 469 |
+
],
|
| 470 |
+
"angle": 0,
|
| 471 |
+
"content": "Due to certain limitations inherent in E2E training, the pursuit of alternative methods to E2E training has gradually garnered attention in recent years [27]. Several efforts have been dedicated to addressing the biologically implausible aspects of E2E training, such as the training methodology of target propagation [4, 23, 25] attempts to train a specialized backward connection by utilizing local reconstruction targets. Additionally, some recent studies have attempted to completely avoid backpropagation in neural networks through forward gradient learning [11, 34]. Meanwhile, the weight transport problem [10]. This has been tackled either by employing distinct feedback connections [1, 26] or by directly disseminating global errors to each hidden unit [8, 31]. While these methods have somewhat alleviated the biological implausibility of E2E training, they still rely on global objectives, which fundamentally differ from biological neural networks that rely on local synapses for information transmission. Furthermore, these methods currently struggle to be effective on large datasets [12]."
|
| 472 |
+
},
|
| 473 |
+
{
|
| 474 |
+
"type": "title",
|
| 475 |
+
"bbox": [
|
| 476 |
+
0.216,
|
| 477 |
+
0.407,
|
| 478 |
+
0.331,
|
| 479 |
+
0.423
|
| 480 |
+
],
|
| 481 |
+
"angle": 0,
|
| 482 |
+
"content": "3 Method"
|
| 483 |
+
},
|
| 484 |
+
{
|
| 485 |
+
"type": "title",
|
| 486 |
+
"bbox": [
|
| 487 |
+
0.215,
|
| 488 |
+
0.438,
|
| 489 |
+
0.373,
|
| 490 |
+
0.453
|
| 491 |
+
],
|
| 492 |
+
"angle": 0,
|
| 493 |
+
"content": "3.1 Preliminaries"
|
| 494 |
+
},
|
| 495 |
+
{
|
| 496 |
+
"type": "text",
|
| 497 |
+
"bbox": [
|
| 498 |
+
0.214,
|
| 499 |
+
0.463,
|
| 500 |
+
0.788,
|
| 501 |
+
0.599
|
| 502 |
+
],
|
| 503 |
+
"angle": 0,
|
| 504 |
+
"content": "To set the stage, we first provide a brief overview of traditional end-to-end supervised learning and backpropagation mechanisms. We denote a data sample as \\( x \\) and its corresponding ground-truth label as \\( y \\). The entire deep network is fragmented into several local blocks. During the forward propagation process, the output from the j-th block serves as the input for the (j+1)-th block, expressed as \\( x_{j+1} = f_{\\theta_j}(x_j) \\). Here, \\( \\theta_j \\) symbolizes the parameters of the j-th local block and \\( f(\\cdot) \\) represents the forward calculation of the block. We evaluate the loss function \\( \\mathcal{L}(\\hat{y}, y) \\) between the output of the last block and the ground truth label, and propagate it back iteratively to the preceding blocks."
|
| 505 |
+
},
|
| 506 |
+
{
|
| 507 |
+
"type": "text",
|
| 508 |
+
"bbox": [
|
| 509 |
+
0.214,
|
| 510 |
+
0.599,
|
| 511 |
+
0.788,
|
| 512 |
+
0.674
|
| 513 |
+
],
|
| 514 |
+
"angle": 0,
|
| 515 |
+
"content": "Supervised local learning strategies [3, 32, 37] integrate auxiliary networks for local supervision. For each local block, an auxiliary network is affixed. The output from a local block is fed to its corresponding auxiliary network, generating the local supervisory signal as \\(\\hat{y}_j = g_{\\gamma_j}(x_{j + 1})\\). Here, \\(\\gamma_{j}\\) denotes the parameters of the j-th auxiliary network."
|
| 516 |
+
},
|
| 517 |
+
{
|
| 518 |
+
"type": "text",
|
| 519 |
+
"bbox": [
|
| 520 |
+
0.214,
|
| 521 |
+
0.675,
|
| 522 |
+
0.788,
|
| 523 |
+
0.707
|
| 524 |
+
],
|
| 525 |
+
"angle": 0,
|
| 526 |
+
"content": "In this setup, we update the parameters of the j-th auxiliary network and local block, \\(\\gamma_{j},\\theta_{j}\\), as follows:"
|
| 527 |
+
},
|
| 528 |
+
{
|
| 529 |
+
"type": "equation",
|
| 530 |
+
"bbox": [
|
| 531 |
+
0.405,
|
| 532 |
+
0.721,
|
| 533 |
+
0.785,
|
| 534 |
+
0.737
|
| 535 |
+
],
|
| 536 |
+
"angle": 0,
|
| 537 |
+
"content": "\\[\n\\gamma_ {j} \\leftarrow \\gamma_ {j} - \\eta_ {a} \\times \\nabla_ {\\gamma_ {j}} \\mathcal {L} (\\hat {y} _ {j}, y) \\tag {1}\n\\]"
|
| 538 |
+
},
|
| 539 |
+
{
|
| 540 |
+
"type": "equation",
|
| 541 |
+
"bbox": [
|
| 542 |
+
0.407,
|
| 543 |
+
0.758,
|
| 544 |
+
0.785,
|
| 545 |
+
0.774
|
| 546 |
+
],
|
| 547 |
+
"angle": 0,
|
| 548 |
+
"content": "\\[\n\\theta_ {j} \\leftarrow \\theta_ {j} - \\eta_ {l} \\times \\nabla_ {\\theta_ {j}} \\mathcal {L} \\left(y _ {\\hat {j}}, y\\right) \\tag {2}\n\\]"
|
| 549 |
+
},
|
| 550 |
+
{
|
| 551 |
+
"type": "text",
|
| 552 |
+
"bbox": [
|
| 553 |
+
0.214,
|
| 554 |
+
0.78,
|
| 555 |
+
0.788,
|
| 556 |
+
0.841
|
| 557 |
+
],
|
| 558 |
+
"angle": 0,
|
| 559 |
+
"content": "where \\(\\eta_{a},\\eta_{l}\\) are the learning rates of the auxiliary networks and local blocks, respectively. By attaching auxiliary networks, each local block becomes gradient-isolated and can be updated with local supervision rather than global backpropagation."
|
| 560 |
+
}
|
| 561 |
+
],
|
| 562 |
+
[
|
| 563 |
+
{
|
| 564 |
+
"type": "page_number",
|
| 565 |
+
"bbox": [
|
| 566 |
+
0.218,
|
| 567 |
+
0.116,
|
| 568 |
+
0.23,
|
| 569 |
+
0.127
|
| 570 |
+
],
|
| 571 |
+
"angle": 0,
|
| 572 |
+
"content": "6"
|
| 573 |
+
},
|
| 574 |
+
{
|
| 575 |
+
"type": "header",
|
| 576 |
+
"bbox": [
|
| 577 |
+
0.272,
|
| 578 |
+
0.115,
|
| 579 |
+
0.384,
|
| 580 |
+
0.128
|
| 581 |
+
],
|
| 582 |
+
"angle": 0,
|
| 583 |
+
"content": "Junhao Su et al."
|
| 584 |
+
},
|
| 585 |
+
{
|
| 586 |
+
"type": "title",
|
| 587 |
+
"bbox": [
|
| 588 |
+
0.216,
|
| 589 |
+
0.147,
|
| 590 |
+
0.519,
|
| 591 |
+
0.163
|
| 592 |
+
],
|
| 593 |
+
"angle": 0,
|
| 594 |
+
"content": "3.2 Momentum Auxiliary Network"
|
| 595 |
+
},
|
| 596 |
+
{
|
| 597 |
+
"type": "text",
|
| 598 |
+
"bbox": [
|
| 599 |
+
0.214,
|
| 600 |
+
0.173,
|
| 601 |
+
0.788,
|
| 602 |
+
0.248
|
| 603 |
+
],
|
| 604 |
+
"angle": 0,
|
| 605 |
+
"content": "Existing techniques incorporate supervision signals into individual local blocks, which allows for parallel parameter updates and reduces memory overhead. However, this approach can lead to a short-sightedness problem, where each local block neglects the information from subsequent blocks, ultimately resulting in suboptimal final accuracy."
|
| 606 |
+
},
|
| 607 |
+
{
|
| 608 |
+
"type": "text",
|
| 609 |
+
"bbox": [
|
| 610 |
+
0.214,
|
| 611 |
+
0.249,
|
| 612 |
+
0.788,
|
| 613 |
+
0.325
|
| 614 |
+
],
|
| 615 |
+
"angle": 0,
|
| 616 |
+
"content": "To address this issue, we propose a comprehensive information exchange module—the Momentum Auxiliary Network (MAN). The MAN employs the Exponential Moving Average (EMA) mechanism [14] as a conduit for transferring information from subsequent blocks to the current block. In the context of MAN, we update the parameters of the j-th auxiliary network and local block as follows:"
|
| 617 |
+
},
|
| 618 |
+
{
|
| 619 |
+
"type": "equation",
|
| 620 |
+
"bbox": [
|
| 621 |
+
0.405,
|
| 622 |
+
0.339,
|
| 623 |
+
0.786,
|
| 624 |
+
0.357
|
| 625 |
+
],
|
| 626 |
+
"angle": 0,
|
| 627 |
+
"content": "\\[\n\\gamma_ {j} \\leftarrow \\gamma_ {j} - \\eta_ {a} \\times \\nabla_ {\\gamma_ {j}} \\mathcal {L} \\left(\\hat {y} _ {j}, y\\right) \\tag {3}\n\\]"
|
| 628 |
+
},
|
| 629 |
+
{
|
| 630 |
+
"type": "equation",
|
| 631 |
+
"bbox": [
|
| 632 |
+
0.425,
|
| 633 |
+
0.378,
|
| 634 |
+
0.787,
|
| 635 |
+
0.395
|
| 636 |
+
],
|
| 637 |
+
"angle": 0,
|
| 638 |
+
"content": "\\[\n\\gamma_ {j} \\leftarrow E M A \\left(\\gamma_ {j}, \\theta_ {j + 1}\\right) \\tag {4}\n\\]"
|
| 639 |
+
},
|
| 640 |
+
{
|
| 641 |
+
"type": "equation",
|
| 642 |
+
"bbox": [
|
| 643 |
+
0.405,
|
| 644 |
+
0.416,
|
| 645 |
+
0.786,
|
| 646 |
+
0.433
|
| 647 |
+
],
|
| 648 |
+
"angle": 0,
|
| 649 |
+
"content": "\\[\n\\theta_ {j} \\leftarrow \\theta_ {j} - \\eta_ {l} \\times \\nabla_ {\\theta_ {j}} \\mathcal {L} (\\hat {y} _ {j}, y) \\tag {5}\n\\]"
|
| 650 |
+
},
|
| 651 |
+
{
|
| 652 |
+
"type": "text",
|
| 653 |
+
"bbox": [
|
| 654 |
+
0.214,
|
| 655 |
+
0.439,
|
| 656 |
+
0.788,
|
| 657 |
+
0.5
|
| 658 |
+
],
|
| 659 |
+
"angle": 0,
|
| 660 |
+
"content": "where \\(\\gamma_{j}\\) represents the parameters of the j-th auxiliary network and \\(\\theta_{j}\\) represents those of the j-th local block. After updating with local gradients, \\(\\gamma_{j}\\) undergoes further refinement by incorporating the parameters of the subsequent local block via the EMA, which is a weighted sum operation."
|
| 661 |
+
},
|
| 662 |
+
{
|
| 663 |
+
"type": "text",
|
| 664 |
+
"bbox": [
|
| 665 |
+
0.214,
|
| 666 |
+
0.5,
|
| 667 |
+
0.487,
|
| 668 |
+
0.741
|
| 669 |
+
],
|
| 670 |
+
"angle": 0,
|
| 671 |
+
"content": "However, through experimental results, it becomes apparent that directly applying EMA parameters to update the auxiliary network has its limitations, and it provides limited improvements to each local block. Upon analyzing, we find that the features learned between each local block vary to a certain degree, which hampers the effectiveness of the EMA update method. As a consequence, a learnable bias is introduced to augment the learning capability of the hidden layers within the local block, while also compensating for the deficiencies of the EMA update method."
|
| 672 |
+
},
|
| 673 |
+
{
|
| 674 |
+
"type": "image",
|
| 675 |
+
"bbox": [
|
| 676 |
+
0.499,
|
| 677 |
+
0.529,
|
| 678 |
+
0.785,
|
| 679 |
+
0.633
|
| 680 |
+
],
|
| 681 |
+
"angle": 0,
|
| 682 |
+
"content": null
|
| 683 |
+
},
|
| 684 |
+
{
|
| 685 |
+
"type": "image_caption",
|
| 686 |
+
"bbox": [
|
| 687 |
+
0.497,
|
| 688 |
+
0.644,
|
| 689 |
+
0.788,
|
| 690 |
+
0.741
|
| 691 |
+
],
|
| 692 |
+
"angle": 0,
|
| 693 |
+
"content": "Fig. 3: Details of the Momentum Auxiliary Network. Local \\((\\mathrm{i} + 1)\\) represents the \\((\\mathrm{i} + 1)\\)-th gradient-isolated local block, which contains layers from layer m to layer \\((\\mathrm{m} + \\mathrm{n})\\), totaling n+1 layers \\((\\mathrm{n} \\geqslant 0)\\). We only use the parameters of the first layer to ensure a balance in GPU memory usage."
|
| 694 |
+
},
|
| 695 |
+
{
|
| 696 |
+
"type": "text",
|
| 697 |
+
"bbox": [
|
| 698 |
+
0.214,
|
| 699 |
+
0.741,
|
| 700 |
+
0.788,
|
| 701 |
+
0.771
|
| 702 |
+
],
|
| 703 |
+
"angle": 0,
|
| 704 |
+
"content": "Therefore, the parameter update method for the j-th auxiliary network can be written as:"
|
| 705 |
+
},
|
| 706 |
+
{
|
| 707 |
+
"type": "equation",
|
| 708 |
+
"bbox": [
|
| 709 |
+
0.356,
|
| 710 |
+
0.786,
|
| 711 |
+
0.787,
|
| 712 |
+
0.806
|
| 713 |
+
],
|
| 714 |
+
"angle": 0,
|
| 715 |
+
"content": "\\[\n\\left(\\gamma_ {j}, b _ {j}\\right) \\leftarrow \\left(\\gamma_ {j}, b _ {j}\\right) - \\eta_ {a} \\times \\nabla_ {\\left(\\gamma_ {j}, b _ {j}\\right)} \\mathcal {L} \\left(\\hat {y} _ {j}, y\\right) \\tag {6}\n\\]"
|
| 716 |
+
},
|
| 717 |
+
{
|
| 718 |
+
"type": "equation",
|
| 719 |
+
"bbox": [
|
| 720 |
+
0.424,
|
| 721 |
+
0.825,
|
| 722 |
+
0.787,
|
| 723 |
+
0.843
|
| 724 |
+
],
|
| 725 |
+
"angle": 0,
|
| 726 |
+
"content": "\\[\n\\gamma_ {j} \\leftarrow E M A \\left(\\gamma_ {j}, \\theta_ {j + 1}\\right) \\tag {7}\n\\]"
|
| 727 |
+
}
|
| 728 |
+
],
|
| 729 |
+
[
|
| 730 |
+
{
|
| 731 |
+
"type": "header",
|
| 732 |
+
"bbox": [
|
| 733 |
+
0.323,
|
| 734 |
+
0.115,
|
| 735 |
+
0.733,
|
| 736 |
+
0.131
|
| 737 |
+
],
|
| 738 |
+
"angle": 0,
|
| 739 |
+
"content": "Momentum Auxiliary Network for Supervised Local Learning"
|
| 740 |
+
},
|
| 741 |
+
{
|
| 742 |
+
"type": "page_number",
|
| 743 |
+
"bbox": [
|
| 744 |
+
0.775,
|
| 745 |
+
0.117,
|
| 746 |
+
0.785,
|
| 747 |
+
0.127
|
| 748 |
+
],
|
| 749 |
+
"angle": 0,
|
| 750 |
+
"content": "7"
|
| 751 |
+
},
|
| 752 |
+
{
|
| 753 |
+
"type": "text",
|
| 754 |
+
"bbox": [
|
| 755 |
+
0.214,
|
| 756 |
+
0.147,
|
| 757 |
+
0.784,
|
| 758 |
+
0.192
|
| 759 |
+
],
|
| 760 |
+
"angle": 0,
|
| 761 |
+
"content": "where \\((\\gamma_j, b_j)\\) represent the parameters of the j-th auxiliary network. During this process, we update the parameters \\(\\gamma_j\\) of the j-th auxiliary network jointly through the learnable bias \\(b_j\\) and EMA."
|
| 762 |
+
},
|
| 763 |
+
{
|
| 764 |
+
"type": "text",
|
| 765 |
+
"bbox": [
|
| 766 |
+
0.214,
|
| 767 |
+
0.192,
|
| 768 |
+
0.784,
|
| 769 |
+
0.282
|
| 770 |
+
],
|
| 771 |
+
"angle": 0,
|
| 772 |
+
"content": "Essentially, the proposed MAN facilitates communication between subsequent blocks through the EMA mechanism [14], effectively resolving the short-sightedness problem inherent in traditional supervised local learning methods. Concurrently, it introduces an independent learnable bias to mitigate information discrepancies caused by feature variations among different local blocks, thereby aligning the output of each block closer to the global target."
|
| 773 |
+
},
|
| 774 |
+
{
|
| 775 |
+
"type": "text",
|
| 776 |
+
"bbox": [
|
| 777 |
+
0.214,
|
| 778 |
+
0.283,
|
| 779 |
+
0.784,
|
| 780 |
+
0.344
|
| 781 |
+
],
|
| 782 |
+
"angle": 0,
|
| 783 |
+
"content": "In addition to the two innovative methods mentioned above, MAN demonstrates strong versatility. It can be seamlessly integrated into existing supervised local learning methods, and it excels across various datasets, reflecting its flexibility and generalizability."
|
| 784 |
+
},
|
| 785 |
+
{
|
| 786 |
+
"type": "title",
|
| 787 |
+
"bbox": [
|
| 788 |
+
0.215,
|
| 789 |
+
0.365,
|
| 790 |
+
0.376,
|
| 791 |
+
0.383
|
| 792 |
+
],
|
| 793 |
+
"angle": 0,
|
| 794 |
+
"content": "4 Experiments"
|
| 795 |
+
},
|
| 796 |
+
{
|
| 797 |
+
"type": "title",
|
| 798 |
+
"bbox": [
|
| 799 |
+
0.215,
|
| 800 |
+
0.395,
|
| 801 |
+
0.429,
|
| 802 |
+
0.411
|
| 803 |
+
],
|
| 804 |
+
"angle": 0,
|
| 805 |
+
"content": "4.1 Experimental Setup"
|
| 806 |
+
},
|
| 807 |
+
{
|
| 808 |
+
"type": "text",
|
| 809 |
+
"bbox": [
|
| 810 |
+
0.214,
|
| 811 |
+
0.418,
|
| 812 |
+
0.784,
|
| 813 |
+
0.463
|
| 814 |
+
],
|
| 815 |
+
"angle": 0,
|
| 816 |
+
"content": "We conduct experiments using four widely adopted datasets: CIFAR-10 [21], SVHN [30], STL-10 [9], and ImageNet [12], with ResNets [15] of varying depths serving as the network architectures."
|
| 817 |
+
},
|
| 818 |
+
{
|
| 819 |
+
"type": "text",
|
| 820 |
+
"bbox": [
|
| 821 |
+
0.214,
|
| 822 |
+
0.464,
|
| 823 |
+
0.785,
|
| 824 |
+
0.585
|
| 825 |
+
],
|
| 826 |
+
"angle": 0,
|
| 827 |
+
"content": "Three state-of-the-art supervised local learning methods are selected for comparison: PredSim [32], DGL [3], and InfoPro [37]. We then partition the network into K local blocks, each containing an approximately equal number of layers. Our proposed Momentum Auxiliary Network is incorporated only into the first K-1 local blocks. The K-th local block does not employ an auxiliary network and is directly connected to the output classifier. We compare these configurations against traditional E2E and original supervised local learning methods to maintain consistent training settings and eliminate confounding variables."
|
| 828 |
+
},
|
| 829 |
+
{
|
| 830 |
+
"type": "title",
|
| 831 |
+
"bbox": [
|
| 832 |
+
0.215,
|
| 833 |
+
0.605,
|
| 834 |
+
0.416,
|
| 835 |
+
0.621
|
| 836 |
+
],
|
| 837 |
+
"angle": 0,
|
| 838 |
+
"content": "4.2 Implement Details"
|
| 839 |
+
},
|
| 840 |
+
{
|
| 841 |
+
"type": "text",
|
| 842 |
+
"bbox": [
|
| 843 |
+
0.214,
|
| 844 |
+
0.629,
|
| 845 |
+
0.784,
|
| 846 |
+
0.72
|
| 847 |
+
],
|
| 848 |
+
"angle": 0,
|
| 849 |
+
"content": "In our experiments on CIFAR-10 [21], SVHN [30], and STL-10 [9] datasets with ResNet-32 [15] and ResNet-110 [15], we utilize the SGD optimizer with Nesterov momentum set at 0.9 and an L2 weight decay factor of 1e-4. We employ batch sizes of 1024 for CIFAR-10 and SVHN and 128 for STL-10. The training duration spans 400 epochs, starting with initial learning rates of 0.8 for CIFAR-10 / SVHN and 0.1 for STL-10, following a cosine annealing scheduler [9]."
|
| 850 |
+
},
|
| 851 |
+
{
|
| 852 |
+
"type": "text",
|
| 853 |
+
"bbox": [
|
| 854 |
+
0.214,
|
| 855 |
+
0.72,
|
| 856 |
+
0.785,
|
| 857 |
+
0.84
|
| 858 |
+
],
|
| 859 |
+
"angle": 0,
|
| 860 |
+
"content": "In our experiments with ImageNet [12], we adopt different training settings for various architectures. We train VGG13 [36] for 90 epochs with an initial learning rate of 0.025. For ResNet-101 [15], ResNet-152 [15], and ResNeXt-101,32 \\(\\times\\) 8d [39], we train them for 90 epochs as well, with initial learning rates of 0.05, 0.05, and 0.025, respectively. We set the batch sizes to 64 for VGG13, 128 for ResNet-101 and ResNet-152, and 64 for ResNeXt-101,32 \\(\\times\\) 8d. We maintain consistency with other training configurations as previously described for CIFAR-10 [21]."
|
| 861 |
+
}
|
| 862 |
+
],
|
| 863 |
+
[
|
| 864 |
+
{
|
| 865 |
+
"type": "page_number",
|
| 866 |
+
"bbox": [
|
| 867 |
+
0.218,
|
| 868 |
+
0.116,
|
| 869 |
+
0.23,
|
| 870 |
+
0.127
|
| 871 |
+
],
|
| 872 |
+
"angle": 0,
|
| 873 |
+
"content": "8"
|
| 874 |
+
},
|
| 875 |
+
{
|
| 876 |
+
"type": "header",
|
| 877 |
+
"bbox": [
|
| 878 |
+
0.272,
|
| 879 |
+
0.115,
|
| 880 |
+
0.383,
|
| 881 |
+
0.128
|
| 882 |
+
],
|
| 883 |
+
"angle": 0,
|
| 884 |
+
"content": "Junhao Su et al."
|
| 885 |
+
},
|
| 886 |
+
{
|
| 887 |
+
"type": "text",
|
| 888 |
+
"bbox": [
|
| 889 |
+
0.214,
|
| 890 |
+
0.147,
|
| 891 |
+
0.788,
|
| 892 |
+
0.208
|
| 893 |
+
],
|
| 894 |
+
"angle": 0,
|
| 895 |
+
"content": "Based on our multiple experimental results, we select the hyperparameter momentum value as 0.995, due to its consistently stable and superior performance. We will provide the experimental results of other momentum hyperparameters in the supplementary materials."
|
| 896 |
+
},
|
| 897 |
+
{
|
| 898 |
+
"type": "title",
|
| 899 |
+
"bbox": [
|
| 900 |
+
0.216,
|
| 901 |
+
0.23,
|
| 902 |
+
0.599,
|
| 903 |
+
0.246
|
| 904 |
+
],
|
| 905 |
+
"angle": 0,
|
| 906 |
+
"content": "4.3 Results on Image Classification Datasets"
|
| 907 |
+
},
|
| 908 |
+
{
|
| 909 |
+
"type": "text",
|
| 910 |
+
"bbox": [
|
| 911 |
+
0.214,
|
| 912 |
+
0.255,
|
| 913 |
+
0.788,
|
| 914 |
+
0.332
|
| 915 |
+
],
|
| 916 |
+
"angle": 0,
|
| 917 |
+
"content": "Results on Image Classification Benchmarks: We start by assessing the accuracy performance of our approach using the CIFAR-10 [21], SVHN [30], and STL-10 [9] datasets. We employ ResNet-32 [15], partitioned into 8 and 16 local blocks, and ResNet-110 [15], divided into 32 and 55 local blocks. As illustrated in Table 1, our MAN significantly bolsters the accuracy of all methods."
|
| 918 |
+
},
|
| 919 |
+
{
|
| 920 |
+
"type": "table_caption",
|
| 921 |
+
"bbox": [
|
| 922 |
+
0.216,
|
| 923 |
+
0.358,
|
| 924 |
+
0.787,
|
| 925 |
+
0.389
|
| 926 |
+
],
|
| 927 |
+
"angle": 0,
|
| 928 |
+
"content": "Table 1: Performance of different networks with varying numbers of local blocks. The average test error is obtained by 5 experiments. The * means addition of our MAN."
|
| 929 |
+
},
|
| 930 |
+
{
|
| 931 |
+
"type": "table",
|
| 932 |
+
"bbox": [
|
| 933 |
+
0.221,
|
| 934 |
+
0.399,
|
| 935 |
+
0.789,
|
| 936 |
+
0.598
|
| 937 |
+
],
|
| 938 |
+
"angle": 0,
|
| 939 |
+
"content": "<table><tr><td rowspan=\"2\">Dataset</td><td rowspan=\"2\">Method</td><td colspan=\"2\">ResNet-32</td><td colspan=\"2\">ResNet-110</td></tr><tr><td>K = 8 (Test Error)</td><td>K = 16 (Test Error)</td><td>K = 32 (Test Error)</td><td>K = 55 (Test Error)</td></tr><tr><td rowspan=\"6\">CIFAR-10(E2E(ResNet-32)=6.37,E2E(ResNet-110)=5.42)</td><td>PredSim [32]</td><td>20.62</td><td>22.71</td><td>22.08</td><td>24.74</td></tr><tr><td>PredSim*</td><td>14.29(↓6.33)</td><td>15.58(↓7.13)</td><td>17.05(↓5.03)</td><td>18.02(↓6.72)</td></tr><tr><td>DGL [3]</td><td>11.63</td><td>14.08</td><td>12.51</td><td>14.45</td></tr><tr><td>DGL*</td><td>8.42(↓3.21)</td><td>9.11(↓4.97)</td><td>9.65(↓2.86)</td><td>9.73(↓4.72)</td></tr><tr><td>InfoPro [37]</td><td>11.51</td><td>12.93</td><td>12.26</td><td>13.22</td></tr><tr><td>InfoPro*</td><td>9.32(↓2.19)</td><td>9.65(↓3.28)</td><td>9.06(↓3.20)</td><td>9.77(↓3.45)</td></tr><tr><td rowspan=\"6\">STL-10(E2E(ResNet-32)=19.35,E2E(ResNet-110)=19.67)</td><td>PredSim [32]</td><td>31.97</td><td>32.90</td><td>32.05</td><td>33.27</td></tr><tr><td>PredSim*</td><td>29.97(↓2.00)</td><td>29.99(↓2.91)</td><td>30.48(↓1.57)</td><td>31.55(↓1.72)</td></tr><tr><td>DGL [3]</td><td>25.05</td><td>27.14</td><td>25.67</td><td>28.16</td></tr><tr><td>DGL*</td><td>20.74(↓4.31)</td><td>21.37(↓5.77)</td><td>22.54(↓3.13)</td><td>22.69(↓5.47)</td></tr><tr><td>InfoPro [37]</td><td>27.32</td><td>29.28</td><td>28.58</td><td>29.20</td></tr><tr><td>InfoPro*</td><td>23.17(↓4.15)</td><td>23.54(↓5.74)</td><td>24.08(↓4.50)</td><td>24.74(↓4.46)</td></tr><tr><td rowspan=\"6\">SVHN(E2E(ResNet-32)=2.99,E2E(ResNet-110)=2.92)</td><td>PredSim [32]</td><td>6.91</td><td>8.08</td><td>9.12</td><td>10.47</td></tr><tr><td>PredSim*</td><td>5.54(↓1.37)</td><td>6.39(↓1.69)</td><td>7.27(↓1.85)</td><td>8.24(↓2.23)</td></tr><tr><td>DGL [3]</td><td>4.83</td><td>5.05</td><td>5.12</td><td>5.36</td></tr><tr><td>DGL*</td><td>3.80(↓1.03)</td><td>4.04(↓1.01)</td><td>4.08(↓1.04)</td><td>4.52(↓0.84)</td></tr><tr><td>InfoPro [37]</td><td>5.61</td><td>5.97</td><td>5.89</td><td>6.11</td></tr><tr><td>InfoPro*</td><td>4.49(↓1.12)</td><td>5.19(↓0.78)</td><td>4.85(↓1.04)</td><td>4.99(↓1.12)</td></tr></table>"
|
| 940 |
+
},
|
| 941 |
+
{
|
| 942 |
+
"type": "text",
|
| 943 |
+
"bbox": [
|
| 944 |
+
0.214,
|
| 945 |
+
0.629,
|
| 946 |
+
0.788,
|
| 947 |
+
0.795
|
| 948 |
+
],
|
| 949 |
+
"angle": 0,
|
| 950 |
+
"content": "On the CIFAR-10 dataset [21], our method exhibits considerable improvements in diminishing test errors across various methods. In the relatively shallower network of ResNet-32 (K=16), where individual layers function as gradient-isolated local blocks, we record a reduction in test errors for PredSim [32], DGL [3], and InfoPro [37], from 22.71, 14.08, and 12.93 to 15.58, 9.11, and 9.65 respectively. This translates to a performance enhancement exceeding \\(25\\%\\) for all methods. Even though the performance across all methods in the comparatively deeper network, ResNet-110 (K=55), is somewhat inferior due to the inherent need for more global information in such networks, our method still delivers exceptional performance. It achieves approximately a \\(20\\%\\) improvement across all methods, underscoring the robust effectiveness of MAN in deeper networks."
|
| 951 |
+
},
|
| 952 |
+
{
|
| 953 |
+
"type": "text",
|
| 954 |
+
"bbox": [
|
| 955 |
+
0.214,
|
| 956 |
+
0.796,
|
| 957 |
+
0.788,
|
| 958 |
+
0.842
|
| 959 |
+
],
|
| 960 |
+
"angle": 0,
|
| 961 |
+
"content": "When applied to other datasets, MAN can also reduce the test error of Pred-Sim [32], DGL [3], and InfoPro [37] by at least \\(5\\%\\), \\(12\\%\\), and \\(16\\%\\) on the STL-10 [9] dataset. On the SVHN [30] dataset, our improvements over the three"
|
| 962 |
+
}
|
| 963 |
+
],
|
| 964 |
+
[
|
| 965 |
+
{
|
| 966 |
+
"type": "header",
|
| 967 |
+
"bbox": [
|
| 968 |
+
0.323,
|
| 969 |
+
0.115,
|
| 970 |
+
0.733,
|
| 971 |
+
0.13
|
| 972 |
+
],
|
| 973 |
+
"angle": 0,
|
| 974 |
+
"content": "Momentum Auxiliary Network for Supervised Local Learning"
|
| 975 |
+
},
|
| 976 |
+
{
|
| 977 |
+
"type": "page_number",
|
| 978 |
+
"bbox": [
|
| 979 |
+
0.775,
|
| 980 |
+
0.117,
|
| 981 |
+
0.785,
|
| 982 |
+
0.127
|
| 983 |
+
],
|
| 984 |
+
"angle": 0,
|
| 985 |
+
"content": "9"
|
| 986 |
+
},
|
| 987 |
+
{
|
| 988 |
+
"type": "text",
|
| 989 |
+
"bbox": [
|
| 990 |
+
0.214,
|
| 991 |
+
0.147,
|
| 992 |
+
0.784,
|
| 993 |
+
0.208
|
| 994 |
+
],
|
| 995 |
+
"angle": 0,
|
| 996 |
+
"content": "methods also surpass \\(20\\%\\), \\(21\\%\\), and \\(13\\%\\). As can be seen, the improvement our MAN introduces to all methods is quite remarkable—comparable even to the accuracy of E2E training—and it significantly mitigates the overwhelming performance issue that has continually plagued supervised local learning."
|
| 997 |
+
},
|
| 998 |
+
{
|
| 999 |
+
"type": "text",
|
| 1000 |
+
"bbox": [
|
| 1001 |
+
0.214,
|
| 1002 |
+
0.21,
|
| 1003 |
+
0.785,
|
| 1004 |
+
0.331
|
| 1005 |
+
],
|
| 1006 |
+
"angle": 0,
|
| 1007 |
+
"content": "Results on ImageNet: We further validate the effectiveness of our approach on ImageNet [12] using four networks of varying depths (ResNets [15] and VGG13 [36]). As depicted in Table 2, when we employ VGG13 as the backbone and divide the network into 10 blocks, DGL [3] achieves merely a Top1-Error of 35.60 and a Top5-Error of 14.2, representing a substantial gap when compared to the E2E method. However, with the introduction of our MAN, the Top1-Error reduces by 3.61 points, and the Top5-Error decreases by 3.36 points for DGL. This significant enhancement brings the performance closer to the E2E method."
|
| 1008 |
+
},
|
| 1009 |
+
{
|
| 1010 |
+
"type": "text",
|
| 1011 |
+
"bbox": [
|
| 1012 |
+
0.214,
|
| 1013 |
+
0.333,
|
| 1014 |
+
0.787,
|
| 1015 |
+
0.44
|
| 1016 |
+
],
|
| 1017 |
+
"angle": 0,
|
| 1018 |
+
"content": "As illustrated in Table 2, when we use ResNet-101 [15], ResNet-152 [15], ResNeXt-101, \\(32 \\times 8\\mathrm{d}\\) [39] as backbones and divide the network into four blocks, the performance of InfoPro [37] is already below that of E2E. After incorporating our MAN, the Top-1 Error of these three backbone networks can be reduced by approximately \\(6\\%\\) compared to the original, surpassing the performance of E2E training. These results underscore the effectiveness of our MAN on the large-scale ImageNet [12] dataset, even when using deeper networks."
|
| 1019 |
+
},
|
| 1020 |
+
{
|
| 1021 |
+
"type": "table_caption",
|
| 1022 |
+
"bbox": [
|
| 1023 |
+
0.329,
|
| 1024 |
+
0.472,
|
| 1025 |
+
0.674,
|
| 1026 |
+
0.487
|
| 1027 |
+
],
|
| 1028 |
+
"angle": 0,
|
| 1029 |
+
"content": "Table 2: Results on the validation set of ImageNet"
|
| 1030 |
+
},
|
| 1031 |
+
{
|
| 1032 |
+
"type": "table",
|
| 1033 |
+
"bbox": [
|
| 1034 |
+
0.235,
|
| 1035 |
+
0.497,
|
| 1036 |
+
0.768,
|
| 1037 |
+
0.698
|
| 1038 |
+
],
|
| 1039 |
+
"angle": 0,
|
| 1040 |
+
"content": "<table><tr><td>Network</td><td>Method</td><td>Top1-Error</td><td>Top5-Error</td></tr><tr><td rowspan=\"5\">ResNet-101</td><td>E2E</td><td>22.03</td><td>5.93</td></tr><tr><td>InfoPro(K=2) [37]</td><td>21.85</td><td>5.89</td></tr><tr><td>InfoPro*(K=2)</td><td>21.65(↓0.20)</td><td>5.49(↓0.40)</td></tr><tr><td>InfoPro(K=4) [37]</td><td>22.81</td><td>6.54</td></tr><tr><td>InfoPro*(K=4)</td><td>21.73(↓1.08)</td><td>5.81(↓0.73)</td></tr><tr><td rowspan=\"5\">ResNet-152</td><td>E2E</td><td>21.60</td><td>5.92</td></tr><tr><td>InfoPro(K=2) [37]</td><td>21.45</td><td>5.84</td></tr><tr><td>InfoPro*(K=2)</td><td>21.23(↓0.22)</td><td>5.53(↓0.31)</td></tr><tr><td>InfoPro(K=4) [37]</td><td>22.93</td><td>6.71</td></tr><tr><td>InfoPro*(K=4)</td><td>21.59(↓1.34)</td><td>5.89(↓0.82)</td></tr><tr><td rowspan=\"5\">ResNeXt-101, 32 × 8d</td><td>E2E</td><td>20.64</td><td>5.40</td></tr><tr><td>InfoPro(K=2) [37]</td><td>20.35</td><td>5.28</td></tr><tr><td>InfoPro*(K=2)</td><td>20.11(↓0.24)</td><td>5.18(↓0.10)</td></tr><tr><td>InfoPro(K=4) [37]</td><td>21.69</td><td>6.11</td></tr><tr><td>InfoPro*(K=4)</td><td>20.37(↓1.32)</td><td>5.34(↓0.77)</td></tr><tr><td rowspan=\"3\">VGG13</td><td>E2E</td><td>28.41</td><td>9.63</td></tr><tr><td>DGL [3]</td><td>35.60</td><td>14.20</td></tr><tr><td>DGL*(K=10)</td><td>31.99(↓3.61)</td><td>10.84(↓3.36)</td></tr></table>"
|
| 1041 |
+
},
|
| 1042 |
+
{
|
| 1043 |
+
"type": "text",
|
| 1044 |
+
"bbox": [
|
| 1045 |
+
0.214,
|
| 1046 |
+
0.735,
|
| 1047 |
+
0.788,
|
| 1048 |
+
0.841
|
| 1049 |
+
],
|
| 1050 |
+
"angle": 0,
|
| 1051 |
+
"content": "Training-Accuracy Curve Analysis: As depicted in the accuracy-epoch curve in Fig. 4, our Momentum Auxiliary Network consistently outperforms the original method in terms of accuracy throughout the entire training process. This underscores its reliability and stability during the training process of classification tasks. Moreover, our MAN achieves a higher accuracy earlier in the later stages of training and attains stability sooner, indicating a faster convergence rate—a critical attribute in large-scale and complex tasks."
|
| 1052 |
+
}
|
| 1053 |
+
],
|
| 1054 |
+
[
|
| 1055 |
+
{
|
| 1056 |
+
"type": "page_number",
|
| 1057 |
+
"bbox": [
|
| 1058 |
+
0.218,
|
| 1059 |
+
0.116,
|
| 1060 |
+
0.236,
|
| 1061 |
+
0.127
|
| 1062 |
+
],
|
| 1063 |
+
"angle": 0,
|
| 1064 |
+
"content": "10"
|
| 1065 |
+
},
|
| 1066 |
+
{
|
| 1067 |
+
"type": "header",
|
| 1068 |
+
"bbox": [
|
| 1069 |
+
0.272,
|
| 1070 |
+
0.115,
|
| 1071 |
+
0.383,
|
| 1072 |
+
0.127
|
| 1073 |
+
],
|
| 1074 |
+
"angle": 0,
|
| 1075 |
+
"content": "Junhao Su et al."
|
| 1076 |
+
},
|
| 1077 |
+
{
|
| 1078 |
+
"type": "image",
|
| 1079 |
+
"bbox": [
|
| 1080 |
+
0.231,
|
| 1081 |
+
0.162,
|
| 1082 |
+
0.481,
|
| 1083 |
+
0.309
|
| 1084 |
+
],
|
| 1085 |
+
"angle": 0,
|
| 1086 |
+
"content": null
|
| 1087 |
+
},
|
| 1088 |
+
{
|
| 1089 |
+
"type": "image",
|
| 1090 |
+
"bbox": [
|
| 1091 |
+
0.513,
|
| 1092 |
+
0.162,
|
| 1093 |
+
0.761,
|
| 1094 |
+
0.308
|
| 1095 |
+
],
|
| 1096 |
+
"angle": 0,
|
| 1097 |
+
"content": null
|
| 1098 |
+
},
|
| 1099 |
+
{
|
| 1100 |
+
"type": "image_caption",
|
| 1101 |
+
"bbox": [
|
| 1102 |
+
0.215,
|
| 1103 |
+
0.321,
|
| 1104 |
+
0.788,
|
| 1105 |
+
0.351
|
| 1106 |
+
],
|
| 1107 |
+
"angle": 0,
|
| 1108 |
+
"content": "Fig. 4: Training-Accuracy curves, the left uses ResNet-32 (K=16) as the backbone, while the right uses ResNet-110 (K=55). Both are utilizing the CIFAR-10 dataset."
|
| 1109 |
+
},
|
| 1110 |
+
{
|
| 1111 |
+
"type": "text",
|
| 1112 |
+
"bbox": [
|
| 1113 |
+
0.214,
|
| 1114 |
+
0.379,
|
| 1115 |
+
0.789,
|
| 1116 |
+
0.487
|
| 1117 |
+
],
|
| 1118 |
+
"angle": 0,
|
| 1119 |
+
"content": "In summary, our Momentum Auxiliary Network significantly enhances the accuracy of traditional supervised local learning methods by promoting information exchange between local blocks. The beneficial information exchange facilitated by Momentum Auxiliary Network significantly mitigates the pervasive issue of shortsightedness within the supervised local learning domain, thereby offering advantages in terms of training stability and convergence speed. This is crucial for training models more effectively and efficiently."
|
| 1120 |
+
},
|
| 1121 |
+
{
|
| 1122 |
+
"type": "table_caption",
|
| 1123 |
+
"bbox": [
|
| 1124 |
+
0.215,
|
| 1125 |
+
0.512,
|
| 1126 |
+
0.789,
|
| 1127 |
+
0.542
|
| 1128 |
+
],
|
| 1129 |
+
"angle": 0,
|
| 1130 |
+
"content": "Table 3: Comparison of GPU memory usage using InfoPro as baseline with different backbones on the ImageNet dataset."
|
| 1131 |
+
},
|
| 1132 |
+
{
|
| 1133 |
+
"type": "table",
|
| 1134 |
+
"bbox": [
|
| 1135 |
+
0.238,
|
| 1136 |
+
0.553,
|
| 1137 |
+
0.763,
|
| 1138 |
+
0.674
|
| 1139 |
+
],
|
| 1140 |
+
"angle": 0,
|
| 1141 |
+
"content": "<table><tr><td>Method</td><td>ResNet-101\nGPU Memory(GB)</td><td>ResNet-152\nGPU Memory(GB)</td><td>ResNeXt-101, 32×8d\nGPU Memory(GB)</td></tr><tr><td>E2E</td><td>19.71</td><td>26.29</td><td>19.22</td></tr><tr><td>InfoPro(K=2) [37]</td><td>12.06(↓ 38.8%)</td><td>15.53(↓ 40.9%)</td><td>11.55(↓ 39.9%)</td></tr><tr><td>InfoPro*(K=2)</td><td>12.32(↓ 37.5%)</td><td>15.93(↓ 39.4%)</td><td>11.73(↓ 38.9%)</td></tr><tr><td>InfoPro(K=4) [37]</td><td>10.37(↓ 47.3%)</td><td>13.48(↓ 48.7%)</td><td>10.24(↓ 46.7%)</td></tr><tr><td>InfoPro*(K=4)</td><td>10.69(↓ 45.8%)</td><td>13.91(↓ 47.1%)</td><td>10.49(↓ 45.5%)</td></tr></table>"
|
| 1142 |
+
},
|
| 1143 |
+
{
|
| 1144 |
+
"type": "text",
|
| 1145 |
+
"bbox": [
|
| 1146 |
+
0.214,
|
| 1147 |
+
0.704,
|
| 1148 |
+
0.789,
|
| 1149 |
+
0.842
|
| 1150 |
+
],
|
| 1151 |
+
"angle": 0,
|
| 1152 |
+
"content": "Results on GPU memory requirement: Supervised local learning markedly conserves GPU memory by limiting gradient propagation to within local blocks, thereby ensuring that backpropagation transpires exclusively within each respective local block and its corresponding auxiliary network. This strategy significantly diminishes the storage demands for activation parameters and gradient information, which would typically disseminate across local blocks. Consequently, this facilitates substantial savings in GPU memory utilization. Our analysis, featuring a comprehensive comparison of GPU memory usage on the ImageNet dataset, illustrates that the implementation of a Momentum Auxiliary Network"
|
| 1153 |
+
}
|
| 1154 |
+
],
|
| 1155 |
+
[
|
| 1156 |
+
{
|
| 1157 |
+
"type": "header",
|
| 1158 |
+
"bbox": [
|
| 1159 |
+
0.323,
|
| 1160 |
+
0.115,
|
| 1161 |
+
0.733,
|
| 1162 |
+
0.13
|
| 1163 |
+
],
|
| 1164 |
+
"angle": 0,
|
| 1165 |
+
"content": "Momentum Auxiliary Network for Supervised Local Learning"
|
| 1166 |
+
},
|
| 1167 |
+
{
|
| 1168 |
+
"type": "page_number",
|
| 1169 |
+
"bbox": [
|
| 1170 |
+
0.769,
|
| 1171 |
+
0.117,
|
| 1172 |
+
0.784,
|
| 1173 |
+
0.127
|
| 1174 |
+
],
|
| 1175 |
+
"angle": 0,
|
| 1176 |
+
"content": "11"
|
| 1177 |
+
},
|
| 1178 |
+
{
|
| 1179 |
+
"type": "table_caption",
|
| 1180 |
+
"bbox": [
|
| 1181 |
+
0.216,
|
| 1182 |
+
0.145,
|
| 1183 |
+
0.784,
|
| 1184 |
+
0.187
|
| 1185 |
+
],
|
| 1186 |
+
"angle": 0,
|
| 1187 |
+
"content": "Table 4: Abalation study of MAN. (a) Using DGL as baseline and ResNet-32 \\((\\mathrm{K} = 16)\\) as backbone on the CIFAR-10 dataset. (b) Using InfoPro as baseline and ResNet-101 \\((\\mathrm{K} = 4)\\) as backbone on the ImageNet dataset. LB stands for Learnable Bias."
|
| 1188 |
+
},
|
| 1189 |
+
{
|
| 1190 |
+
"type": "table",
|
| 1191 |
+
"bbox": [
|
| 1192 |
+
0.218,
|
| 1193 |
+
0.198,
|
| 1194 |
+
0.391,
|
| 1195 |
+
0.273
|
| 1196 |
+
],
|
| 1197 |
+
"angle": 0,
|
| 1198 |
+
"content": "<table><tr><td>EMA</td><td>LB</td><td>Test Error</td></tr><tr><td>×</td><td>×</td><td>14.08</td></tr><tr><td>✓</td><td>×</td><td>11.07(↓3.01)</td></tr><tr><td>✓</td><td>✓</td><td>9.11(↓4.97)</td></tr><tr><td colspan=\"3\">(a)</td></tr></table>"
|
| 1199 |
+
},
|
| 1200 |
+
{
|
| 1201 |
+
"type": "table",
|
| 1202 |
+
"bbox": [
|
| 1203 |
+
0.5,
|
| 1204 |
+
0.199,
|
| 1205 |
+
0.765,
|
| 1206 |
+
0.273
|
| 1207 |
+
],
|
| 1208 |
+
"angle": 0,
|
| 1209 |
+
"content": "<table><tr><td>EMA</td><td>LB</td><td>Top1-Error</td><td>Top5-Error</td></tr><tr><td>×</td><td>×</td><td>22.81</td><td>6.54</td></tr><tr><td>✓</td><td>×</td><td>22.09(↓ 0.72)</td><td>6.07(↓ 0.47)</td></tr><tr><td>✓</td><td>✓</td><td>21.73(↓ 1.08)</td><td>5.81(↓ 0.73)</td></tr><tr><td colspan=\"4\">(b)</td></tr></table>"
|
| 1210 |
+
},
|
| 1211 |
+
{
|
| 1212 |
+
"type": "text",
|
| 1213 |
+
"bbox": [
|
| 1214 |
+
0.214,
|
| 1215 |
+
0.303,
|
| 1216 |
+
0.784,
|
| 1217 |
+
0.333
|
| 1218 |
+
],
|
| 1219 |
+
"angle": 0,
|
| 1220 |
+
"content": "can elicit notable performance enhancements while incurring only a minimal uptick in GPU memory consumption."
|
| 1221 |
+
},
|
| 1222 |
+
{
|
| 1223 |
+
"type": "text",
|
| 1224 |
+
"bbox": [
|
| 1225 |
+
0.214,
|
| 1226 |
+
0.334,
|
| 1227 |
+
0.785,
|
| 1228 |
+
0.515
|
| 1229 |
+
],
|
| 1230 |
+
"angle": 0,
|
| 1231 |
+
"content": "As shown in Table 3, it indicates the substantial GPU memory reduction on the ImageNet [12]. When applying our method to InfoPro [37] on ResNet-101 [15], ResNet-152 [15], and ResNeXt-101, \\(32 \\times 8d\\) [39], dividing the backbone into two local blocks results in a GPU memory saving of \\(37.5\\%\\) to \\(39.4\\%\\). When the network is divided into four local blocks, the degree of GPU memory savings is further enhanced, exceeding \\(45\\%\\). Combined with Table 2, it can be seen that we achieve better performance with almost half the GPU memory required for end-to-end training. Further analysis of the memory usage comparison with the original method shows that our approach only increases the GPU memory by about \\(1\\%\\) over the original method, while achieving a performance improvement of over \\(5\\%\\). These results highlight the excellent balance that the MAN method achieves in terms of GPU memory usage and performance enhancement."
|
| 1232 |
+
},
|
| 1233 |
+
{
|
| 1234 |
+
"type": "title",
|
| 1235 |
+
"bbox": [
|
| 1236 |
+
0.216,
|
| 1237 |
+
0.54,
|
| 1238 |
+
0.401,
|
| 1239 |
+
0.554
|
| 1240 |
+
],
|
| 1241 |
+
"angle": 0,
|
| 1242 |
+
"content": "4.4 Ablation Studies"
|
| 1243 |
+
},
|
| 1244 |
+
{
|
| 1245 |
+
"type": "text",
|
| 1246 |
+
"bbox": [
|
| 1247 |
+
0.214,
|
| 1248 |
+
0.567,
|
| 1249 |
+
0.785,
|
| 1250 |
+
0.627
|
| 1251 |
+
],
|
| 1252 |
+
"angle": 0,
|
| 1253 |
+
"content": "We conduct an ablation study on the CIFAR-10 [21] dataset and ImageNet dataset to assess the impact of the EMA method [14] and learnable bias in the MAN on performance. For this analysis, we use ResNet-32 (K=16) [15] as the backbone and the original DGL [3] method as a comparison baseline."
|
| 1254 |
+
},
|
| 1255 |
+
{
|
| 1256 |
+
"type": "text",
|
| 1257 |
+
"bbox": [
|
| 1258 |
+
0.214,
|
| 1259 |
+
0.629,
|
| 1260 |
+
0.785,
|
| 1261 |
+
0.734
|
| 1262 |
+
],
|
| 1263 |
+
"angle": 0,
|
| 1264 |
+
"content": "As shown in Table 4(a), when we only use the EMA [14] in MAN to promote information exchange with subsequent blocks, without adding a learnable bias, the test error decreases from 14.08 to 11.07. When we further add a learnable bias, the test error decreases from 11.07 to 9.11, it is evident that both the EMA method and learnable bias contribute to performance improvement. We hypothesize that the EMA and learnable bias are somewhat complementary in terms of the features they learn. To validate this, we conduct feature visualization."
|
| 1265 |
+
},
|
| 1266 |
+
{
|
| 1267 |
+
"type": "text",
|
| 1268 |
+
"bbox": [
|
| 1269 |
+
0.214,
|
| 1270 |
+
0.735,
|
| 1271 |
+
0.787,
|
| 1272 |
+
0.84
|
| 1273 |
+
],
|
| 1274 |
+
"angle": 0,
|
| 1275 |
+
"content": "As depicted in Fig. 5(a), we can observe that the original method without the addition of MAN learns very limited and chaotic features. After adding the EMA method alone, the features it learns are clearly more characteristic and defined, indicating that it indeed receives some global features from the information in the subsequent block. However, there are still many blurry features, which may be due to the EMA parameter update method [14], causing information imbalance due to feature differences between different local blocks. When we"
|
| 1276 |
+
}
|
| 1277 |
+
],
|
| 1278 |
+
[
|
| 1279 |
+
{
|
| 1280 |
+
"type": "page_number",
|
| 1281 |
+
"bbox": [
|
| 1282 |
+
0.218,
|
| 1283 |
+
0.116,
|
| 1284 |
+
0.236,
|
| 1285 |
+
0.127
|
| 1286 |
+
],
|
| 1287 |
+
"angle": 0,
|
| 1288 |
+
"content": "12"
|
| 1289 |
+
},
|
| 1290 |
+
{
|
| 1291 |
+
"type": "header",
|
| 1292 |
+
"bbox": [
|
| 1293 |
+
0.272,
|
| 1294 |
+
0.115,
|
| 1295 |
+
0.382,
|
| 1296 |
+
0.127
|
| 1297 |
+
],
|
| 1298 |
+
"angle": 0,
|
| 1299 |
+
"content": "Junhao Su et al."
|
| 1300 |
+
},
|
| 1301 |
+
{
|
| 1302 |
+
"type": "text",
|
| 1303 |
+
"bbox": [
|
| 1304 |
+
0.214,
|
| 1305 |
+
0.147,
|
| 1306 |
+
0.788,
|
| 1307 |
+
0.284
|
| 1308 |
+
],
|
| 1309 |
+
"angle": 0,
|
| 1310 |
+
"content": "add learnable bias to the original method alone, the features it learns are more concentrated and clear, indicating that the learnable bias significantly enhances the learning ability of the current local block's hidden layer, but it still lacks some details because it does not receive more global information from the subsequent block. When we add the EMA method and learnable bias simultaneously, the learned features are not only clear and specific, but also more comprehensive in detail, proving that their learning abilities are complementary, and greatly improving the performance of the original method. The feature visualization in Fig. 5 verifies our previous thoughts."
|
| 1311 |
+
},
|
| 1312 |
+
{
|
| 1313 |
+
"type": "image",
|
| 1314 |
+
"bbox": [
|
| 1315 |
+
0.245,
|
| 1316 |
+
0.317,
|
| 1317 |
+
0.337,
|
| 1318 |
+
0.387
|
| 1319 |
+
],
|
| 1320 |
+
"angle": 0,
|
| 1321 |
+
"content": null
|
| 1322 |
+
},
|
| 1323 |
+
{
|
| 1324 |
+
"type": "image_caption",
|
| 1325 |
+
"bbox": [
|
| 1326 |
+
0.281,
|
| 1327 |
+
0.394,
|
| 1328 |
+
0.302,
|
| 1329 |
+
0.408
|
| 1330 |
+
],
|
| 1331 |
+
"angle": 0,
|
| 1332 |
+
"content": "(a)"
|
| 1333 |
+
},
|
| 1334 |
+
{
|
| 1335 |
+
"type": "image",
|
| 1336 |
+
"bbox": [
|
| 1337 |
+
0.384,
|
| 1338 |
+
0.318,
|
| 1339 |
+
0.477,
|
| 1340 |
+
0.387
|
| 1341 |
+
],
|
| 1342 |
+
"angle": 0,
|
| 1343 |
+
"content": null
|
| 1344 |
+
},
|
| 1345 |
+
{
|
| 1346 |
+
"type": "image_caption",
|
| 1347 |
+
"bbox": [
|
| 1348 |
+
0.421,
|
| 1349 |
+
0.394,
|
| 1350 |
+
0.442,
|
| 1351 |
+
0.408
|
| 1352 |
+
],
|
| 1353 |
+
"angle": 0,
|
| 1354 |
+
"content": "(b)"
|
| 1355 |
+
},
|
| 1356 |
+
{
|
| 1357 |
+
"type": "image",
|
| 1358 |
+
"bbox": [
|
| 1359 |
+
0.525,
|
| 1360 |
+
0.318,
|
| 1361 |
+
0.617,
|
| 1362 |
+
0.387
|
| 1363 |
+
],
|
| 1364 |
+
"angle": 0,
|
| 1365 |
+
"content": null
|
| 1366 |
+
},
|
| 1367 |
+
{
|
| 1368 |
+
"type": "image_caption",
|
| 1369 |
+
"bbox": [
|
| 1370 |
+
0.563,
|
| 1371 |
+
0.394,
|
| 1372 |
+
0.583,
|
| 1373 |
+
0.408
|
| 1374 |
+
],
|
| 1375 |
+
"angle": 0,
|
| 1376 |
+
"content": "(c)"
|
| 1377 |
+
},
|
| 1378 |
+
{
|
| 1379 |
+
"type": "image",
|
| 1380 |
+
"bbox": [
|
| 1381 |
+
0.668,
|
| 1382 |
+
0.318,
|
| 1383 |
+
0.757,
|
| 1384 |
+
0.387
|
| 1385 |
+
],
|
| 1386 |
+
"angle": 0,
|
| 1387 |
+
"content": null
|
| 1388 |
+
},
|
| 1389 |
+
{
|
| 1390 |
+
"type": "image_caption",
|
| 1391 |
+
"bbox": [
|
| 1392 |
+
0.702,
|
| 1393 |
+
0.393,
|
| 1394 |
+
0.724,
|
| 1395 |
+
0.408
|
| 1396 |
+
],
|
| 1397 |
+
"angle": 0,
|
| 1398 |
+
"content": "(d)"
|
| 1399 |
+
},
|
| 1400 |
+
{
|
| 1401 |
+
"type": "image_caption",
|
| 1402 |
+
"bbox": [
|
| 1403 |
+
0.214,
|
| 1404 |
+
0.419,
|
| 1405 |
+
0.788,
|
| 1406 |
+
0.462
|
| 1407 |
+
],
|
| 1408 |
+
"angle": 0,
|
| 1409 |
+
"content": "Fig.5: Feature maps comparison of (a) Original Method without MAN, (b) MAN with only EMA, (c) MAN with only learnable bias and (d) MAN with both EMA and learnable bias. The feature map are obtained using ResNet-32 \\((\\mathrm{K} = 16)\\) on CIFAR-10."
|
| 1410 |
+
},
|
| 1411 |
+
{
|
| 1412 |
+
"type": "title",
|
| 1413 |
+
"bbox": [
|
| 1414 |
+
0.215,
|
| 1415 |
+
0.493,
|
| 1416 |
+
0.478,
|
| 1417 |
+
0.508
|
| 1418 |
+
],
|
| 1419 |
+
"angle": 0,
|
| 1420 |
+
"content": "4.5 The Effectiveness of EMA"
|
| 1421 |
+
},
|
| 1422 |
+
{
|
| 1423 |
+
"type": "text",
|
| 1424 |
+
"bbox": [
|
| 1425 |
+
0.214,
|
| 1426 |
+
0.518,
|
| 1427 |
+
0.788,
|
| 1428 |
+
0.671
|
| 1429 |
+
],
|
| 1430 |
+
"angle": 0,
|
| 1431 |
+
"content": "To verify the effectiveness of the EMA update method in MAN, we conduct further ablation experiments. As shown in Table 5, using only the parameters of the next layer in MAN results in very limited improvement. In contrast, employing the EMA method to utilize the parameters of the next layer leads to more significant enhancements. This is because the subsequent local block and the current local block learn different features and have different learning objectives, leading to minimal improvement. However, using the EMA method can gradually promote information exchange between local blocks and reduce parameter fluctuations during training, leading to smoother parameter updates. This beneficial interaction brings significant improvements to MAN."
|
| 1432 |
+
},
|
| 1433 |
+
{
|
| 1434 |
+
"type": "table_caption",
|
| 1435 |
+
"bbox": [
|
| 1436 |
+
0.214,
|
| 1437 |
+
0.696,
|
| 1438 |
+
0.788,
|
| 1439 |
+
0.753
|
| 1440 |
+
],
|
| 1441 |
+
"angle": 0,
|
| 1442 |
+
"content": "Table 5: Abalation study of EMA. (a) Using DGL as baseline and ResNet-32 \\((\\mathrm{K} = 16)\\) as backbone on the CIFAR-10 dataset. (b) Using InfoPro as baseline and ResNet-101 \\((\\mathrm{K} = 4)\\) as backbone on the ImageNet dataset. Parameter signifies whether each local block utilizes the parameters of the adjacent next layer connected to it."
|
| 1443 |
+
},
|
| 1444 |
+
{
|
| 1445 |
+
"type": "table",
|
| 1446 |
+
"bbox": [
|
| 1447 |
+
0.218,
|
| 1448 |
+
0.763,
|
| 1449 |
+
0.411,
|
| 1450 |
+
0.824
|
| 1451 |
+
],
|
| 1452 |
+
"angle": 0,
|
| 1453 |
+
"content": "<table><tr><td>Parameter</td><td>EMA</td><td>Test</td><td>Error</td></tr><tr><td>×</td><td>×</td><td>14.08</td><td></td></tr><tr><td>✓</td><td>×</td><td>12.93</td><td></td></tr><tr><td>✓</td><td>✓</td><td>11.07</td><td></td></tr></table>"
|
| 1454 |
+
},
|
| 1455 |
+
{
|
| 1456 |
+
"type": "table_caption",
|
| 1457 |
+
"bbox": [
|
| 1458 |
+
0.304,
|
| 1459 |
+
0.824,
|
| 1460 |
+
0.323,
|
| 1461 |
+
0.836
|
| 1462 |
+
],
|
| 1463 |
+
"angle": 0,
|
| 1464 |
+
"content": "(a)"
|
| 1465 |
+
},
|
| 1466 |
+
{
|
| 1467 |
+
"type": "table",
|
| 1468 |
+
"bbox": [
|
| 1469 |
+
0.5,
|
| 1470 |
+
0.763,
|
| 1471 |
+
0.774,
|
| 1472 |
+
0.824
|
| 1473 |
+
],
|
| 1474 |
+
"angle": 0,
|
| 1475 |
+
"content": "<table><tr><td>Parameter</td><td>EMA</td><td>Top1-Error</td><td>Top5-Error</td></tr><tr><td>×</td><td>×</td><td>22.81</td><td>6.54</td></tr><tr><td>✓</td><td>×</td><td>22.64</td><td>6.41</td></tr><tr><td>✓</td><td>✓</td><td>22.09</td><td>6.07</td></tr></table>"
|
| 1476 |
+
},
|
| 1477 |
+
{
|
| 1478 |
+
"type": "table_caption",
|
| 1479 |
+
"bbox": [
|
| 1480 |
+
0.626,
|
| 1481 |
+
0.824,
|
| 1482 |
+
0.647,
|
| 1483 |
+
0.836
|
| 1484 |
+
],
|
| 1485 |
+
"angle": 0,
|
| 1486 |
+
"content": "(b)"
|
| 1487 |
+
}
|
| 1488 |
+
],
|
| 1489 |
+
[
|
| 1490 |
+
{
|
| 1491 |
+
"type": "header",
|
| 1492 |
+
"bbox": [
|
| 1493 |
+
0.323,
|
| 1494 |
+
0.115,
|
| 1495 |
+
0.732,
|
| 1496 |
+
0.129
|
| 1497 |
+
],
|
| 1498 |
+
"angle": 0,
|
| 1499 |
+
"content": "Momentum Auxiliary Network for Supervised Local Learning"
|
| 1500 |
+
},
|
| 1501 |
+
{
|
| 1502 |
+
"type": "page_number",
|
| 1503 |
+
"bbox": [
|
| 1504 |
+
0.769,
|
| 1505 |
+
0.117,
|
| 1506 |
+
0.785,
|
| 1507 |
+
0.127
|
| 1508 |
+
],
|
| 1509 |
+
"angle": 0,
|
| 1510 |
+
"content": "13"
|
| 1511 |
+
},
|
| 1512 |
+
{
|
| 1513 |
+
"type": "title",
|
| 1514 |
+
"bbox": [
|
| 1515 |
+
0.217,
|
| 1516 |
+
0.147,
|
| 1517 |
+
0.495,
|
| 1518 |
+
0.163
|
| 1519 |
+
],
|
| 1520 |
+
"angle": 0,
|
| 1521 |
+
"content": "4.6 Linear Separability Analysis"
|
| 1522 |
+
},
|
| 1523 |
+
{
|
| 1524 |
+
"type": "text",
|
| 1525 |
+
"bbox": [
|
| 1526 |
+
0.214,
|
| 1527 |
+
0.172,
|
| 1528 |
+
0.788,
|
| 1529 |
+
0.369
|
| 1530 |
+
],
|
| 1531 |
+
"angle": 0,
|
| 1532 |
+
"content": "To demonstrate the effectiveness of our Momentum Auxiliary Network, we freeze the parameters of the main network and train a classifier for each local block to obtain the classification accuracy of each local block. As depicted in Fig. 6, we use DGL [3] as a baseline, and after integrating our method, the accuracy of the earlier layers decreases, while the accuracy of the middle and later layers significantly improve. The decrease in accuracy of the earlier layers suggests that they have learned more generalized features. Although these features are not beneficial for optimizing local objectives, they are beneficial from a global perspective. The middle and later layers, receiving these globally beneficial features, have seen a significant increase in accuracy. These results illustrate that our proposed method has facilitated information interaction between gradient-isolated local blocks. This addresses the shortsightedness issue existing in the current supervised local learning field and significantly enhances their performance."
|
| 1533 |
+
},
|
| 1534 |
+
{
|
| 1535 |
+
"type": "image",
|
| 1536 |
+
"bbox": [
|
| 1537 |
+
0.248,
|
| 1538 |
+
0.396,
|
| 1539 |
+
0.499,
|
| 1540 |
+
0.561
|
| 1541 |
+
],
|
| 1542 |
+
"angle": 0,
|
| 1543 |
+
"content": null
|
| 1544 |
+
},
|
| 1545 |
+
{
|
| 1546 |
+
"type": "image",
|
| 1547 |
+
"bbox": [
|
| 1548 |
+
0.506,
|
| 1549 |
+
0.397,
|
| 1550 |
+
0.756,
|
| 1551 |
+
0.561
|
| 1552 |
+
],
|
| 1553 |
+
"angle": 0,
|
| 1554 |
+
"content": null
|
| 1555 |
+
},
|
| 1556 |
+
{
|
| 1557 |
+
"type": "image_caption",
|
| 1558 |
+
"bbox": [
|
| 1559 |
+
0.214,
|
| 1560 |
+
0.573,
|
| 1561 |
+
0.788,
|
| 1562 |
+
0.601
|
| 1563 |
+
],
|
| 1564 |
+
"angle": 0,
|
| 1565 |
+
"content": "Fig. 6: Comparison of layer-wise linear separability across different learning rules on ResNet-32 and ResNet-110."
|
| 1566 |
+
},
|
| 1567 |
+
{
|
| 1568 |
+
"type": "title",
|
| 1569 |
+
"bbox": [
|
| 1570 |
+
0.216,
|
| 1571 |
+
0.634,
|
| 1572 |
+
0.55,
|
| 1573 |
+
0.65
|
| 1574 |
+
],
|
| 1575 |
+
"angle": 0,
|
| 1576 |
+
"content": "4.7 Representation Similarity Analysis"
|
| 1577 |
+
},
|
| 1578 |
+
{
|
| 1579 |
+
"type": "text",
|
| 1580 |
+
"bbox": [
|
| 1581 |
+
0.214,
|
| 1582 |
+
0.659,
|
| 1583 |
+
0.788,
|
| 1584 |
+
0.841
|
| 1585 |
+
],
|
| 1586 |
+
"angle": 0,
|
| 1587 |
+
"content": "To further validate the efficacy of MAN, we conduct a Centered Kernel Alignment (CKA) experiment [20]. CKA serves as a metric to assess the similarity between feature representations. If a method's CKA score is closer to 1 in relation to the E2E training method, it indicates that the method's feature learning process is more aligned with that of E2E training. As depicted in Fig. 7, we use DGL [3] as a baseline and incorporate our method. It can be observed that whether adding the EMA method or learnable bias alone, the CKA scores significantly improve compared to the original method. When both the EMA and learnable bias are used in MAN, the CKA score further improves and performs more stably. Notably, the most significant increase in the CKA score occurs in the early and late layers. In conjunction with our previous analysis of the linear separability experiment, this is because if the early layers' learning method is closer"
|
| 1588 |
+
}
|
| 1589 |
+
],
|
| 1590 |
+
[
|
| 1591 |
+
{
|
| 1592 |
+
"type": "page_number",
|
| 1593 |
+
"bbox": [
|
| 1594 |
+
0.218,
|
| 1595 |
+
0.116,
|
| 1596 |
+
0.236,
|
| 1597 |
+
0.127
|
| 1598 |
+
],
|
| 1599 |
+
"angle": 0,
|
| 1600 |
+
"content": "14"
|
| 1601 |
+
},
|
| 1602 |
+
{
|
| 1603 |
+
"type": "header",
|
| 1604 |
+
"bbox": [
|
| 1605 |
+
0.272,
|
| 1606 |
+
0.115,
|
| 1607 |
+
0.382,
|
| 1608 |
+
0.127
|
| 1609 |
+
],
|
| 1610 |
+
"angle": 0,
|
| 1611 |
+
"content": "Junhao Su et al."
|
| 1612 |
+
},
|
| 1613 |
+
{
|
| 1614 |
+
"type": "text",
|
| 1615 |
+
"bbox": [
|
| 1616 |
+
0.214,
|
| 1617 |
+
0.147,
|
| 1618 |
+
0.788,
|
| 1619 |
+
0.253
|
| 1620 |
+
],
|
| 1621 |
+
"angle": 0,
|
| 1622 |
+
"content": "to E2E training, they will focus more on learning general features to optimize the global objective, rather than focusing narrowly on local objectives. While this may result in poorer classification capabilities in the early layers, it greatly contributes to the overall performance improvement of the network. Through the analysis of images and experimental results, it can be demonstrated that MAN enables information interaction between gradient-isolated local blocks, solving the myopia problem present in current supervised local learning methods."
|
| 1623 |
+
},
|
| 1624 |
+
{
|
| 1625 |
+
"type": "image",
|
| 1626 |
+
"bbox": [
|
| 1627 |
+
0.248,
|
| 1628 |
+
0.281,
|
| 1629 |
+
0.501,
|
| 1630 |
+
0.446
|
| 1631 |
+
],
|
| 1632 |
+
"angle": 0,
|
| 1633 |
+
"content": null
|
| 1634 |
+
},
|
| 1635 |
+
{
|
| 1636 |
+
"type": "image",
|
| 1637 |
+
"bbox": [
|
| 1638 |
+
0.503,
|
| 1639 |
+
0.281,
|
| 1640 |
+
0.758,
|
| 1641 |
+
0.446
|
| 1642 |
+
],
|
| 1643 |
+
"angle": 0,
|
| 1644 |
+
"content": null
|
| 1645 |
+
},
|
| 1646 |
+
{
|
| 1647 |
+
"type": "image_caption",
|
| 1648 |
+
"bbox": [
|
| 1649 |
+
0.214,
|
| 1650 |
+
0.458,
|
| 1651 |
+
0.788,
|
| 1652 |
+
0.501
|
| 1653 |
+
],
|
| 1654 |
+
"angle": 0,
|
| 1655 |
+
"content": "Fig. 7: Assessment of Similarity in Layer-wise Representations. We use Centered Kernel Alignment (CKA) [20] to quantify the degree of similarity in the layer-wise representations between the E2E backpropagation (BP) and our proposed MAN."
|
| 1656 |
+
},
|
| 1657 |
+
{
|
| 1658 |
+
"type": "title",
|
| 1659 |
+
"bbox": [
|
| 1660 |
+
0.216,
|
| 1661 |
+
0.535,
|
| 1662 |
+
0.36,
|
| 1663 |
+
0.55
|
| 1664 |
+
],
|
| 1665 |
+
"angle": 0,
|
| 1666 |
+
"content": "5 Conclusion"
|
| 1667 |
+
},
|
| 1668 |
+
{
|
| 1669 |
+
"type": "text",
|
| 1670 |
+
"bbox": [
|
| 1671 |
+
0.214,
|
| 1672 |
+
0.568,
|
| 1673 |
+
0.788,
|
| 1674 |
+
0.749
|
| 1675 |
+
],
|
| 1676 |
+
"angle": 0,
|
| 1677 |
+
"content": "This study addresses the performance disparities between traditional supervised local learning and end-to-end (E2E) methods in deep learning. We introduce a versatile Momentum Auxiliary Network to tackle the short-sighted problem in the early optimization work related to supervised local learning, facilitating information exchange between gradient-isolated local blocks. We integrate our Momentum Auxiliary Network into three advanced supervised local learning approaches and evaluate their performance across network architectures with varying depths on four widely adopted datasets. The results demonstrate our method's ability to significantly enhance the ultimate output performance of original supervised local learning methods. Particularly when combined with InfoPro [37], our method significantly reduces GPU memory usage while consistently maintaining performance levels closely aligned with E2E approaches."
|
| 1678 |
+
},
|
| 1679 |
+
{
|
| 1680 |
+
"type": "text",
|
| 1681 |
+
"bbox": [
|
| 1682 |
+
0.214,
|
| 1683 |
+
0.75,
|
| 1684 |
+
0.788,
|
| 1685 |
+
0.84
|
| 1686 |
+
],
|
| 1687 |
+
"angle": 0,
|
| 1688 |
+
"content": "Limitations and future works: Despite the superior performance on large-scale problems like ImageNet [12], our method still performs less accurately than E2E on some conventional image classification datasets. This may be due to the MAN using too few information interaction layers when the network is divided into a larger number of local blocks. In future work, we could explore deepening these information interaction layers to achieve better precision performance."
|
| 1689 |
+
}
|
| 1690 |
+
],
|
| 1691 |
+
[
|
| 1692 |
+
{
|
| 1693 |
+
"type": "header",
|
| 1694 |
+
"bbox": [
|
| 1695 |
+
0.323,
|
| 1696 |
+
0.115,
|
| 1697 |
+
0.732,
|
| 1698 |
+
0.129
|
| 1699 |
+
],
|
| 1700 |
+
"angle": 0,
|
| 1701 |
+
"content": "Momentum Auxiliary Network for Supervised Local Learning"
|
| 1702 |
+
},
|
| 1703 |
+
{
|
| 1704 |
+
"type": "page_number",
|
| 1705 |
+
"bbox": [
|
| 1706 |
+
0.769,
|
| 1707 |
+
0.117,
|
| 1708 |
+
0.785,
|
| 1709 |
+
0.127
|
| 1710 |
+
],
|
| 1711 |
+
"angle": 0,
|
| 1712 |
+
"content": "15"
|
| 1713 |
+
},
|
| 1714 |
+
{
|
| 1715 |
+
"type": "title",
|
| 1716 |
+
"bbox": [
|
| 1717 |
+
0.218,
|
| 1718 |
+
0.145,
|
| 1719 |
+
0.393,
|
| 1720 |
+
0.163
|
| 1721 |
+
],
|
| 1722 |
+
"angle": 0,
|
| 1723 |
+
"content": "Acknowledgement"
|
| 1724 |
+
},
|
| 1725 |
+
{
|
| 1726 |
+
"type": "text",
|
| 1727 |
+
"bbox": [
|
| 1728 |
+
0.217,
|
| 1729 |
+
0.176,
|
| 1730 |
+
0.785,
|
| 1731 |
+
0.236
|
| 1732 |
+
],
|
| 1733 |
+
"angle": 0,
|
| 1734 |
+
"content": "This work was supported by the National Natural Science Foundation of China (52278154), the Natural Science Foundation of Jiangsu (BK20231429), the Fundamental Research Funds for the Central Universities (2242024RCB0008), and assist from the program of Zhishan Young Scholar of Southeast University."
|
| 1735 |
+
},
|
| 1736 |
+
{
|
| 1737 |
+
"type": "title",
|
| 1738 |
+
"bbox": [
|
| 1739 |
+
0.218,
|
| 1740 |
+
0.259,
|
| 1741 |
+
0.323,
|
| 1742 |
+
0.274
|
| 1743 |
+
],
|
| 1744 |
+
"angle": 0,
|
| 1745 |
+
"content": "References"
|
| 1746 |
+
},
|
| 1747 |
+
{
|
| 1748 |
+
"type": "ref_text",
|
| 1749 |
+
"bbox": [
|
| 1750 |
+
0.226,
|
| 1751 |
+
0.289,
|
| 1752 |
+
0.786,
|
| 1753 |
+
0.331
|
| 1754 |
+
],
|
| 1755 |
+
"angle": 0,
|
| 1756 |
+
"content": "1. Akrout, M., Wilson, C., Humphreys, P., Lillicrap, T., Tweed, D.B.: Deep learning without weight transport. Advances in neural information processing systems 32 (2019)"
|
| 1757 |
+
},
|
| 1758 |
+
{
|
| 1759 |
+
"type": "ref_text",
|
| 1760 |
+
"bbox": [
|
| 1761 |
+
0.226,
|
| 1762 |
+
0.332,
|
| 1763 |
+
0.786,
|
| 1764 |
+
0.372
|
| 1765 |
+
],
|
| 1766 |
+
"angle": 0,
|
| 1767 |
+
"content": "2. Belilovsky, E., Eickenberg, M., Oyallon, E.: Greedy layerwise learning can scale to imagenet. In: International conference on machine learning. pp. 583-593. PMLR (2019)"
|
| 1768 |
+
},
|
| 1769 |
+
{
|
| 1770 |
+
"type": "ref_text",
|
| 1771 |
+
"bbox": [
|
| 1772 |
+
0.226,
|
| 1773 |
+
0.373,
|
| 1774 |
+
0.786,
|
| 1775 |
+
0.4
|
| 1776 |
+
],
|
| 1777 |
+
"angle": 0,
|
| 1778 |
+
"content": "3. Belilovsky, E., Eickenberg, M., Oyallon, E.: Decoupled greedy learning of cnns. In: International Conference on Machine Learning. pp. 736-745. PMLR (2020)"
|
| 1779 |
+
},
|
| 1780 |
+
{
|
| 1781 |
+
"type": "ref_text",
|
| 1782 |
+
"bbox": [
|
| 1783 |
+
0.226,
|
| 1784 |
+
0.401,
|
| 1785 |
+
0.786,
|
| 1786 |
+
0.428
|
| 1787 |
+
],
|
| 1788 |
+
"angle": 0,
|
| 1789 |
+
"content": "4. Bengio, Y.: How auto-encoders could provide credit assignment in deep networks via target propagation. arXiv preprint arXiv:1407.7906 (2014)"
|
| 1790 |
+
},
|
| 1791 |
+
{
|
| 1792 |
+
"type": "ref_text",
|
| 1793 |
+
"bbox": [
|
| 1794 |
+
0.226,
|
| 1795 |
+
0.429,
|
| 1796 |
+
0.786,
|
| 1797 |
+
0.455
|
| 1798 |
+
],
|
| 1799 |
+
"angle": 0,
|
| 1800 |
+
"content": "5. Bengio, Y., Lamblin, P., Popovici, D., Larochelle, H.: Greedy layer-wise training of deep networks. Advances in neural information processing systems 19 (2006)"
|
| 1801 |
+
},
|
| 1802 |
+
{
|
| 1803 |
+
"type": "ref_text",
|
| 1804 |
+
"bbox": [
|
| 1805 |
+
0.226,
|
| 1806 |
+
0.457,
|
| 1807 |
+
0.786,
|
| 1808 |
+
0.482
|
| 1809 |
+
],
|
| 1810 |
+
"angle": 0,
|
| 1811 |
+
"content": "6. Bengio, Y., Lee, D.H., Bornschein, J., Mesnard, T., Lin, Z.: Towards biologically plausible deep learning. arXiv preprint arXiv:1502.04156 (2015)"
|
| 1812 |
+
},
|
| 1813 |
+
{
|
| 1814 |
+
"type": "ref_text",
|
| 1815 |
+
"bbox": [
|
| 1816 |
+
0.226,
|
| 1817 |
+
0.484,
|
| 1818 |
+
0.786,
|
| 1819 |
+
0.51
|
| 1820 |
+
],
|
| 1821 |
+
"angle": 0,
|
| 1822 |
+
"content": "7. Caporale, N., Dan, Y.: Spike timing-dependent plasticity: a hebbian learning rule. Annu. Rev. Neurosci. 31, 25-46 (2008)"
|
| 1823 |
+
},
|
| 1824 |
+
{
|
| 1825 |
+
"type": "ref_text",
|
| 1826 |
+
"bbox": [
|
| 1827 |
+
0.226,
|
| 1828 |
+
0.511,
|
| 1829 |
+
0.786,
|
| 1830 |
+
0.551
|
| 1831 |
+
],
|
| 1832 |
+
"angle": 0,
|
| 1833 |
+
"content": "8. Clark, D., Abbott, L., Chung, S.: Credit assignment through broadcasting a global error vector. Advances in Neural Information Processing Systems 34, 10053-10066 (2021)"
|
| 1834 |
+
},
|
| 1835 |
+
{
|
| 1836 |
+
"type": "ref_text",
|
| 1837 |
+
"bbox": [
|
| 1838 |
+
0.226,
|
| 1839 |
+
0.552,
|
| 1840 |
+
0.786,
|
| 1841 |
+
0.607
|
| 1842 |
+
],
|
| 1843 |
+
"angle": 0,
|
| 1844 |
+
"content": "9. Coates, A., Ng, A., Lee, H.: An analysis of single-layer networks in unsupervised feature learning. In: Proceedings of the fourteenth international conference on artificial intelligence and statistics. pp. 215-223. JMLR Workshop and Conference Proceedings (2011)"
|
| 1845 |
+
},
|
| 1846 |
+
{
|
| 1847 |
+
"type": "ref_text",
|
| 1848 |
+
"bbox": [
|
| 1849 |
+
0.226,
|
| 1850 |
+
0.608,
|
| 1851 |
+
0.786,
|
| 1852 |
+
0.634
|
| 1853 |
+
],
|
| 1854 |
+
"angle": 0,
|
| 1855 |
+
"content": "0. Crick, F.: The recent excitement about neural networks. Nature 337(6203), 129-132 (1989)"
|
| 1856 |
+
},
|
| 1857 |
+
{
|
| 1858 |
+
"type": "ref_text",
|
| 1859 |
+
"bbox": [
|
| 1860 |
+
0.226,
|
| 1861 |
+
0.635,
|
| 1862 |
+
0.786,
|
| 1863 |
+
0.675
|
| 1864 |
+
],
|
| 1865 |
+
"angle": 0,
|
| 1866 |
+
"content": "1. Dellafererra, G., Kreiman, G.: Error-driven input modulation: solving the credit assignment problem without a backward pass. In: International Conference on Machine Learning. pp. 4937-4955. PMLR (2022)"
|
| 1867 |
+
},
|
| 1868 |
+
{
|
| 1869 |
+
"type": "ref_text",
|
| 1870 |
+
"bbox": [
|
| 1871 |
+
0.226,
|
| 1872 |
+
0.676,
|
| 1873 |
+
0.786,
|
| 1874 |
+
0.716
|
| 1875 |
+
],
|
| 1876 |
+
"angle": 0,
|
| 1877 |
+
"content": "2. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: 2009 IEEE conference on computer vision and pattern recognition. pp. 248-255. IEEE (2009)"
|
| 1878 |
+
},
|
| 1879 |
+
{
|
| 1880 |
+
"type": "ref_text",
|
| 1881 |
+
"bbox": [
|
| 1882 |
+
0.226,
|
| 1883 |
+
0.717,
|
| 1884 |
+
0.786,
|
| 1885 |
+
0.758
|
| 1886 |
+
],
|
| 1887 |
+
"angle": 0,
|
| 1888 |
+
"content": "3. Gomez, A.N., Key, O., Perlin, K., Gou, S., Frosst, N., Dean, J., Gal, Y.: Interlocking backpropagation: Improving depthwise model-parallelism. The Journal of Machine Learning Research 23(1), 7714-7741 (2022)"
|
| 1889 |
+
},
|
| 1890 |
+
{
|
| 1891 |
+
"type": "ref_text",
|
| 1892 |
+
"bbox": [
|
| 1893 |
+
0.226,
|
| 1894 |
+
0.759,
|
| 1895 |
+
0.786,
|
| 1896 |
+
0.799
|
| 1897 |
+
],
|
| 1898 |
+
"angle": 0,
|
| 1899 |
+
"content": "4. He, K., Fan, H., Wu, Y., Xie, S., Girshick, R.: Momentum contrast for unsupervised visual representation learning. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 9729-9738 (2020)"
|
| 1900 |
+
},
|
| 1901 |
+
{
|
| 1902 |
+
"type": "ref_text",
|
| 1903 |
+
"bbox": [
|
| 1904 |
+
0.226,
|
| 1905 |
+
0.8,
|
| 1906 |
+
0.786,
|
| 1907 |
+
0.84
|
| 1908 |
+
],
|
| 1909 |
+
"angle": 0,
|
| 1910 |
+
"content": "5. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016)"
|
| 1911 |
+
},
|
| 1912 |
+
{
|
| 1913 |
+
"type": "list",
|
| 1914 |
+
"bbox": [
|
| 1915 |
+
0.226,
|
| 1916 |
+
0.289,
|
| 1917 |
+
0.786,
|
| 1918 |
+
0.84
|
| 1919 |
+
],
|
| 1920 |
+
"angle": 0,
|
| 1921 |
+
"content": null
|
| 1922 |
+
}
|
| 1923 |
+
],
|
| 1924 |
+
[
|
| 1925 |
+
{
|
| 1926 |
+
"type": "page_number",
|
| 1927 |
+
"bbox": [
|
| 1928 |
+
0.218,
|
| 1929 |
+
0.116,
|
| 1930 |
+
0.236,
|
| 1931 |
+
0.127
|
| 1932 |
+
],
|
| 1933 |
+
"angle": 0,
|
| 1934 |
+
"content": "16"
|
| 1935 |
+
},
|
| 1936 |
+
{
|
| 1937 |
+
"type": "header",
|
| 1938 |
+
"bbox": [
|
| 1939 |
+
0.272,
|
| 1940 |
+
0.115,
|
| 1941 |
+
0.382,
|
| 1942 |
+
0.127
|
| 1943 |
+
],
|
| 1944 |
+
"angle": 0,
|
| 1945 |
+
"content": "Junhao Su et al."
|
| 1946 |
+
},
|
| 1947 |
+
{
|
| 1948 |
+
"type": "ref_text",
|
| 1949 |
+
"bbox": [
|
| 1950 |
+
0.218,
|
| 1951 |
+
0.147,
|
| 1952 |
+
0.787,
|
| 1953 |
+
0.175
|
| 1954 |
+
],
|
| 1955 |
+
"angle": 0,
|
| 1956 |
+
"content": "16. Hinton, G.E., Osindero, S., Teh, Y.W.: A fast learning algorithm for deep belief nets. Neural computation 18(7), 1527-1554 (2006)"
|
| 1957 |
+
},
|
| 1958 |
+
{
|
| 1959 |
+
"type": "ref_text",
|
| 1960 |
+
"bbox": [
|
| 1961 |
+
0.218,
|
| 1962 |
+
0.176,
|
| 1963 |
+
0.786,
|
| 1964 |
+
0.217
|
| 1965 |
+
],
|
| 1966 |
+
"angle": 0,
|
| 1967 |
+
"content": "17. Huang, G., Liu, Z., Van Der Maaten, L., Weinberger, K.Q.: Densely connected convolutional networks. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4700-4708 (2017)"
|
| 1968 |
+
},
|
| 1969 |
+
{
|
| 1970 |
+
"type": "ref_text",
|
| 1971 |
+
"bbox": [
|
| 1972 |
+
0.218,
|
| 1973 |
+
0.218,
|
| 1974 |
+
0.786,
|
| 1975 |
+
0.259
|
| 1976 |
+
],
|
| 1977 |
+
"angle": 0,
|
| 1978 |
+
"content": "18. Illing, B., Ventura, J., Bellec, G., Gerstner, W.: Local plasticity rules can learn deep representations using self-supervised contrastive predictions. Advances in Neural Information Processing Systems 34, 30365-30379 (2021)"
|
| 1979 |
+
},
|
| 1980 |
+
{
|
| 1981 |
+
"type": "ref_text",
|
| 1982 |
+
"bbox": [
|
| 1983 |
+
0.217,
|
| 1984 |
+
0.26,
|
| 1985 |
+
0.787,
|
| 1986 |
+
0.3
|
| 1987 |
+
],
|
| 1988 |
+
"angle": 0,
|
| 1989 |
+
"content": "19. Jaderberg, M., Czarnecki, W.M., Osindero, S., Vinyals, O., Graves, A., Silver, D., Kavukcuoglu, K.: Decoupled neural interfaces using synthetic gradients. In: International conference on machine learning. pp. 1627-1635. PMLR (2017)"
|
| 1990 |
+
},
|
| 1991 |
+
{
|
| 1992 |
+
"type": "ref_text",
|
| 1993 |
+
"bbox": [
|
| 1994 |
+
0.217,
|
| 1995 |
+
0.301,
|
| 1996 |
+
0.787,
|
| 1997 |
+
0.342
|
| 1998 |
+
],
|
| 1999 |
+
"angle": 0,
|
| 2000 |
+
"content": "20. Kornblith, S., Norouzi, M., Lee, H., Hinton, G.: Similarity of neural network representations revisited. In: International conference on machine learning. pp. 3519-3529. PMLR (2019)"
|
| 2001 |
+
},
|
| 2002 |
+
{
|
| 2003 |
+
"type": "ref_text",
|
| 2004 |
+
"bbox": [
|
| 2005 |
+
0.217,
|
| 2006 |
+
0.343,
|
| 2007 |
+
0.787,
|
| 2008 |
+
0.37
|
| 2009 |
+
],
|
| 2010 |
+
"angle": 0,
|
| 2011 |
+
"content": "21. Krizhevsky, A., Hinton, G., et al.: Learning multiple layers of features from tiny images (2009)"
|
| 2012 |
+
},
|
| 2013 |
+
{
|
| 2014 |
+
"type": "ref_text",
|
| 2015 |
+
"bbox": [
|
| 2016 |
+
0.217,
|
| 2017 |
+
0.371,
|
| 2018 |
+
0.787,
|
| 2019 |
+
0.412
|
| 2020 |
+
],
|
| 2021 |
+
"angle": 0,
|
| 2022 |
+
"content": "22. Laskin, M., Metz, L., Nabarro, S., Saroufim, M., Noune, B., Luschi, C., Sohl-Dickstein, J., Abbeel, P.: Parallel training of deep networks with local updates. arXiv preprint arXiv:2012.03837 (2020)"
|
| 2023 |
+
},
|
| 2024 |
+
{
|
| 2025 |
+
"type": "ref_text",
|
| 2026 |
+
"bbox": [
|
| 2027 |
+
0.217,
|
| 2028 |
+
0.412,
|
| 2029 |
+
0.787,
|
| 2030 |
+
0.439
|
| 2031 |
+
],
|
| 2032 |
+
"angle": 0,
|
| 2033 |
+
"content": "23. Le Cun, Y.: Learning process in an asymmetric threshold network. In: Disordered systems and biological organization, pp. 233-240. Springer (1986)"
|
| 2034 |
+
},
|
| 2035 |
+
{
|
| 2036 |
+
"type": "ref_text",
|
| 2037 |
+
"bbox": [
|
| 2038 |
+
0.217,
|
| 2039 |
+
0.44,
|
| 2040 |
+
0.787,
|
| 2041 |
+
0.466
|
| 2042 |
+
],
|
| 2043 |
+
"angle": 0,
|
| 2044 |
+
"content": "24. LeCun, Y., Bengio, Y., Hinton, G.: Deep learning. nature 521(7553), 436-444 (2015)"
|
| 2045 |
+
},
|
| 2046 |
+
{
|
| 2047 |
+
"type": "ref_text",
|
| 2048 |
+
"bbox": [
|
| 2049 |
+
0.217,
|
| 2050 |
+
0.467,
|
| 2051 |
+
0.787,
|
| 2052 |
+
0.522
|
| 2053 |
+
],
|
| 2054 |
+
"angle": 0,
|
| 2055 |
+
"content": "25. Lee, D.H., Zhang, S., Fischer, A., Bengio, Y.: Difference target propagation. In: Machine Learning and Knowledge Discovery in Databases: European Conference, ECML PKDD 2015, Porto, Portugal, September 7-11, 2015, Proceedings, Part I 15. pp. 498-515. Springer (2015)"
|
| 2056 |
+
},
|
| 2057 |
+
{
|
| 2058 |
+
"type": "ref_text",
|
| 2059 |
+
"bbox": [
|
| 2060 |
+
0.217,
|
| 2061 |
+
0.523,
|
| 2062 |
+
0.787,
|
| 2063 |
+
0.563
|
| 2064 |
+
],
|
| 2065 |
+
"angle": 0,
|
| 2066 |
+
"content": "26. Lillicrap, T.P., Cownden, D., Tweed, D.B., Akerman, C.J.: Random synaptic feedback weights support error backpropagation for deep learning. Nature communications 7(1), 13276 (2016)"
|
| 2067 |
+
},
|
| 2068 |
+
{
|
| 2069 |
+
"type": "ref_text",
|
| 2070 |
+
"bbox": [
|
| 2071 |
+
0.217,
|
| 2072 |
+
0.565,
|
| 2073 |
+
0.787,
|
| 2074 |
+
0.592
|
| 2075 |
+
],
|
| 2076 |
+
"angle": 0,
|
| 2077 |
+
"content": "27. Lillicrap, T.P., Santoro, A., Marris, L., Akerman, C.J., Hinton, G.: Backpropagation and the brain. Nature Reviews Neuroscience 21(6), 335-346 (2020)"
|
| 2078 |
+
},
|
| 2079 |
+
{
|
| 2080 |
+
"type": "ref_text",
|
| 2081 |
+
"bbox": [
|
| 2082 |
+
0.217,
|
| 2083 |
+
0.592,
|
| 2084 |
+
0.787,
|
| 2085 |
+
0.633
|
| 2086 |
+
],
|
| 2087 |
+
"angle": 0,
|
| 2088 |
+
"content": "28. Löwe, S., O'Connor, P., Veeling, B.: Putting an end to end-to-end: Gradient-isolated learning of representations. Advances in neural information processing systems 32 (2019)"
|
| 2089 |
+
},
|
| 2090 |
+
{
|
| 2091 |
+
"type": "ref_text",
|
| 2092 |
+
"bbox": [
|
| 2093 |
+
0.217,
|
| 2094 |
+
0.634,
|
| 2095 |
+
0.787,
|
| 2096 |
+
0.66
|
| 2097 |
+
],
|
| 2098 |
+
"angle": 0,
|
| 2099 |
+
"content": "29. Mostafa, H., Ramesh, V., Cauwenberghs, G.: Deep supervised learning using local errors. Frontiers in neuroscience 12, 608 (2018)"
|
| 2100 |
+
},
|
| 2101 |
+
{
|
| 2102 |
+
"type": "ref_text",
|
| 2103 |
+
"bbox": [
|
| 2104 |
+
0.217,
|
| 2105 |
+
0.661,
|
| 2106 |
+
0.787,
|
| 2107 |
+
0.688
|
| 2108 |
+
],
|
| 2109 |
+
"angle": 0,
|
| 2110 |
+
"content": "30. Netzer, Y., Wang, T., Coates, A., Bissacco, A., Wu, B., Ng, A.Y.: Reading digits in natural images with unsupervised feature learning (2011)"
|
| 2111 |
+
},
|
| 2112 |
+
{
|
| 2113 |
+
"type": "ref_text",
|
| 2114 |
+
"bbox": [
|
| 2115 |
+
0.217,
|
| 2116 |
+
0.689,
|
| 2117 |
+
0.787,
|
| 2118 |
+
0.716
|
| 2119 |
+
],
|
| 2120 |
+
"angle": 0,
|
| 2121 |
+
"content": "31. Nokland, A.: Direct feedback alignment provides learning in deep neural networks. Advances in neural information processing systems 29 (2016)"
|
| 2122 |
+
},
|
| 2123 |
+
{
|
| 2124 |
+
"type": "ref_text",
|
| 2125 |
+
"bbox": [
|
| 2126 |
+
0.217,
|
| 2127 |
+
0.717,
|
| 2128 |
+
0.787,
|
| 2129 |
+
0.744
|
| 2130 |
+
],
|
| 2131 |
+
"angle": 0,
|
| 2132 |
+
"content": "32. Nokland, A., Eidnes, L.H.: Training neural networks with local error signals. In: International conference on machine learning. pp. 4839-4850. PMLR (2019)"
|
| 2133 |
+
},
|
| 2134 |
+
{
|
| 2135 |
+
"type": "ref_text",
|
| 2136 |
+
"bbox": [
|
| 2137 |
+
0.217,
|
| 2138 |
+
0.745,
|
| 2139 |
+
0.787,
|
| 2140 |
+
0.785
|
| 2141 |
+
],
|
| 2142 |
+
"angle": 0,
|
| 2143 |
+
"content": "33. Pyeon, M., Moon, J., Hahn, T., Kim, G.: Sedona: Search for decoupled neural networks toward greedy block-wise learning. In: International Conference on Learning Representations (2020)"
|
| 2144 |
+
},
|
| 2145 |
+
{
|
| 2146 |
+
"type": "ref_text",
|
| 2147 |
+
"bbox": [
|
| 2148 |
+
0.217,
|
| 2149 |
+
0.786,
|
| 2150 |
+
0.787,
|
| 2151 |
+
0.813
|
| 2152 |
+
],
|
| 2153 |
+
"angle": 0,
|
| 2154 |
+
"content": "34. Ren, M., Kornblith, S., Liao, R., Hinton, G.: Scaling forward gradient with local losses. arXiv preprint arXiv:2210.03310 (2022)"
|
| 2155 |
+
},
|
| 2156 |
+
{
|
| 2157 |
+
"type": "ref_text",
|
| 2158 |
+
"bbox": [
|
| 2159 |
+
0.217,
|
| 2160 |
+
0.814,
|
| 2161 |
+
0.787,
|
| 2162 |
+
0.841
|
| 2163 |
+
],
|
| 2164 |
+
"angle": 0,
|
| 2165 |
+
"content": "35. Rumelhart, D.E., Hinton, G.E., Williams, R.J.: Learning representations by backpropagating errors. nature 323(6088), 533-536 (1986)"
|
| 2166 |
+
},
|
| 2167 |
+
{
|
| 2168 |
+
"type": "list",
|
| 2169 |
+
"bbox": [
|
| 2170 |
+
0.217,
|
| 2171 |
+
0.147,
|
| 2172 |
+
0.787,
|
| 2173 |
+
0.841
|
| 2174 |
+
],
|
| 2175 |
+
"angle": 0,
|
| 2176 |
+
"content": null
|
| 2177 |
+
}
|
| 2178 |
+
],
|
| 2179 |
+
[
|
| 2180 |
+
{
|
| 2181 |
+
"type": "header",
|
| 2182 |
+
"bbox": [
|
| 2183 |
+
0.323,
|
| 2184 |
+
0.115,
|
| 2185 |
+
0.732,
|
| 2186 |
+
0.129
|
| 2187 |
+
],
|
| 2188 |
+
"angle": 0,
|
| 2189 |
+
"content": "Momentum Auxiliary Network for Supervised Local Learning"
|
| 2190 |
+
},
|
| 2191 |
+
{
|
| 2192 |
+
"type": "page_number",
|
| 2193 |
+
"bbox": [
|
| 2194 |
+
0.769,
|
| 2195 |
+
0.117,
|
| 2196 |
+
0.785,
|
| 2197 |
+
0.127
|
| 2198 |
+
],
|
| 2199 |
+
"angle": 0,
|
| 2200 |
+
"content": "17"
|
| 2201 |
+
},
|
| 2202 |
+
{
|
| 2203 |
+
"type": "ref_text",
|
| 2204 |
+
"bbox": [
|
| 2205 |
+
0.214,
|
| 2206 |
+
0.147,
|
| 2207 |
+
0.788,
|
| 2208 |
+
0.175
|
| 2209 |
+
],
|
| 2210 |
+
"angle": 0,
|
| 2211 |
+
"content": "36. Simonyan, K., Zisserman, A.: Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556 (2014)"
|
| 2212 |
+
},
|
| 2213 |
+
{
|
| 2214 |
+
"type": "ref_text",
|
| 2215 |
+
"bbox": [
|
| 2216 |
+
0.214,
|
| 2217 |
+
0.175,
|
| 2218 |
+
0.788,
|
| 2219 |
+
0.217
|
| 2220 |
+
],
|
| 2221 |
+
"angle": 0,
|
| 2222 |
+
"content": "37. Wang, Y., Ni, Z., Song, S., Yang, L., Huang, G.: Revisiting locally supervised learning: an alternative to end-to-end training. arXiv preprint arXiv:2101.10832 (2021)"
|
| 2223 |
+
},
|
| 2224 |
+
{
|
| 2225 |
+
"type": "ref_text",
|
| 2226 |
+
"bbox": [
|
| 2227 |
+
0.214,
|
| 2228 |
+
0.218,
|
| 2229 |
+
0.788,
|
| 2230 |
+
0.272
|
| 2231 |
+
],
|
| 2232 |
+
"angle": 0,
|
| 2233 |
+
"content": "38. Wu, B., Nair, S., Martin-Martin, R., Fei-Fei, L., Finn, C.: Greedy hierarchical variational autoencoders for large-scale video prediction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2318-2328 (2021)"
|
| 2234 |
+
},
|
| 2235 |
+
{
|
| 2236 |
+
"type": "ref_text",
|
| 2237 |
+
"bbox": [
|
| 2238 |
+
0.214,
|
| 2239 |
+
0.273,
|
| 2240 |
+
0.788,
|
| 2241 |
+
0.314
|
| 2242 |
+
],
|
| 2243 |
+
"angle": 0,
|
| 2244 |
+
"content": "39. Xie, S., Girshick, R., Dollar, P., Tu, Z., He, K.: Aggregated residual transformations for deep neural networks. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1492-1500 (2017)"
|
| 2245 |
+
},
|
| 2246 |
+
{
|
| 2247 |
+
"type": "ref_text",
|
| 2248 |
+
"bbox": [
|
| 2249 |
+
0.214,
|
| 2250 |
+
0.315,
|
| 2251 |
+
0.788,
|
| 2252 |
+
0.342
|
| 2253 |
+
],
|
| 2254 |
+
"angle": 0,
|
| 2255 |
+
"content": "40. Xie, X., Seung, H.S.: Equivalence of backpropagation and contrastive hebbian learning in a layered network. Neural computation 15(2), 441-454 (2003)"
|
| 2256 |
+
},
|
| 2257 |
+
{
|
| 2258 |
+
"type": "ref_text",
|
| 2259 |
+
"bbox": [
|
| 2260 |
+
0.214,
|
| 2261 |
+
0.343,
|
| 2262 |
+
0.788,
|
| 2263 |
+
0.371
|
| 2264 |
+
],
|
| 2265 |
+
"angle": 0,
|
| 2266 |
+
"content": "41. Xiong, Y., Ren, M., Urtasun, R.: Loco: Local contrastive representation learning. Advances in neural information processing systems 33, 11142-11153 (2020)"
|
| 2267 |
+
},
|
| 2268 |
+
{
|
| 2269 |
+
"type": "list",
|
| 2270 |
+
"bbox": [
|
| 2271 |
+
0.214,
|
| 2272 |
+
0.147,
|
| 2273 |
+
0.788,
|
| 2274 |
+
0.371
|
| 2275 |
+
],
|
| 2276 |
+
"angle": 0,
|
| 2277 |
+
"content": null
|
| 2278 |
+
}
|
| 2279 |
+
]
|
| 2280 |
+
]
|
2024/Momentum Auxiliary Network for Supervised Local Learning/eb53acf2-10ec-44f2-923d-fc539ed31362_origin.pdf
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:fd662352e3197cb1ed844eeec1fac4224ea191df1ec672dce3ec19afc1d47a03
|
| 3 |
+
size 4177036
|
2024/Momentum Auxiliary Network for Supervised Local Learning/full.md
ADDED
|
@@ -0,0 +1,295 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Momentum Auxiliary Network for Supervised Local Learning
|
| 2 |
+
|
| 3 |
+
Junhao Su $^{1*}$ , Changpeng Cai $^{1*}$ , Feiyu Zhu $^{2*}$ , Chenghao He $^{4}$
|
| 4 |
+
|
| 5 |
+
Xiaojie $\mathrm{Xu}^5$ , Dongzhi Guan $^{1\boxtimes}$ , and Chenyang Si $^{3\boxtimes}$
|
| 6 |
+
|
| 7 |
+
$^{1}$ Southeast University
|
| 8 |
+
|
| 9 |
+
2 University of Shanghai for Science and Technology
|
| 10 |
+
|
| 11 |
+
<sup>3</sup> Nanyang Technological University
|
| 12 |
+
|
| 13 |
+
$^{4}$ East China University of Science and Technology
|
| 14 |
+
|
| 15 |
+
<sup>5</sup> The Hong Kong University of Science and Technology
|
| 16 |
+
|
| 17 |
+
Abstract. Deep neural networks conventionally employ end-to-end backpropagation for their training process, which lacks biological credibility and triggers a locking dilemma during network parameter updates, leading to significant GPU memory use. Supervised local learning, which segments the network into multiple local blocks updated by independent auxiliary networks. However, these methods cannot replace end-to-end training due to lower accuracy, as gradients only propagate within their local block, creating a lack of information exchange between blocks. To address this issue and establish information transfer across blocks, we propose a Momentum Auxiliary Network (MAN) that establishes a dynamic interaction mechanism. The MAN leverages an exponential moving average (EMA) of the parameters from adjacent local blocks to enhance information flow. This auxiliary network, updated through EMA, helps bridge the informational gap between blocks. Nevertheless, we observe that directly applying EMA parameters has certain limitations due to feature discrepancies among local blocks. To overcome this, we introduce learnable biases, further boosting performance. We have validated our method on four image classification datasets (CIFAR-10, STL-10, SVHN, ImageNet), attaining superior performance and substantial memory savings. Notably, our method can reduce GPU memory usage by more than $45\%$ on the ImageNet dataset compared to end-to-end training, while achieving higher performance. The Momentum Auxiliary Network thus offers a new perspective for supervised local learning. Our code is available at: https://github.com/JunhaoSuO/MAN.
|
| 18 |
+
|
| 19 |
+
Keywords: Local Learning $\cdot$ Image Classification $\cdot$ Momentum Auxiliary Network
|
| 20 |
+
|
| 21 |
+

|
| 22 |
+
Fig. 1: Comparison between different methods with MAN and the original methods in terms of accuracy. Results are obtained using ResNet-110 (K=55) on the various datasets.
|
| 23 |
+
|
| 24 |
+
# 1 Introduction
|
| 25 |
+
|
| 26 |
+
In deep learning, the prevalent use of end-to-end backpropagation [35] is essential for training complex neural networks, which requires heavy computational work for loss evaluation and successive gradient descent through network layers to refine parameters [17, 24, 29]. This method starkly contrasts with the local signal processing of biological synapses [6, 7, 40] and imposes a 'locking' constraint [19] that delays parameter updates until the entire forward and backward pass is complete. Such constraints exacerbate challenges like reduced parallelism and increased GPU memory usage [28, 38], compromising training efficiency and scalability. An emerging alternative to traditional end-to-end (E2E) training is Local Learning [2, 3, 5, 18, 19, 29, 32, 37, 41], which promises to mitigate the drawbacks of traditional methods. These methods divide the network into discrete, gradient-isolated blocks, each of which is updated by its dedicated auxiliary network according to distinct local objectives [3, 32, 37]. This training strategy enables immediate parameter updates upon local error reception, circumventing the sequential update bottleneck of E2E training, thereby significantly enhancing the efficiency of parallel training [13, 22]. Moreover, local learning reduces the demand on GPU memory by only retaining gradients for the local and auxiliary networks, thus eliminating the need to store extensive global gradient information and saving computational resources [28, 38]. However, even with these advantages, there is still a significant performance gap between local learning and traditional E2E methods, especially as the network is divided into more local blocks. Current local learning techniques mainly focus on improving the design of auxiliary network structures [3] and making local loss functions better to close this performance gap [32, 37]. Yet, these improvements do not completely address the inherent short-sightedness of local learning: the separation into blocks can make each part of the network only focus on its local objectives, possibly ig-
|
| 27 |
+
|
| 28 |
+
noring the overall objectives of the network. This can lead to the discarding of globally beneficial information due to the lack of inter-block communication.
|
| 29 |
+
|
| 30 |
+
In this paper, we introduce the Momentum Auxiliary Network (MAN), a novel network architecture designed to mitigate the inherent limitations of supervised local learning by enhancing inter-block communication. Specifically, MAN not only accepts the current local block as input but also absorbs the parameters of its next block. Upon completion of a forward pass, parameter updates are performed using local gradients refined through Exponential Moving Average (EMA) techniques [14]. This novel approach enables each local block to integrate information from the following block, thereby extending the operational perspective of each block beyond its immediate objectives and further aligning with the global objective of the network. This addresses the critical limitations in local learning frameworks, ensuring a more cohesive alignment with the overarching network objectives. Furthermore, we identify that directly applying EMA parameters is constrained by the feature incongruities across gradient-isolated blocks. To mitigate this, we implement additional learnable bias terms to each auxiliary block, enhancing their ability to share information effectively. The proposed MAN requires only a minimal increase in memory use but significantly improves performance by enhancing the information sharing ability of local blocks. The efficacy of the MAN approach is validated on a suite of benchmark image classification datasets, including CIFAR-10 [21], STL-10 [9], SVHN [30], and ImageNet [12]. The results demonstrate the effectiveness of our method in surpassing the limitations of traditional supervised local learning.
|
| 31 |
+
|
| 32 |
+
The contributions of this paper could be summarized as follows:
|
| 33 |
+
|
| 34 |
+
- We propose Momentum Auxiliary Network (MAN), designed to facilitate inter-block communication with Exponential Moving Average (EMA), mitigating the short-sightedness issue in conventional supervised local learning techniques and culminating in an elevated overall performance of the network.
|
| 35 |
+
- MAN is a versatile, plug-and-play approach that can seamlessly integrate with any supervised local learning technique, markedly amplifying their efficacy while requiring only a negligible increase in memory usage.
|
| 36 |
+
- MAN has demonstrated its effectiveness through experiments on benchmark image classification datasets, achieving state-of-the-art performance. Notably, on the ImageNet [12] dataset, it outperforms E2E training while using significantly less memory.
|
| 37 |
+
|
| 38 |
+
# 2 Related Work
|
| 39 |
+
|
| 40 |
+
# 2.1 Local Learning
|
| 41 |
+
|
| 42 |
+
Local learning is first proposed as an innovative deep learning algorithm with the intention of utilizing memory more efficiently and adhering more closely to principles of biological plausibility [10]. This approach emerges as a response to
|
| 43 |
+
|
| 44 |
+

|
| 45 |
+
Fig. 2: Comparison of (a) end-to-end backpropagation, (b) other supervised local learning methods, and (c) our proposed method. Unlike E2E, supervised local learning separates the network into K gradient-isolated local blocks.
|
| 46 |
+
|
| 47 |
+
the limitations presented by global E2E training [16], prompting the development of alternative supervised local learning methodologies. Examples of these methodologies include a differentiable search algorithm, which decouples network blocks for block-level learning and selects a manually assisted network for each local block [33], as well as a self-supervised contrasting loss function that leverages its local learning rules [18, 41]. In supervised local learning, training primarily relies on the design of reasonable supervised local loss functions or the manual construction of auxiliary networks.
|
| 48 |
+
|
| 49 |
+
Supervised Local Loss: InfoPro [37] presents a method that involves the reconstruction loss for each hidden layer, coupled with cross-entropy loss. This approach forces the hidden layers to preserve vital input information, thereby preventing the early layers from discarding information relevant to the task. PredSim [32] employs layer-wise loss functions for network training. It makes use of two distinct supervised loss functions to generate local error signals for hidden layers. The optimization of these losses takes place during supervised local learning, thus eliminating the necessity to propagate global errors back to the hidden layers. These methods have achieved advanced performance in the field of supervised local learning, but there is still a significant gap compared to the performance of E2E training.
|
| 50 |
+
|
| 51 |
+
Auxiliary Networks: InfoPro [37] incorporates two auxiliary networks into its approach. One network consists of a single convolutional layer followed by two fully-connected layers, which are employed for the cross-entropy loss. The other network utilizes two convolutional layers with upsampling for the reconstruction loss. Alternatively, DGL [3] designs a structure that uses three consecutive convolutional layers connected to a global pooling layer, followed by three consecutive fully connected layers. Moreover, DGL restricts the parameter of the auxiliary network to only $5\%$ of the main network capacity, offering advantages in terms of speed and memory savings. However, these method still perform poorly when the network is divided into a large number of local blocks.
|
| 52 |
+
|
| 53 |
+
# 2.2 Alternative Learning Rules to E2E Training
|
| 54 |
+
|
| 55 |
+
Due to certain limitations inherent in E2E training, the pursuit of alternative methods to E2E training has gradually garnered attention in recent years [27]. Several efforts have been dedicated to addressing the biologically implausible aspects of E2E training, such as the training methodology of target propagation [4, 23, 25] attempts to train a specialized backward connection by utilizing local reconstruction targets. Additionally, some recent studies have attempted to completely avoid backpropagation in neural networks through forward gradient learning [11, 34]. Meanwhile, the weight transport problem [10]. This has been tackled either by employing distinct feedback connections [1, 26] or by directly disseminating global errors to each hidden unit [8, 31]. While these methods have somewhat alleviated the biological implausibility of E2E training, they still rely on global objectives, which fundamentally differ from biological neural networks that rely on local synapses for information transmission. Furthermore, these methods currently struggle to be effective on large datasets [12].
|
| 56 |
+
|
| 57 |
+
# 3 Method
|
| 58 |
+
|
| 59 |
+
# 3.1 Preliminaries
|
| 60 |
+
|
| 61 |
+
To set the stage, we first provide a brief overview of traditional end-to-end supervised learning and backpropagation mechanisms. We denote a data sample as $x$ and its corresponding ground-truth label as $y$ . The entire deep network is fragmented into several local blocks. During the forward propagation process, the output from the j-th block serves as the input for the (j+1)-th block, expressed as $x_{j+1} = f_{\theta_j}(x_j)$ . Here, $\theta_j$ symbolizes the parameters of the j-th local block and $f(\cdot)$ represents the forward calculation of the block. We evaluate the loss function $\mathcal{L}(\hat{y}, y)$ between the output of the last block and the ground truth label, and propagate it back iteratively to the preceding blocks.
|
| 62 |
+
|
| 63 |
+
Supervised local learning strategies [3, 32, 37] integrate auxiliary networks for local supervision. For each local block, an auxiliary network is affixed. The output from a local block is fed to its corresponding auxiliary network, generating the local supervisory signal as $\hat{y}_j = g_{\gamma_j}(x_{j + 1})$ . Here, $\gamma_{j}$ denotes the parameters of the j-th auxiliary network.
|
| 64 |
+
|
| 65 |
+
In this setup, we update the parameters of the j-th auxiliary network and local block, $\gamma_{j},\theta_{j}$ , as follows:
|
| 66 |
+
|
| 67 |
+
$$
|
| 68 |
+
\gamma_ {j} \leftarrow \gamma_ {j} - \eta_ {a} \times \nabla_ {\gamma_ {j}} \mathcal {L} (\hat {y} _ {j}, y) \tag {1}
|
| 69 |
+
$$
|
| 70 |
+
|
| 71 |
+
$$
|
| 72 |
+
\theta_ {j} \leftarrow \theta_ {j} - \eta_ {l} \times \nabla_ {\theta_ {j}} \mathcal {L} \left(y _ {\hat {j}}, y\right) \tag {2}
|
| 73 |
+
$$
|
| 74 |
+
|
| 75 |
+
where $\eta_{a},\eta_{l}$ are the learning rates of the auxiliary networks and local blocks, respectively. By attaching auxiliary networks, each local block becomes gradient-isolated and can be updated with local supervision rather than global backpropagation.
|
| 76 |
+
|
| 77 |
+
# 3.2 Momentum Auxiliary Network
|
| 78 |
+
|
| 79 |
+
Existing techniques incorporate supervision signals into individual local blocks, which allows for parallel parameter updates and reduces memory overhead. However, this approach can lead to a short-sightedness problem, where each local block neglects the information from subsequent blocks, ultimately resulting in suboptimal final accuracy.
|
| 80 |
+
|
| 81 |
+
To address this issue, we propose a comprehensive information exchange module—the Momentum Auxiliary Network (MAN). The MAN employs the Exponential Moving Average (EMA) mechanism [14] as a conduit for transferring information from subsequent blocks to the current block. In the context of MAN, we update the parameters of the j-th auxiliary network and local block as follows:
|
| 82 |
+
|
| 83 |
+
$$
|
| 84 |
+
\gamma_ {j} \leftarrow \gamma_ {j} - \eta_ {a} \times \nabla_ {\gamma_ {j}} \mathcal {L} \left(\hat {y} _ {j}, y\right) \tag {3}
|
| 85 |
+
$$
|
| 86 |
+
|
| 87 |
+
$$
|
| 88 |
+
\gamma_ {j} \leftarrow E M A \left(\gamma_ {j}, \theta_ {j + 1}\right) \tag {4}
|
| 89 |
+
$$
|
| 90 |
+
|
| 91 |
+
$$
|
| 92 |
+
\theta_ {j} \leftarrow \theta_ {j} - \eta_ {l} \times \nabla_ {\theta_ {j}} \mathcal {L} (\hat {y} _ {j}, y) \tag {5}
|
| 93 |
+
$$
|
| 94 |
+
|
| 95 |
+
where $\gamma_{j}$ represents the parameters of the j-th auxiliary network and $\theta_{j}$ represents those of the j-th local block. After updating with local gradients, $\gamma_{j}$ undergoes further refinement by incorporating the parameters of the subsequent local block via the EMA, which is a weighted sum operation.
|
| 96 |
+
|
| 97 |
+
However, through experimental results, it becomes apparent that directly applying EMA parameters to update the auxiliary network has its limitations, and it provides limited improvements to each local block. Upon analyzing, we find that the features learned between each local block vary to a certain degree, which hampers the effectiveness of the EMA update method. As a consequence, a learnable bias is introduced to augment the learning capability of the hidden layers within the local block, while also compensating for the deficiencies of the EMA update method.
|
| 98 |
+
|
| 99 |
+

|
| 100 |
+
Fig. 3: Details of the Momentum Auxiliary Network. Local $(\mathrm{i} + 1)$ represents the $(\mathrm{i} + 1)$ -th gradient-isolated local block, which contains layers from layer m to layer $(\mathrm{m} + \mathrm{n})$ , totaling n+1 layers $(\mathrm{n} \geqslant 0)$ . We only use the parameters of the first layer to ensure a balance in GPU memory usage.
|
| 101 |
+
|
| 102 |
+
Therefore, the parameter update method for the j-th auxiliary network can be written as:
|
| 103 |
+
|
| 104 |
+
$$
|
| 105 |
+
\left(\gamma_ {j}, b _ {j}\right) \leftarrow \left(\gamma_ {j}, b _ {j}\right) - \eta_ {a} \times \nabla_ {\left(\gamma_ {j}, b _ {j}\right)} \mathcal {L} \left(\hat {y} _ {j}, y\right) \tag {6}
|
| 106 |
+
$$
|
| 107 |
+
|
| 108 |
+
$$
|
| 109 |
+
\gamma_ {j} \leftarrow E M A \left(\gamma_ {j}, \theta_ {j + 1}\right) \tag {7}
|
| 110 |
+
$$
|
| 111 |
+
|
| 112 |
+
where $(\gamma_j, b_j)$ represent the parameters of the j-th auxiliary network. During this process, we update the parameters $\gamma_j$ of the j-th auxiliary network jointly through the learnable bias $b_j$ and EMA.
|
| 113 |
+
|
| 114 |
+
Essentially, the proposed MAN facilitates communication between subsequent blocks through the EMA mechanism [14], effectively resolving the short-sightedness problem inherent in traditional supervised local learning methods. Concurrently, it introduces an independent learnable bias to mitigate information discrepancies caused by feature variations among different local blocks, thereby aligning the output of each block closer to the global target.
|
| 115 |
+
|
| 116 |
+
In addition to the two innovative methods mentioned above, MAN demonstrates strong versatility. It can be seamlessly integrated into existing supervised local learning methods, and it excels across various datasets, reflecting its flexibility and generalizability.
|
| 117 |
+
|
| 118 |
+
# 4 Experiments
|
| 119 |
+
|
| 120 |
+
# 4.1 Experimental Setup
|
| 121 |
+
|
| 122 |
+
We conduct experiments using four widely adopted datasets: CIFAR-10 [21], SVHN [30], STL-10 [9], and ImageNet [12], with ResNets [15] of varying depths serving as the network architectures.
|
| 123 |
+
|
| 124 |
+
Three state-of-the-art supervised local learning methods are selected for comparison: PredSim [32], DGL [3], and InfoPro [37]. We then partition the network into K local blocks, each containing an approximately equal number of layers. Our proposed Momentum Auxiliary Network is incorporated only into the first K-1 local blocks. The K-th local block does not employ an auxiliary network and is directly connected to the output classifier. We compare these configurations against traditional E2E and original supervised local learning methods to maintain consistent training settings and eliminate confounding variables.
|
| 125 |
+
|
| 126 |
+
# 4.2 Implement Details
|
| 127 |
+
|
| 128 |
+
In our experiments on CIFAR-10 [21], SVHN [30], and STL-10 [9] datasets with ResNet-32 [15] and ResNet-110 [15], we utilize the SGD optimizer with Nesterov momentum set at 0.9 and an L2 weight decay factor of 1e-4. We employ batch sizes of 1024 for CIFAR-10 and SVHN and 128 for STL-10. The training duration spans 400 epochs, starting with initial learning rates of 0.8 for CIFAR-10 / SVHN and 0.1 for STL-10, following a cosine annealing scheduler [9].
|
| 129 |
+
|
| 130 |
+
In our experiments with ImageNet [12], we adopt different training settings for various architectures. We train VGG13 [36] for 90 epochs with an initial learning rate of 0.025. For ResNet-101 [15], ResNet-152 [15], and ResNeXt-101,32 $\times$ 8d [39], we train them for 90 epochs as well, with initial learning rates of 0.05, 0.05, and 0.025, respectively. We set the batch sizes to 64 for VGG13, 128 for ResNet-101 and ResNet-152, and 64 for ResNeXt-101,32 $\times$ 8d. We maintain consistency with other training configurations as previously described for CIFAR-10 [21].
|
| 131 |
+
|
| 132 |
+
Based on our multiple experimental results, we select the hyperparameter momentum value as 0.995, due to its consistently stable and superior performance. We will provide the experimental results of other momentum hyperparameters in the supplementary materials.
|
| 133 |
+
|
| 134 |
+
# 4.3 Results on Image Classification Datasets
|
| 135 |
+
|
| 136 |
+
Results on Image Classification Benchmarks: We start by assessing the accuracy performance of our approach using the CIFAR-10 [21], SVHN [30], and STL-10 [9] datasets. We employ ResNet-32 [15], partitioned into 8 and 16 local blocks, and ResNet-110 [15], divided into 32 and 55 local blocks. As illustrated in Table 1, our MAN significantly bolsters the accuracy of all methods.
|
| 137 |
+
|
| 138 |
+
Table 1: Performance of different networks with varying numbers of local blocks. The average test error is obtained by 5 experiments. The * means addition of our MAN.
|
| 139 |
+
|
| 140 |
+
<table><tr><td rowspan="2">Dataset</td><td rowspan="2">Method</td><td colspan="2">ResNet-32</td><td colspan="2">ResNet-110</td></tr><tr><td>K = 8 (Test Error)</td><td>K = 16 (Test Error)</td><td>K = 32 (Test Error)</td><td>K = 55 (Test Error)</td></tr><tr><td rowspan="6">CIFAR-10(E2E(ResNet-32)=6.37,E2E(ResNet-110)=5.42)</td><td>PredSim [32]</td><td>20.62</td><td>22.71</td><td>22.08</td><td>24.74</td></tr><tr><td>PredSim*</td><td>14.29(↓6.33)</td><td>15.58(↓7.13)</td><td>17.05(↓5.03)</td><td>18.02(↓6.72)</td></tr><tr><td>DGL [3]</td><td>11.63</td><td>14.08</td><td>12.51</td><td>14.45</td></tr><tr><td>DGL*</td><td>8.42(↓3.21)</td><td>9.11(↓4.97)</td><td>9.65(↓2.86)</td><td>9.73(↓4.72)</td></tr><tr><td>InfoPro [37]</td><td>11.51</td><td>12.93</td><td>12.26</td><td>13.22</td></tr><tr><td>InfoPro*</td><td>9.32(↓2.19)</td><td>9.65(↓3.28)</td><td>9.06(↓3.20)</td><td>9.77(↓3.45)</td></tr><tr><td rowspan="6">STL-10(E2E(ResNet-32)=19.35,E2E(ResNet-110)=19.67)</td><td>PredSim [32]</td><td>31.97</td><td>32.90</td><td>32.05</td><td>33.27</td></tr><tr><td>PredSim*</td><td>29.97(↓2.00)</td><td>29.99(↓2.91)</td><td>30.48(↓1.57)</td><td>31.55(↓1.72)</td></tr><tr><td>DGL [3]</td><td>25.05</td><td>27.14</td><td>25.67</td><td>28.16</td></tr><tr><td>DGL*</td><td>20.74(↓4.31)</td><td>21.37(↓5.77)</td><td>22.54(↓3.13)</td><td>22.69(↓5.47)</td></tr><tr><td>InfoPro [37]</td><td>27.32</td><td>29.28</td><td>28.58</td><td>29.20</td></tr><tr><td>InfoPro*</td><td>23.17(↓4.15)</td><td>23.54(↓5.74)</td><td>24.08(↓4.50)</td><td>24.74(↓4.46)</td></tr><tr><td rowspan="6">SVHN(E2E(ResNet-32)=2.99,E2E(ResNet-110)=2.92)</td><td>PredSim [32]</td><td>6.91</td><td>8.08</td><td>9.12</td><td>10.47</td></tr><tr><td>PredSim*</td><td>5.54(↓1.37)</td><td>6.39(↓1.69)</td><td>7.27(↓1.85)</td><td>8.24(↓2.23)</td></tr><tr><td>DGL [3]</td><td>4.83</td><td>5.05</td><td>5.12</td><td>5.36</td></tr><tr><td>DGL*</td><td>3.80(↓1.03)</td><td>4.04(↓1.01)</td><td>4.08(↓1.04)</td><td>4.52(↓0.84)</td></tr><tr><td>InfoPro [37]</td><td>5.61</td><td>5.97</td><td>5.89</td><td>6.11</td></tr><tr><td>InfoPro*</td><td>4.49(↓1.12)</td><td>5.19(↓0.78)</td><td>4.85(↓1.04)</td><td>4.99(↓1.12)</td></tr></table>
|
| 141 |
+
|
| 142 |
+
On the CIFAR-10 dataset [21], our method exhibits considerable improvements in diminishing test errors across various methods. In the relatively shallower network of ResNet-32 (K=16), where individual layers function as gradient-isolated local blocks, we record a reduction in test errors for PredSim [32], DGL [3], and InfoPro [37], from 22.71, 14.08, and 12.93 to 15.58, 9.11, and 9.65 respectively. This translates to a performance enhancement exceeding $25\%$ for all methods. Even though the performance across all methods in the comparatively deeper network, ResNet-110 (K=55), is somewhat inferior due to the inherent need for more global information in such networks, our method still delivers exceptional performance. It achieves approximately a $20\%$ improvement across all methods, underscoring the robust effectiveness of MAN in deeper networks.
|
| 143 |
+
|
| 144 |
+
When applied to other datasets, MAN can also reduce the test error of Pred-Sim [32], DGL [3], and InfoPro [37] by at least $5\%$ , $12\%$ , and $16\%$ on the STL-10 [9] dataset. On the SVHN [30] dataset, our improvements over the three
|
| 145 |
+
|
| 146 |
+
methods also surpass $20\%$ , $21\%$ , and $13\%$ . As can be seen, the improvement our MAN introduces to all methods is quite remarkable—comparable even to the accuracy of E2E training—and it significantly mitigates the overwhelming performance issue that has continually plagued supervised local learning.
|
| 147 |
+
|
| 148 |
+
Results on ImageNet: We further validate the effectiveness of our approach on ImageNet [12] using four networks of varying depths (ResNets [15] and VGG13 [36]). As depicted in Table 2, when we employ VGG13 as the backbone and divide the network into 10 blocks, DGL [3] achieves merely a Top1-Error of 35.60 and a Top5-Error of 14.2, representing a substantial gap when compared to the E2E method. However, with the introduction of our MAN, the Top1-Error reduces by 3.61 points, and the Top5-Error decreases by 3.36 points for DGL. This significant enhancement brings the performance closer to the E2E method.
|
| 149 |
+
|
| 150 |
+
As illustrated in Table 2, when we use ResNet-101 [15], ResNet-152 [15], ResNeXt-101, $32 \times 8\mathrm{d}$ [39] as backbones and divide the network into four blocks, the performance of InfoPro [37] is already below that of E2E. After incorporating our MAN, the Top-1 Error of these three backbone networks can be reduced by approximately $6\%$ compared to the original, surpassing the performance of E2E training. These results underscore the effectiveness of our MAN on the large-scale ImageNet [12] dataset, even when using deeper networks.
|
| 151 |
+
|
| 152 |
+
Table 2: Results on the validation set of ImageNet
|
| 153 |
+
|
| 154 |
+
<table><tr><td>Network</td><td>Method</td><td>Top1-Error</td><td>Top5-Error</td></tr><tr><td rowspan="5">ResNet-101</td><td>E2E</td><td>22.03</td><td>5.93</td></tr><tr><td>InfoPro(K=2) [37]</td><td>21.85</td><td>5.89</td></tr><tr><td>InfoPro*(K=2)</td><td>21.65(↓0.20)</td><td>5.49(↓0.40)</td></tr><tr><td>InfoPro(K=4) [37]</td><td>22.81</td><td>6.54</td></tr><tr><td>InfoPro*(K=4)</td><td>21.73(↓1.08)</td><td>5.81(↓0.73)</td></tr><tr><td rowspan="5">ResNet-152</td><td>E2E</td><td>21.60</td><td>5.92</td></tr><tr><td>InfoPro(K=2) [37]</td><td>21.45</td><td>5.84</td></tr><tr><td>InfoPro*(K=2)</td><td>21.23(↓0.22)</td><td>5.53(↓0.31)</td></tr><tr><td>InfoPro(K=4) [37]</td><td>22.93</td><td>6.71</td></tr><tr><td>InfoPro*(K=4)</td><td>21.59(↓1.34)</td><td>5.89(↓0.82)</td></tr><tr><td rowspan="5">ResNeXt-101, 32 × 8d</td><td>E2E</td><td>20.64</td><td>5.40</td></tr><tr><td>InfoPro(K=2) [37]</td><td>20.35</td><td>5.28</td></tr><tr><td>InfoPro*(K=2)</td><td>20.11(↓0.24)</td><td>5.18(↓0.10)</td></tr><tr><td>InfoPro(K=4) [37]</td><td>21.69</td><td>6.11</td></tr><tr><td>InfoPro*(K=4)</td><td>20.37(↓1.32)</td><td>5.34(↓0.77)</td></tr><tr><td rowspan="3">VGG13</td><td>E2E</td><td>28.41</td><td>9.63</td></tr><tr><td>DGL [3]</td><td>35.60</td><td>14.20</td></tr><tr><td>DGL*(K=10)</td><td>31.99(↓3.61)</td><td>10.84(↓3.36)</td></tr></table>
|
| 155 |
+
|
| 156 |
+
Training-Accuracy Curve Analysis: As depicted in the accuracy-epoch curve in Fig. 4, our Momentum Auxiliary Network consistently outperforms the original method in terms of accuracy throughout the entire training process. This underscores its reliability and stability during the training process of classification tasks. Moreover, our MAN achieves a higher accuracy earlier in the later stages of training and attains stability sooner, indicating a faster convergence rate—a critical attribute in large-scale and complex tasks.
|
| 157 |
+
|
| 158 |
+

|
| 159 |
+
Fig. 4: Training-Accuracy curves, the left uses ResNet-32 (K=16) as the backbone, while the right uses ResNet-110 (K=55). Both are utilizing the CIFAR-10 dataset.
|
| 160 |
+
|
| 161 |
+

|
| 162 |
+
|
| 163 |
+
In summary, our Momentum Auxiliary Network significantly enhances the accuracy of traditional supervised local learning methods by promoting information exchange between local blocks. The beneficial information exchange facilitated by Momentum Auxiliary Network significantly mitigates the pervasive issue of shortsightedness within the supervised local learning domain, thereby offering advantages in terms of training stability and convergence speed. This is crucial for training models more effectively and efficiently.
|
| 164 |
+
|
| 165 |
+
Table 3: Comparison of GPU memory usage using InfoPro as baseline with different backbones on the ImageNet dataset.
|
| 166 |
+
|
| 167 |
+
<table><tr><td>Method</td><td>ResNet-101
|
| 168 |
+
GPU Memory(GB)</td><td>ResNet-152
|
| 169 |
+
GPU Memory(GB)</td><td>ResNeXt-101, 32×8d
|
| 170 |
+
GPU Memory(GB)</td></tr><tr><td>E2E</td><td>19.71</td><td>26.29</td><td>19.22</td></tr><tr><td>InfoPro(K=2) [37]</td><td>12.06(↓ 38.8%)</td><td>15.53(↓ 40.9%)</td><td>11.55(↓ 39.9%)</td></tr><tr><td>InfoPro*(K=2)</td><td>12.32(↓ 37.5%)</td><td>15.93(↓ 39.4%)</td><td>11.73(↓ 38.9%)</td></tr><tr><td>InfoPro(K=4) [37]</td><td>10.37(↓ 47.3%)</td><td>13.48(↓ 48.7%)</td><td>10.24(↓ 46.7%)</td></tr><tr><td>InfoPro*(K=4)</td><td>10.69(↓ 45.8%)</td><td>13.91(↓ 47.1%)</td><td>10.49(↓ 45.5%)</td></tr></table>
|
| 171 |
+
|
| 172 |
+
Results on GPU memory requirement: Supervised local learning markedly conserves GPU memory by limiting gradient propagation to within local blocks, thereby ensuring that backpropagation transpires exclusively within each respective local block and its corresponding auxiliary network. This strategy significantly diminishes the storage demands for activation parameters and gradient information, which would typically disseminate across local blocks. Consequently, this facilitates substantial savings in GPU memory utilization. Our analysis, featuring a comprehensive comparison of GPU memory usage on the ImageNet dataset, illustrates that the implementation of a Momentum Auxiliary Network
|
| 173 |
+
|
| 174 |
+
Table 4: Abalation study of MAN. (a) Using DGL as baseline and ResNet-32 $(\mathrm{K} = 16)$ as backbone on the CIFAR-10 dataset. (b) Using InfoPro as baseline and ResNet-101 $(\mathrm{K} = 4)$ as backbone on the ImageNet dataset. LB stands for Learnable Bias.
|
| 175 |
+
|
| 176 |
+
<table><tr><td>EMA</td><td>LB</td><td>Test Error</td></tr><tr><td>×</td><td>×</td><td>14.08</td></tr><tr><td>✓</td><td>×</td><td>11.07(↓3.01)</td></tr><tr><td>✓</td><td>✓</td><td>9.11(↓4.97)</td></tr><tr><td colspan="3">(a)</td></tr></table>
|
| 177 |
+
|
| 178 |
+
<table><tr><td>EMA</td><td>LB</td><td>Top1-Error</td><td>Top5-Error</td></tr><tr><td>×</td><td>×</td><td>22.81</td><td>6.54</td></tr><tr><td>✓</td><td>×</td><td>22.09(↓ 0.72)</td><td>6.07(↓ 0.47)</td></tr><tr><td>✓</td><td>✓</td><td>21.73(↓ 1.08)</td><td>5.81(↓ 0.73)</td></tr><tr><td colspan="4">(b)</td></tr></table>
|
| 179 |
+
|
| 180 |
+
can elicit notable performance enhancements while incurring only a minimal uptick in GPU memory consumption.
|
| 181 |
+
|
| 182 |
+
As shown in Table 3, it indicates the substantial GPU memory reduction on the ImageNet [12]. When applying our method to InfoPro [37] on ResNet-101 [15], ResNet-152 [15], and ResNeXt-101, $32 \times 8d$ [39], dividing the backbone into two local blocks results in a GPU memory saving of $37.5\%$ to $39.4\%$ . When the network is divided into four local blocks, the degree of GPU memory savings is further enhanced, exceeding $45\%$ . Combined with Table 2, it can be seen that we achieve better performance with almost half the GPU memory required for end-to-end training. Further analysis of the memory usage comparison with the original method shows that our approach only increases the GPU memory by about $1\%$ over the original method, while achieving a performance improvement of over $5\%$ . These results highlight the excellent balance that the MAN method achieves in terms of GPU memory usage and performance enhancement.
|
| 183 |
+
|
| 184 |
+
# 4.4 Ablation Studies
|
| 185 |
+
|
| 186 |
+
We conduct an ablation study on the CIFAR-10 [21] dataset and ImageNet dataset to assess the impact of the EMA method [14] and learnable bias in the MAN on performance. For this analysis, we use ResNet-32 (K=16) [15] as the backbone and the original DGL [3] method as a comparison baseline.
|
| 187 |
+
|
| 188 |
+
As shown in Table 4(a), when we only use the EMA [14] in MAN to promote information exchange with subsequent blocks, without adding a learnable bias, the test error decreases from 14.08 to 11.07. When we further add a learnable bias, the test error decreases from 11.07 to 9.11, it is evident that both the EMA method and learnable bias contribute to performance improvement. We hypothesize that the EMA and learnable bias are somewhat complementary in terms of the features they learn. To validate this, we conduct feature visualization.
|
| 189 |
+
|
| 190 |
+
As depicted in Fig. 5(a), we can observe that the original method without the addition of MAN learns very limited and chaotic features. After adding the EMA method alone, the features it learns are clearly more characteristic and defined, indicating that it indeed receives some global features from the information in the subsequent block. However, there are still many blurry features, which may be due to the EMA parameter update method [14], causing information imbalance due to feature differences between different local blocks. When we
|
| 191 |
+
|
| 192 |
+
add learnable bias to the original method alone, the features it learns are more concentrated and clear, indicating that the learnable bias significantly enhances the learning ability of the current local block's hidden layer, but it still lacks some details because it does not receive more global information from the subsequent block. When we add the EMA method and learnable bias simultaneously, the learned features are not only clear and specific, but also more comprehensive in detail, proving that their learning abilities are complementary, and greatly improving the performance of the original method. The feature visualization in Fig. 5 verifies our previous thoughts.
|
| 193 |
+
|
| 194 |
+

|
| 195 |
+
(a)
|
| 196 |
+
|
| 197 |
+

|
| 198 |
+
(b)
|
| 199 |
+
Fig.5: Feature maps comparison of (a) Original Method without MAN, (b) MAN with only EMA, (c) MAN with only learnable bias and (d) MAN with both EMA and learnable bias. The feature map are obtained using ResNet-32 $(\mathrm{K} = 16)$ on CIFAR-10.
|
| 200 |
+
|
| 201 |
+

|
| 202 |
+
(c)
|
| 203 |
+
|
| 204 |
+

|
| 205 |
+
(d)
|
| 206 |
+
|
| 207 |
+
# 4.5 The Effectiveness of EMA
|
| 208 |
+
|
| 209 |
+
To verify the effectiveness of the EMA update method in MAN, we conduct further ablation experiments. As shown in Table 5, using only the parameters of the next layer in MAN results in very limited improvement. In contrast, employing the EMA method to utilize the parameters of the next layer leads to more significant enhancements. This is because the subsequent local block and the current local block learn different features and have different learning objectives, leading to minimal improvement. However, using the EMA method can gradually promote information exchange between local blocks and reduce parameter fluctuations during training, leading to smoother parameter updates. This beneficial interaction brings significant improvements to MAN.
|
| 210 |
+
|
| 211 |
+
Table 5: Abalation study of EMA. (a) Using DGL as baseline and ResNet-32 $(\mathrm{K} = 16)$ as backbone on the CIFAR-10 dataset. (b) Using InfoPro as baseline and ResNet-101 $(\mathrm{K} = 4)$ as backbone on the ImageNet dataset. Parameter signifies whether each local block utilizes the parameters of the adjacent next layer connected to it.
|
| 212 |
+
|
| 213 |
+
<table><tr><td>Parameter</td><td>EMA</td><td>Test</td><td>Error</td></tr><tr><td>×</td><td>×</td><td>14.08</td><td></td></tr><tr><td>✓</td><td>×</td><td>12.93</td><td></td></tr><tr><td>✓</td><td>✓</td><td>11.07</td><td></td></tr></table>
|
| 214 |
+
|
| 215 |
+
(a)
|
| 216 |
+
|
| 217 |
+
<table><tr><td>Parameter</td><td>EMA</td><td>Top1-Error</td><td>Top5-Error</td></tr><tr><td>×</td><td>×</td><td>22.81</td><td>6.54</td></tr><tr><td>✓</td><td>×</td><td>22.64</td><td>6.41</td></tr><tr><td>✓</td><td>✓</td><td>22.09</td><td>6.07</td></tr></table>
|
| 218 |
+
|
| 219 |
+
(b)
|
| 220 |
+
|
| 221 |
+
# 4.6 Linear Separability Analysis
|
| 222 |
+
|
| 223 |
+
To demonstrate the effectiveness of our Momentum Auxiliary Network, we freeze the parameters of the main network and train a classifier for each local block to obtain the classification accuracy of each local block. As depicted in Fig. 6, we use DGL [3] as a baseline, and after integrating our method, the accuracy of the earlier layers decreases, while the accuracy of the middle and later layers significantly improve. The decrease in accuracy of the earlier layers suggests that they have learned more generalized features. Although these features are not beneficial for optimizing local objectives, they are beneficial from a global perspective. The middle and later layers, receiving these globally beneficial features, have seen a significant increase in accuracy. These results illustrate that our proposed method has facilitated information interaction between gradient-isolated local blocks. This addresses the shortsightedness issue existing in the current supervised local learning field and significantly enhances their performance.
|
| 224 |
+
|
| 225 |
+

|
| 226 |
+
Fig. 6: Comparison of layer-wise linear separability across different learning rules on ResNet-32 and ResNet-110.
|
| 227 |
+
|
| 228 |
+

|
| 229 |
+
|
| 230 |
+
# 4.7 Representation Similarity Analysis
|
| 231 |
+
|
| 232 |
+
To further validate the efficacy of MAN, we conduct a Centered Kernel Alignment (CKA) experiment [20]. CKA serves as a metric to assess the similarity between feature representations. If a method's CKA score is closer to 1 in relation to the E2E training method, it indicates that the method's feature learning process is more aligned with that of E2E training. As depicted in Fig. 7, we use DGL [3] as a baseline and incorporate our method. It can be observed that whether adding the EMA method or learnable bias alone, the CKA scores significantly improve compared to the original method. When both the EMA and learnable bias are used in MAN, the CKA score further improves and performs more stably. Notably, the most significant increase in the CKA score occurs in the early and late layers. In conjunction with our previous analysis of the linear separability experiment, this is because if the early layers' learning method is closer
|
| 233 |
+
|
| 234 |
+
to E2E training, they will focus more on learning general features to optimize the global objective, rather than focusing narrowly on local objectives. While this may result in poorer classification capabilities in the early layers, it greatly contributes to the overall performance improvement of the network. Through the analysis of images and experimental results, it can be demonstrated that MAN enables information interaction between gradient-isolated local blocks, solving the myopia problem present in current supervised local learning methods.
|
| 235 |
+
|
| 236 |
+

|
| 237 |
+
Fig. 7: Assessment of Similarity in Layer-wise Representations. We use Centered Kernel Alignment (CKA) [20] to quantify the degree of similarity in the layer-wise representations between the E2E backpropagation (BP) and our proposed MAN.
|
| 238 |
+
|
| 239 |
+

|
| 240 |
+
|
| 241 |
+
# 5 Conclusion
|
| 242 |
+
|
| 243 |
+
This study addresses the performance disparities between traditional supervised local learning and end-to-end (E2E) methods in deep learning. We introduce a versatile Momentum Auxiliary Network to tackle the short-sighted problem in the early optimization work related to supervised local learning, facilitating information exchange between gradient-isolated local blocks. We integrate our Momentum Auxiliary Network into three advanced supervised local learning approaches and evaluate their performance across network architectures with varying depths on four widely adopted datasets. The results demonstrate our method's ability to significantly enhance the ultimate output performance of original supervised local learning methods. Particularly when combined with InfoPro [37], our method significantly reduces GPU memory usage while consistently maintaining performance levels closely aligned with E2E approaches.
|
| 244 |
+
|
| 245 |
+
Limitations and future works: Despite the superior performance on large-scale problems like ImageNet [12], our method still performs less accurately than E2E on some conventional image classification datasets. This may be due to the MAN using too few information interaction layers when the network is divided into a larger number of local blocks. In future work, we could explore deepening these information interaction layers to achieve better precision performance.
|
| 246 |
+
|
| 247 |
+
# Acknowledgement
|
| 248 |
+
|
| 249 |
+
This work was supported by the National Natural Science Foundation of China (52278154), the Natural Science Foundation of Jiangsu (BK20231429), the Fundamental Research Funds for the Central Universities (2242024RCB0008), and assist from the program of Zhishan Young Scholar of Southeast University.
|
| 250 |
+
|
| 251 |
+
# References
|
| 252 |
+
|
| 253 |
+
1. Akrout, M., Wilson, C., Humphreys, P., Lillicrap, T., Tweed, D.B.: Deep learning without weight transport. Advances in neural information processing systems 32 (2019)
|
| 254 |
+
2. Belilovsky, E., Eickenberg, M., Oyallon, E.: Greedy layerwise learning can scale to imagenet. In: International conference on machine learning. pp. 583-593. PMLR (2019)
|
| 255 |
+
3. Belilovsky, E., Eickenberg, M., Oyallon, E.: Decoupled greedy learning of cnns. In: International Conference on Machine Learning. pp. 736-745. PMLR (2020)
|
| 256 |
+
4. Bengio, Y.: How auto-encoders could provide credit assignment in deep networks via target propagation. arXiv preprint arXiv:1407.7906 (2014)
|
| 257 |
+
5. Bengio, Y., Lamblin, P., Popovici, D., Larochelle, H.: Greedy layer-wise training of deep networks. Advances in neural information processing systems 19 (2006)
|
| 258 |
+
6. Bengio, Y., Lee, D.H., Bornschein, J., Mesnard, T., Lin, Z.: Towards biologically plausible deep learning. arXiv preprint arXiv:1502.04156 (2015)
|
| 259 |
+
7. Caporale, N., Dan, Y.: Spike timing-dependent plasticity: a hebbian learning rule. Annu. Rev. Neurosci. 31, 25-46 (2008)
|
| 260 |
+
8. Clark, D., Abbott, L., Chung, S.: Credit assignment through broadcasting a global error vector. Advances in Neural Information Processing Systems 34, 10053-10066 (2021)
|
| 261 |
+
9. Coates, A., Ng, A., Lee, H.: An analysis of single-layer networks in unsupervised feature learning. In: Proceedings of the fourteenth international conference on artificial intelligence and statistics. pp. 215-223. JMLR Workshop and Conference Proceedings (2011)
|
| 262 |
+
0. Crick, F.: The recent excitement about neural networks. Nature 337(6203), 129-132 (1989)
|
| 263 |
+
1. Dellafererra, G., Kreiman, G.: Error-driven input modulation: solving the credit assignment problem without a backward pass. In: International Conference on Machine Learning. pp. 4937-4955. PMLR (2022)
|
| 264 |
+
2. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: 2009 IEEE conference on computer vision and pattern recognition. pp. 248-255. IEEE (2009)
|
| 265 |
+
3. Gomez, A.N., Key, O., Perlin, K., Gou, S., Frosst, N., Dean, J., Gal, Y.: Interlocking backpropagation: Improving depthwise model-parallelism. The Journal of Machine Learning Research 23(1), 7714-7741 (2022)
|
| 266 |
+
4. He, K., Fan, H., Wu, Y., Xie, S., Girshick, R.: Momentum contrast for unsupervised visual representation learning. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 9729-9738 (2020)
|
| 267 |
+
5. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016)
|
| 268 |
+
|
| 269 |
+
16. Hinton, G.E., Osindero, S., Teh, Y.W.: A fast learning algorithm for deep belief nets. Neural computation 18(7), 1527-1554 (2006)
|
| 270 |
+
17. Huang, G., Liu, Z., Van Der Maaten, L., Weinberger, K.Q.: Densely connected convolutional networks. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4700-4708 (2017)
|
| 271 |
+
18. Illing, B., Ventura, J., Bellec, G., Gerstner, W.: Local plasticity rules can learn deep representations using self-supervised contrastive predictions. Advances in Neural Information Processing Systems 34, 30365-30379 (2021)
|
| 272 |
+
19. Jaderberg, M., Czarnecki, W.M., Osindero, S., Vinyals, O., Graves, A., Silver, D., Kavukcuoglu, K.: Decoupled neural interfaces using synthetic gradients. In: International conference on machine learning. pp. 1627-1635. PMLR (2017)
|
| 273 |
+
20. Kornblith, S., Norouzi, M., Lee, H., Hinton, G.: Similarity of neural network representations revisited. In: International conference on machine learning. pp. 3519-3529. PMLR (2019)
|
| 274 |
+
21. Krizhevsky, A., Hinton, G., et al.: Learning multiple layers of features from tiny images (2009)
|
| 275 |
+
22. Laskin, M., Metz, L., Nabarro, S., Saroufim, M., Noune, B., Luschi, C., Sohl-Dickstein, J., Abbeel, P.: Parallel training of deep networks with local updates. arXiv preprint arXiv:2012.03837 (2020)
|
| 276 |
+
23. Le Cun, Y.: Learning process in an asymmetric threshold network. In: Disordered systems and biological organization, pp. 233-240. Springer (1986)
|
| 277 |
+
24. LeCun, Y., Bengio, Y., Hinton, G.: Deep learning. nature 521(7553), 436-444 (2015)
|
| 278 |
+
25. Lee, D.H., Zhang, S., Fischer, A., Bengio, Y.: Difference target propagation. In: Machine Learning and Knowledge Discovery in Databases: European Conference, ECML PKDD 2015, Porto, Portugal, September 7-11, 2015, Proceedings, Part I 15. pp. 498-515. Springer (2015)
|
| 279 |
+
26. Lillicrap, T.P., Cownden, D., Tweed, D.B., Akerman, C.J.: Random synaptic feedback weights support error backpropagation for deep learning. Nature communications 7(1), 13276 (2016)
|
| 280 |
+
27. Lillicrap, T.P., Santoro, A., Marris, L., Akerman, C.J., Hinton, G.: Backpropagation and the brain. Nature Reviews Neuroscience 21(6), 335-346 (2020)
|
| 281 |
+
28. Löwe, S., O'Connor, P., Veeling, B.: Putting an end to end-to-end: Gradient-isolated learning of representations. Advances in neural information processing systems 32 (2019)
|
| 282 |
+
29. Mostafa, H., Ramesh, V., Cauwenberghs, G.: Deep supervised learning using local errors. Frontiers in neuroscience 12, 608 (2018)
|
| 283 |
+
30. Netzer, Y., Wang, T., Coates, A., Bissacco, A., Wu, B., Ng, A.Y.: Reading digits in natural images with unsupervised feature learning (2011)
|
| 284 |
+
31. Nokland, A.: Direct feedback alignment provides learning in deep neural networks. Advances in neural information processing systems 29 (2016)
|
| 285 |
+
32. Nokland, A., Eidnes, L.H.: Training neural networks with local error signals. In: International conference on machine learning. pp. 4839-4850. PMLR (2019)
|
| 286 |
+
33. Pyeon, M., Moon, J., Hahn, T., Kim, G.: Sedona: Search for decoupled neural networks toward greedy block-wise learning. In: International Conference on Learning Representations (2020)
|
| 287 |
+
34. Ren, M., Kornblith, S., Liao, R., Hinton, G.: Scaling forward gradient with local losses. arXiv preprint arXiv:2210.03310 (2022)
|
| 288 |
+
35. Rumelhart, D.E., Hinton, G.E., Williams, R.J.: Learning representations by backpropagating errors. nature 323(6088), 533-536 (1986)
|
| 289 |
+
|
| 290 |
+
36. Simonyan, K., Zisserman, A.: Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556 (2014)
|
| 291 |
+
37. Wang, Y., Ni, Z., Song, S., Yang, L., Huang, G.: Revisiting locally supervised learning: an alternative to end-to-end training. arXiv preprint arXiv:2101.10832 (2021)
|
| 292 |
+
38. Wu, B., Nair, S., Martin-Martin, R., Fei-Fei, L., Finn, C.: Greedy hierarchical variational autoencoders for large-scale video prediction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2318-2328 (2021)
|
| 293 |
+
39. Xie, S., Girshick, R., Dollar, P., Tu, Z., He, K.: Aggregated residual transformations for deep neural networks. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1492-1500 (2017)
|
| 294 |
+
40. Xie, X., Seung, H.S.: Equivalence of backpropagation and contrastive hebbian learning in a layered network. Neural computation 15(2), 441-454 (2003)
|
| 295 |
+
41. Xiong, Y., Ren, M., Urtasun, R.: Loco: Local contrastive representation learning. Advances in neural information processing systems 33, 11142-11153 (2020)
|
2024/Momentum Auxiliary Network for Supervised Local Learning/images.zip
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:acea2424a4a0d06b5376c20db28bbaecb9311844a0ca13724560ced29ec1fe86
|
| 3 |
+
size 471586
|
2024/Momentum Auxiliary Network for Supervised Local Learning/layout.json
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
2024/Mono-ViFI_ A Unified Learning Framework for Self-supervised Single- and Multi-frame Monocular Depth Estimation/47dcf543-78f7-47b3-955c-778904eede62_content_list.json
ADDED
|
@@ -0,0 +1,1893 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
{
|
| 3 |
+
"type": "text",
|
| 4 |
+
"text": "Mono-ViFi: A Unified Learning Framework for Self-supervised Single- and Multi-frame Monocular Depth Estimation",
|
| 5 |
+
"text_level": 1,
|
| 6 |
+
"bbox": [
|
| 7 |
+
233,
|
| 8 |
+
140,
|
| 9 |
+
769,
|
| 10 |
+
209
|
| 11 |
+
],
|
| 12 |
+
"page_idx": 0
|
| 13 |
+
},
|
| 14 |
+
{
|
| 15 |
+
"type": "text",
|
| 16 |
+
"text": "Jinfeng Liu, Lingtong Kong\\*, Bo Li, Zerong Wang, Hong Gu, and Jinwei Chen",
|
| 17 |
+
"bbox": [
|
| 18 |
+
215,
|
| 19 |
+
234,
|
| 20 |
+
785,
|
| 21 |
+
251
|
| 22 |
+
],
|
| 23 |
+
"page_idx": 0
|
| 24 |
+
},
|
| 25 |
+
{
|
| 26 |
+
"type": "text",
|
| 27 |
+
"text": "vivo Mobile Communication Co., Ltd, China",
|
| 28 |
+
"bbox": [
|
| 29 |
+
349,
|
| 30 |
+
262,
|
| 31 |
+
651,
|
| 32 |
+
276
|
| 33 |
+
],
|
| 34 |
+
"page_idx": 0
|
| 35 |
+
},
|
| 36 |
+
{
|
| 37 |
+
"type": "text",
|
| 38 |
+
"text": "{liujinfeng, ltkong, libra, wangzerong, guhong, jinwei.chen}@vivo.com",
|
| 39 |
+
"bbox": [
|
| 40 |
+
254,
|
| 41 |
+
277,
|
| 42 |
+
750,
|
| 43 |
+
291
|
| 44 |
+
],
|
| 45 |
+
"page_idx": 0
|
| 46 |
+
},
|
| 47 |
+
{
|
| 48 |
+
"type": "text",
|
| 49 |
+
"text": "Abstract. Self-supervised monocular depth estimation has gathered notable interest since it can liberate training from dependency on depth annotations. In monocular video training case, recent methods only conduct view synthesis between existing camera views, leading to insufficient guidance. To tackle this, we try to synthesize more virtual camera views by flow-based video frame interpolation (VFI), termed as temporal augmentation. For multi-frame inference, to sidestep the problem of dynamic objects encountered by explicit geometry-based methods like ManyDepth, we return to the feature fusion paradigm and design a VFI-assisted multi-frame fusion module to align and aggregate multi-frame features, using motion and occlusion information obtained by the flow-based VFI model. Finally, we construct a unified self-supervised learning framework, named Mono-ViFi, to bilaterally connect single- and multiframe depth. In this framework, spatial data augmentation through image affine transformation is incorporated for data diversity, along with a triplet depth consistency loss for regularization. The single- and multiframe models can share weights, making our framework compact and memory-efficient. Extensive experiments demonstrate that our method can bring significant improvements to current advanced architectures. Source code is available at https://github.com/LiuJF1226/Mono-ViFi.",
|
| 50 |
+
"bbox": [
|
| 51 |
+
261,
|
| 52 |
+
327,
|
| 53 |
+
740,
|
| 54 |
+
604
|
| 55 |
+
],
|
| 56 |
+
"page_idx": 0
|
| 57 |
+
},
|
| 58 |
+
{
|
| 59 |
+
"type": "text",
|
| 60 |
+
"text": "Keywords: Monocular depth estimation $\\cdot$ Self-supervised learning $\\cdot$ Data augmentation $\\cdot$ Video frame interpolation",
|
| 61 |
+
"bbox": [
|
| 62 |
+
261,
|
| 63 |
+
618,
|
| 64 |
+
745,
|
| 65 |
+
646
|
| 66 |
+
],
|
| 67 |
+
"page_idx": 0
|
| 68 |
+
},
|
| 69 |
+
{
|
| 70 |
+
"type": "text",
|
| 71 |
+
"text": "1 Introduction",
|
| 72 |
+
"text_level": 1,
|
| 73 |
+
"bbox": [
|
| 74 |
+
217,
|
| 75 |
+
672,
|
| 76 |
+
375,
|
| 77 |
+
688
|
| 78 |
+
],
|
| 79 |
+
"page_idx": 0
|
| 80 |
+
},
|
| 81 |
+
{
|
| 82 |
+
"type": "text",
|
| 83 |
+
"text": "Monocular depth estimation (MDE) is a fundamental 3D vision task, which finds applications in robotics, autonomous driving and VR/AR. To reduce the dependence on depth annotations, self-supervised MDE [1,2,15,17,46,63,66] has risen as a convenient and practical solution, which uses monocular videos, stereo pairs, or even both to generate supervision through view synthesis, where an image photometric reconstruction loss is minimized. Compared to stereo training case [15,17,59] with known camera pose, learning from monocular videos [18,58, 66] is more challenging since it requires an extra pose network to estimate camera ego-motion. Even so, it is much easier to collect monocular videos compared to",
|
| 84 |
+
"bbox": [
|
| 85 |
+
212,
|
| 86 |
+
703,
|
| 87 |
+
787,
|
| 88 |
+
840
|
| 89 |
+
],
|
| 90 |
+
"page_idx": 0
|
| 91 |
+
},
|
| 92 |
+
{
|
| 93 |
+
"type": "text",
|
| 94 |
+
"text": "stereo image pairs, making monocular training more attractive. Therefore, this paper also focuses on self-supervised training with unlabeled monocular videos.",
|
| 95 |
+
"bbox": [
|
| 96 |
+
212,
|
| 97 |
+
146,
|
| 98 |
+
782,
|
| 99 |
+
176
|
| 100 |
+
],
|
| 101 |
+
"page_idx": 1
|
| 102 |
+
},
|
| 103 |
+
{
|
| 104 |
+
"type": "text",
|
| 105 |
+
"text": "Recent works have made various attempts to improve accuracy, like introducing masking techniques [18, 66] and leveraging segmentation priors [28, 31]. However, they only conduct view synthesis between existing fixed camera views within video datasets, leading to insufficient guidance especially in invisible regions by occlusion or disappearance. Considering that video data possesses temporal continuity, an intuitive idea is to generate novel frames, i.e., virtual camera views in the temporal dimension. We term it as temporal augmentation, which can be achieved by video frame interpolation (VFI), as shown in Fig. 1(a).",
|
| 106 |
+
"bbox": [
|
| 107 |
+
212,
|
| 108 |
+
181,
|
| 109 |
+
785,
|
| 110 |
+
303
|
| 111 |
+
],
|
| 112 |
+
"page_idx": 1
|
| 113 |
+
},
|
| 114 |
+
{
|
| 115 |
+
"type": "text",
|
| 116 |
+
"text": "From another perspective, multi-frame reasoning paradigm have recently been researched because multiple frames can be available for reasoning in many application scenarios. ManyDepth [58] proposes to use cost volume constructed from multi-frame matching features. It explicitly reasons about geometry based on static scene assumption, which can be easily affected by dynamic objects and occlusion. DynamicDepth [14] attempts to tackle dynamic scenes, but relies on depth priors and segmentation masks of moving objects when reasoning. To sidestep the issue of dynamic objects in multi-frame inference encountered by these methods, we return to the traditional paradigm [42, 49], which aggregates multi-frame features directly. However, direct feature fusion without alignment harms the learning of temporal representations, since feature embeddings at the same location in neighbor frames are usually irrelevant, making it difficult for models to leverage the correlation of multiple frames. Hence, we regard feature alignment as a reasonable and necessary step before aggregation. As revealed in several video enhancement tasks like video super-resolution [6, 7] and frame interpolation [26, 29, 34, 36], inter-frame feature alignment can be implemented by warping through optical flow. This inspires us to employ a flow-based VFI model to concurrently conduct temporal augmentation and provide explicit correspondence for feature alignment. Moreover, the VFI model can reason out occlusion mask between neighbor frames, helping to alleviate inter-frame occlusion problem in feature fusion. Based on above analysis, we design a VFI-assisted multi-frame fusion module for multi-frame depth estimation, as depicted in Fig. 2.",
|
| 117 |
+
"bbox": [
|
| 118 |
+
212,
|
| 119 |
+
306,
|
| 120 |
+
787,
|
| 121 |
+
638
|
| 122 |
+
],
|
| 123 |
+
"page_idx": 1
|
| 124 |
+
},
|
| 125 |
+
{
|
| 126 |
+
"type": "text",
|
| 127 |
+
"text": "Finally, we build a unified framework to bilaterally connect single- and multiframe depth, as demonstrated in Fig. 1(b). The direct way is by depth consistency, usually for geometric regularization or distillation in self-supervised MDE. ManyDepth conducts unilateral distillation from single-frame model to multiframe one. DynamicDepth extends it to a bilateral mutual reinforcement, but only in the regions of moving objects. Differently, our framework applies the bilateral consistency on the whole image. For an original target image, we use a standard view depth consistency (SVDC) loss between its two depth maps predicted from the single- and multi-frame models. We also include classic data augmentation strategies, categorized as spatial augmentation. Specifically, we combine resizing-cropping with image rotation, forming a generic affine transformation. Then, we give the formulation of rectified camera pose for affine-augmented view synthesis using the cue of relative object size [4, 55]. For the",
|
| 128 |
+
"bbox": [
|
| 129 |
+
212,
|
| 130 |
+
643,
|
| 131 |
+
787,
|
| 132 |
+
840
|
| 133 |
+
],
|
| 134 |
+
"page_idx": 1
|
| 135 |
+
},
|
| 136 |
+
{
|
| 137 |
+
"type": "page_number",
|
| 138 |
+
"text": "2",
|
| 139 |
+
"bbox": [
|
| 140 |
+
217,
|
| 141 |
+
114,
|
| 142 |
+
228,
|
| 143 |
+
126
|
| 144 |
+
],
|
| 145 |
+
"page_idx": 1
|
| 146 |
+
},
|
| 147 |
+
{
|
| 148 |
+
"type": "header",
|
| 149 |
+
"text": "J. Liu et al.",
|
| 150 |
+
"bbox": [
|
| 151 |
+
271,
|
| 152 |
+
114,
|
| 153 |
+
351,
|
| 154 |
+
127
|
| 155 |
+
],
|
| 156 |
+
"page_idx": 1
|
| 157 |
+
},
|
| 158 |
+
{
|
| 159 |
+
"type": "image",
|
| 160 |
+
"img_path": "images/9a3d1e5d9591e39c21ee0f4c6d3c34d41e74bc997a12e6f691af7d204581947b.jpg",
|
| 161 |
+
"image_caption": [
|
| 162 |
+
"(a) Temporal Augmentation with VFI"
|
| 163 |
+
],
|
| 164 |
+
"image_footnote": [],
|
| 165 |
+
"bbox": [
|
| 166 |
+
217,
|
| 167 |
+
141,
|
| 168 |
+
521,
|
| 169 |
+
294
|
| 170 |
+
],
|
| 171 |
+
"page_idx": 2
|
| 172 |
+
},
|
| 173 |
+
{
|
| 174 |
+
"type": "image",
|
| 175 |
+
"img_path": "images/ae377703be20437e014938cb17aad97ac308c67f058d7004bde80aa93374d895.jpg",
|
| 176 |
+
"image_caption": [
|
| 177 |
+
"(b) Learning Framework $(I_{t}$ for example",
|
| 178 |
+
"Fig. 1: Overview of Mono-ViFI. (a) We achieve temporal augmentation by a flow-based VFI model $\\mathcal{H}$ , which also reasons out intermediate optical flow and occlusion mask for multi-frame inference. (b) For target image $I_{t}$ , we obtain its single-frame depth $D_{t}$ and multi-frame depth $D_{t}^{m}$ . Also, $I_{t}$ is augmented to $\\widetilde{I}_{t}$ through affine transformation $\\mathcal{A}$ . We then calculate the depth $\\widehat{D}_t$ of $\\widetilde{I}_t$ and inversely convert it to the original view, generating $\\widehat{D}_t$ . Finally, we enforce a triplet depth consistency loss $L_{tc}(t)$ among the three depth maps, including a standard view depth consistency loss $L_{sv}(t)$ and two scale-aware depth consistency losses, $L_{sa}(t)$ and $L_{sa}^{m}(t)$ . Note that each depth map also corresponds to a photometric loss and a smoothness loss, which are omitted here."
|
| 179 |
+
],
|
| 180 |
+
"image_footnote": [],
|
| 181 |
+
"bbox": [
|
| 182 |
+
522,
|
| 183 |
+
142,
|
| 184 |
+
784,
|
| 185 |
+
292
|
| 186 |
+
],
|
| 187 |
+
"page_idx": 2
|
| 188 |
+
},
|
| 189 |
+
{
|
| 190 |
+
"type": "text",
|
| 191 |
+
"text": "depth map from augmented target view, we enforce two scale-aware depth consistency (SADC) losses between it and above two standard view depths, respectively. SADC stems from the relative size cue, which assumes that when an image is zoomed up by a factor, the depth decreases by the same factor. It guides depth model to learn consistent geometric scale relationship from scene variations. The SVDC loss and two SADC losses are termed as a triplet depth consistency loss.",
|
| 192 |
+
"bbox": [
|
| 193 |
+
212,
|
| 194 |
+
476,
|
| 195 |
+
787,
|
| 196 |
+
566
|
| 197 |
+
],
|
| 198 |
+
"page_idx": 2
|
| 199 |
+
},
|
| 200 |
+
{
|
| 201 |
+
"type": "text",
|
| 202 |
+
"text": "Our framework is named Mono-ViFI, since the self-supervised Monocular depth is largely boosted by Video Frame Interpolation. To our best knowledge, we are the first to leverage VFI for the improvement of self-supervised MDE. The main contributions of this paper are summarized as follows:",
|
| 203 |
+
"bbox": [
|
| 204 |
+
212,
|
| 205 |
+
566,
|
| 206 |
+
787,
|
| 207 |
+
628
|
| 208 |
+
],
|
| 209 |
+
"page_idx": 2
|
| 210 |
+
},
|
| 211 |
+
{
|
| 212 |
+
"type": "list",
|
| 213 |
+
"sub_type": "text",
|
| 214 |
+
"list_items": [
|
| 215 |
+
"- We propose Mono-ViFi, a unified learning framework to bilaterally connect self-supervised single- and multi-frame MDE. In Mono-ViFi, the single- and multi-frame models can share weights, making the unified framework more compact and memory-efficient.",
|
| 216 |
+
"- We conduct temporal augmentation through flow-based VFI, and develop a novel VFI-assisted multi-frame fusion module, which utilizes motion and occlusion from flow-based VFI model to improve multi-frame depth.",
|
| 217 |
+
"- We incorporate image affine transformation as spatial augmentation, along with a triplet depth consistency loss for regularization and distillation.",
|
| 218 |
+
"- Experiments show that Mono-ViFI can achieve state-of-the-art performance. Moreover, our approach is orthogonal to architecture design, allowing current advanced depth networks to be further improved by seamlessly integrating them with Mono-ViFI."
|
| 219 |
+
],
|
| 220 |
+
"bbox": [
|
| 221 |
+
223,
|
| 222 |
+
641,
|
| 223 |
+
785,
|
| 224 |
+
839
|
| 225 |
+
],
|
| 226 |
+
"page_idx": 2
|
| 227 |
+
},
|
| 228 |
+
{
|
| 229 |
+
"type": "header",
|
| 230 |
+
"text": "Mono-ViFi: A Unified Framework for Self-supervised Monocular Depth",
|
| 231 |
+
"bbox": [
|
| 232 |
+
256,
|
| 233 |
+
113,
|
| 234 |
+
732,
|
| 235 |
+
130
|
| 236 |
+
],
|
| 237 |
+
"page_idx": 2
|
| 238 |
+
},
|
| 239 |
+
{
|
| 240 |
+
"type": "page_number",
|
| 241 |
+
"text": "3",
|
| 242 |
+
"bbox": [
|
| 243 |
+
774,
|
| 244 |
+
114,
|
| 245 |
+
785,
|
| 246 |
+
127
|
| 247 |
+
],
|
| 248 |
+
"page_idx": 2
|
| 249 |
+
},
|
| 250 |
+
{
|
| 251 |
+
"type": "text",
|
| 252 |
+
"text": "2 Related Work",
|
| 253 |
+
"text_level": 1,
|
| 254 |
+
"bbox": [
|
| 255 |
+
215,
|
| 256 |
+
143,
|
| 257 |
+
387,
|
| 258 |
+
160
|
| 259 |
+
],
|
| 260 |
+
"page_idx": 3
|
| 261 |
+
},
|
| 262 |
+
{
|
| 263 |
+
"type": "text",
|
| 264 |
+
"text": "2.1 Self-supervised Monocular Depth Estimation",
|
| 265 |
+
"text_level": 1,
|
| 266 |
+
"bbox": [
|
| 267 |
+
215,
|
| 268 |
+
172,
|
| 269 |
+
633,
|
| 270 |
+
188
|
| 271 |
+
],
|
| 272 |
+
"page_idx": 3
|
| 273 |
+
},
|
| 274 |
+
{
|
| 275 |
+
"type": "text",
|
| 276 |
+
"text": "Self-supervised learning paradigm for monocular depth estimation can be categorized into stereo training and monocular training according to the data source.",
|
| 277 |
+
"bbox": [
|
| 278 |
+
212,
|
| 279 |
+
194,
|
| 280 |
+
782,
|
| 281 |
+
223
|
| 282 |
+
],
|
| 283 |
+
"page_idx": 3
|
| 284 |
+
},
|
| 285 |
+
{
|
| 286 |
+
"type": "text",
|
| 287 |
+
"text": "Stereo training leverages the left-right consistency in stereo image pairs. Garg et al. [15] first consider the training as a view synthesis problem with a photometric consistency loss between stereo pairs. Building upon this idea, Godard et al. [17] introduce a left-right disparity consistency loss. Then, it is further extended by using depth hints [59], data augmentation [43, 55], knowledge distillation [43, 45, 55] and plane-based scene representation [3, 55].",
|
| 288 |
+
"bbox": [
|
| 289 |
+
212,
|
| 290 |
+
224,
|
| 291 |
+
784,
|
| 292 |
+
314
|
| 293 |
+
],
|
| 294 |
+
"page_idx": 3
|
| 295 |
+
},
|
| 296 |
+
{
|
| 297 |
+
"type": "text",
|
| 298 |
+
"text": "Monocular training leverages the temporal consistency in monocular video frames. Zhou et al. [66] propose a pioneering method which jointly learns depth and camera ego-motion from monocular videos in a self-supervised way. Then, Bian et al. [5] propose a scale-consistent depth and ego-motion estimation approach by adding a depth consistency loss, which reduces scale drift of estimated depth but decreases the accuracy. Later, Godard et al. [18] propose Monodepth2, a strong baseline for most follow-up works, which employs a minimum reprojection loss to mitigate occlusion issues and an auto-masking loss to filter out moving objects with similar velocity as the camera. Furthermore, some works introduce multi-task learning to combine depth estimation with optical flow [9, 32, 47, 60], semantic segmentation [27, 28, 40] and instance segmentation [31]. Others [21-23, 35, 37, 39, 62, 64, 65] try to improve the network architectures for depth estimation. However, these methods only utilize the geometric constrains between existing fixed camera views in video datasets. This insufficient guidance may block current depth models from unleashing their full capabilities. To address it, we conduct temporal augmentation through VFI in this paper.",
|
| 299 |
+
"bbox": [
|
| 300 |
+
212,
|
| 301 |
+
316,
|
| 302 |
+
785,
|
| 303 |
+
556
|
| 304 |
+
],
|
| 305 |
+
"page_idx": 3
|
| 306 |
+
},
|
| 307 |
+
{
|
| 308 |
+
"type": "text",
|
| 309 |
+
"text": "2.2 Multi-frame Monocular Depth",
|
| 310 |
+
"text_level": 1,
|
| 311 |
+
"bbox": [
|
| 312 |
+
215,
|
| 313 |
+
575,
|
| 314 |
+
513,
|
| 315 |
+
590
|
| 316 |
+
],
|
| 317 |
+
"page_idx": 3
|
| 318 |
+
},
|
| 319 |
+
{
|
| 320 |
+
"type": "text",
|
| 321 |
+
"text": "Above single-frame depth models do not leverage any temporal information for inference, which limits the performance to some degree. Regarding multi-frame depth prediction, early works adopt test-time refinement [9, 30, 38], use spatial-temporal attention [49], or combine traditional depth networks with recurrent layers [54, 61] to process sequence information. Another broad group of methods [14, 20, 56, 58] utilize cost volume from stereo matching, which explicitly encode and reason from geometric constraints by combining a multi-frame cost volume with single-frame features. On the one hand, this type of methods based on ManyDepth [58] usually meet with the problem of dynamic scenes. To sidestep this issue, we develop a VFI-assisted multi-frame fusion depth model to align and aggregate multi-frame features by optical flow and occlusion masks. On the other hand, ManyDepth uses single-frame depth prior to boost multi-frame depth estimation. However, this improvement is unidirectional. DynamicDepth [14] extends it to a mutual reinforcement, but only targets on dynamic objects. Differently, we propose a unified framework to bilaterally connect single- and multi-frame depth in a more compact and memory-efficient manner.",
|
| 322 |
+
"bbox": [
|
| 323 |
+
212,
|
| 324 |
+
598,
|
| 325 |
+
787,
|
| 326 |
+
840
|
| 327 |
+
],
|
| 328 |
+
"page_idx": 3
|
| 329 |
+
},
|
| 330 |
+
{
|
| 331 |
+
"type": "page_number",
|
| 332 |
+
"text": "4",
|
| 333 |
+
"bbox": [
|
| 334 |
+
217,
|
| 335 |
+
114,
|
| 336 |
+
230,
|
| 337 |
+
126
|
| 338 |
+
],
|
| 339 |
+
"page_idx": 3
|
| 340 |
+
},
|
| 341 |
+
{
|
| 342 |
+
"type": "header",
|
| 343 |
+
"text": "J. Liu et al.",
|
| 344 |
+
"bbox": [
|
| 345 |
+
271,
|
| 346 |
+
114,
|
| 347 |
+
351,
|
| 348 |
+
127
|
| 349 |
+
],
|
| 350 |
+
"page_idx": 3
|
| 351 |
+
},
|
| 352 |
+
{
|
| 353 |
+
"type": "text",
|
| 354 |
+
"text": "3 Problem Setup",
|
| 355 |
+
"text_level": 1,
|
| 356 |
+
"bbox": [
|
| 357 |
+
215,
|
| 358 |
+
143,
|
| 359 |
+
398,
|
| 360 |
+
162
|
| 361 |
+
],
|
| 362 |
+
"page_idx": 4
|
| 363 |
+
},
|
| 364 |
+
{
|
| 365 |
+
"type": "text",
|
| 366 |
+
"text": "Self-supervised MDE in the monocular training case leverages the geometric constraints from video sequences. Specifically, for a training triplet $\\{I_{t-2}, I_t, I_{t+2}\\}$ which contains three consecutive frames, we use $I_t$ as the target view and the other two as the source views $\\{I_s\\}_{s \\in \\{t-2, t+2\\}}$ to implement view synthesis. Here the temporal notation $t \\pm 2$ instead of $t \\pm 1$ is for the convenience of following description. Also, a pose network is utilized to estimate the camera pose from the target view to the source view, denoted as $\\pmb{T}_{t \\to s} = [\\pmb{R}_{t \\to s} | \\pmb{t}_{t \\to s}]$ , where $\\pmb{R}_{t \\to s}$ and $\\pmb{t}_{t \\to s}$ represent the rotation and translation components. Formally, the target depth $D_t$ and the camera pose $\\pmb{T}_{t \\to s}$ can be derived from:",
|
| 367 |
+
"bbox": [
|
| 368 |
+
212,
|
| 369 |
+
176,
|
| 370 |
+
787,
|
| 371 |
+
313
|
| 372 |
+
],
|
| 373 |
+
"page_idx": 4
|
| 374 |
+
},
|
| 375 |
+
{
|
| 376 |
+
"type": "equation",
|
| 377 |
+
"text": "\n$$\nD _ {t} = \\mathcal {D} \\left(\\mathcal {E} \\left(I _ {t}\\right)\\right), \\quad \\boldsymbol {T} _ {t \\rightarrow s} = \\mathcal {P} \\left(I _ {t}, I _ {s}\\right), \\tag {1}\n$$\n",
|
| 378 |
+
"text_format": "latex",
|
| 379 |
+
"bbox": [
|
| 380 |
+
377,
|
| 381 |
+
321,
|
| 382 |
+
784,
|
| 383 |
+
335
|
| 384 |
+
],
|
| 385 |
+
"page_idx": 4
|
| 386 |
+
},
|
| 387 |
+
{
|
| 388 |
+
"type": "text",
|
| 389 |
+
"text": "where $\\mathcal{E}$ and $\\mathcal{D}$ represent the encoder and decoder, respectively, of a single-frame depth network, while $\\mathcal{P}$ denotes the pose network. Then, the reconstructed view of $I_{t}$ from the source view $I_{s}$ can be obtained by:",
|
| 390 |
+
"bbox": [
|
| 391 |
+
212,
|
| 392 |
+
342,
|
| 393 |
+
784,
|
| 394 |
+
387
|
| 395 |
+
],
|
| 396 |
+
"page_idx": 4
|
| 397 |
+
},
|
| 398 |
+
{
|
| 399 |
+
"type": "equation",
|
| 400 |
+
"text": "\n$$\nI _ {s \\rightarrow t} = I _ {s} \\left\\langle \\operatorname {p r o j} \\left(D _ {t}, \\boldsymbol {T} _ {t \\rightarrow s}, \\boldsymbol {K}\\right)\\right\\rangle , \\tag {2}\n$$\n",
|
| 401 |
+
"text_format": "latex",
|
| 402 |
+
"bbox": [
|
| 403 |
+
390,
|
| 404 |
+
395,
|
| 405 |
+
784,
|
| 406 |
+
411
|
| 407 |
+
],
|
| 408 |
+
"page_idx": 4
|
| 409 |
+
},
|
| 410 |
+
{
|
| 411 |
+
"type": "text",
|
| 412 |
+
"text": "where $I_{s\\rightarrow t}$ is the reconstructed target image, $\\langle \\rangle$ means the differentiable bilinear sampling operator, $proj()$ returns the 2D coordinates of the pixels in $I_{t}$ after they are reprojected to the viewpoint of $I_{s}$ , and $K$ is the camera intrinsic matrix assumed to be known. Finally, the per-pixel photometric loss between $I_{s\\rightarrow t}$ and $I_{t}$ is used to optimize the depth and pose networks in an end-to-end manner. We follow Monodepth2 [18] to adopt the combination of L1 distance and SSIM [57] as the photometric error function and take the minimum across all the source views at each pixel to handle occlusion, which is formulated as:",
|
| 413 |
+
"bbox": [
|
| 414 |
+
212,
|
| 415 |
+
417,
|
| 416 |
+
787,
|
| 417 |
+
537
|
| 418 |
+
],
|
| 419 |
+
"page_idx": 4
|
| 420 |
+
},
|
| 421 |
+
{
|
| 422 |
+
"type": "equation",
|
| 423 |
+
"text": "\n$$\nL _ {p e} (t) = \\min _ {s} \\frac {\\alpha}{2} \\left(1 - \\operatorname {S S I M} \\left(I _ {t}, I _ {s \\rightarrow t}\\right)\\right) + (1 - \\alpha) \\| I _ {t}, I _ {s \\rightarrow t} \\| _ {1}, \\tag {3}\n$$\n",
|
| 424 |
+
"text_format": "latex",
|
| 425 |
+
"bbox": [
|
| 426 |
+
295,
|
| 427 |
+
544,
|
| 428 |
+
784,
|
| 429 |
+
568
|
| 430 |
+
],
|
| 431 |
+
"page_idx": 4
|
| 432 |
+
},
|
| 433 |
+
{
|
| 434 |
+
"type": "text",
|
| 435 |
+
"text": "where $\\alpha = 0.85$ . Besides, the edge-aware smoothness loss is employed to cope with depth discontinuities:",
|
| 436 |
+
"bbox": [
|
| 437 |
+
212,
|
| 438 |
+
573,
|
| 439 |
+
784,
|
| 440 |
+
602
|
| 441 |
+
],
|
| 442 |
+
"page_idx": 4
|
| 443 |
+
},
|
| 444 |
+
{
|
| 445 |
+
"type": "equation",
|
| 446 |
+
"text": "\n$$\nL _ {s m} (t) = \\left| \\partial_ {x} d _ {t} ^ {*} \\right| e ^ {- \\left| \\partial_ {x} I _ {t} \\right|} + \\left| \\partial_ {y} d _ {t} ^ {*} \\right| e ^ {- \\left| \\partial_ {y} I _ {t} \\right|}, \\tag {4}\n$$\n",
|
| 447 |
+
"text_format": "latex",
|
| 448 |
+
"bbox": [
|
| 449 |
+
361,
|
| 450 |
+
609,
|
| 451 |
+
784,
|
| 452 |
+
626
|
| 453 |
+
],
|
| 454 |
+
"page_idx": 4
|
| 455 |
+
},
|
| 456 |
+
{
|
| 457 |
+
"type": "text",
|
| 458 |
+
"text": "where $d_t^* = d_t / \\overline{d_t}$ is the mean-normalized inverse depth to discourage shrinking [53]. In summary, the standard self-supervised loss on $I_{t}$ is:",
|
| 459 |
+
"bbox": [
|
| 460 |
+
212,
|
| 461 |
+
632,
|
| 462 |
+
782,
|
| 463 |
+
662
|
| 464 |
+
],
|
| 465 |
+
"page_idx": 4
|
| 466 |
+
},
|
| 467 |
+
{
|
| 468 |
+
"type": "equation",
|
| 469 |
+
"text": "\n$$\nL _ {s s} (t) = \\mathcal {L} \\left(D _ {t}, \\boldsymbol {T} _ {t \\rightarrow s}, I _ {t}, I _ {s}\\right) = \\mu L _ {p e} (t) + \\gamma L _ {s m} (t), \\tag {5}\n$$\n",
|
| 470 |
+
"text_format": "latex",
|
| 471 |
+
"bbox": [
|
| 472 |
+
326,
|
| 473 |
+
671,
|
| 474 |
+
784,
|
| 475 |
+
686
|
| 476 |
+
],
|
| 477 |
+
"page_idx": 4
|
| 478 |
+
},
|
| 479 |
+
{
|
| 480 |
+
"type": "text",
|
| 481 |
+
"text": "where $\\mu$ is the auto-masking in [18] and $\\gamma = 0.001$ .",
|
| 482 |
+
"bbox": [
|
| 483 |
+
214,
|
| 484 |
+
691,
|
| 485 |
+
583,
|
| 486 |
+
708
|
| 487 |
+
],
|
| 488 |
+
"page_idx": 4
|
| 489 |
+
},
|
| 490 |
+
{
|
| 491 |
+
"type": "text",
|
| 492 |
+
"text": "4 Methodology",
|
| 493 |
+
"text_level": 1,
|
| 494 |
+
"bbox": [
|
| 495 |
+
215,
|
| 496 |
+
731,
|
| 497 |
+
380,
|
| 498 |
+
750
|
| 499 |
+
],
|
| 500 |
+
"page_idx": 4
|
| 501 |
+
},
|
| 502 |
+
{
|
| 503 |
+
"type": "text",
|
| 504 |
+
"text": "In this paper, we propose novel Mono-ViFI for self-supervised monocular training, which bilaterally connects single- and multi-frame depth models into a unified learning framework. With flow-based VFI model, we introduce temporal augmentation and develop VFI-assisted multi-frame fusion depth, which are presented in Sec. 4.1 and Sec. 4.2, respectively. Besides, our framework incorporates",
|
| 505 |
+
"bbox": [
|
| 506 |
+
212,
|
| 507 |
+
763,
|
| 508 |
+
787,
|
| 509 |
+
840
|
| 510 |
+
],
|
| 511 |
+
"page_idx": 4
|
| 512 |
+
},
|
| 513 |
+
{
|
| 514 |
+
"type": "header",
|
| 515 |
+
"text": "Mono-ViFi: A Unified Framework for Self-supervised Monocular Depth",
|
| 516 |
+
"bbox": [
|
| 517 |
+
256,
|
| 518 |
+
113,
|
| 519 |
+
732,
|
| 520 |
+
130
|
| 521 |
+
],
|
| 522 |
+
"page_idx": 4
|
| 523 |
+
},
|
| 524 |
+
{
|
| 525 |
+
"type": "page_number",
|
| 526 |
+
"text": "5",
|
| 527 |
+
"bbox": [
|
| 528 |
+
774,
|
| 529 |
+
116,
|
| 530 |
+
784,
|
| 531 |
+
126
|
| 532 |
+
],
|
| 533 |
+
"page_idx": 4
|
| 534 |
+
},
|
| 535 |
+
{
|
| 536 |
+
"type": "text",
|
| 537 |
+
"text": "spatial data augmentation and a triplet depth consistency loss, which are separately described in Sec. 4.3 and Sec. 4.4. Finally, we give the overall training objective in Sec. 4.5. The overview of our Mono-ViFi is depicted in Fig. 1.",
|
| 538 |
+
"bbox": [
|
| 539 |
+
212,
|
| 540 |
+
146,
|
| 541 |
+
787,
|
| 542 |
+
194
|
| 543 |
+
],
|
| 544 |
+
"page_idx": 5
|
| 545 |
+
},
|
| 546 |
+
{
|
| 547 |
+
"type": "text",
|
| 548 |
+
"text": "4.1 Temporal Augmentation with VFI",
|
| 549 |
+
"text_level": 1,
|
| 550 |
+
"bbox": [
|
| 551 |
+
215,
|
| 552 |
+
210,
|
| 553 |
+
547,
|
| 554 |
+
227
|
| 555 |
+
],
|
| 556 |
+
"page_idx": 5
|
| 557 |
+
},
|
| 558 |
+
{
|
| 559 |
+
"type": "text",
|
| 560 |
+
"text": "Firstly, we generate more new frames in the temporal dimension based on existing video frames for temporal augmentation. To achieve this, we adopt a pretrained flow-based VFI model, such as IFRNet [29], which takes two temporally adjacent frames $\\{I_0,I_2\\}$ as input to synthesize an intermediate frame $I_{1}$ . The frame interpolation process can be formulated as:",
|
| 561 |
+
"bbox": [
|
| 562 |
+
212,
|
| 563 |
+
233,
|
| 564 |
+
787,
|
| 565 |
+
309
|
| 566 |
+
],
|
| 567 |
+
"page_idx": 5
|
| 568 |
+
},
|
| 569 |
+
{
|
| 570 |
+
"type": "equation",
|
| 571 |
+
"text": "\n$$\nF _ {1 \\rightarrow 0}, F _ {1 \\rightarrow 2}, M _ {1} = \\mathcal {H} \\left(I _ {0}, I _ {2}\\right), \\tag {6}\n$$\n",
|
| 572 |
+
"text_format": "latex",
|
| 573 |
+
"bbox": [
|
| 574 |
+
395,
|
| 575 |
+
316,
|
| 576 |
+
784,
|
| 577 |
+
333
|
| 578 |
+
],
|
| 579 |
+
"page_idx": 5
|
| 580 |
+
},
|
| 581 |
+
{
|
| 582 |
+
"type": "equation",
|
| 583 |
+
"text": "\n$$\nI _ {0 \\rightarrow 1} = \\omega \\left(I _ {0}, F _ {1 \\rightarrow 0}\\right), \\quad I _ {2 \\rightarrow 1} = \\omega \\left(I _ {2}, F _ {1 \\rightarrow 2}\\right), \\tag {7}\n$$\n",
|
| 584 |
+
"text_format": "latex",
|
| 585 |
+
"bbox": [
|
| 586 |
+
346,
|
| 587 |
+
335,
|
| 588 |
+
784,
|
| 589 |
+
351
|
| 590 |
+
],
|
| 591 |
+
"page_idx": 5
|
| 592 |
+
},
|
| 593 |
+
{
|
| 594 |
+
"type": "equation",
|
| 595 |
+
"text": "\n$$\nI _ {1} = M _ {1} \\odot I _ {0 \\rightarrow 1} + (1 - M _ {1}) \\odot I _ {2 \\rightarrow 1}, \\tag {8}\n$$\n",
|
| 596 |
+
"text_format": "latex",
|
| 597 |
+
"bbox": [
|
| 598 |
+
372,
|
| 599 |
+
354,
|
| 600 |
+
784,
|
| 601 |
+
369
|
| 602 |
+
],
|
| 603 |
+
"page_idx": 5
|
| 604 |
+
},
|
| 605 |
+
{
|
| 606 |
+
"type": "text",
|
| 607 |
+
"text": "where $\\mathcal{H}$ is the VFI model, $F_{1\\rightarrow 0}$ and $F_{1\\rightarrow 2}$ are the intermediate optical flow, $\\omega$ represents backward warping. $I_{0\\rightarrow 1}$ and $I_{2\\rightarrow 1}$ denote the warped images to the intermediate position, while $M_{1}$ ranging from 0 to 1 is the merge mask adjusting the mixing ratio according to bidirectional occlusion information. Here, $\\odot$ indicates the element-wise multiplication. Considering a triplet $\\{I_{t - 2},I_t,I_{t + 2}\\}$ with three consecutive frames as described in Sec. 3, we use the VFI model to synthesize $I_{t - 1}$ from $\\{I_{t - 2},I_t\\}$ and $I_{t + 1}$ from $\\{I_t,I_{t + 2}\\}$ , as shown in Fig. 1(a). Then, we regard $I_{t - 1}$ and $I_{t + 1}$ as another two target frames. All three target frames have common sources of $\\{I_{t - 2},I_{t + 2}\\}$ . The following operations and losses are the same for each of them, and we only take $I_{t}$ for instance.",
|
| 608 |
+
"bbox": [
|
| 609 |
+
212,
|
| 610 |
+
376,
|
| 611 |
+
787,
|
| 612 |
+
527
|
| 613 |
+
],
|
| 614 |
+
"page_idx": 5
|
| 615 |
+
},
|
| 616 |
+
{
|
| 617 |
+
"type": "text",
|
| 618 |
+
"text": "4.2 VFI-assisted Multi-frame Fusion Depth",
|
| 619 |
+
"text_level": 1,
|
| 620 |
+
"bbox": [
|
| 621 |
+
215,
|
| 622 |
+
547,
|
| 623 |
+
588,
|
| 624 |
+
561
|
| 625 |
+
],
|
| 626 |
+
"page_idx": 5
|
| 627 |
+
},
|
| 628 |
+
{
|
| 629 |
+
"type": "text",
|
| 630 |
+
"text": "Apart from temporal augmentation, our employed VFI model can also provide other inter-frame information, such as optical flow and occlusion merge masks. This enables us to design a novel VFI-assisted multi-frame fusion module for multi-frame depth prediction, as demonstrated in Fig. 2. The fusion module consists of two parts, i.e., motion-aware feature alignment (MAFA) and occlusion-alarviated feature fusion (OAFF). Taking original triplet $\\{I_{t - 2},I_t,I_{t + 2}\\}$ for example, we implement forward process of the VFI model to interpolate the middle position $t$ of $\\{I_{t - 2},I_{t + 2}\\}$ in Eq. (9), yielding intermediate flow $F_{t\\rightarrow (t - 2)},F_{t\\rightarrow (t + 2)}$ and merge mask $M_{t}$ to align and aggregate multi-frame features.",
|
| 631 |
+
"bbox": [
|
| 632 |
+
212,
|
| 633 |
+
569,
|
| 634 |
+
787,
|
| 635 |
+
705
|
| 636 |
+
],
|
| 637 |
+
"page_idx": 5
|
| 638 |
+
},
|
| 639 |
+
{
|
| 640 |
+
"type": "equation",
|
| 641 |
+
"text": "\n$$\nF _ {t \\rightarrow (t - 2)}, F _ {t \\rightarrow (t + 2)}, M _ {t} = \\mathcal {H} \\left(I _ {t - 2}, I _ {t + 2}\\right). \\tag {9}\n$$\n",
|
| 642 |
+
"text_format": "latex",
|
| 643 |
+
"bbox": [
|
| 644 |
+
362,
|
| 645 |
+
713,
|
| 646 |
+
784,
|
| 647 |
+
729
|
| 648 |
+
],
|
| 649 |
+
"page_idx": 5
|
| 650 |
+
},
|
| 651 |
+
{
|
| 652 |
+
"type": "text",
|
| 653 |
+
"text": "Multi-frame Feature Extraction. Our fusion module accepts multi-frame features extracted by the same encoder, which is shared with the single-frame depth encoder $\\mathcal{E}$ . The extraction process can be written as:",
|
| 654 |
+
"bbox": [
|
| 655 |
+
212,
|
| 656 |
+
734,
|
| 657 |
+
787,
|
| 658 |
+
780
|
| 659 |
+
],
|
| 660 |
+
"page_idx": 5
|
| 661 |
+
},
|
| 662 |
+
{
|
| 663 |
+
"type": "equation",
|
| 664 |
+
"text": "\n$$\n\\left\\{\\phi_ {t - 2} ^ {k}, \\phi_ {t} ^ {k}, \\phi_ {t + 2} ^ {k} \\right\\} _ {k = 1} ^ {N} = \\left\\{\\mathcal {E} \\left(I _ {t - 2}\\right), \\mathcal {E} \\left(I _ {t}\\right), \\mathcal {E} \\left(I _ {t + 2}\\right) \\right\\}, \\tag {10}\n$$\n",
|
| 665 |
+
"text_format": "latex",
|
| 666 |
+
"bbox": [
|
| 667 |
+
334,
|
| 668 |
+
786,
|
| 669 |
+
784,
|
| 670 |
+
803
|
| 671 |
+
],
|
| 672 |
+
"page_idx": 5
|
| 673 |
+
},
|
| 674 |
+
{
|
| 675 |
+
"type": "text",
|
| 676 |
+
"text": "where $\\{\\phi_{t - 2}^k,\\phi_t^k,\\phi_{t + 2}^k\\}_{k = 1}^N$ are the multi-level features of the three frames and $N$ is the number of feature levels.",
|
| 677 |
+
"bbox": [
|
| 678 |
+
212,
|
| 679 |
+
808,
|
| 680 |
+
787,
|
| 681 |
+
839
|
| 682 |
+
],
|
| 683 |
+
"page_idx": 5
|
| 684 |
+
},
|
| 685 |
+
{
|
| 686 |
+
"type": "page_number",
|
| 687 |
+
"text": "6",
|
| 688 |
+
"bbox": [
|
| 689 |
+
217,
|
| 690 |
+
114,
|
| 691 |
+
228,
|
| 692 |
+
126
|
| 693 |
+
],
|
| 694 |
+
"page_idx": 5
|
| 695 |
+
},
|
| 696 |
+
{
|
| 697 |
+
"type": "header",
|
| 698 |
+
"text": "J. Liu et al.",
|
| 699 |
+
"bbox": [
|
| 700 |
+
271,
|
| 701 |
+
114,
|
| 702 |
+
351,
|
| 703 |
+
127
|
| 704 |
+
],
|
| 705 |
+
"page_idx": 5
|
| 706 |
+
},
|
| 707 |
+
{
|
| 708 |
+
"type": "image",
|
| 709 |
+
"img_path": "images/5a8c29e39bf5ffda6cab944146caef336883dae0319f1957a820946db98e53dc.jpg",
|
| 710 |
+
"image_caption": [
|
| 711 |
+
"Fig. 2: Our VFI-assisted multi-frame fusion depth model. Motion and occlusion provided by VFI can be used to align and fuse multi-frame features in an explicit manner."
|
| 712 |
+
],
|
| 713 |
+
"image_footnote": [],
|
| 714 |
+
"bbox": [
|
| 715 |
+
215,
|
| 716 |
+
145,
|
| 717 |
+
787,
|
| 718 |
+
309
|
| 719 |
+
],
|
| 720 |
+
"page_idx": 6
|
| 721 |
+
},
|
| 722 |
+
{
|
| 723 |
+
"type": "text",
|
| 724 |
+
"text": "Motion-aware Feature Alignment (MAFA). For the $k$ -th level, we use bidirectional optical flow to align the neighbor features $\\phi_{t-2}^k$ and $\\phi_{t+2}^k$ to the target position $t$ , generating $\\phi_{(t-2) \\to t}^k$ and $\\phi_{(t+2) \\to t}^k$ , respectively, formulated as:",
|
| 725 |
+
"bbox": [
|
| 726 |
+
214,
|
| 727 |
+
385,
|
| 728 |
+
782,
|
| 729 |
+
434
|
| 730 |
+
],
|
| 731 |
+
"page_idx": 6
|
| 732 |
+
},
|
| 733 |
+
{
|
| 734 |
+
"type": "equation",
|
| 735 |
+
"text": "\n$$\n\\begin{array}{l} \\phi_ {(t - 2) \\rightarrow t} ^ {k} = \\omega \\left(\\phi_ {t - 2} ^ {k}, F _ {t \\rightarrow (t - 2)} ^ {k}\\right), \\\\ \\phi_ {(t + 2) \\rightarrow t} ^ {k} = \\omega \\left(\\phi_ {t + 2} ^ {k}, F _ {t \\rightarrow (t + 2)} ^ {k}\\right), \\tag {11} \\\\ \\end{array}\n$$\n",
|
| 736 |
+
"text_format": "latex",
|
| 737 |
+
"bbox": [
|
| 738 |
+
388,
|
| 739 |
+
439,
|
| 740 |
+
784,
|
| 741 |
+
481
|
| 742 |
+
],
|
| 743 |
+
"page_idx": 6
|
| 744 |
+
},
|
| 745 |
+
{
|
| 746 |
+
"type": "text",
|
| 747 |
+
"text": "where $F_{t\\rightarrow (t - 2)}^k$ and $F_{t\\rightarrow (t + 2)}^k$ denote the corresponding flows of the $k$ -th level, downsampled from $F_{t\\rightarrow (t - 2)}$ and $F_{t\\rightarrow (t + 2)}$ , respectively. Inspired by recent 2D implicit neural representation [8], we also concatenate the aligned features with encoded optical flow, which can be denoted as:",
|
| 748 |
+
"bbox": [
|
| 749 |
+
214,
|
| 750 |
+
487,
|
| 751 |
+
782,
|
| 752 |
+
547
|
| 753 |
+
],
|
| 754 |
+
"page_idx": 6
|
| 755 |
+
},
|
| 756 |
+
{
|
| 757 |
+
"type": "equation",
|
| 758 |
+
"text": "\n$$\n\\begin{array}{l} \\varphi_ {(t - 2) \\rightarrow t} ^ {k} = \\operatorname {C o n c a t} \\left(\\phi_ {(t - 2) \\rightarrow t} ^ {k}, g \\left(F _ {t \\rightarrow (t - 2)} ^ {k}\\right)\\right), \\\\ \\varphi_ {(t + 2) \\rightarrow t} ^ {k} = \\operatorname {C o n c a t} \\left(\\phi_ {(t + 2) \\rightarrow t} ^ {k}, g \\left(F _ {t \\rightarrow (t + 2)} ^ {k}\\right)\\right), \\tag {12} \\\\ \\varphi_ {t} ^ {k} = \\operatorname {C o n c a t} \\left(\\phi_ {t} ^ {k}, g \\left(F _ {0} ^ {k}\\right)\\right), \\\\ \\end{array}\n$$\n",
|
| 759 |
+
"text_format": "latex",
|
| 760 |
+
"bbox": [
|
| 761 |
+
344,
|
| 762 |
+
553,
|
| 763 |
+
784,
|
| 764 |
+
614
|
| 765 |
+
],
|
| 766 |
+
"page_idx": 6
|
| 767 |
+
},
|
| 768 |
+
{
|
| 769 |
+
"type": "text",
|
| 770 |
+
"text": "where Concat means feature concatenation, $F_0^k$ denotes zero flow and $g$ is fourier positional encoding [41,51] to transform optical flow to high-frequency signal. Actually, optical flow indicates the relative coordinate of a feature vector in the warped feature map w.r.t. the original one. When incorporating it, multi-frame model can adaptively adjust feature learning through motion-aware alignment.",
|
| 771 |
+
"bbox": [
|
| 772 |
+
214,
|
| 773 |
+
623,
|
| 774 |
+
782,
|
| 775 |
+
698
|
| 776 |
+
],
|
| 777 |
+
"page_idx": 6
|
| 778 |
+
},
|
| 779 |
+
{
|
| 780 |
+
"type": "text",
|
| 781 |
+
"text": "Occlusion-alleviated Feature Fusion (OAFF). Due to the ubiquitous interframe occlusion, aggregating the aligned features directly through summation or concatenation can deteriorate the performance. Therefore, we use the $k$ -th level merge mask $M_t^k$ downsampled from $M_t$ to mitigate this issue, just like the image merging process of VFI in Eq. (8):",
|
| 782 |
+
"bbox": [
|
| 783 |
+
214,
|
| 784 |
+
702,
|
| 785 |
+
782,
|
| 786 |
+
777
|
| 787 |
+
],
|
| 788 |
+
"page_idx": 6
|
| 789 |
+
},
|
| 790 |
+
{
|
| 791 |
+
"type": "equation",
|
| 792 |
+
"text": "\n$$\n\\chi_ {t} ^ {k} = M _ {t} ^ {k} \\odot \\varphi_ {(t - 2) \\rightarrow t} ^ {k} + (1 - M _ {t} ^ {k}) \\odot \\varphi_ {(t + 2) \\rightarrow t} ^ {k}, \\tag {13}\n$$\n",
|
| 793 |
+
"text_format": "latex",
|
| 794 |
+
"bbox": [
|
| 795 |
+
336,
|
| 796 |
+
782,
|
| 797 |
+
784,
|
| 798 |
+
801
|
| 799 |
+
],
|
| 800 |
+
"page_idx": 6
|
| 801 |
+
},
|
| 802 |
+
{
|
| 803 |
+
"type": "text",
|
| 804 |
+
"text": "where $\\chi_t^k$ is the occlusion-alleviated feature merged from the temporal positions $t - 2$ and $t + 2$ . It is then concatenated with the feature $\\varphi_t^k$ in the position $t$ .",
|
| 805 |
+
"bbox": [
|
| 806 |
+
214,
|
| 807 |
+
809,
|
| 808 |
+
782,
|
| 809 |
+
839
|
| 810 |
+
],
|
| 811 |
+
"page_idx": 6
|
| 812 |
+
},
|
| 813 |
+
{
|
| 814 |
+
"type": "header",
|
| 815 |
+
"text": "Mono-ViFi: A Unified Framework for Self-supervised Monocular Depth",
|
| 816 |
+
"bbox": [
|
| 817 |
+
256,
|
| 818 |
+
113,
|
| 819 |
+
732,
|
| 820 |
+
130
|
| 821 |
+
],
|
| 822 |
+
"page_idx": 6
|
| 823 |
+
},
|
| 824 |
+
{
|
| 825 |
+
"type": "page_number",
|
| 826 |
+
"text": "7",
|
| 827 |
+
"bbox": [
|
| 828 |
+
774,
|
| 829 |
+
114,
|
| 830 |
+
784,
|
| 831 |
+
126
|
| 832 |
+
],
|
| 833 |
+
"page_idx": 6
|
| 834 |
+
},
|
| 835 |
+
{
|
| 836 |
+
"type": "text",
|
| 837 |
+
"text": "And we can obtain the final feature after fusion by:",
|
| 838 |
+
"bbox": [
|
| 839 |
+
215,
|
| 840 |
+
146,
|
| 841 |
+
584,
|
| 842 |
+
161
|
| 843 |
+
],
|
| 844 |
+
"page_idx": 7
|
| 845 |
+
},
|
| 846 |
+
{
|
| 847 |
+
"type": "equation",
|
| 848 |
+
"text": "\n$$\n\\psi_ {t} ^ {k} = \\operatorname {C o n v} _ {1 \\times 1} \\left(\\operatorname {C o n c a t} \\left(\\varphi_ {t} ^ {k}, \\chi_ {t} ^ {k}\\right)\\right), \\tag {14}\n$$\n",
|
| 849 |
+
"text_format": "latex",
|
| 850 |
+
"bbox": [
|
| 851 |
+
385,
|
| 852 |
+
167,
|
| 853 |
+
785,
|
| 854 |
+
185
|
| 855 |
+
],
|
| 856 |
+
"page_idx": 7
|
| 857 |
+
},
|
| 858 |
+
{
|
| 859 |
+
"type": "text",
|
| 860 |
+
"text": "where $\\mathrm{Conv}_{1\\times 1}$ is a $1\\times 1$ convolutional layer to transform the feature into the same channel as the original feature $\\phi_t^k$ , enabling the multi-frame depth decoder to employ the same architecture as the single-frame one.",
|
| 861 |
+
"bbox": [
|
| 862 |
+
212,
|
| 863 |
+
190,
|
| 864 |
+
782,
|
| 865 |
+
236
|
| 866 |
+
],
|
| 867 |
+
"page_idx": 7
|
| 868 |
+
},
|
| 869 |
+
{
|
| 870 |
+
"type": "text",
|
| 871 |
+
"text": "Multi-frame Depth Decoder. To translate the multi-level fusion features into a depth map, a decoder is required. Concretely, we use the same architecture as the single-frame depth decoder, but the parameters are not shared. This encourages the single- and multi-frame models to decode the features in their individual ways without conflict. Finally, the fusion features $\\{\\psi_t^k\\}_{k=1}^N$ are processed by a multi-frame depth decoder $\\mathcal{D}_m$ to generate the fusion depth $D_t^m$ of the target frame $I_t$ , which can be written as:",
|
| 872 |
+
"bbox": [
|
| 873 |
+
212,
|
| 874 |
+
238,
|
| 875 |
+
784,
|
| 876 |
+
342
|
| 877 |
+
],
|
| 878 |
+
"page_idx": 7
|
| 879 |
+
},
|
| 880 |
+
{
|
| 881 |
+
"type": "equation",
|
| 882 |
+
"text": "\n$$\nD _ {t} ^ {m} = \\mathcal {D} _ {m} \\left(\\left\\{\\psi_ {t} ^ {k} \\right\\} _ {k = 1} ^ {N}\\right). \\tag {15}\n$$\n",
|
| 883 |
+
"text_format": "latex",
|
| 884 |
+
"bbox": [
|
| 885 |
+
426,
|
| 886 |
+
349,
|
| 887 |
+
785,
|
| 888 |
+
367
|
| 889 |
+
],
|
| 890 |
+
"page_idx": 7
|
| 891 |
+
},
|
| 892 |
+
{
|
| 893 |
+
"type": "text",
|
| 894 |
+
"text": "Note that the multi-frame depth $D_{t}^{m}$ also corresponds to a self-supervised loss like Eq. (5), which is denoted as:",
|
| 895 |
+
"bbox": [
|
| 896 |
+
212,
|
| 897 |
+
373,
|
| 898 |
+
782,
|
| 899 |
+
402
|
| 900 |
+
],
|
| 901 |
+
"page_idx": 7
|
| 902 |
+
},
|
| 903 |
+
{
|
| 904 |
+
"type": "equation",
|
| 905 |
+
"text": "\n$$\nL _ {s s} ^ {m} (t) = \\mathcal {L} \\left(D _ {t} ^ {m}, \\boldsymbol {T} _ {t \\rightarrow s}, I _ {t}, I _ {s}\\right). \\tag {16}\n$$\n",
|
| 906 |
+
"text_format": "latex",
|
| 907 |
+
"bbox": [
|
| 908 |
+
398,
|
| 909 |
+
411,
|
| 910 |
+
785,
|
| 911 |
+
426
|
| 912 |
+
],
|
| 913 |
+
"page_idx": 7
|
| 914 |
+
},
|
| 915 |
+
{
|
| 916 |
+
"type": "text",
|
| 917 |
+
"text": "4.3 Spatial Augmentation",
|
| 918 |
+
"text_level": 1,
|
| 919 |
+
"bbox": [
|
| 920 |
+
215,
|
| 921 |
+
455,
|
| 922 |
+
442,
|
| 923 |
+
472
|
| 924 |
+
],
|
| 925 |
+
"page_idx": 7
|
| 926 |
+
},
|
| 927 |
+
{
|
| 928 |
+
"type": "text",
|
| 929 |
+
"text": "Orthogonal to temporal augmentation, we also include classic data augmentation to spatially expand the data diversity, which is categorized as spatial augmentation in this paper. Resizing-cropping is a commonly used and effective augmentation approach. Based on this operation, we introduce image rotation, extending it to a generic affine transformation (cropping is actually image translation). Specifically, we follow [55] to consider that the augmented image and the original image are captured by the same camera system whose intrinsics are fixed. For resizing operation, it is assumed that when an image is zoomed up by a factor $f_{s} \\geq 1$ , the relative size of an object increases by the same factor $f_{s}$ and its depth decreases by $f_{s}$ simultaneously. This is inspired from a vital monocular cue [4] that the closer an object is, the larger its relative size is. Our affine transformation $\\mathcal{A}$ is applied to both the target image $I_{t}$ and the source image $I_{s}$ , generating $\\widetilde{I}_{t} = \\mathcal{A}(I_{t})$ and $\\widetilde{I}_{s} = \\mathcal{A}(I_{s})$ , respectively. For the augmented target view, we can obtain its depth through the single-frame model by $\\widetilde{D}_{t} = \\mathcal{D}(\\mathcal{E}(\\widetilde{I}_{t}))$ . To conduct view synthesis on augmented images, the original camera pose $T_{t\\rightarrow s} = [\\pmb{R}_{t\\rightarrow s}|\\pmb{t}_{t\\rightarrow s}]$ from target to source should be rectified to $\\widetilde{T}_{t\\rightarrow s} = [\\widetilde{\\pmb{R}}_{t\\rightarrow s}|\\widetilde{\\pmb{t}}_{t\\rightarrow s}]$ as done in [55]:",
|
| 930 |
+
"bbox": [
|
| 931 |
+
212,
|
| 932 |
+
489,
|
| 933 |
+
787,
|
| 934 |
+
750
|
| 935 |
+
],
|
| 936 |
+
"page_idx": 7
|
| 937 |
+
},
|
| 938 |
+
{
|
| 939 |
+
"type": "equation",
|
| 940 |
+
"text": "\n$$\n\\tilde {\\boldsymbol {R}} _ {t \\rightarrow s} = \\boldsymbol {R} _ {c} \\boldsymbol {R} _ {t \\rightarrow s} \\boldsymbol {R} _ {c} ^ {- 1}, \\quad \\tilde {\\boldsymbol {t}} _ {t \\rightarrow s} = \\boldsymbol {R} _ {c} \\boldsymbol {t} _ {t \\rightarrow s}, \\tag {17}\n$$\n",
|
| 941 |
+
"text_format": "latex",
|
| 942 |
+
"bbox": [
|
| 943 |
+
354,
|
| 944 |
+
756,
|
| 945 |
+
785,
|
| 946 |
+
773
|
| 947 |
+
],
|
| 948 |
+
"page_idx": 7
|
| 949 |
+
},
|
| 950 |
+
{
|
| 951 |
+
"type": "text",
|
| 952 |
+
"text": "where $\\mathbf{R}_c$ is the rectification matrix determined only by the camera intrinsics and the augmentation parameters. Different from [55] calculating $\\mathbf{R}_c$ under resizing-cropping, we give its formulation and derivation under a more general affine transformation, which can be seen in our supplementary. Therefore, we can derive",
|
| 953 |
+
"bbox": [
|
| 954 |
+
212,
|
| 955 |
+
779,
|
| 956 |
+
787,
|
| 957 |
+
840
|
| 958 |
+
],
|
| 959 |
+
"page_idx": 7
|
| 960 |
+
},
|
| 961 |
+
{
|
| 962 |
+
"type": "page_number",
|
| 963 |
+
"text": "8",
|
| 964 |
+
"bbox": [
|
| 965 |
+
217,
|
| 966 |
+
114,
|
| 967 |
+
228,
|
| 968 |
+
126
|
| 969 |
+
],
|
| 970 |
+
"page_idx": 7
|
| 971 |
+
},
|
| 972 |
+
{
|
| 973 |
+
"type": "header",
|
| 974 |
+
"text": "J. Liu et al.",
|
| 975 |
+
"bbox": [
|
| 976 |
+
271,
|
| 977 |
+
114,
|
| 978 |
+
351,
|
| 979 |
+
127
|
| 980 |
+
],
|
| 981 |
+
"page_idx": 7
|
| 982 |
+
},
|
| 983 |
+
{
|
| 984 |
+
"type": "text",
|
| 985 |
+
"text": "the self-supervised loss $\\widetilde{L}_{ss}(t)$ on the augmented images similarly to Eq. (5):",
|
| 986 |
+
"bbox": [
|
| 987 |
+
215,
|
| 988 |
+
145,
|
| 989 |
+
763,
|
| 990 |
+
162
|
| 991 |
+
],
|
| 992 |
+
"page_idx": 8
|
| 993 |
+
},
|
| 994 |
+
{
|
| 995 |
+
"type": "equation",
|
| 996 |
+
"text": "\n$$\n\\widetilde {L} _ {s s} (t) = \\mathcal {L} (\\widetilde {D} _ {t}, \\widetilde {\\boldsymbol {T}} _ {t \\rightarrow s}, \\widetilde {I} _ {t}, \\widetilde {I} _ {s}). \\tag {18}\n$$\n",
|
| 997 |
+
"text_format": "latex",
|
| 998 |
+
"bbox": [
|
| 999 |
+
401,
|
| 1000 |
+
167,
|
| 1001 |
+
784,
|
| 1002 |
+
186
|
| 1003 |
+
],
|
| 1004 |
+
"page_idx": 8
|
| 1005 |
+
},
|
| 1006 |
+
{
|
| 1007 |
+
"type": "text",
|
| 1008 |
+
"text": "Note that there is a mask to exclude the possible invalid pixels in $\\widetilde{I}_t$ when calculating $\\widetilde{L}_{ss}(t)$ due to image rotation.",
|
| 1009 |
+
"bbox": [
|
| 1010 |
+
214,
|
| 1011 |
+
193,
|
| 1012 |
+
785,
|
| 1013 |
+
224
|
| 1014 |
+
],
|
| 1015 |
+
"page_idx": 8
|
| 1016 |
+
},
|
| 1017 |
+
{
|
| 1018 |
+
"type": "text",
|
| 1019 |
+
"text": "4.4 Triplet Depth Consistency",
|
| 1020 |
+
"text_level": 1,
|
| 1021 |
+
"bbox": [
|
| 1022 |
+
215,
|
| 1023 |
+
243,
|
| 1024 |
+
480,
|
| 1025 |
+
258
|
| 1026 |
+
],
|
| 1027 |
+
"page_idx": 8
|
| 1028 |
+
},
|
| 1029 |
+
{
|
| 1030 |
+
"type": "text",
|
| 1031 |
+
"text": "Now we have single-frame depths $D_{t}$ and $\\widetilde{D}_{t}$ of the original image $I_{t}$ and the augmented image $\\widetilde{I}_{t}$ , along with multi-frame depth $D_{t}^{m}$ of $I_{t}$ . For better knowledge distillation and regularization, our unified learning framework applies a triplet depth consistency loss among the three depth maps, including a standard view depth consistency (SVDC) loss and two scale-aware depth consistency (SADC) losses, which are shown in Fig. 1(b). The SVDC loss $L_{sv}(t)$ conducts mutual distillation between single- and multi-frame models, promoting their mutual reinforcement, which can be denoted as:",
|
| 1032 |
+
"bbox": [
|
| 1033 |
+
214,
|
| 1034 |
+
266,
|
| 1035 |
+
787,
|
| 1036 |
+
387
|
| 1037 |
+
],
|
| 1038 |
+
"page_idx": 8
|
| 1039 |
+
},
|
| 1040 |
+
{
|
| 1041 |
+
"type": "equation",
|
| 1042 |
+
"text": "\n$$\nL _ {s v} (t) = \\operatorname {S I} \\left(D _ {t} ^ {m}, D _ {t}\\right), \\tag {19}\n$$\n",
|
| 1043 |
+
"text_format": "latex",
|
| 1044 |
+
"bbox": [
|
| 1045 |
+
426,
|
| 1046 |
+
395,
|
| 1047 |
+
784,
|
| 1048 |
+
411
|
| 1049 |
+
],
|
| 1050 |
+
"page_idx": 8
|
| 1051 |
+
},
|
| 1052 |
+
{
|
| 1053 |
+
"type": "text",
|
| 1054 |
+
"text": "where $\\mathrm{SI}()$ is the widely used scale-invariant error function [13]. Besides, we use depth $\\widetilde{D}_t$ under spatial augmentation to enforce two SADC losses, i.e., $L_{sa}(t)$ between $\\widetilde{D}_t$ and $D_t$ , and $L_{sa}^m(t)$ between $\\widetilde{D}_t$ and $D_t^m$ . To this end, we first restore $\\widetilde{D}_t$ to the original image view, generating $\\widehat{D}_t = \\mathcal{A}^{-1}(\\widetilde{D}_t)$ in Fig. 1(b), where $\\mathcal{A}^{-1}$ is the inverse transformation of affine augmentation $\\mathcal{A}$ . To exclude the invalid pixels in $\\widehat{D}_t$ , we adopt a mask $M_{sa}$ and calculate $L_{sa}(t)$ and $L_{sa}^m(t)$ through:",
|
| 1055 |
+
"bbox": [
|
| 1056 |
+
214,
|
| 1057 |
+
416,
|
| 1058 |
+
787,
|
| 1059 |
+
513
|
| 1060 |
+
],
|
| 1061 |
+
"page_idx": 8
|
| 1062 |
+
},
|
| 1063 |
+
{
|
| 1064 |
+
"type": "equation",
|
| 1065 |
+
"text": "\n$$\nL _ {s a} (t) = \\operatorname {S I} \\left(M _ {s a} \\odot D _ {t}, M _ {s a} \\odot f _ {s} \\widehat {D} _ {t}\\right), \\tag {20}\n$$\n",
|
| 1066 |
+
"text_format": "latex",
|
| 1067 |
+
"bbox": [
|
| 1068 |
+
372,
|
| 1069 |
+
532,
|
| 1070 |
+
784,
|
| 1071 |
+
550
|
| 1072 |
+
],
|
| 1073 |
+
"page_idx": 8
|
| 1074 |
+
},
|
| 1075 |
+
{
|
| 1076 |
+
"type": "equation",
|
| 1077 |
+
"text": "\n$$\nL _ {s a} ^ {m} (t) = \\operatorname {S I} \\left(M _ {s a} \\odot D _ {t} ^ {m}, M _ {s a} \\odot f _ {s} \\widehat {D} _ {t}\\right), \\tag {21}\n$$\n",
|
| 1078 |
+
"text_format": "latex",
|
| 1079 |
+
"bbox": [
|
| 1080 |
+
369,
|
| 1081 |
+
555,
|
| 1082 |
+
784,
|
| 1083 |
+
571
|
| 1084 |
+
],
|
| 1085 |
+
"page_idx": 8
|
| 1086 |
+
},
|
| 1087 |
+
{
|
| 1088 |
+
"type": "text",
|
| 1089 |
+
"text": "where the factor $f_{s}$ is due to the assumption in Sec. 4.3 that the depth of an object in $I_{t}$ is $f_{s}$ times of that in $\\widetilde{I}_{t}$ . Our SADC constrains the 3D geometric relationship between the scenes of $I_{t}$ and $\\widetilde{I}_{t}$ , providing scale-aware regularization. Combining SVDC and SADC, we have the triplet depth consistency loss as:",
|
| 1090 |
+
"bbox": [
|
| 1091 |
+
214,
|
| 1092 |
+
578,
|
| 1093 |
+
787,
|
| 1094 |
+
638
|
| 1095 |
+
],
|
| 1096 |
+
"page_idx": 8
|
| 1097 |
+
},
|
| 1098 |
+
{
|
| 1099 |
+
"type": "equation",
|
| 1100 |
+
"text": "\n$$\nL _ {t c} (t) = L _ {s v} (t) + L _ {s a} (t) + L _ {s a} ^ {m} (t). \\tag {22}\n$$\n",
|
| 1101 |
+
"text_format": "latex",
|
| 1102 |
+
"bbox": [
|
| 1103 |
+
382,
|
| 1104 |
+
646,
|
| 1105 |
+
784,
|
| 1106 |
+
662
|
| 1107 |
+
],
|
| 1108 |
+
"page_idx": 8
|
| 1109 |
+
},
|
| 1110 |
+
{
|
| 1111 |
+
"type": "text",
|
| 1112 |
+
"text": "4.5 Training Objective",
|
| 1113 |
+
"text_level": 1,
|
| 1114 |
+
"bbox": [
|
| 1115 |
+
215,
|
| 1116 |
+
680,
|
| 1117 |
+
418,
|
| 1118 |
+
696
|
| 1119 |
+
],
|
| 1120 |
+
"page_idx": 8
|
| 1121 |
+
},
|
| 1122 |
+
{
|
| 1123 |
+
"type": "text",
|
| 1124 |
+
"text": "For target position $t$ , we have already obtained the self-supervised losses $L_{ss}(t)$ , $L_{ss}^{m}(t)$ and $\\widetilde{L}_{ss}(t)$ in Eqs. (5), (16) and (18), respectively. Besides, our triplet depth consistency loss $L_{tc}(t)$ is acquired in Eq. (22). Therefore, the total loss on position $t$ can be summed as:",
|
| 1125 |
+
"bbox": [
|
| 1126 |
+
214,
|
| 1127 |
+
703,
|
| 1128 |
+
785,
|
| 1129 |
+
765
|
| 1130 |
+
],
|
| 1131 |
+
"page_idx": 8
|
| 1132 |
+
},
|
| 1133 |
+
{
|
| 1134 |
+
"type": "equation",
|
| 1135 |
+
"text": "\n$$\nL (t) = L _ {s s} (t) + L _ {s s} ^ {m} (t) + \\widetilde {L} _ {s s} (t) + \\lambda L _ {t c} (t), \\tag {23}\n$$\n",
|
| 1136 |
+
"text_format": "latex",
|
| 1137 |
+
"bbox": [
|
| 1138 |
+
354,
|
| 1139 |
+
770,
|
| 1140 |
+
784,
|
| 1141 |
+
787
|
| 1142 |
+
],
|
| 1143 |
+
"page_idx": 8
|
| 1144 |
+
},
|
| 1145 |
+
{
|
| 1146 |
+
"type": "text",
|
| 1147 |
+
"text": "where $\\lambda$ is a hyper-parameter set to 0.2 in this paper. For $t - 1$ and $t + 1$ , we can derive their losses $L(t - 1)$ and $L(t + 1)$ in the same way. In summary, the overall loss function of our Mono-ViFI is $L = L(t - 1) + L(t) + L(t + 1)$ .",
|
| 1148 |
+
"bbox": [
|
| 1149 |
+
214,
|
| 1150 |
+
794,
|
| 1151 |
+
787,
|
| 1152 |
+
840
|
| 1153 |
+
],
|
| 1154 |
+
"page_idx": 8
|
| 1155 |
+
},
|
| 1156 |
+
{
|
| 1157 |
+
"type": "header",
|
| 1158 |
+
"text": "Mono-ViFi: A Unified Framework for Self-supervised Monocular Depth",
|
| 1159 |
+
"bbox": [
|
| 1160 |
+
254,
|
| 1161 |
+
113,
|
| 1162 |
+
730,
|
| 1163 |
+
128
|
| 1164 |
+
],
|
| 1165 |
+
"page_idx": 8
|
| 1166 |
+
},
|
| 1167 |
+
{
|
| 1168 |
+
"type": "page_number",
|
| 1169 |
+
"text": "9",
|
| 1170 |
+
"bbox": [
|
| 1171 |
+
774,
|
| 1172 |
+
114,
|
| 1173 |
+
784,
|
| 1174 |
+
126
|
| 1175 |
+
],
|
| 1176 |
+
"page_idx": 8
|
| 1177 |
+
},
|
| 1178 |
+
{
|
| 1179 |
+
"type": "text",
|
| 1180 |
+
"text": "Table 1: Performance comparison on KITTI [16] benchmark using the raw ground truth, with a resolution of $640 \\times 192$ . The best results in each subsection are in bold; second best are underlined. For MACs of multi-frame models, we use the formulation A/B. A denotes the MACs when predicting for a single frame; B means the average MACs per frame when processing a long video sequence. #Params and MACs of our multi-frame models include the VFI network.",
|
| 1181 |
+
"bbox": [
|
| 1182 |
+
212,
|
| 1183 |
+
143,
|
| 1184 |
+
787,
|
| 1185 |
+
227
|
| 1186 |
+
],
|
| 1187 |
+
"page_idx": 9
|
| 1188 |
+
},
|
| 1189 |
+
{
|
| 1190 |
+
"type": "table",
|
| 1191 |
+
"img_path": "images/5be4c37c825e8a0cf365029b8fd0f56453e49802107cdf6d98fe3832c9b17e38.jpg",
|
| 1192 |
+
"table_caption": [],
|
| 1193 |
+
"table_footnote": [],
|
| 1194 |
+
"table_body": "<table><tr><td>Method</td><td>Test frames</td><td>#Params</td><td>MACs</td><td>Abs Rel</td><td>Sq Rel</td><td>RMSE</td><td>RMSE log</td><td>δ1</td><td>δ2</td><td>δ3</td></tr><tr><td>PackNet-SfM [21]</td><td>1</td><td>128M</td><td>205G</td><td>0.111</td><td>0.785</td><td>4.601</td><td>0.189</td><td>0.878</td><td>0.960</td><td>0.982</td></tr><tr><td>HR-Depth [39]</td><td>1</td><td>16.9M</td><td>31.6G</td><td>0.109</td><td>0.792</td><td>4.632</td><td>0.185</td><td>0.884</td><td>0.962</td><td>0.983</td></tr><tr><td>R-MSFM6 [67]</td><td>1</td><td>3.8M</td><td>31.2G</td><td>0.112</td><td>0.806</td><td>4.704</td><td>0.191</td><td>0.878</td><td>0.960</td><td>0.981</td></tr><tr><td>BRNet [22]</td><td>1</td><td>19.1M</td><td>33.7G</td><td>0.105</td><td>0.698</td><td>4.462</td><td>0.179</td><td>0.890</td><td>0.965</td><td>0.984</td></tr><tr><td>MonoViT [64]</td><td>1</td><td>27M</td><td>-</td><td>0.099</td><td>0.708</td><td>4.372</td><td>0.175</td><td>0.900</td><td>0.967</td><td>0.984</td></tr><tr><td>DIFFNet [65]</td><td>1</td><td>12.2M</td><td>23.4G</td><td>0.102</td><td>0.764</td><td>4.483</td><td>0.180</td><td>0.896</td><td>0.965</td><td>0.983</td></tr><tr><td>SD-SSMDE [44]</td><td>1</td><td>-</td><td>-</td><td>0.100</td><td>0.661</td><td>4.264</td><td>0.172</td><td>0.896</td><td>0.967</td><td>0.985</td></tr><tr><td>MUSTNet [48]</td><td>1</td><td>-</td><td>-</td><td>0.106</td><td>0.763</td><td>4.562</td><td>0.182</td><td>0.888</td><td>0.963</td><td>0.983</td></tr><tr><td>DepthSegNet [40]</td><td>1</td><td>30.3M</td><td>-</td><td>0.104</td><td>0.690</td><td>4.473</td><td>0.179</td><td>0.886</td><td>0.965</td><td>0.984</td></tr><tr><td>DualRefine [2]</td><td>1</td><td>-</td><td>-</td><td>0.103</td><td>0.776</td><td>4.491</td><td>0.181</td><td>0.894</td><td>0.965</td><td>0.983</td></tr><tr><td>DaCNN [23]</td><td>1</td><td>13M</td><td>-</td><td>0.099</td><td>0.661</td><td>4.316</td><td>0.173</td><td>0.897</td><td>0.967</td><td>0.985</td></tr><tr><td>Monodepth2 (ResNet18) [18]</td><td>1</td><td>14.3M</td><td>8G</td><td>0.115</td><td>0.903</td><td>4.863</td><td>0.193</td><td>0.877</td><td>0.959</td><td>0.981</td></tr><tr><td>Mono-ViFI (ResNet18)</td><td>1</td><td>14.3M</td><td>8G</td><td>0.105</td><td>0.708</td><td>4.446</td><td>0.179</td><td>0.887</td><td>0.965</td><td>0.984</td></tr><tr><td>Lite-Mono [62]</td><td>1</td><td>3.1M</td><td>5.1G</td><td>0.107</td><td>0.765</td><td>4.561</td><td>0.183</td><td>0.886</td><td>0.963</td><td>0.983</td></tr><tr><td>Mono-ViFI (Lite-Mono)</td><td>1</td><td>3.1M</td><td>5.1G</td><td>0.103</td><td>0.680</td><td>4.349</td><td>0.176</td><td>0.891</td><td>0.966</td><td>0.985</td></tr><tr><td>D-HRNet [25]</td><td>1</td><td>10M</td><td>10.8G</td><td>0.102</td><td>0.760</td><td>4.479</td><td>0.179</td><td>0.897</td><td>0.965</td><td>0.983</td></tr><tr><td>RA-Depth (D-HRNet) [25]</td><td>1</td><td>10M</td><td>10.8G</td><td>0.096</td><td>0.632</td><td>4.216</td><td>0.171</td><td>0.903</td><td>0.968</td><td>0.985</td></tr><tr><td>Mono-ViFI (D-HRNet)</td><td>1</td><td>10M</td><td>10.8G</td><td>0.096</td><td>0.627</td><td>4.179</td><td>0.170</td><td>0.903</td><td>0.969</td><td>0.985</td></tr><tr><td>Patil et al. [42]</td><td>N</td><td>-</td><td>-/-</td><td>0.111</td><td>0.821</td><td>4.650</td><td>0.187</td><td>0.883</td><td>0.961</td><td>0.982</td></tr><tr><td>ManyDepth (ResNet18) [58]</td><td>2 (-1, 0)</td><td>26.9M</td><td>15.1G/13.7G</td><td>0.098</td><td>0.770</td><td>4.459</td><td>0.176</td><td>0.900</td><td>0.965</td><td>0.983</td></tr><tr><td>DynamicDepth [14]</td><td>2 (-1, 0)</td><td>-</td><td>-/-</td><td>0.096</td><td>0.720</td><td>4.458</td><td>0.175</td><td>0.897</td><td>0.964</td><td>0.984</td></tr><tr><td>DepthFormer [20]</td><td>2 (-1, 0)</td><td>28.7M</td><td>175G/-</td><td>0.090</td><td>0.661</td><td>4.149</td><td>0.175</td><td>0.905</td><td>0.967</td><td>0.984</td></tr><tr><td>MOVEDepth [56]</td><td>2 (-1, 0)</td><td>28.2M</td><td>20.2G/19.2G</td><td>0.094</td><td>0.704</td><td>4.389</td><td>0.175</td><td>0.902</td><td>0.965</td><td>0.983</td></tr><tr><td>DualRefine [2]</td><td>2 (-1, 0)</td><td>-</td><td>-/-</td><td>0.087</td><td>0.698</td><td>4.234</td><td>0.170</td><td>0.914</td><td>0.967</td><td>0.983</td></tr><tr><td>TC-Depth [49]</td><td>3 (-1, 0, 1)</td><td>-</td><td>-/-</td><td>0.103</td><td>0.746</td><td>4.483</td><td>0.185</td><td>0.894</td><td>-</td><td>0.983</td></tr><tr><td>Mono-ViFI (ResNet18)</td><td>3 (-1, 0, 1)</td><td>17.9M</td><td>25.8G/16.9G</td><td>0.099</td><td>0.661</td><td>4.321</td><td>0.174</td><td>0.898</td><td>0.966</td><td>0.985</td></tr><tr><td>Mono-ViFI (Lite-Mono)</td><td>3 (-1, 0, 1)</td><td>5.9M</td><td>21.9G/13.3G</td><td>0.099</td><td>0.648</td><td>4.256</td><td>0.172</td><td>0.898</td><td>0.967</td><td>0.985</td></tr><tr><td>Mono-ViFI (D-HRNet)</td><td>3 (-1, 0, 1)</td><td>12.9M</td><td>35.7G/19.4G</td><td>0.091</td><td>0.589</td><td>4.088</td><td>0.166</td><td>0.912</td><td>0.969</td><td>0.986</td></tr></table>",
|
| 1195 |
+
"bbox": [
|
| 1196 |
+
220,
|
| 1197 |
+
250,
|
| 1198 |
+
787,
|
| 1199 |
+
537
|
| 1200 |
+
],
|
| 1201 |
+
"page_idx": 9
|
| 1202 |
+
},
|
| 1203 |
+
{
|
| 1204 |
+
"type": "text",
|
| 1205 |
+
"text": "5 Experiments",
|
| 1206 |
+
"text_level": 1,
|
| 1207 |
+
"bbox": [
|
| 1208 |
+
215,
|
| 1209 |
+
574,
|
| 1210 |
+
375,
|
| 1211 |
+
593
|
| 1212 |
+
],
|
| 1213 |
+
"page_idx": 9
|
| 1214 |
+
},
|
| 1215 |
+
{
|
| 1216 |
+
"type": "text",
|
| 1217 |
+
"text": "5.1 Implementation Details",
|
| 1218 |
+
"text_level": 1,
|
| 1219 |
+
"bbox": [
|
| 1220 |
+
215,
|
| 1221 |
+
604,
|
| 1222 |
+
457,
|
| 1223 |
+
619
|
| 1224 |
+
],
|
| 1225 |
+
"page_idx": 9
|
| 1226 |
+
},
|
| 1227 |
+
{
|
| 1228 |
+
"type": "text",
|
| 1229 |
+
"text": "Datasets. We train and evaluate the models on two widely used outdoor urban datasets of KITTI [16] and Cityscapes [10]. The depth labels of KITTI are collected with LiDARs. For fair comparison, we use the Eigen split [12] and follow the pre-processing operation in [18], generating 39,810 monocular triplets for training and 697 frames for evaluation. To reduce the influence of noise in the sparse depth from Velodyne, we also evaluate using $93\\%$ of the Eigen split with the improved ground truth from [52], which contains 652 test frames. In regard to Cityscapes, following [58, 66], we preprocess the monocular sequences into triples, producing 69,731 training images. And we evaluate on 1,525 test images using the ground truth provided by [58]. For generalization comparison, KITTI models are evaluated on Cityscapes and Make3D [50]. Make3D is a small outdoor dataset containing 134 test images with aligned depth information.",
|
| 1230 |
+
"bbox": [
|
| 1231 |
+
212,
|
| 1232 |
+
628,
|
| 1233 |
+
787,
|
| 1234 |
+
809
|
| 1235 |
+
],
|
| 1236 |
+
"page_idx": 9
|
| 1237 |
+
},
|
| 1238 |
+
{
|
| 1239 |
+
"type": "text",
|
| 1240 |
+
"text": "Flow-based VFI Model. We choose IFRNet [29] as our VFI model due to its efficiency. In practice, we use the large model of IFRNet for training and replace",
|
| 1241 |
+
"bbox": [
|
| 1242 |
+
212,
|
| 1243 |
+
809,
|
| 1244 |
+
785,
|
| 1245 |
+
839
|
| 1246 |
+
],
|
| 1247 |
+
"page_idx": 9
|
| 1248 |
+
},
|
| 1249 |
+
{
|
| 1250 |
+
"type": "page_number",
|
| 1251 |
+
"text": "10",
|
| 1252 |
+
"bbox": [
|
| 1253 |
+
217,
|
| 1254 |
+
114,
|
| 1255 |
+
235,
|
| 1256 |
+
126
|
| 1257 |
+
],
|
| 1258 |
+
"page_idx": 9
|
| 1259 |
+
},
|
| 1260 |
+
{
|
| 1261 |
+
"type": "header",
|
| 1262 |
+
"text": "J. Liu et al.",
|
| 1263 |
+
"bbox": [
|
| 1264 |
+
271,
|
| 1265 |
+
114,
|
| 1266 |
+
351,
|
| 1267 |
+
126
|
| 1268 |
+
],
|
| 1269 |
+
"page_idx": 9
|
| 1270 |
+
},
|
| 1271 |
+
{
|
| 1272 |
+
"type": "image",
|
| 1273 |
+
"img_path": "images/99e79450c6426335bf4c0e9386f8b6e7c1850d1ad49ca9df9568c42be55984ec.jpg",
|
| 1274 |
+
"image_caption": [
|
| 1275 |
+
"Fig. 3: Qualitative results on KITTI. Error maps in columns 2 and 4 show the Abs Rel error compared to the improved ground truth [52] from good (blue) to bad (red)."
|
| 1276 |
+
],
|
| 1277 |
+
"image_footnote": [],
|
| 1278 |
+
"bbox": [
|
| 1279 |
+
217,
|
| 1280 |
+
145,
|
| 1281 |
+
395,
|
| 1282 |
+
305
|
| 1283 |
+
],
|
| 1284 |
+
"page_idx": 10
|
| 1285 |
+
},
|
| 1286 |
+
{
|
| 1287 |
+
"type": "image",
|
| 1288 |
+
"img_path": "images/0da88bfe2f0d99496838e7378941a00260474aa275d552af0216217e93a6d7e8.jpg",
|
| 1289 |
+
"image_caption": [],
|
| 1290 |
+
"image_footnote": [],
|
| 1291 |
+
"bbox": [
|
| 1292 |
+
395,
|
| 1293 |
+
146,
|
| 1294 |
+
517,
|
| 1295 |
+
304
|
| 1296 |
+
],
|
| 1297 |
+
"page_idx": 10
|
| 1298 |
+
},
|
| 1299 |
+
{
|
| 1300 |
+
"type": "image",
|
| 1301 |
+
"img_path": "images/3c94a3e3f9f98c4879c9225deb6ca36e4b8fe91f56f484a1376167f9532d1156.jpg",
|
| 1302 |
+
"image_caption": [],
|
| 1303 |
+
"image_footnote": [],
|
| 1304 |
+
"bbox": [
|
| 1305 |
+
519,
|
| 1306 |
+
146,
|
| 1307 |
+
651,
|
| 1308 |
+
304
|
| 1309 |
+
],
|
| 1310 |
+
"page_idx": 10
|
| 1311 |
+
},
|
| 1312 |
+
{
|
| 1313 |
+
"type": "image",
|
| 1314 |
+
"img_path": "images/283d98377f6cc2f608e38498ec70a8b43581e3270bbefc364a23cbbba0a2309c.jpg",
|
| 1315 |
+
"image_caption": [],
|
| 1316 |
+
"image_footnote": [],
|
| 1317 |
+
"bbox": [
|
| 1318 |
+
651,
|
| 1319 |
+
148,
|
| 1320 |
+
782,
|
| 1321 |
+
304
|
| 1322 |
+
],
|
| 1323 |
+
"page_idx": 10
|
| 1324 |
+
},
|
| 1325 |
+
{
|
| 1326 |
+
"type": "text",
|
| 1327 |
+
"text": "it with the small model when evaluating for multi-frame fusion depth. Please see ablation about their gap in supplementary. Both large and small IFRNets are pretrained on KITTI/Cityscapes and frozen in Mono-ViFI training. The pretraining data split is the same as self-supervised depth training. Each original triplet in depth training is directly used for VFI training. Besides, we employ another two flow-based VFI networks, RIFE [26] and AMT [34], to verify the robustness of our method to different VFI models.",
|
| 1328 |
+
"bbox": [
|
| 1329 |
+
212,
|
| 1330 |
+
372,
|
| 1331 |
+
787,
|
| 1332 |
+
477
|
| 1333 |
+
],
|
| 1334 |
+
"page_idx": 10
|
| 1335 |
+
},
|
| 1336 |
+
{
|
| 1337 |
+
"type": "text",
|
| 1338 |
+
"text": "Depth and Pose Networks. We experiment with three depth architectures, including the ResNet18 [24] based network in Monodepth2 [18], medium LiteMono [62], and D-HRNet proposed in RA-Depth [25]. For the pose network, we follow Monodepth2 to use a ResNet18 based model which takes a pair of images (six channels) as input and predict a 6-DoF relative pose. Encoders of all the networks are pretrained on ImageNet [11].",
|
| 1339 |
+
"bbox": [
|
| 1340 |
+
212,
|
| 1341 |
+
479,
|
| 1342 |
+
787,
|
| 1343 |
+
570
|
| 1344 |
+
],
|
| 1345 |
+
"page_idx": 10
|
| 1346 |
+
},
|
| 1347 |
+
{
|
| 1348 |
+
"type": "text",
|
| 1349 |
+
"text": "5.2 Performance Comparison",
|
| 1350 |
+
"text_level": 1,
|
| 1351 |
+
"bbox": [
|
| 1352 |
+
214,
|
| 1353 |
+
589,
|
| 1354 |
+
472,
|
| 1355 |
+
604
|
| 1356 |
+
],
|
| 1357 |
+
"page_idx": 10
|
| 1358 |
+
},
|
| 1359 |
+
{
|
| 1360 |
+
"type": "text",
|
| 1361 |
+
"text": "KITTI Results. Quantitative comparison results on KITTI under $640 \\times 192$ resolution are listed in Table 1 and Table 2, using raw ground truth and improved ground truth, respectively. It is obvious that our method with D-HRNet can achieve state-of-the-art performance. Besides, our Mono-ViFI brings notable improvements to the three depth architectures without increasing model complexity (see each subsection separated by dotted lines). Despite that the difference between Mono-ViFI (D-HRNet) and RA-Depth [25] under single-frame mode is slight in Table 1, it is more significant in Table 2 when using improved ground truth. For multi-frame inference, Mono-ViFI with ResNet18 has already shown advantages over ManyDepth [58] and DynamicDepth [14] on some metrics including Sq Rel, RMSE, RMSE log, $\\delta_{2}$ and $\\delta_{3}$ . When changing to D-HRNet backbone, proposed Mono-ViFI can totally outperform them. Although our models have more multiply-add computations (MACs) when predicting for a single frame, the average MACs per frame are comparable when processing a long video sequence. Fig. 3 depicts two visual comparison results on KITTI, where",
|
| 1362 |
+
"bbox": [
|
| 1363 |
+
212,
|
| 1364 |
+
613,
|
| 1365 |
+
787,
|
| 1366 |
+
840
|
| 1367 |
+
],
|
| 1368 |
+
"page_idx": 10
|
| 1369 |
+
},
|
| 1370 |
+
{
|
| 1371 |
+
"type": "header",
|
| 1372 |
+
"text": "Mono-ViFi: A Unified Framework for Self-supervised Monocular Depth",
|
| 1373 |
+
"bbox": [
|
| 1374 |
+
256,
|
| 1375 |
+
113,
|
| 1376 |
+
730,
|
| 1377 |
+
128
|
| 1378 |
+
],
|
| 1379 |
+
"page_idx": 10
|
| 1380 |
+
},
|
| 1381 |
+
{
|
| 1382 |
+
"type": "page_number",
|
| 1383 |
+
"text": "11",
|
| 1384 |
+
"bbox": [
|
| 1385 |
+
767,
|
| 1386 |
+
114,
|
| 1387 |
+
782,
|
| 1388 |
+
126
|
| 1389 |
+
],
|
| 1390 |
+
"page_idx": 10
|
| 1391 |
+
},
|
| 1392 |
+
{
|
| 1393 |
+
"type": "table",
|
| 1394 |
+
"img_path": "images/435e95fff156baa792e511883d37b23e8e8a2f35072e483c0c45f816e4ef97a9.jpg",
|
| 1395 |
+
"table_caption": [
|
| 1396 |
+
"Table 2: Performance comparison on Table 3: Performance comparison on KITTI [16] using improved ground truth Cityscapes [10] with a resolution of $416 \\times$ from [52], with a resolution of $640 \\times 192$ . 128 unless stated."
|
| 1397 |
+
],
|
| 1398 |
+
"table_footnote": [],
|
| 1399 |
+
"table_body": "<table><tr><td>Method</td><td>Abs Rel</td><td>Sq Rel</td><td>RMSE</td><td>RMSE log</td><td>δ1</td><td>δ2</td><td>δ3</td></tr><tr><td colspan=\"8\">Single-frame Inference</td></tr><tr><td>PackNet-SIM [21]</td><td>0.078</td><td>0.420</td><td>3.485</td><td>0.121</td><td>0.931</td><td>0.986</td><td>0.996</td></tr><tr><td>R-MSF M6 [67]</td><td>0.088</td><td>0.492</td><td>3.837</td><td>0.135</td><td>0.915</td><td>0.983</td><td>0.995</td></tr><tr><td>DIFFNet [65]</td><td>0.076</td><td>0.414</td><td>3.405</td><td>0.119</td><td>0.936</td><td>0.988</td><td>0.996</td></tr><tr><td>Mono-ViF [64]</td><td>0.074</td><td>0.388</td><td>3.414</td><td>0.115</td><td>0.938</td><td>0.989</td><td>0.997</td></tr><tr><td>Monodepth2 (ResNet18) [18]</td><td>0.090</td><td>0.545</td><td>3.942</td><td>0.137</td><td>0.914</td><td>0.983</td><td>0.995</td></tr><tr><td>Mono-ViF (ResNet18)</td><td>0.084</td><td>0.431</td><td>3.618</td><td>0.126</td><td>0.924</td><td>0.986</td><td>0.997</td></tr><tr><td>Lite-Mono [62]</td><td>0.082</td><td>0.455</td><td>3.685</td><td>0.127</td><td>0.923</td><td>0.985</td><td>0.996</td></tr><tr><td>Mono-ViF (Lite-Mono)</td><td>0.080</td><td>0.400</td><td>3.497</td><td>0.121</td><td>0.930</td><td>0.987</td><td>0.997</td></tr><tr><td>D-HRNet [25]</td><td>0.077</td><td>0.423</td><td>3.396</td><td>0.119</td><td>0.935</td><td>0.987</td><td>0.996</td></tr><tr><td>RA-Depth (D-HRNet) [25]</td><td>0.074</td><td>0.363</td><td>3.349</td><td>0.114</td><td>0.940</td><td>0.990</td><td>0.997</td></tr><tr><td>Mono-ViF (D-HRNet)</td><td>0.072</td><td>0.342</td><td>3.254</td><td>0.111</td><td>0.941</td><td>0.991</td><td>0.998</td></tr><tr><td colspan=\"8\">Multi-frame Inference</td></tr><tr><td>ManyDepth (ResNet18) [58]</td><td>0.070</td><td>0.399</td><td>3.455</td><td>0.113</td><td>0.941</td><td>0.989</td><td>0.997</td></tr><tr><td>MOVEDepth [56]</td><td>0.065</td><td>0.377</td><td>3.449</td><td>0.112</td><td>0.942</td><td>0.988</td><td>0.996</td></tr><tr><td>Mono-ViF (ResNet18)</td><td>0.074</td><td>0.374</td><td>3.381</td><td>0.113</td><td>0.941</td><td>0.990</td><td>0.997</td></tr><tr><td>Mono-ViF (Lite-Mono)</td><td>0.074</td><td>0.354</td><td>3.372</td><td>0.114</td><td>0.939</td><td>0.989</td><td>0.997</td></tr><tr><td>Mono-ViF (D-HRNet)</td><td>0.066</td><td>0.294</td><td>3.099</td><td>0.102</td><td>0.952</td><td>0.992</td><td>0.998</td></tr></table>",
|
| 1400 |
+
"bbox": [
|
| 1401 |
+
225,
|
| 1402 |
+
227,
|
| 1403 |
+
496,
|
| 1404 |
+
358
|
| 1405 |
+
],
|
| 1406 |
+
"page_idx": 11
|
| 1407 |
+
},
|
| 1408 |
+
{
|
| 1409 |
+
"type": "table",
|
| 1410 |
+
"img_path": "images/e9bc3f513579ccdbf6b94249042eb8622a3c7da720244114eb709a8573d7254c.jpg",
|
| 1411 |
+
"table_caption": [],
|
| 1412 |
+
"table_footnote": [],
|
| 1413 |
+
"table_body": "<table><tr><td>Method</td><td>Abs Rel</td><td>Sq Rel</td><td>RMSE</td><td>RMSE log</td><td>δ1</td><td>δ2</td><td>δ3</td></tr><tr><td colspan=\"8\">Single-frame Inference</td></tr><tr><td>Videos in the wild [19]</td><td>0.127</td><td>1.330</td><td>6.960</td><td>0.195</td><td>0.830</td><td>0.947</td><td>0.981</td></tr><tr><td>Li et al. [33]</td><td>0.119</td><td>1.290</td><td>6.980</td><td>0.190</td><td>0.846</td><td>0.952</td><td>0.982</td></tr><tr><td>InstDM [31] (832 × 256)</td><td>0.111</td><td>1.158</td><td>6.437</td><td>0.182</td><td>0.868</td><td>0.961</td><td>0.983</td></tr><tr><td>SD-SSMDE [44]</td><td>0.110</td><td>0.988</td><td>5.953</td><td>0.165</td><td>-</td><td>-</td><td>-</td></tr><tr><td>Monodepth2 (ResNet18) [18]</td><td>0.129</td><td>1.569</td><td>6.376</td><td>0.197</td><td>0.849</td><td>0.957</td><td>0.983</td></tr><tr><td>Mono-ViFi (ResNet18)</td><td>0.112</td><td>1.003</td><td>5.899</td><td>0.165</td><td>0.872</td><td>0.966</td><td>0.990</td></tr><tr><td>Lite-Mono [62]</td><td>0.114</td><td>1.194</td><td>6.328</td><td>0.172</td><td>0.871</td><td>0.967</td><td>0.989</td></tr><tr><td>Mono-ViFi (Lite-Mono)</td><td>0.107</td><td>0.940</td><td>5.794</td><td>0.159</td><td>0.876</td><td>0.968</td><td>0.991</td></tr><tr><td>D-HRNet [25]</td><td>0.111</td><td>1.142</td><td>6.205</td><td>0.169</td><td>0.879</td><td>0.968</td><td>0.989</td></tr><tr><td>Mono-ViFi (D-HRNet)</td><td>0.105</td><td>0.880</td><td>5.550</td><td>0.156</td><td>0.881</td><td>0.969</td><td>0.992</td></tr><tr><td colspan=\"8\">Multi-frame Inference</td></tr><tr><td>ManyDepth (ResNet18) [58]</td><td>0.114</td><td>1.193</td><td>6.223</td><td>0.170</td><td>0.875</td><td>0.967</td><td>0.989</td></tr><tr><td>DynamicDepth [14]</td><td>0.104</td><td>1.011</td><td>5.987</td><td>0.159</td><td>0.889</td><td>0.972</td><td>0.991</td></tr><tr><td>TC-Depth [49]</td><td>0.110</td><td>0.958</td><td>5.820</td><td>-</td><td>0.867</td><td>-</td><td>0.991</td></tr><tr><td>Mono-ViFi (ResNet18)</td><td>0.107</td><td>0.934</td><td>5.688</td><td>0.159</td><td>0.880</td><td>0.969</td><td>0.990</td></tr><tr><td>Mono-ViFi (Lite-Mono)</td><td>0.105</td><td>0.904</td><td>5.646</td><td>0.157</td><td>0.880</td><td>0.970</td><td>0.991</td></tr><tr><td>Mono-ViFi (D-HRNet)</td><td>0.102</td><td>0.836</td><td>5.395</td><td>0.153</td><td>0.885</td><td>0.971</td><td>0.992</td></tr></table>",
|
| 1414 |
+
"bbox": [
|
| 1415 |
+
506,
|
| 1416 |
+
227,
|
| 1417 |
+
779,
|
| 1418 |
+
358
|
| 1419 |
+
],
|
| 1420 |
+
"page_idx": 11
|
| 1421 |
+
},
|
| 1422 |
+
{
|
| 1423 |
+
"type": "table",
|
| 1424 |
+
"img_path": "images/20f3189acd7c8cdae6073b6e3aed93ff243e1048e4d833e70a54d3eb35192845.jpg",
|
| 1425 |
+
"table_caption": [
|
| 1426 |
+
"Table 4: Evaluation results on dynamic objects and static regions of Cityscapes [10] dataset under $416\\times 128$ resolution. Note that DynamicDepth relies on depth prior and segmentation masks of dynamic objects when reasoning."
|
| 1427 |
+
],
|
| 1428 |
+
"table_footnote": [],
|
| 1429 |
+
"table_body": "<table><tr><td>Method</td><td>Test Area</td><td>Abs Rel</td><td>Sq Rel</td><td>RMSE</td><td>RMSE log</td><td>δ1</td><td>δ2</td><td>δ3</td></tr><tr><td colspan=\"9\">Multi-frame Inference</td></tr><tr><td rowspan=\"3\">ManyDepth (ResNet18) [58]DynamicDepth [14]Mono-ViFI (ResNet18)</td><td rowspan=\"3\">Dynamic Objects</td><td>0.172</td><td>2.206</td><td>5.924</td><td>0.216</td><td>0.792</td><td>0.917</td><td>0.967</td></tr><tr><td>0.144</td><td>1.554</td><td>5.238</td><td>0.185</td><td>0.831</td><td>0.950</td><td>0.982</td></tr><tr><td>0.162</td><td>1.453</td><td>5.114</td><td>0.199</td><td>0.794</td><td>0.924</td><td>0.977</td></tr><tr><td rowspan=\"3\">ManyDepth (ResNet18) [58]DynamicDepth [14]Mono-ViFI (ResNet18)</td><td rowspan=\"3\">Static Regions</td><td>0.107</td><td>1.090</td><td>6.314</td><td>0.160</td><td>0.882</td><td>0.972</td><td>0.991</td></tr><tr><td>0.100</td><td>0.977</td><td>6.155</td><td>0.155</td><td>0.892</td><td>0.974</td><td>0.992</td></tr><tr><td>0.099</td><td>0.907</td><td>5.835</td><td>0.149</td><td>0.890</td><td>0.975</td><td>0.992</td></tr></table>",
|
| 1430 |
+
"bbox": [
|
| 1431 |
+
261,
|
| 1432 |
+
440,
|
| 1433 |
+
743,
|
| 1434 |
+
523
|
| 1435 |
+
],
|
| 1436 |
+
"page_idx": 11
|
| 1437 |
+
},
|
| 1438 |
+
{
|
| 1439 |
+
"type": "text",
|
| 1440 |
+
"text": "we choose D-HRNet as the baseline depth model. It can be seen that depth maps look qualitatively similar among all four methods. However, the error maps reveal mistakes of D-HRNet baseline and RA-Depth in some ambiguous regions like bushes (see the red boxes). In contrast, our Mono-ViFi have dramatically lower error in these regions, especially under multi-frame reasoning.",
|
| 1441 |
+
"bbox": [
|
| 1442 |
+
212,
|
| 1443 |
+
566,
|
| 1444 |
+
784,
|
| 1445 |
+
642
|
| 1446 |
+
],
|
| 1447 |
+
"page_idx": 11
|
| 1448 |
+
},
|
| 1449 |
+
{
|
| 1450 |
+
"type": "text",
|
| 1451 |
+
"text": "Cityscapes Results. Results of Cityscapes in Table 3 demonstrate the same trend as KITTI. Moreover, proposed multi-frame model with ResNet18 can beat ManyDepth on all metrics, while our Mono-ViFI is better than DynamicDepth only on partial indexes. It is because Cityscapes with more dynamic scenes is more challenging than KITTI, and DynamicDepth leverages additional information to cope with dynamic objects. For fine-grained analysis, we separately evaluate on the dynamic objects and static regions, using the segmentation masks of dynamic objects provided in [14], as listed in Table 4. On dynamic objects, Mono-ViFI with ResNet18 surpasses ManyDepth by a large margin, demonstrating that our multi-frame inference paradigm can sidestep the issue of dynamic objects to a greater extent compared to ManyDepth. DynamicDepth utilizes depth prior and segmentation masks of dynamic objects to tackle this problem for multi-frame inference. Even though, our method still excels it on Sq Rel and",
|
| 1452 |
+
"bbox": [
|
| 1453 |
+
212,
|
| 1454 |
+
643,
|
| 1455 |
+
787,
|
| 1456 |
+
840
|
| 1457 |
+
],
|
| 1458 |
+
"page_idx": 11
|
| 1459 |
+
},
|
| 1460 |
+
{
|
| 1461 |
+
"type": "page_number",
|
| 1462 |
+
"text": "12",
|
| 1463 |
+
"bbox": [
|
| 1464 |
+
217,
|
| 1465 |
+
114,
|
| 1466 |
+
235,
|
| 1467 |
+
126
|
| 1468 |
+
],
|
| 1469 |
+
"page_idx": 11
|
| 1470 |
+
},
|
| 1471 |
+
{
|
| 1472 |
+
"type": "header",
|
| 1473 |
+
"text": "J. Liu et al.",
|
| 1474 |
+
"bbox": [
|
| 1475 |
+
271,
|
| 1476 |
+
114,
|
| 1477 |
+
351,
|
| 1478 |
+
126
|
| 1479 |
+
],
|
| 1480 |
+
"page_idx": 11
|
| 1481 |
+
},
|
| 1482 |
+
{
|
| 1483 |
+
"type": "image",
|
| 1484 |
+
"img_path": "images/b055474670ed0ea897e8507ce5a1717e48236a09eacf6bed41360cfcb8c2708b.jpg",
|
| 1485 |
+
"image_caption": [
|
| 1486 |
+
"Fig. 4: Qualitative generalization results of a YouTube video example. All depth maps are predicted under $640 \\times 360$ resolution using KITTI models trained under $640 \\times 192$ ."
|
| 1487 |
+
],
|
| 1488 |
+
"image_footnote": [],
|
| 1489 |
+
"bbox": [
|
| 1490 |
+
254,
|
| 1491 |
+
146,
|
| 1492 |
+
750,
|
| 1493 |
+
308
|
| 1494 |
+
],
|
| 1495 |
+
"page_idx": 12
|
| 1496 |
+
},
|
| 1497 |
+
{
|
| 1498 |
+
"type": "table",
|
| 1499 |
+
"img_path": "images/473ae66e2f84105a645934f03f578f72612f2632967d3f61d15e1e3511764b58.jpg",
|
| 1500 |
+
"table_caption": [
|
| 1501 |
+
"Table 5: Ablation results of each proposed component on Cityscapes [10]. SA is spatial augmentation. TA is temporal augmentation. SVDC denotes standard view depth consistency. SADC denotes scale-aware depth consistency. FA is feature alignment. MAFA is motion-aware feature alignment. OAFF means occlusion-alleviated feature fusion."
|
| 1502 |
+
],
|
| 1503 |
+
"table_footnote": [],
|
| 1504 |
+
"table_body": "<table><tr><td>Method</td><td>SA</td><td>TA</td><td>SVDC</td><td>SADC</td><td>FA</td><td>MAFA</td><td>OAFF</td><td>Abs Rel</td><td>Sq Rel</td><td>RMSE</td><td>δ1</td></tr><tr><td colspan=\"12\">Single-frame Inference</td></tr><tr><td>Baseline (Monodepth2 [18])</td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td>0.129</td><td>1.569</td><td>6.876</td><td>0.849</td></tr><tr><td>Baseline + TA</td><td></td><td>✓</td><td></td><td></td><td></td><td></td><td></td><td>0.119</td><td>1.254</td><td>6.523</td><td>0.857</td></tr><tr><td>Baseline + SVDC</td><td></td><td></td><td>✓</td><td></td><td></td><td></td><td></td><td>0.117</td><td>1.186</td><td>6.492</td><td>0.862</td></tr><tr><td>Baseline + SA</td><td>✓</td><td></td><td></td><td></td><td></td><td></td><td></td><td>0.122</td><td>1.386</td><td>6.557</td><td>0.856</td></tr><tr><td>Baseline + SA + SADC</td><td>✓</td><td></td><td></td><td>✓</td><td></td><td></td><td></td><td>0.116</td><td>1.079</td><td>6.189</td><td>0.868</td></tr><tr><td>Mono-ViFI (Single-frame)</td><td>✓</td><td>✓</td><td>✓</td><td>✓</td><td></td><td></td><td></td><td>0.112</td><td>1.003</td><td>5.899</td><td>0.872</td></tr><tr><td colspan=\"12\">Multi-frame Inference</td></tr><tr><td>Baseline (direct feature fusion)</td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td>0.120</td><td>1.360</td><td>6.491</td><td>0.869</td></tr><tr><td>FA (not motion-aware)</td><td></td><td></td><td></td><td></td><td>✓</td><td></td><td></td><td>0.116</td><td>1.217</td><td>6.387</td><td>0.871</td></tr><tr><td>MAFA</td><td></td><td></td><td></td><td></td><td></td><td>✓</td><td></td><td>0.114</td><td>1.150</td><td>6.115</td><td>0.874</td></tr><tr><td>MAFA + OAFF</td><td></td><td></td><td></td><td></td><td></td><td>✓</td><td>✓</td><td>0.111</td><td>1.107</td><td>6.012</td><td>0.876</td></tr><tr><td>MAFA + OAFF + TA</td><td></td><td>✓</td><td></td><td></td><td></td><td>✓</td><td>✓</td><td>0.109</td><td>1.023</td><td>5.892</td><td>0.877</td></tr><tr><td>Mono-ViFI (Multi-frame)</td><td>✓</td><td>✓</td><td>✓</td><td>✓</td><td></td><td>✓</td><td>✓</td><td>0.107</td><td>0.934</td><td>5.688</td><td>0.880</td></tr></table>",
|
| 1505 |
+
"bbox": [
|
| 1506 |
+
263,
|
| 1507 |
+
431,
|
| 1508 |
+
743,
|
| 1509 |
+
594
|
| 1510 |
+
],
|
| 1511 |
+
"page_idx": 12
|
| 1512 |
+
},
|
| 1513 |
+
{
|
| 1514 |
+
"type": "text",
|
| 1515 |
+
"text": "RMSE. When evaluating on static regions, proposed Mono-ViFI shows absolute superiority over ManyDepth and DynamicDepth.",
|
| 1516 |
+
"bbox": [
|
| 1517 |
+
212,
|
| 1518 |
+
637,
|
| 1519 |
+
784,
|
| 1520 |
+
666
|
| 1521 |
+
],
|
| 1522 |
+
"page_idx": 12
|
| 1523 |
+
},
|
| 1524 |
+
{
|
| 1525 |
+
"type": "text",
|
| 1526 |
+
"text": "Generalization Results. We evaluate the KITTI models on Cityscapes and Make3D without fine-tuning, results of which are put in the supplementary. For qualitative comparison, we give a visualization example in Fig. 4, where KITTI models are used to predict the multi-frame depth of a YouTube video frame downloaded from the Wind Walk Travel Videos channel. This indicates that our Mono-ViFI has better generalization ability than existing methods.",
|
| 1527 |
+
"bbox": [
|
| 1528 |
+
212,
|
| 1529 |
+
667,
|
| 1530 |
+
785,
|
| 1531 |
+
758
|
| 1532 |
+
],
|
| 1533 |
+
"page_idx": 12
|
| 1534 |
+
},
|
| 1535 |
+
{
|
| 1536 |
+
"type": "text",
|
| 1537 |
+
"text": "5.3 Ablation Study",
|
| 1538 |
+
"text_level": 1,
|
| 1539 |
+
"bbox": [
|
| 1540 |
+
214,
|
| 1541 |
+
782,
|
| 1542 |
+
388,
|
| 1543 |
+
797
|
| 1544 |
+
],
|
| 1545 |
+
"page_idx": 12
|
| 1546 |
+
},
|
| 1547 |
+
{
|
| 1548 |
+
"type": "text",
|
| 1549 |
+
"text": "In this part, we explore the influence of each component in Mono-ViFi. Our ablations are conducted on Cityscapes or KITTI with ResNet18 backbone.",
|
| 1550 |
+
"bbox": [
|
| 1551 |
+
212,
|
| 1552 |
+
809,
|
| 1553 |
+
784,
|
| 1554 |
+
839
|
| 1555 |
+
],
|
| 1556 |
+
"page_idx": 12
|
| 1557 |
+
},
|
| 1558 |
+
{
|
| 1559 |
+
"type": "header",
|
| 1560 |
+
"text": "Mono-ViFi: A Unified Framework for Self-supervised Monocular Depth",
|
| 1561 |
+
"bbox": [
|
| 1562 |
+
256,
|
| 1563 |
+
114,
|
| 1564 |
+
730,
|
| 1565 |
+
128
|
| 1566 |
+
],
|
| 1567 |
+
"page_idx": 12
|
| 1568 |
+
},
|
| 1569 |
+
{
|
| 1570 |
+
"type": "page_number",
|
| 1571 |
+
"text": "13",
|
| 1572 |
+
"bbox": [
|
| 1573 |
+
767,
|
| 1574 |
+
116,
|
| 1575 |
+
784,
|
| 1576 |
+
126
|
| 1577 |
+
],
|
| 1578 |
+
"page_idx": 12
|
| 1579 |
+
},
|
| 1580 |
+
{
|
| 1581 |
+
"type": "table",
|
| 1582 |
+
"img_path": "images/93df679f27216709e4f6f97dd794cad2c50611cd94ee31e2e2b70395ae90721c.jpg",
|
| 1583 |
+
"table_caption": [
|
| 1584 |
+
"Table 6: Ablation of different flow-based VFI models on KITTI [16]."
|
| 1585 |
+
],
|
| 1586 |
+
"table_footnote": [],
|
| 1587 |
+
"table_body": "<table><tr><td>VFI Model</td><td>Abs Rel</td><td>Sq Rel</td><td>RMSE</td><td>RMSE log</td><td>δ1</td><td>δ2</td><td>δ3</td></tr><tr><td colspan=\"8\">Multi-frame Inference</td></tr><tr><td>IFRNet [29] (Ours)</td><td>0.099</td><td>0.661</td><td>4.321</td><td>0.174</td><td>0.898</td><td>0.966</td><td>0.985</td></tr><tr><td>RIFE [26]</td><td>0.100</td><td>0.672</td><td>4.352</td><td>0.175</td><td>0.896</td><td>0.965</td><td>0.985</td></tr><tr><td>AMT [34]</td><td>0.099</td><td>0.653</td><td>4.335</td><td>0.174</td><td>0.897</td><td>0.966</td><td>0.985</td></tr></table>",
|
| 1588 |
+
"bbox": [
|
| 1589 |
+
305,
|
| 1590 |
+
171,
|
| 1591 |
+
702,
|
| 1592 |
+
238
|
| 1593 |
+
],
|
| 1594 |
+
"page_idx": 13
|
| 1595 |
+
},
|
| 1596 |
+
{
|
| 1597 |
+
"type": "text",
|
| 1598 |
+
"text": "Single-frame Ablation. We first ablate on the single-frame model. As listed in the top half part of Table 5, the standard view depth consistency and spatial augmentation can bring gains individually. Note that scale-aware depth consistency $L_{sa}$ should be based on spatial augmentation. When adding $L_{sa}$ to spatial augmentation, the accuracy can be further improved. And this leads to the most significant enhancement among all components, since $L_{sa}$ can enforce the model to learn scale relationship from scene variations. In orthogonal dimension, our proposed temporal augmentation can also reduce the prediction error by providing more guidance from newly synthesized camera views. When integrating all above elements, our single-frame model can achieve the best performance.",
|
| 1599 |
+
"bbox": [
|
| 1600 |
+
212,
|
| 1601 |
+
276,
|
| 1602 |
+
785,
|
| 1603 |
+
426
|
| 1604 |
+
],
|
| 1605 |
+
"page_idx": 13
|
| 1606 |
+
},
|
| 1607 |
+
{
|
| 1608 |
+
"type": "text",
|
| 1609 |
+
"text": "Multi-frame Ablation. To probe the effect of each component in our VFI-assisted multi-frame fusion module, we conduct ablation experiments under multi-frame inference, whose results are summarized in the bottom half part of Table 5. The baseline is stacking the raw feature of three frames directly, i.e., direct feature fusion. Then, we gradually apply components of the proposed fusion module, including motion-aware feature alignment and occlusion-alleviated feature fusion. The results demonstrate that these strategies can bring cumulative improvements. Furthermore, when incorporating spatial augmentation, temporal augmentation and relevant depth consistency losses, our multi-frame model can be further promoted like the single-frame setting.",
|
| 1610 |
+
"bbox": [
|
| 1611 |
+
212,
|
| 1612 |
+
426,
|
| 1613 |
+
785,
|
| 1614 |
+
577
|
| 1615 |
+
],
|
| 1616 |
+
"page_idx": 13
|
| 1617 |
+
},
|
| 1618 |
+
{
|
| 1619 |
+
"type": "text",
|
| 1620 |
+
"text": "Ablation of different VFI models. Apart from IFRNet [29], we additionally experiment with another two flow-based VFI models, RIFE [26] and AMT [34]. The results are depicted in Table 6. It can be observed that there is only a negligible performance difference among the three VFI networks, which reveals that our method is robust to different VFI models.",
|
| 1621 |
+
"bbox": [
|
| 1622 |
+
212,
|
| 1623 |
+
578,
|
| 1624 |
+
785,
|
| 1625 |
+
652
|
| 1626 |
+
],
|
| 1627 |
+
"page_idx": 13
|
| 1628 |
+
},
|
| 1629 |
+
{
|
| 1630 |
+
"type": "text",
|
| 1631 |
+
"text": "6 Conclusion",
|
| 1632 |
+
"text_level": 1,
|
| 1633 |
+
"bbox": [
|
| 1634 |
+
215,
|
| 1635 |
+
674,
|
| 1636 |
+
359,
|
| 1637 |
+
690
|
| 1638 |
+
],
|
| 1639 |
+
"page_idx": 13
|
| 1640 |
+
},
|
| 1641 |
+
{
|
| 1642 |
+
"type": "text",
|
| 1643 |
+
"text": "In conclusion, this paper presents a novel self-supervised learning framework, named Mono-ViFI, to bilaterally connect single- and multi-frame depth estimation in a unified manner. First, a flow-based VFI model is employed for temporal augmentation. Then, we propose a VFI-assisted multi-frame fusion module for more accurate multi-frame depth. Besides, resizing-cropping is extended to affine transformation as spatial augmentation for enriching data diversity. Finally, we develop a new triplet depth consistency loss for better regularization. Experiments show that our approach can unleash the potential of existing depth models to a greater extent, but without increasing inference complexity.",
|
| 1644 |
+
"bbox": [
|
| 1645 |
+
212,
|
| 1646 |
+
704,
|
| 1647 |
+
785,
|
| 1648 |
+
839
|
| 1649 |
+
],
|
| 1650 |
+
"page_idx": 13
|
| 1651 |
+
},
|
| 1652 |
+
{
|
| 1653 |
+
"type": "page_number",
|
| 1654 |
+
"text": "14",
|
| 1655 |
+
"bbox": [
|
| 1656 |
+
217,
|
| 1657 |
+
114,
|
| 1658 |
+
235,
|
| 1659 |
+
126
|
| 1660 |
+
],
|
| 1661 |
+
"page_idx": 13
|
| 1662 |
+
},
|
| 1663 |
+
{
|
| 1664 |
+
"type": "header",
|
| 1665 |
+
"text": "J. Liu et al.",
|
| 1666 |
+
"bbox": [
|
| 1667 |
+
271,
|
| 1668 |
+
114,
|
| 1669 |
+
351,
|
| 1670 |
+
126
|
| 1671 |
+
],
|
| 1672 |
+
"page_idx": 13
|
| 1673 |
+
},
|
| 1674 |
+
{
|
| 1675 |
+
"type": "text",
|
| 1676 |
+
"text": "References",
|
| 1677 |
+
"text_level": 1,
|
| 1678 |
+
"bbox": [
|
| 1679 |
+
215,
|
| 1680 |
+
143,
|
| 1681 |
+
321,
|
| 1682 |
+
159
|
| 1683 |
+
],
|
| 1684 |
+
"page_idx": 14
|
| 1685 |
+
},
|
| 1686 |
+
{
|
| 1687 |
+
"type": "list",
|
| 1688 |
+
"sub_type": "ref_text",
|
| 1689 |
+
"list_items": [
|
| 1690 |
+
"1. Bae, J., Moon, S., Im, S.: Deep digging into the generalization of self-supervised monocular depth estimation. In: AAAI. pp. 187-196 (2023)",
|
| 1691 |
+
"2. Bangunharcana, A., Magd, A., Kim, K.S.: Dualrefine: Self-supervised depth and pose estimation through iterative epipolar sampling and refinement toward equilibrium. In: CVPR. pp. 726-738 (2023)",
|
| 1692 |
+
"3. Bello, J.L.G., Kim, M.: Forget about the lidar: Self-supervised depth estimators with med probability volumes. In: NeurIPS. pp. 12626-12637 (2020)",
|
| 1693 |
+
"4. Bello, J.L.G., Moon, J., Kim, M.: Positional information is all you need: A novel pipeline for self-supervised svde from videos. arXiv:2205.08851 (2022)",
|
| 1694 |
+
"5. Bian, J., Li, Z., Wang, N., Zhan, H., Shen, C., Cheng, M.M., Reid, I.: Unsupervised scale-consistent depth and ego-motion learning from monocular video. In: NeurIPS. pp. 35-45 (2019)",
|
| 1695 |
+
"6. Chan, K.C., Wang, X., Yu, K., Dong, C., Loy, C.C.: Basicvsr: The search for essential components in video super-resolution and beyond. In: CVPR. pp. 4945-4954 (2021)",
|
| 1696 |
+
"7. Chan, K.C., Zhou, S., Xu, X., Loy, C.C.: Basicvsr++: Improving video superresolution with enhanced propagation and alignment. In: CVPR. pp. 5962-5971 (2022)",
|
| 1697 |
+
"8. Chen, Y., Liu, S., Wang, X.: Learning continuous image representation with local implicit image function. In: CVPR. pp. 8624-8634 (2021)",
|
| 1698 |
+
"9. Chen, Y., Schmid, C., Sminchisescu, C.: Self-supervised learning with geometric constraints in monocular video: Connecting flow, depth, and camera. In: ICCV. pp. 7063-7072 (2019)",
|
| 1699 |
+
"10. Cordts, M., Omran, M., Ramos, S., Rehfeld, T., Enzweiler, M., Benenson, R., Franke, U., Roth, S., Schiele, B.: The cityscapes dataset for semantic urban scene understanding. In: CVPR. pp. 3213-3223 (2016)",
|
| 1700 |
+
"1. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: CVPR. pp. 248-255 (2009)",
|
| 1701 |
+
"2. Eigen, D., Fergus, R.: Predicting depth, surface normals and semantic labels with a common multi-scale convolutional architecture. In: ICCV. pp. 2650-2658 (2015)",
|
| 1702 |
+
"3. Eigen, D., Puhrsch, C., Fergus, R.: Depth map prediction from a single image using a multi-scale deep network. In: NeurIPS. pp. 2366-2374 (2014)",
|
| 1703 |
+
"4. Feng, Z., Yang, L., Jing, L., Wang, H., Tian, Y., Li, B.: Disentangling object motion and occlusion for unsupervised multi-frame monocular depth. In: ECCV. pp. 228-244 (2022)",
|
| 1704 |
+
"5. Garg, R., Kumar, B.G.V., Carneiro, G., Reid, I.: Unsupervised cnn for single view depth estimation: Geometry to the rescue. In: ECCV. pp. 740-756 (2016)",
|
| 1705 |
+
"6. Geiger, A., Lenz, P., Stiller, C., Urtasun, R.: Vision meets robotics: The kitti dataset. *IJRR* **32**(11), 1231–1237 (2013)",
|
| 1706 |
+
"7. Godard, C., Aodha, O.M., Brostow, G.J.: Unsupervised monocular depth estimation with left-right consistency. In: CVPR. pp. 6602-6611 (2017)",
|
| 1707 |
+
"8. Godard, C., Aodha, O.M., Firman, M., Brostow, G.J.: Digging into self-supervised monocular depth estimation. In: CVPR. pp. 3828-3838 (2019)",
|
| 1708 |
+
"9. Gordon, A., Li, H., Jonschkowski, R., Angelova, A.: Depth from videos in the wild: Unsupervised monocular depth learning from unknown cameras. In: ICCV. pp. 8976-8985 (2019)",
|
| 1709 |
+
"20. Guizilini, V., Ambrus, R., Chen, D., Zakharov, S., Gaidon, A.: Multi-frame self-supervised depth with transformers. In: CVPR. pp. 160-170 (2022)"
|
| 1710 |
+
],
|
| 1711 |
+
"bbox": [
|
| 1712 |
+
225,
|
| 1713 |
+
176,
|
| 1714 |
+
784,
|
| 1715 |
+
839
|
| 1716 |
+
],
|
| 1717 |
+
"page_idx": 14
|
| 1718 |
+
},
|
| 1719 |
+
{
|
| 1720 |
+
"type": "header",
|
| 1721 |
+
"text": "Mono-ViFi: A Unified Framework for Self-supervised Monocular Depth",
|
| 1722 |
+
"bbox": [
|
| 1723 |
+
256,
|
| 1724 |
+
114,
|
| 1725 |
+
730,
|
| 1726 |
+
128
|
| 1727 |
+
],
|
| 1728 |
+
"page_idx": 14
|
| 1729 |
+
},
|
| 1730 |
+
{
|
| 1731 |
+
"type": "page_number",
|
| 1732 |
+
"text": "15",
|
| 1733 |
+
"bbox": [
|
| 1734 |
+
767,
|
| 1735 |
+
116,
|
| 1736 |
+
784,
|
| 1737 |
+
126
|
| 1738 |
+
],
|
| 1739 |
+
"page_idx": 14
|
| 1740 |
+
},
|
| 1741 |
+
{
|
| 1742 |
+
"type": "list",
|
| 1743 |
+
"sub_type": "ref_text",
|
| 1744 |
+
"list_items": [
|
| 1745 |
+
"21. Guizilini, V., Ambrus, R., Pillai, S., Gaidon, A.: 3d packing for self-supervised monocular depth estimation. In: CVPR. pp. 2482-2491 (2020)",
|
| 1746 |
+
"22. Han, W., Yin, J., Jin, X., Dai, X., Shen, J.: Brnet: Exploring comprehensive features for monocular depth estimation. In: ECCV. pp. 586-602 (2022)",
|
| 1747 |
+
"23. Han, W., Yin, J., Shen, J.: Self-supervised monocular depth estimation by direction-aware cumulative convolution network. In: ICCV (2023)",
|
| 1748 |
+
"24. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: CVPR. pp. 770-778 (2016)",
|
| 1749 |
+
"25. He, M., Hui, L., Bian, Y., Ren, J., Xie, J., Yang, J.: Ra-depth: Resolution adaptive self-supervised monocular depth estimation. In: ECCV. pp. 565-581 (2022)",
|
| 1750 |
+
"26. Huang, Z., Zhang, T., Heng, W., Shi, B., Zhou, S.: Real-time intermediate flow estimation for video frame interpolation. In: ECCV. pp. 624-642 (2022)",
|
| 1751 |
+
"27. Jung, H., Park, E., Yoo, S.: Fine-grained semantics-aware representation enhancement for self-supervised monocular depth estimation. In: ICCV. pp. 12622-12632 (2021)",
|
| 1752 |
+
"28. Klingner, M., Termöhlen, J.A., Mikolajczyk, J., Fingscheidt, T.: Self-supervised monocular depth estimation: Solving the dynamic object problem by semantic guidance. In: ECCV. pp. 582-600 (2020)",
|
| 1753 |
+
"29. Kong, L., Jiang, B., Luo, D., Chu, W., Huang, X., Tai, Y., Wang, C., Yang, J.: Ifrnet: Intermediate feature refine network for efficient frame interpolation. In: CVPR. pp. 1959-1968 (2022)",
|
| 1754 |
+
"30. Kuznietsov, Y., Proesmans, M., Gool, L.V.: Comoda: Continuous monocular depth adaptation using past experiences. In: WACV. pp. 2906-2916 (2021)",
|
| 1755 |
+
"31. Lee, S., Im, S., Lin, S., Kweon, I.S.: Learning monocular depth in dynamic scenes via instance-aware projection consistency. In: AAAI. pp. 1863-1872 (2021)",
|
| 1756 |
+
"32. Lee, S., Rameau, F., Pan, F., Kweon, I.S.: Attentive and contrastive learning for joint depth and motion field estimation. In: ICCV. pp. 4862-4871 (2021)",
|
| 1757 |
+
"33. Li, H., Gordon, A., Zhao, H., Casser, V., Angelova, A.: Unsupervised monocular depth learning in dynamic scenes. In: CoRL. pp. 1908-1917 (2020)",
|
| 1758 |
+
"34. Li, Z., Zhu, Z.L., Han, L.H., Hou, Q., Guo, C.L., Cheng, M.M.: Amt: All-pairs multi-field transforms for efficient frame interpolation. In: CVPR. pp. 9801-9810 (2023)",
|
| 1759 |
+
"35. Liu, J., Kong, L., Yang, J.: Designing and searching for lightweight monocular depth network. In: ICONIP. pp. 477-488 (2021)",
|
| 1760 |
+
"36. Liu, J., Kong, L., Yang, J.: Atca: An arc trajectory based model with curvature attention for video frame interpolation. In: ICIP. pp. 1486-1490 (2022)",
|
| 1761 |
+
"37. Liu, J., Kong, L., Yang, J., Liu, W.: Towards better data exploitation in self-supervised monocular depth estimation. IEEE Robotics and Automation Letters 9(1), 763-770 (2024)",
|
| 1762 |
+
"38. Luo, X., Huang, J.B., Szeliski, R., Matzen, K., Kopf, J.: Consistent video depth estimation. TOG 39(71), 1-13 (2020)",
|
| 1763 |
+
"39. Lyu, X., Liu, L., Wang, M., Kong, X., Liu, L., Liu, Y., Chen, X., Yuan, Y.: Hrdepth: High resolution self-supervised monocular depth estimation. In: AAAI. pp. 2294-2301 (2021)",
|
| 1764 |
+
"40. Ma, J., Lei, X., Liu, N., Zhao, X., Pu, S.: Towards comprehensive representation enhancement in semantics-guided self-supervised monocular depth estimation. In: ECCV. pp. 304-321 (2022)",
|
| 1765 |
+
"41. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. In: ECCV. pp. 405-421 (2020)"
|
| 1766 |
+
],
|
| 1767 |
+
"bbox": [
|
| 1768 |
+
215,
|
| 1769 |
+
146,
|
| 1770 |
+
784,
|
| 1771 |
+
840
|
| 1772 |
+
],
|
| 1773 |
+
"page_idx": 15
|
| 1774 |
+
},
|
| 1775 |
+
{
|
| 1776 |
+
"type": "page_number",
|
| 1777 |
+
"text": "16",
|
| 1778 |
+
"bbox": [
|
| 1779 |
+
217,
|
| 1780 |
+
114,
|
| 1781 |
+
235,
|
| 1782 |
+
126
|
| 1783 |
+
],
|
| 1784 |
+
"page_idx": 15
|
| 1785 |
+
},
|
| 1786 |
+
{
|
| 1787 |
+
"type": "header",
|
| 1788 |
+
"text": "J. Liu et al.",
|
| 1789 |
+
"bbox": [
|
| 1790 |
+
271,
|
| 1791 |
+
114,
|
| 1792 |
+
351,
|
| 1793 |
+
126
|
| 1794 |
+
],
|
| 1795 |
+
"page_idx": 15
|
| 1796 |
+
},
|
| 1797 |
+
{
|
| 1798 |
+
"type": "list",
|
| 1799 |
+
"sub_type": "ref_text",
|
| 1800 |
+
"list_items": [
|
| 1801 |
+
"42. Patil, V., Gansbeke, W.V., Dai, D., Gool, L.V.: Don't forget the past: Recurrent depth estimation from monocular video. IEEE Robotics and Automation Letters 5(4), 6813-6820 (2020)",
|
| 1802 |
+
"43. Peng, R., Wang, R., Lai, Y., Tang, L., Cai, Y.: Excavating the potential capacity of selfsupervised monocular depth estimation. In: ICCV. pp. 15560-15569 (2021)",
|
| 1803 |
+
"44. Petrovai, A., Nedevschi, S.: Exploiting pseudo labels in a self-supervised learning framework for improved monocular depth estimation. In: CVPR. pp. 1568-1578 (2022)",
|
| 1804 |
+
"45. Pilzer, A., Lathuilière, S., Sebe, N., Ricci, E.: Refine and distill: Exploiting cycle-inconsistency and knowledge distillation for unsupervised monocular depth estimation. In: CVPR. pp. 9760-9769 (2019)",
|
| 1805 |
+
"46. Poggi, M., Aleotti, F., Tosi, F., Mattoccia, S.: On the uncertainty of self-supervised monocular depth estimation. In: CVPR. pp. 3224-3234 (2020)",
|
| 1806 |
+
"47. Ranjan, A., Jampani, V., Balles, L., Kim, K., Sun, D., Wulff, J., Black, M.J.: Competitive collaboration: Joint unsupervised learning of depth, camera motion, optical flow and motion segmentation. In: CVPR. pp. 12240-12249 (2019)",
|
| 1807 |
+
"48. Ren, W., Wang, L., Piao, Y., Zhang, M., Lu, H., Liu, T.: Adaptive co-teaching for unsupervised monocular depth estimation. In: ECCV. pp. 89-105 (2022)",
|
| 1808 |
+
"49. Ruhkamp, P., Gao, D., Chen, H., Navab, N., Busam, B.: Attention meets geometry: Geometry guided spatial-temporal attention for consistent self-supervised monococular depth estimation. In: 3DV. pp. 837-847 (2021)",
|
| 1809 |
+
"50. Saxena, A., Sun, M., Ng, A.Y.: Make3d: Learning 3d scene structure from a single still image. PAMI 31(5), 824-840 (2009)",
|
| 1810 |
+
"51. Tancik, M., Srinivasan, P.P., Mildenhall, B., Fridovich-Keil, S., Raghavan, N., Singhal, U., Ramamoorthi, R., Barron, J.T., Ng, R.: Fourier features let networks learn high frequency functions in low dimensional domains. In: NeurIPS. pp. 7537-7547 (2020)",
|
| 1811 |
+
"52. Uhrig, J., Schneider, N., Schneider, L., Franke, U., Brox, T., Geiger, A.: Sparsity invariant cnns. In: 3DV. pp. 11-20 (2017)",
|
| 1812 |
+
"53. Wang, C., Buenaposada, J.M., Zhu, R., Lucey, S.: Learning depth from monocular videos using direct methods. In: CVPR. pp. 2022-2030 (2018)",
|
| 1813 |
+
"54. Wang, R., Pizer, S.M., Frahm, J.M.: Recurrent neural network for (un-)supervised learning of monocular video visual odometry and depth. In: CVPR. pp. 5555-5564 (2019)",
|
| 1814 |
+
"55. Wang, R., Yu, Z., Gao, S.: Planedepth: Self-supervised depth estimation via orthogonal planes. In: CVPR. pp. 21425-21434 (2023)",
|
| 1815 |
+
"56. Wang, X., Zhu, Z., Huang, G., Chi, X., Ye, Y., Chen, Z., Wang, X.: Drafting monocular cues and velocity guidance for self-supervised multi-frame depth learning. In: AAAI. pp. 2689-2697 (2023)",
|
| 1816 |
+
"57. Wang, Z., Bovik, A.C., Sheikh, H.R., Simoncelli, E.P.: Image quality assessment: From error visibility to structural similarity. IEEE TIP 13(4), 600-612 (2004)",
|
| 1817 |
+
"58. Watson, J., Aodha, O.M., Prisacariu, V.A., Brostow, G.J., Firman, M.: The temporal opportunist: Self-supervised multi-frame monocular depth. In: CVPR. pp. 1164-1174 (2021)",
|
| 1818 |
+
"59. Watson, J., Firman, M., Brostow, G.J., Turmukhambetov, D.: Self-supervised monocular depth hints. In: ICCV. pp. 2162-2171 (2019)",
|
| 1819 |
+
"60. Yin, Z., Shi, J.: Geonet: Unsupervised learning of dense depth, optical flow and camera pose. In: CVPR. pp. 1983-1992 (2018)",
|
| 1820 |
+
"61. Zhang, H., Li, Y., Cao, Y., Liu, Y., Shen, C., Yan, Y.: Exploiting temporal consistency for real-time video depth estimation. In: ICCV. pp. 1725-1734 (2019)"
|
| 1821 |
+
],
|
| 1822 |
+
"bbox": [
|
| 1823 |
+
215,
|
| 1824 |
+
146,
|
| 1825 |
+
785,
|
| 1826 |
+
840
|
| 1827 |
+
],
|
| 1828 |
+
"page_idx": 16
|
| 1829 |
+
},
|
| 1830 |
+
{
|
| 1831 |
+
"type": "header",
|
| 1832 |
+
"text": "Mono-ViFi: A Unified Framework for Self-supervised Monocular Depth",
|
| 1833 |
+
"bbox": [
|
| 1834 |
+
256,
|
| 1835 |
+
114,
|
| 1836 |
+
730,
|
| 1837 |
+
128
|
| 1838 |
+
],
|
| 1839 |
+
"page_idx": 16
|
| 1840 |
+
},
|
| 1841 |
+
{
|
| 1842 |
+
"type": "page_number",
|
| 1843 |
+
"text": "17",
|
| 1844 |
+
"bbox": [
|
| 1845 |
+
767,
|
| 1846 |
+
114,
|
| 1847 |
+
785,
|
| 1848 |
+
126
|
| 1849 |
+
],
|
| 1850 |
+
"page_idx": 16
|
| 1851 |
+
},
|
| 1852 |
+
{
|
| 1853 |
+
"type": "list",
|
| 1854 |
+
"sub_type": "ref_text",
|
| 1855 |
+
"list_items": [
|
| 1856 |
+
"62. Zhang, N., Nex, F., Vosselman, G., Kerle, N.: Lite-mono: A lightweight cnn and transformer architecture for self-supervised monocular depth estimation. In: CVPR. pp. 18537-18546 (2023)",
|
| 1857 |
+
"63. Zhao, C., Poggi, M., Tosi, F., Zhou, L., Sun, Q., Tanga, Y., Mattoccia, S.: Gasmono: Geometry-aided self-supervised monocular depth estimation for indoor scenes. In: ICCV. pp. 16163-16174 (2023)",
|
| 1858 |
+
"64. Zhao, C., Zhang, Y., Poggi, M., Tosi, F., Guo, X., Zhu, Z., Huang, G., Tang, Y., Mattoccia, S.: Monovit: Self-supervised monocular depth estimation with a vision transformer. In: 3DV. pp. 668-678 (2022)",
|
| 1859 |
+
"65. Zhou, H., Greenwood, D., Taylor, S.: Self-supervised monocular depth estimation with internal feature fusion. In: BMVC (2021)",
|
| 1860 |
+
"66. Zhou, T., Brown, M., Snavely, N., Lowe, D.G.: Unsupervised learning of depth and ego-motion from video. In: CVPR. pp. 6612-6619 (2017)",
|
| 1861 |
+
"67. Zhou, Z., Fan, X., Shi, P., Xin, Y.: R-msfm: Recurrent multi-scale feature modulation for monocular depth estimating. In: ICCV. pp. 12757-12766 (2021)"
|
| 1862 |
+
],
|
| 1863 |
+
"bbox": [
|
| 1864 |
+
215,
|
| 1865 |
+
146,
|
| 1866 |
+
784,
|
| 1867 |
+
354
|
| 1868 |
+
],
|
| 1869 |
+
"page_idx": 17
|
| 1870 |
+
},
|
| 1871 |
+
{
|
| 1872 |
+
"type": "page_number",
|
| 1873 |
+
"text": "18",
|
| 1874 |
+
"bbox": [
|
| 1875 |
+
217,
|
| 1876 |
+
114,
|
| 1877 |
+
235,
|
| 1878 |
+
126
|
| 1879 |
+
],
|
| 1880 |
+
"page_idx": 17
|
| 1881 |
+
},
|
| 1882 |
+
{
|
| 1883 |
+
"type": "header",
|
| 1884 |
+
"text": "J. Liu et al.",
|
| 1885 |
+
"bbox": [
|
| 1886 |
+
271,
|
| 1887 |
+
114,
|
| 1888 |
+
351,
|
| 1889 |
+
126
|
| 1890 |
+
],
|
| 1891 |
+
"page_idx": 17
|
| 1892 |
+
}
|
| 1893 |
+
]
|
2024/Mono-ViFI_ A Unified Learning Framework for Self-supervised Single- and Multi-frame Monocular Depth Estimation/47dcf543-78f7-47b3-955c-778904eede62_model.json
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
2024/Mono-ViFI_ A Unified Learning Framework for Self-supervised Single- and Multi-frame Monocular Depth Estimation/47dcf543-78f7-47b3-955c-778904eede62_origin.pdf
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:9bf5b6ab3ed2b8499ae8b7d4bda67e4fe164537d53d2966e2fb4d597ce9d7107
|
| 3 |
+
size 1590008
|
2024/Mono-ViFI_ A Unified Learning Framework for Self-supervised Single- and Multi-frame Monocular Depth Estimation/full.md
ADDED
|
@@ -0,0 +1,369 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Mono-ViFi: A Unified Learning Framework for Self-supervised Single- and Multi-frame Monocular Depth Estimation
|
| 2 |
+
|
| 3 |
+
Jinfeng Liu, Lingtong Kong\*, Bo Li, Zerong Wang, Hong Gu, and Jinwei Chen
|
| 4 |
+
|
| 5 |
+
vivo Mobile Communication Co., Ltd, China
|
| 6 |
+
|
| 7 |
+
{liujinfeng, ltkong, libra, wangzerong, guhong, jinwei.chen}@vivo.com
|
| 8 |
+
|
| 9 |
+
Abstract. Self-supervised monocular depth estimation has gathered notable interest since it can liberate training from dependency on depth annotations. In monocular video training case, recent methods only conduct view synthesis between existing camera views, leading to insufficient guidance. To tackle this, we try to synthesize more virtual camera views by flow-based video frame interpolation (VFI), termed as temporal augmentation. For multi-frame inference, to sidestep the problem of dynamic objects encountered by explicit geometry-based methods like ManyDepth, we return to the feature fusion paradigm and design a VFI-assisted multi-frame fusion module to align and aggregate multi-frame features, using motion and occlusion information obtained by the flow-based VFI model. Finally, we construct a unified self-supervised learning framework, named Mono-ViFi, to bilaterally connect single- and multiframe depth. In this framework, spatial data augmentation through image affine transformation is incorporated for data diversity, along with a triplet depth consistency loss for regularization. The single- and multiframe models can share weights, making our framework compact and memory-efficient. Extensive experiments demonstrate that our method can bring significant improvements to current advanced architectures. Source code is available at https://github.com/LiuJF1226/Mono-ViFi.
|
| 10 |
+
|
| 11 |
+
Keywords: Monocular depth estimation $\cdot$ Self-supervised learning $\cdot$ Data augmentation $\cdot$ Video frame interpolation
|
| 12 |
+
|
| 13 |
+
# 1 Introduction
|
| 14 |
+
|
| 15 |
+
Monocular depth estimation (MDE) is a fundamental 3D vision task, which finds applications in robotics, autonomous driving and VR/AR. To reduce the dependence on depth annotations, self-supervised MDE [1,2,15,17,46,63,66] has risen as a convenient and practical solution, which uses monocular videos, stereo pairs, or even both to generate supervision through view synthesis, where an image photometric reconstruction loss is minimized. Compared to stereo training case [15,17,59] with known camera pose, learning from monocular videos [18,58, 66] is more challenging since it requires an extra pose network to estimate camera ego-motion. Even so, it is much easier to collect monocular videos compared to
|
| 16 |
+
|
| 17 |
+
stereo image pairs, making monocular training more attractive. Therefore, this paper also focuses on self-supervised training with unlabeled monocular videos.
|
| 18 |
+
|
| 19 |
+
Recent works have made various attempts to improve accuracy, like introducing masking techniques [18, 66] and leveraging segmentation priors [28, 31]. However, they only conduct view synthesis between existing fixed camera views within video datasets, leading to insufficient guidance especially in invisible regions by occlusion or disappearance. Considering that video data possesses temporal continuity, an intuitive idea is to generate novel frames, i.e., virtual camera views in the temporal dimension. We term it as temporal augmentation, which can be achieved by video frame interpolation (VFI), as shown in Fig. 1(a).
|
| 20 |
+
|
| 21 |
+
From another perspective, multi-frame reasoning paradigm have recently been researched because multiple frames can be available for reasoning in many application scenarios. ManyDepth [58] proposes to use cost volume constructed from multi-frame matching features. It explicitly reasons about geometry based on static scene assumption, which can be easily affected by dynamic objects and occlusion. DynamicDepth [14] attempts to tackle dynamic scenes, but relies on depth priors and segmentation masks of moving objects when reasoning. To sidestep the issue of dynamic objects in multi-frame inference encountered by these methods, we return to the traditional paradigm [42, 49], which aggregates multi-frame features directly. However, direct feature fusion without alignment harms the learning of temporal representations, since feature embeddings at the same location in neighbor frames are usually irrelevant, making it difficult for models to leverage the correlation of multiple frames. Hence, we regard feature alignment as a reasonable and necessary step before aggregation. As revealed in several video enhancement tasks like video super-resolution [6, 7] and frame interpolation [26, 29, 34, 36], inter-frame feature alignment can be implemented by warping through optical flow. This inspires us to employ a flow-based VFI model to concurrently conduct temporal augmentation and provide explicit correspondence for feature alignment. Moreover, the VFI model can reason out occlusion mask between neighbor frames, helping to alleviate inter-frame occlusion problem in feature fusion. Based on above analysis, we design a VFI-assisted multi-frame fusion module for multi-frame depth estimation, as depicted in Fig. 2.
|
| 22 |
+
|
| 23 |
+
Finally, we build a unified framework to bilaterally connect single- and multiframe depth, as demonstrated in Fig. 1(b). The direct way is by depth consistency, usually for geometric regularization or distillation in self-supervised MDE. ManyDepth conducts unilateral distillation from single-frame model to multiframe one. DynamicDepth extends it to a bilateral mutual reinforcement, but only in the regions of moving objects. Differently, our framework applies the bilateral consistency on the whole image. For an original target image, we use a standard view depth consistency (SVDC) loss between its two depth maps predicted from the single- and multi-frame models. We also include classic data augmentation strategies, categorized as spatial augmentation. Specifically, we combine resizing-cropping with image rotation, forming a generic affine transformation. Then, we give the formulation of rectified camera pose for affine-augmented view synthesis using the cue of relative object size [4, 55]. For the
|
| 24 |
+
|
| 25 |
+

|
| 26 |
+
(a) Temporal Augmentation with VFI
|
| 27 |
+
|
| 28 |
+

|
| 29 |
+
(b) Learning Framework $(I_{t}$ for example
|
| 30 |
+
Fig. 1: Overview of Mono-ViFI. (a) We achieve temporal augmentation by a flow-based VFI model $\mathcal{H}$ , which also reasons out intermediate optical flow and occlusion mask for multi-frame inference. (b) For target image $I_{t}$ , we obtain its single-frame depth $D_{t}$ and multi-frame depth $D_{t}^{m}$ . Also, $I_{t}$ is augmented to $\widetilde{I}_{t}$ through affine transformation $\mathcal{A}$ . We then calculate the depth $\widehat{D}_t$ of $\widetilde{I}_t$ and inversely convert it to the original view, generating $\widehat{D}_t$ . Finally, we enforce a triplet depth consistency loss $L_{tc}(t)$ among the three depth maps, including a standard view depth consistency loss $L_{sv}(t)$ and two scale-aware depth consistency losses, $L_{sa}(t)$ and $L_{sa}^{m}(t)$ . Note that each depth map also corresponds to a photometric loss and a smoothness loss, which are omitted here.
|
| 31 |
+
|
| 32 |
+
depth map from augmented target view, we enforce two scale-aware depth consistency (SADC) losses between it and above two standard view depths, respectively. SADC stems from the relative size cue, which assumes that when an image is zoomed up by a factor, the depth decreases by the same factor. It guides depth model to learn consistent geometric scale relationship from scene variations. The SVDC loss and two SADC losses are termed as a triplet depth consistency loss.
|
| 33 |
+
|
| 34 |
+
Our framework is named Mono-ViFI, since the self-supervised Monocular depth is largely boosted by Video Frame Interpolation. To our best knowledge, we are the first to leverage VFI for the improvement of self-supervised MDE. The main contributions of this paper are summarized as follows:
|
| 35 |
+
|
| 36 |
+
- We propose Mono-ViFi, a unified learning framework to bilaterally connect self-supervised single- and multi-frame MDE. In Mono-ViFi, the single- and multi-frame models can share weights, making the unified framework more compact and memory-efficient.
|
| 37 |
+
- We conduct temporal augmentation through flow-based VFI, and develop a novel VFI-assisted multi-frame fusion module, which utilizes motion and occlusion from flow-based VFI model to improve multi-frame depth.
|
| 38 |
+
- We incorporate image affine transformation as spatial augmentation, along with a triplet depth consistency loss for regularization and distillation.
|
| 39 |
+
- Experiments show that Mono-ViFI can achieve state-of-the-art performance. Moreover, our approach is orthogonal to architecture design, allowing current advanced depth networks to be further improved by seamlessly integrating them with Mono-ViFI.
|
| 40 |
+
|
| 41 |
+
# 2 Related Work
|
| 42 |
+
|
| 43 |
+
# 2.1 Self-supervised Monocular Depth Estimation
|
| 44 |
+
|
| 45 |
+
Self-supervised learning paradigm for monocular depth estimation can be categorized into stereo training and monocular training according to the data source.
|
| 46 |
+
|
| 47 |
+
Stereo training leverages the left-right consistency in stereo image pairs. Garg et al. [15] first consider the training as a view synthesis problem with a photometric consistency loss between stereo pairs. Building upon this idea, Godard et al. [17] introduce a left-right disparity consistency loss. Then, it is further extended by using depth hints [59], data augmentation [43, 55], knowledge distillation [43, 45, 55] and plane-based scene representation [3, 55].
|
| 48 |
+
|
| 49 |
+
Monocular training leverages the temporal consistency in monocular video frames. Zhou et al. [66] propose a pioneering method which jointly learns depth and camera ego-motion from monocular videos in a self-supervised way. Then, Bian et al. [5] propose a scale-consistent depth and ego-motion estimation approach by adding a depth consistency loss, which reduces scale drift of estimated depth but decreases the accuracy. Later, Godard et al. [18] propose Monodepth2, a strong baseline for most follow-up works, which employs a minimum reprojection loss to mitigate occlusion issues and an auto-masking loss to filter out moving objects with similar velocity as the camera. Furthermore, some works introduce multi-task learning to combine depth estimation with optical flow [9, 32, 47, 60], semantic segmentation [27, 28, 40] and instance segmentation [31]. Others [21-23, 35, 37, 39, 62, 64, 65] try to improve the network architectures for depth estimation. However, these methods only utilize the geometric constrains between existing fixed camera views in video datasets. This insufficient guidance may block current depth models from unleashing their full capabilities. To address it, we conduct temporal augmentation through VFI in this paper.
|
| 50 |
+
|
| 51 |
+
# 2.2 Multi-frame Monocular Depth
|
| 52 |
+
|
| 53 |
+
Above single-frame depth models do not leverage any temporal information for inference, which limits the performance to some degree. Regarding multi-frame depth prediction, early works adopt test-time refinement [9, 30, 38], use spatial-temporal attention [49], or combine traditional depth networks with recurrent layers [54, 61] to process sequence information. Another broad group of methods [14, 20, 56, 58] utilize cost volume from stereo matching, which explicitly encode and reason from geometric constraints by combining a multi-frame cost volume with single-frame features. On the one hand, this type of methods based on ManyDepth [58] usually meet with the problem of dynamic scenes. To sidestep this issue, we develop a VFI-assisted multi-frame fusion depth model to align and aggregate multi-frame features by optical flow and occlusion masks. On the other hand, ManyDepth uses single-frame depth prior to boost multi-frame depth estimation. However, this improvement is unidirectional. DynamicDepth [14] extends it to a mutual reinforcement, but only targets on dynamic objects. Differently, we propose a unified framework to bilaterally connect single- and multi-frame depth in a more compact and memory-efficient manner.
|
| 54 |
+
|
| 55 |
+
# 3 Problem Setup
|
| 56 |
+
|
| 57 |
+
Self-supervised MDE in the monocular training case leverages the geometric constraints from video sequences. Specifically, for a training triplet $\{I_{t-2}, I_t, I_{t+2}\}$ which contains three consecutive frames, we use $I_t$ as the target view and the other two as the source views $\{I_s\}_{s \in \{t-2, t+2\}}$ to implement view synthesis. Here the temporal notation $t \pm 2$ instead of $t \pm 1$ is for the convenience of following description. Also, a pose network is utilized to estimate the camera pose from the target view to the source view, denoted as $\pmb{T}_{t \to s} = [\pmb{R}_{t \to s} | \pmb{t}_{t \to s}]$ , where $\pmb{R}_{t \to s}$ and $\pmb{t}_{t \to s}$ represent the rotation and translation components. Formally, the target depth $D_t$ and the camera pose $\pmb{T}_{t \to s}$ can be derived from:
|
| 58 |
+
|
| 59 |
+
$$
|
| 60 |
+
D _ {t} = \mathcal {D} \left(\mathcal {E} \left(I _ {t}\right)\right), \quad \boldsymbol {T} _ {t \rightarrow s} = \mathcal {P} \left(I _ {t}, I _ {s}\right), \tag {1}
|
| 61 |
+
$$
|
| 62 |
+
|
| 63 |
+
where $\mathcal{E}$ and $\mathcal{D}$ represent the encoder and decoder, respectively, of a single-frame depth network, while $\mathcal{P}$ denotes the pose network. Then, the reconstructed view of $I_{t}$ from the source view $I_{s}$ can be obtained by:
|
| 64 |
+
|
| 65 |
+
$$
|
| 66 |
+
I _ {s \rightarrow t} = I _ {s} \left\langle \operatorname {p r o j} \left(D _ {t}, \boldsymbol {T} _ {t \rightarrow s}, \boldsymbol {K}\right)\right\rangle , \tag {2}
|
| 67 |
+
$$
|
| 68 |
+
|
| 69 |
+
where $I_{s\rightarrow t}$ is the reconstructed target image, $\langle \rangle$ means the differentiable bilinear sampling operator, $proj()$ returns the 2D coordinates of the pixels in $I_{t}$ after they are reprojected to the viewpoint of $I_{s}$ , and $K$ is the camera intrinsic matrix assumed to be known. Finally, the per-pixel photometric loss between $I_{s\rightarrow t}$ and $I_{t}$ is used to optimize the depth and pose networks in an end-to-end manner. We follow Monodepth2 [18] to adopt the combination of L1 distance and SSIM [57] as the photometric error function and take the minimum across all the source views at each pixel to handle occlusion, which is formulated as:
|
| 70 |
+
|
| 71 |
+
$$
|
| 72 |
+
L _ {p e} (t) = \min _ {s} \frac {\alpha}{2} \left(1 - \operatorname {S S I M} \left(I _ {t}, I _ {s \rightarrow t}\right)\right) + (1 - \alpha) \| I _ {t}, I _ {s \rightarrow t} \| _ {1}, \tag {3}
|
| 73 |
+
$$
|
| 74 |
+
|
| 75 |
+
where $\alpha = 0.85$ . Besides, the edge-aware smoothness loss is employed to cope with depth discontinuities:
|
| 76 |
+
|
| 77 |
+
$$
|
| 78 |
+
L _ {s m} (t) = \left| \partial_ {x} d _ {t} ^ {*} \right| e ^ {- \left| \partial_ {x} I _ {t} \right|} + \left| \partial_ {y} d _ {t} ^ {*} \right| e ^ {- \left| \partial_ {y} I _ {t} \right|}, \tag {4}
|
| 79 |
+
$$
|
| 80 |
+
|
| 81 |
+
where $d_t^* = d_t / \overline{d_t}$ is the mean-normalized inverse depth to discourage shrinking [53]. In summary, the standard self-supervised loss on $I_{t}$ is:
|
| 82 |
+
|
| 83 |
+
$$
|
| 84 |
+
L _ {s s} (t) = \mathcal {L} \left(D _ {t}, \boldsymbol {T} _ {t \rightarrow s}, I _ {t}, I _ {s}\right) = \mu L _ {p e} (t) + \gamma L _ {s m} (t), \tag {5}
|
| 85 |
+
$$
|
| 86 |
+
|
| 87 |
+
where $\mu$ is the auto-masking in [18] and $\gamma = 0.001$ .
|
| 88 |
+
|
| 89 |
+
# 4 Methodology
|
| 90 |
+
|
| 91 |
+
In this paper, we propose novel Mono-ViFI for self-supervised monocular training, which bilaterally connects single- and multi-frame depth models into a unified learning framework. With flow-based VFI model, we introduce temporal augmentation and develop VFI-assisted multi-frame fusion depth, which are presented in Sec. 4.1 and Sec. 4.2, respectively. Besides, our framework incorporates
|
| 92 |
+
|
| 93 |
+
spatial data augmentation and a triplet depth consistency loss, which are separately described in Sec. 4.3 and Sec. 4.4. Finally, we give the overall training objective in Sec. 4.5. The overview of our Mono-ViFi is depicted in Fig. 1.
|
| 94 |
+
|
| 95 |
+
# 4.1 Temporal Augmentation with VFI
|
| 96 |
+
|
| 97 |
+
Firstly, we generate more new frames in the temporal dimension based on existing video frames for temporal augmentation. To achieve this, we adopt a pretrained flow-based VFI model, such as IFRNet [29], which takes two temporally adjacent frames $\{I_0,I_2\}$ as input to synthesize an intermediate frame $I_{1}$ . The frame interpolation process can be formulated as:
|
| 98 |
+
|
| 99 |
+
$$
|
| 100 |
+
F _ {1 \rightarrow 0}, F _ {1 \rightarrow 2}, M _ {1} = \mathcal {H} \left(I _ {0}, I _ {2}\right), \tag {6}
|
| 101 |
+
$$
|
| 102 |
+
|
| 103 |
+
$$
|
| 104 |
+
I _ {0 \rightarrow 1} = \omega \left(I _ {0}, F _ {1 \rightarrow 0}\right), \quad I _ {2 \rightarrow 1} = \omega \left(I _ {2}, F _ {1 \rightarrow 2}\right), \tag {7}
|
| 105 |
+
$$
|
| 106 |
+
|
| 107 |
+
$$
|
| 108 |
+
I _ {1} = M _ {1} \odot I _ {0 \rightarrow 1} + (1 - M _ {1}) \odot I _ {2 \rightarrow 1}, \tag {8}
|
| 109 |
+
$$
|
| 110 |
+
|
| 111 |
+
where $\mathcal{H}$ is the VFI model, $F_{1\rightarrow 0}$ and $F_{1\rightarrow 2}$ are the intermediate optical flow, $\omega$ represents backward warping. $I_{0\rightarrow 1}$ and $I_{2\rightarrow 1}$ denote the warped images to the intermediate position, while $M_{1}$ ranging from 0 to 1 is the merge mask adjusting the mixing ratio according to bidirectional occlusion information. Here, $\odot$ indicates the element-wise multiplication. Considering a triplet $\{I_{t - 2},I_t,I_{t + 2}\}$ with three consecutive frames as described in Sec. 3, we use the VFI model to synthesize $I_{t - 1}$ from $\{I_{t - 2},I_t\}$ and $I_{t + 1}$ from $\{I_t,I_{t + 2}\}$ , as shown in Fig. 1(a). Then, we regard $I_{t - 1}$ and $I_{t + 1}$ as another two target frames. All three target frames have common sources of $\{I_{t - 2},I_{t + 2}\}$ . The following operations and losses are the same for each of them, and we only take $I_{t}$ for instance.
|
| 112 |
+
|
| 113 |
+
# 4.2 VFI-assisted Multi-frame Fusion Depth
|
| 114 |
+
|
| 115 |
+
Apart from temporal augmentation, our employed VFI model can also provide other inter-frame information, such as optical flow and occlusion merge masks. This enables us to design a novel VFI-assisted multi-frame fusion module for multi-frame depth prediction, as demonstrated in Fig. 2. The fusion module consists of two parts, i.e., motion-aware feature alignment (MAFA) and occlusion-alarviated feature fusion (OAFF). Taking original triplet $\{I_{t - 2},I_t,I_{t + 2}\}$ for example, we implement forward process of the VFI model to interpolate the middle position $t$ of $\{I_{t - 2},I_{t + 2}\}$ in Eq. (9), yielding intermediate flow $F_{t\rightarrow (t - 2)},F_{t\rightarrow (t + 2)}$ and merge mask $M_{t}$ to align and aggregate multi-frame features.
|
| 116 |
+
|
| 117 |
+
$$
|
| 118 |
+
F _ {t \rightarrow (t - 2)}, F _ {t \rightarrow (t + 2)}, M _ {t} = \mathcal {H} \left(I _ {t - 2}, I _ {t + 2}\right). \tag {9}
|
| 119 |
+
$$
|
| 120 |
+
|
| 121 |
+
Multi-frame Feature Extraction. Our fusion module accepts multi-frame features extracted by the same encoder, which is shared with the single-frame depth encoder $\mathcal{E}$ . The extraction process can be written as:
|
| 122 |
+
|
| 123 |
+
$$
|
| 124 |
+
\left\{\phi_ {t - 2} ^ {k}, \phi_ {t} ^ {k}, \phi_ {t + 2} ^ {k} \right\} _ {k = 1} ^ {N} = \left\{\mathcal {E} \left(I _ {t - 2}\right), \mathcal {E} \left(I _ {t}\right), \mathcal {E} \left(I _ {t + 2}\right) \right\}, \tag {10}
|
| 125 |
+
$$
|
| 126 |
+
|
| 127 |
+
where $\{\phi_{t - 2}^k,\phi_t^k,\phi_{t + 2}^k\}_{k = 1}^N$ are the multi-level features of the three frames and $N$ is the number of feature levels.
|
| 128 |
+
|
| 129 |
+

|
| 130 |
+
Fig. 2: Our VFI-assisted multi-frame fusion depth model. Motion and occlusion provided by VFI can be used to align and fuse multi-frame features in an explicit manner.
|
| 131 |
+
|
| 132 |
+
Motion-aware Feature Alignment (MAFA). For the $k$ -th level, we use bidirectional optical flow to align the neighbor features $\phi_{t-2}^k$ and $\phi_{t+2}^k$ to the target position $t$ , generating $\phi_{(t-2) \to t}^k$ and $\phi_{(t+2) \to t}^k$ , respectively, formulated as:
|
| 133 |
+
|
| 134 |
+
$$
|
| 135 |
+
\begin{array}{l} \phi_ {(t - 2) \rightarrow t} ^ {k} = \omega \left(\phi_ {t - 2} ^ {k}, F _ {t \rightarrow (t - 2)} ^ {k}\right), \\ \phi_ {(t + 2) \rightarrow t} ^ {k} = \omega \left(\phi_ {t + 2} ^ {k}, F _ {t \rightarrow (t + 2)} ^ {k}\right), \tag {11} \\ \end{array}
|
| 136 |
+
$$
|
| 137 |
+
|
| 138 |
+
where $F_{t\rightarrow (t - 2)}^k$ and $F_{t\rightarrow (t + 2)}^k$ denote the corresponding flows of the $k$ -th level, downsampled from $F_{t\rightarrow (t - 2)}$ and $F_{t\rightarrow (t + 2)}$ , respectively. Inspired by recent 2D implicit neural representation [8], we also concatenate the aligned features with encoded optical flow, which can be denoted as:
|
| 139 |
+
|
| 140 |
+
$$
|
| 141 |
+
\begin{array}{l} \varphi_ {(t - 2) \rightarrow t} ^ {k} = \operatorname {C o n c a t} \left(\phi_ {(t - 2) \rightarrow t} ^ {k}, g \left(F _ {t \rightarrow (t - 2)} ^ {k}\right)\right), \\ \varphi_ {(t + 2) \rightarrow t} ^ {k} = \operatorname {C o n c a t} \left(\phi_ {(t + 2) \rightarrow t} ^ {k}, g \left(F _ {t \rightarrow (t + 2)} ^ {k}\right)\right), \tag {12} \\ \varphi_ {t} ^ {k} = \operatorname {C o n c a t} \left(\phi_ {t} ^ {k}, g \left(F _ {0} ^ {k}\right)\right), \\ \end{array}
|
| 142 |
+
$$
|
| 143 |
+
|
| 144 |
+
where Concat means feature concatenation, $F_0^k$ denotes zero flow and $g$ is fourier positional encoding [41,51] to transform optical flow to high-frequency signal. Actually, optical flow indicates the relative coordinate of a feature vector in the warped feature map w.r.t. the original one. When incorporating it, multi-frame model can adaptively adjust feature learning through motion-aware alignment.
|
| 145 |
+
|
| 146 |
+
Occlusion-alleviated Feature Fusion (OAFF). Due to the ubiquitous interframe occlusion, aggregating the aligned features directly through summation or concatenation can deteriorate the performance. Therefore, we use the $k$ -th level merge mask $M_t^k$ downsampled from $M_t$ to mitigate this issue, just like the image merging process of VFI in Eq. (8):
|
| 147 |
+
|
| 148 |
+
$$
|
| 149 |
+
\chi_ {t} ^ {k} = M _ {t} ^ {k} \odot \varphi_ {(t - 2) \rightarrow t} ^ {k} + (1 - M _ {t} ^ {k}) \odot \varphi_ {(t + 2) \rightarrow t} ^ {k}, \tag {13}
|
| 150 |
+
$$
|
| 151 |
+
|
| 152 |
+
where $\chi_t^k$ is the occlusion-alleviated feature merged from the temporal positions $t - 2$ and $t + 2$ . It is then concatenated with the feature $\varphi_t^k$ in the position $t$ .
|
| 153 |
+
|
| 154 |
+
And we can obtain the final feature after fusion by:
|
| 155 |
+
|
| 156 |
+
$$
|
| 157 |
+
\psi_ {t} ^ {k} = \operatorname {C o n v} _ {1 \times 1} \left(\operatorname {C o n c a t} \left(\varphi_ {t} ^ {k}, \chi_ {t} ^ {k}\right)\right), \tag {14}
|
| 158 |
+
$$
|
| 159 |
+
|
| 160 |
+
where $\mathrm{Conv}_{1\times 1}$ is a $1\times 1$ convolutional layer to transform the feature into the same channel as the original feature $\phi_t^k$ , enabling the multi-frame depth decoder to employ the same architecture as the single-frame one.
|
| 161 |
+
|
| 162 |
+
Multi-frame Depth Decoder. To translate the multi-level fusion features into a depth map, a decoder is required. Concretely, we use the same architecture as the single-frame depth decoder, but the parameters are not shared. This encourages the single- and multi-frame models to decode the features in their individual ways without conflict. Finally, the fusion features $\{\psi_t^k\}_{k=1}^N$ are processed by a multi-frame depth decoder $\mathcal{D}_m$ to generate the fusion depth $D_t^m$ of the target frame $I_t$ , which can be written as:
|
| 163 |
+
|
| 164 |
+
$$
|
| 165 |
+
D _ {t} ^ {m} = \mathcal {D} _ {m} \left(\left\{\psi_ {t} ^ {k} \right\} _ {k = 1} ^ {N}\right). \tag {15}
|
| 166 |
+
$$
|
| 167 |
+
|
| 168 |
+
Note that the multi-frame depth $D_{t}^{m}$ also corresponds to a self-supervised loss like Eq. (5), which is denoted as:
|
| 169 |
+
|
| 170 |
+
$$
|
| 171 |
+
L _ {s s} ^ {m} (t) = \mathcal {L} \left(D _ {t} ^ {m}, \boldsymbol {T} _ {t \rightarrow s}, I _ {t}, I _ {s}\right). \tag {16}
|
| 172 |
+
$$
|
| 173 |
+
|
| 174 |
+
# 4.3 Spatial Augmentation
|
| 175 |
+
|
| 176 |
+
Orthogonal to temporal augmentation, we also include classic data augmentation to spatially expand the data diversity, which is categorized as spatial augmentation in this paper. Resizing-cropping is a commonly used and effective augmentation approach. Based on this operation, we introduce image rotation, extending it to a generic affine transformation (cropping is actually image translation). Specifically, we follow [55] to consider that the augmented image and the original image are captured by the same camera system whose intrinsics are fixed. For resizing operation, it is assumed that when an image is zoomed up by a factor $f_{s} \geq 1$ , the relative size of an object increases by the same factor $f_{s}$ and its depth decreases by $f_{s}$ simultaneously. This is inspired from a vital monocular cue [4] that the closer an object is, the larger its relative size is. Our affine transformation $\mathcal{A}$ is applied to both the target image $I_{t}$ and the source image $I_{s}$ , generating $\widetilde{I}_{t} = \mathcal{A}(I_{t})$ and $\widetilde{I}_{s} = \mathcal{A}(I_{s})$ , respectively. For the augmented target view, we can obtain its depth through the single-frame model by $\widetilde{D}_{t} = \mathcal{D}(\mathcal{E}(\widetilde{I}_{t}))$ . To conduct view synthesis on augmented images, the original camera pose $T_{t\rightarrow s} = [\pmb{R}_{t\rightarrow s}|\pmb{t}_{t\rightarrow s}]$ from target to source should be rectified to $\widetilde{T}_{t\rightarrow s} = [\widetilde{\pmb{R}}_{t\rightarrow s}|\widetilde{\pmb{t}}_{t\rightarrow s}]$ as done in [55]:
|
| 177 |
+
|
| 178 |
+
$$
|
| 179 |
+
\tilde {\boldsymbol {R}} _ {t \rightarrow s} = \boldsymbol {R} _ {c} \boldsymbol {R} _ {t \rightarrow s} \boldsymbol {R} _ {c} ^ {- 1}, \quad \tilde {\boldsymbol {t}} _ {t \rightarrow s} = \boldsymbol {R} _ {c} \boldsymbol {t} _ {t \rightarrow s}, \tag {17}
|
| 180 |
+
$$
|
| 181 |
+
|
| 182 |
+
where $\mathbf{R}_c$ is the rectification matrix determined only by the camera intrinsics and the augmentation parameters. Different from [55] calculating $\mathbf{R}_c$ under resizing-cropping, we give its formulation and derivation under a more general affine transformation, which can be seen in our supplementary. Therefore, we can derive
|
| 183 |
+
|
| 184 |
+
the self-supervised loss $\widetilde{L}_{ss}(t)$ on the augmented images similarly to Eq. (5):
|
| 185 |
+
|
| 186 |
+
$$
|
| 187 |
+
\widetilde {L} _ {s s} (t) = \mathcal {L} (\widetilde {D} _ {t}, \widetilde {\boldsymbol {T}} _ {t \rightarrow s}, \widetilde {I} _ {t}, \widetilde {I} _ {s}). \tag {18}
|
| 188 |
+
$$
|
| 189 |
+
|
| 190 |
+
Note that there is a mask to exclude the possible invalid pixels in $\widetilde{I}_t$ when calculating $\widetilde{L}_{ss}(t)$ due to image rotation.
|
| 191 |
+
|
| 192 |
+
# 4.4 Triplet Depth Consistency
|
| 193 |
+
|
| 194 |
+
Now we have single-frame depths $D_{t}$ and $\widetilde{D}_{t}$ of the original image $I_{t}$ and the augmented image $\widetilde{I}_{t}$ , along with multi-frame depth $D_{t}^{m}$ of $I_{t}$ . For better knowledge distillation and regularization, our unified learning framework applies a triplet depth consistency loss among the three depth maps, including a standard view depth consistency (SVDC) loss and two scale-aware depth consistency (SADC) losses, which are shown in Fig. 1(b). The SVDC loss $L_{sv}(t)$ conducts mutual distillation between single- and multi-frame models, promoting their mutual reinforcement, which can be denoted as:
|
| 195 |
+
|
| 196 |
+
$$
|
| 197 |
+
L _ {s v} (t) = \operatorname {S I} \left(D _ {t} ^ {m}, D _ {t}\right), \tag {19}
|
| 198 |
+
$$
|
| 199 |
+
|
| 200 |
+
where $\mathrm{SI}()$ is the widely used scale-invariant error function [13]. Besides, we use depth $\widetilde{D}_t$ under spatial augmentation to enforce two SADC losses, i.e., $L_{sa}(t)$ between $\widetilde{D}_t$ and $D_t$ , and $L_{sa}^m(t)$ between $\widetilde{D}_t$ and $D_t^m$ . To this end, we first restore $\widetilde{D}_t$ to the original image view, generating $\widehat{D}_t = \mathcal{A}^{-1}(\widetilde{D}_t)$ in Fig. 1(b), where $\mathcal{A}^{-1}$ is the inverse transformation of affine augmentation $\mathcal{A}$ . To exclude the invalid pixels in $\widehat{D}_t$ , we adopt a mask $M_{sa}$ and calculate $L_{sa}(t)$ and $L_{sa}^m(t)$ through:
|
| 201 |
+
|
| 202 |
+
$$
|
| 203 |
+
L _ {s a} (t) = \operatorname {S I} \left(M _ {s a} \odot D _ {t}, M _ {s a} \odot f _ {s} \widehat {D} _ {t}\right), \tag {20}
|
| 204 |
+
$$
|
| 205 |
+
|
| 206 |
+
$$
|
| 207 |
+
L _ {s a} ^ {m} (t) = \operatorname {S I} \left(M _ {s a} \odot D _ {t} ^ {m}, M _ {s a} \odot f _ {s} \widehat {D} _ {t}\right), \tag {21}
|
| 208 |
+
$$
|
| 209 |
+
|
| 210 |
+
where the factor $f_{s}$ is due to the assumption in Sec. 4.3 that the depth of an object in $I_{t}$ is $f_{s}$ times of that in $\widetilde{I}_{t}$ . Our SADC constrains the 3D geometric relationship between the scenes of $I_{t}$ and $\widetilde{I}_{t}$ , providing scale-aware regularization. Combining SVDC and SADC, we have the triplet depth consistency loss as:
|
| 211 |
+
|
| 212 |
+
$$
|
| 213 |
+
L _ {t c} (t) = L _ {s v} (t) + L _ {s a} (t) + L _ {s a} ^ {m} (t). \tag {22}
|
| 214 |
+
$$
|
| 215 |
+
|
| 216 |
+
# 4.5 Training Objective
|
| 217 |
+
|
| 218 |
+
For target position $t$ , we have already obtained the self-supervised losses $L_{ss}(t)$ , $L_{ss}^{m}(t)$ and $\widetilde{L}_{ss}(t)$ in Eqs. (5), (16) and (18), respectively. Besides, our triplet depth consistency loss $L_{tc}(t)$ is acquired in Eq. (22). Therefore, the total loss on position $t$ can be summed as:
|
| 219 |
+
|
| 220 |
+
$$
|
| 221 |
+
L (t) = L _ {s s} (t) + L _ {s s} ^ {m} (t) + \widetilde {L} _ {s s} (t) + \lambda L _ {t c} (t), \tag {23}
|
| 222 |
+
$$
|
| 223 |
+
|
| 224 |
+
where $\lambda$ is a hyper-parameter set to 0.2 in this paper. For $t - 1$ and $t + 1$ , we can derive their losses $L(t - 1)$ and $L(t + 1)$ in the same way. In summary, the overall loss function of our Mono-ViFI is $L = L(t - 1) + L(t) + L(t + 1)$ .
|
| 225 |
+
|
| 226 |
+
Table 1: Performance comparison on KITTI [16] benchmark using the raw ground truth, with a resolution of $640 \times 192$ . The best results in each subsection are in bold; second best are underlined. For MACs of multi-frame models, we use the formulation A/B. A denotes the MACs when predicting for a single frame; B means the average MACs per frame when processing a long video sequence. #Params and MACs of our multi-frame models include the VFI network.
|
| 227 |
+
|
| 228 |
+
<table><tr><td>Method</td><td>Test frames</td><td>#Params</td><td>MACs</td><td>Abs Rel</td><td>Sq Rel</td><td>RMSE</td><td>RMSE log</td><td>δ1</td><td>δ2</td><td>δ3</td></tr><tr><td>PackNet-SfM [21]</td><td>1</td><td>128M</td><td>205G</td><td>0.111</td><td>0.785</td><td>4.601</td><td>0.189</td><td>0.878</td><td>0.960</td><td>0.982</td></tr><tr><td>HR-Depth [39]</td><td>1</td><td>16.9M</td><td>31.6G</td><td>0.109</td><td>0.792</td><td>4.632</td><td>0.185</td><td>0.884</td><td>0.962</td><td>0.983</td></tr><tr><td>R-MSFM6 [67]</td><td>1</td><td>3.8M</td><td>31.2G</td><td>0.112</td><td>0.806</td><td>4.704</td><td>0.191</td><td>0.878</td><td>0.960</td><td>0.981</td></tr><tr><td>BRNet [22]</td><td>1</td><td>19.1M</td><td>33.7G</td><td>0.105</td><td>0.698</td><td>4.462</td><td>0.179</td><td>0.890</td><td>0.965</td><td>0.984</td></tr><tr><td>MonoViT [64]</td><td>1</td><td>27M</td><td>-</td><td>0.099</td><td>0.708</td><td>4.372</td><td>0.175</td><td>0.900</td><td>0.967</td><td>0.984</td></tr><tr><td>DIFFNet [65]</td><td>1</td><td>12.2M</td><td>23.4G</td><td>0.102</td><td>0.764</td><td>4.483</td><td>0.180</td><td>0.896</td><td>0.965</td><td>0.983</td></tr><tr><td>SD-SSMDE [44]</td><td>1</td><td>-</td><td>-</td><td>0.100</td><td>0.661</td><td>4.264</td><td>0.172</td><td>0.896</td><td>0.967</td><td>0.985</td></tr><tr><td>MUSTNet [48]</td><td>1</td><td>-</td><td>-</td><td>0.106</td><td>0.763</td><td>4.562</td><td>0.182</td><td>0.888</td><td>0.963</td><td>0.983</td></tr><tr><td>DepthSegNet [40]</td><td>1</td><td>30.3M</td><td>-</td><td>0.104</td><td>0.690</td><td>4.473</td><td>0.179</td><td>0.886</td><td>0.965</td><td>0.984</td></tr><tr><td>DualRefine [2]</td><td>1</td><td>-</td><td>-</td><td>0.103</td><td>0.776</td><td>4.491</td><td>0.181</td><td>0.894</td><td>0.965</td><td>0.983</td></tr><tr><td>DaCNN [23]</td><td>1</td><td>13M</td><td>-</td><td>0.099</td><td>0.661</td><td>4.316</td><td>0.173</td><td>0.897</td><td>0.967</td><td>0.985</td></tr><tr><td>Monodepth2 (ResNet18) [18]</td><td>1</td><td>14.3M</td><td>8G</td><td>0.115</td><td>0.903</td><td>4.863</td><td>0.193</td><td>0.877</td><td>0.959</td><td>0.981</td></tr><tr><td>Mono-ViFI (ResNet18)</td><td>1</td><td>14.3M</td><td>8G</td><td>0.105</td><td>0.708</td><td>4.446</td><td>0.179</td><td>0.887</td><td>0.965</td><td>0.984</td></tr><tr><td>Lite-Mono [62]</td><td>1</td><td>3.1M</td><td>5.1G</td><td>0.107</td><td>0.765</td><td>4.561</td><td>0.183</td><td>0.886</td><td>0.963</td><td>0.983</td></tr><tr><td>Mono-ViFI (Lite-Mono)</td><td>1</td><td>3.1M</td><td>5.1G</td><td>0.103</td><td>0.680</td><td>4.349</td><td>0.176</td><td>0.891</td><td>0.966</td><td>0.985</td></tr><tr><td>D-HRNet [25]</td><td>1</td><td>10M</td><td>10.8G</td><td>0.102</td><td>0.760</td><td>4.479</td><td>0.179</td><td>0.897</td><td>0.965</td><td>0.983</td></tr><tr><td>RA-Depth (D-HRNet) [25]</td><td>1</td><td>10M</td><td>10.8G</td><td>0.096</td><td>0.632</td><td>4.216</td><td>0.171</td><td>0.903</td><td>0.968</td><td>0.985</td></tr><tr><td>Mono-ViFI (D-HRNet)</td><td>1</td><td>10M</td><td>10.8G</td><td>0.096</td><td>0.627</td><td>4.179</td><td>0.170</td><td>0.903</td><td>0.969</td><td>0.985</td></tr><tr><td>Patil et al. [42]</td><td>N</td><td>-</td><td>-/-</td><td>0.111</td><td>0.821</td><td>4.650</td><td>0.187</td><td>0.883</td><td>0.961</td><td>0.982</td></tr><tr><td>ManyDepth (ResNet18) [58]</td><td>2 (-1, 0)</td><td>26.9M</td><td>15.1G/13.7G</td><td>0.098</td><td>0.770</td><td>4.459</td><td>0.176</td><td>0.900</td><td>0.965</td><td>0.983</td></tr><tr><td>DynamicDepth [14]</td><td>2 (-1, 0)</td><td>-</td><td>-/-</td><td>0.096</td><td>0.720</td><td>4.458</td><td>0.175</td><td>0.897</td><td>0.964</td><td>0.984</td></tr><tr><td>DepthFormer [20]</td><td>2 (-1, 0)</td><td>28.7M</td><td>175G/-</td><td>0.090</td><td>0.661</td><td>4.149</td><td>0.175</td><td>0.905</td><td>0.967</td><td>0.984</td></tr><tr><td>MOVEDepth [56]</td><td>2 (-1, 0)</td><td>28.2M</td><td>20.2G/19.2G</td><td>0.094</td><td>0.704</td><td>4.389</td><td>0.175</td><td>0.902</td><td>0.965</td><td>0.983</td></tr><tr><td>DualRefine [2]</td><td>2 (-1, 0)</td><td>-</td><td>-/-</td><td>0.087</td><td>0.698</td><td>4.234</td><td>0.170</td><td>0.914</td><td>0.967</td><td>0.983</td></tr><tr><td>TC-Depth [49]</td><td>3 (-1, 0, 1)</td><td>-</td><td>-/-</td><td>0.103</td><td>0.746</td><td>4.483</td><td>0.185</td><td>0.894</td><td>-</td><td>0.983</td></tr><tr><td>Mono-ViFI (ResNet18)</td><td>3 (-1, 0, 1)</td><td>17.9M</td><td>25.8G/16.9G</td><td>0.099</td><td>0.661</td><td>4.321</td><td>0.174</td><td>0.898</td><td>0.966</td><td>0.985</td></tr><tr><td>Mono-ViFI (Lite-Mono)</td><td>3 (-1, 0, 1)</td><td>5.9M</td><td>21.9G/13.3G</td><td>0.099</td><td>0.648</td><td>4.256</td><td>0.172</td><td>0.898</td><td>0.967</td><td>0.985</td></tr><tr><td>Mono-ViFI (D-HRNet)</td><td>3 (-1, 0, 1)</td><td>12.9M</td><td>35.7G/19.4G</td><td>0.091</td><td>0.589</td><td>4.088</td><td>0.166</td><td>0.912</td><td>0.969</td><td>0.986</td></tr></table>
|
| 229 |
+
|
| 230 |
+
# 5 Experiments
|
| 231 |
+
|
| 232 |
+
# 5.1 Implementation Details
|
| 233 |
+
|
| 234 |
+
Datasets. We train and evaluate the models on two widely used outdoor urban datasets of KITTI [16] and Cityscapes [10]. The depth labels of KITTI are collected with LiDARs. For fair comparison, we use the Eigen split [12] and follow the pre-processing operation in [18], generating 39,810 monocular triplets for training and 697 frames for evaluation. To reduce the influence of noise in the sparse depth from Velodyne, we also evaluate using $93\%$ of the Eigen split with the improved ground truth from [52], which contains 652 test frames. In regard to Cityscapes, following [58, 66], we preprocess the monocular sequences into triples, producing 69,731 training images. And we evaluate on 1,525 test images using the ground truth provided by [58]. For generalization comparison, KITTI models are evaluated on Cityscapes and Make3D [50]. Make3D is a small outdoor dataset containing 134 test images with aligned depth information.
|
| 235 |
+
|
| 236 |
+
Flow-based VFI Model. We choose IFRNet [29] as our VFI model due to its efficiency. In practice, we use the large model of IFRNet for training and replace
|
| 237 |
+
|
| 238 |
+

|
| 239 |
+
Fig. 3: Qualitative results on KITTI. Error maps in columns 2 and 4 show the Abs Rel error compared to the improved ground truth [52] from good (blue) to bad (red).
|
| 240 |
+
|
| 241 |
+

|
| 242 |
+
|
| 243 |
+

|
| 244 |
+
|
| 245 |
+

|
| 246 |
+
|
| 247 |
+
it with the small model when evaluating for multi-frame fusion depth. Please see ablation about their gap in supplementary. Both large and small IFRNets are pretrained on KITTI/Cityscapes and frozen in Mono-ViFI training. The pretraining data split is the same as self-supervised depth training. Each original triplet in depth training is directly used for VFI training. Besides, we employ another two flow-based VFI networks, RIFE [26] and AMT [34], to verify the robustness of our method to different VFI models.
|
| 248 |
+
|
| 249 |
+
Depth and Pose Networks. We experiment with three depth architectures, including the ResNet18 [24] based network in Monodepth2 [18], medium LiteMono [62], and D-HRNet proposed in RA-Depth [25]. For the pose network, we follow Monodepth2 to use a ResNet18 based model which takes a pair of images (six channels) as input and predict a 6-DoF relative pose. Encoders of all the networks are pretrained on ImageNet [11].
|
| 250 |
+
|
| 251 |
+
# 5.2 Performance Comparison
|
| 252 |
+
|
| 253 |
+
KITTI Results. Quantitative comparison results on KITTI under $640 \times 192$ resolution are listed in Table 1 and Table 2, using raw ground truth and improved ground truth, respectively. It is obvious that our method with D-HRNet can achieve state-of-the-art performance. Besides, our Mono-ViFI brings notable improvements to the three depth architectures without increasing model complexity (see each subsection separated by dotted lines). Despite that the difference between Mono-ViFI (D-HRNet) and RA-Depth [25] under single-frame mode is slight in Table 1, it is more significant in Table 2 when using improved ground truth. For multi-frame inference, Mono-ViFI with ResNet18 has already shown advantages over ManyDepth [58] and DynamicDepth [14] on some metrics including Sq Rel, RMSE, RMSE log, $\delta_{2}$ and $\delta_{3}$ . When changing to D-HRNet backbone, proposed Mono-ViFI can totally outperform them. Although our models have more multiply-add computations (MACs) when predicting for a single frame, the average MACs per frame are comparable when processing a long video sequence. Fig. 3 depicts two visual comparison results on KITTI, where
|
| 254 |
+
|
| 255 |
+
Table 2: Performance comparison on Table 3: Performance comparison on KITTI [16] using improved ground truth Cityscapes [10] with a resolution of $416 \times$ from [52], with a resolution of $640 \times 192$ . 128 unless stated.
|
| 256 |
+
|
| 257 |
+
<table><tr><td>Method</td><td>Abs Rel</td><td>Sq Rel</td><td>RMSE</td><td>RMSE log</td><td>δ1</td><td>δ2</td><td>δ3</td></tr><tr><td colspan="8">Single-frame Inference</td></tr><tr><td>PackNet-SIM [21]</td><td>0.078</td><td>0.420</td><td>3.485</td><td>0.121</td><td>0.931</td><td>0.986</td><td>0.996</td></tr><tr><td>R-MSF M6 [67]</td><td>0.088</td><td>0.492</td><td>3.837</td><td>0.135</td><td>0.915</td><td>0.983</td><td>0.995</td></tr><tr><td>DIFFNet [65]</td><td>0.076</td><td>0.414</td><td>3.405</td><td>0.119</td><td>0.936</td><td>0.988</td><td>0.996</td></tr><tr><td>Mono-ViF [64]</td><td>0.074</td><td>0.388</td><td>3.414</td><td>0.115</td><td>0.938</td><td>0.989</td><td>0.997</td></tr><tr><td>Monodepth2 (ResNet18) [18]</td><td>0.090</td><td>0.545</td><td>3.942</td><td>0.137</td><td>0.914</td><td>0.983</td><td>0.995</td></tr><tr><td>Mono-ViF (ResNet18)</td><td>0.084</td><td>0.431</td><td>3.618</td><td>0.126</td><td>0.924</td><td>0.986</td><td>0.997</td></tr><tr><td>Lite-Mono [62]</td><td>0.082</td><td>0.455</td><td>3.685</td><td>0.127</td><td>0.923</td><td>0.985</td><td>0.996</td></tr><tr><td>Mono-ViF (Lite-Mono)</td><td>0.080</td><td>0.400</td><td>3.497</td><td>0.121</td><td>0.930</td><td>0.987</td><td>0.997</td></tr><tr><td>D-HRNet [25]</td><td>0.077</td><td>0.423</td><td>3.396</td><td>0.119</td><td>0.935</td><td>0.987</td><td>0.996</td></tr><tr><td>RA-Depth (D-HRNet) [25]</td><td>0.074</td><td>0.363</td><td>3.349</td><td>0.114</td><td>0.940</td><td>0.990</td><td>0.997</td></tr><tr><td>Mono-ViF (D-HRNet)</td><td>0.072</td><td>0.342</td><td>3.254</td><td>0.111</td><td>0.941</td><td>0.991</td><td>0.998</td></tr><tr><td colspan="8">Multi-frame Inference</td></tr><tr><td>ManyDepth (ResNet18) [58]</td><td>0.070</td><td>0.399</td><td>3.455</td><td>0.113</td><td>0.941</td><td>0.989</td><td>0.997</td></tr><tr><td>MOVEDepth [56]</td><td>0.065</td><td>0.377</td><td>3.449</td><td>0.112</td><td>0.942</td><td>0.988</td><td>0.996</td></tr><tr><td>Mono-ViF (ResNet18)</td><td>0.074</td><td>0.374</td><td>3.381</td><td>0.113</td><td>0.941</td><td>0.990</td><td>0.997</td></tr><tr><td>Mono-ViF (Lite-Mono)</td><td>0.074</td><td>0.354</td><td>3.372</td><td>0.114</td><td>0.939</td><td>0.989</td><td>0.997</td></tr><tr><td>Mono-ViF (D-HRNet)</td><td>0.066</td><td>0.294</td><td>3.099</td><td>0.102</td><td>0.952</td><td>0.992</td><td>0.998</td></tr></table>
|
| 258 |
+
|
| 259 |
+
<table><tr><td>Method</td><td>Abs Rel</td><td>Sq Rel</td><td>RMSE</td><td>RMSE log</td><td>δ1</td><td>δ2</td><td>δ3</td></tr><tr><td colspan="8">Single-frame Inference</td></tr><tr><td>Videos in the wild [19]</td><td>0.127</td><td>1.330</td><td>6.960</td><td>0.195</td><td>0.830</td><td>0.947</td><td>0.981</td></tr><tr><td>Li et al. [33]</td><td>0.119</td><td>1.290</td><td>6.980</td><td>0.190</td><td>0.846</td><td>0.952</td><td>0.982</td></tr><tr><td>InstDM [31] (832 × 256)</td><td>0.111</td><td>1.158</td><td>6.437</td><td>0.182</td><td>0.868</td><td>0.961</td><td>0.983</td></tr><tr><td>SD-SSMDE [44]</td><td>0.110</td><td>0.988</td><td>5.953</td><td>0.165</td><td>-</td><td>-</td><td>-</td></tr><tr><td>Monodepth2 (ResNet18) [18]</td><td>0.129</td><td>1.569</td><td>6.376</td><td>0.197</td><td>0.849</td><td>0.957</td><td>0.983</td></tr><tr><td>Mono-ViFi (ResNet18)</td><td>0.112</td><td>1.003</td><td>5.899</td><td>0.165</td><td>0.872</td><td>0.966</td><td>0.990</td></tr><tr><td>Lite-Mono [62]</td><td>0.114</td><td>1.194</td><td>6.328</td><td>0.172</td><td>0.871</td><td>0.967</td><td>0.989</td></tr><tr><td>Mono-ViFi (Lite-Mono)</td><td>0.107</td><td>0.940</td><td>5.794</td><td>0.159</td><td>0.876</td><td>0.968</td><td>0.991</td></tr><tr><td>D-HRNet [25]</td><td>0.111</td><td>1.142</td><td>6.205</td><td>0.169</td><td>0.879</td><td>0.968</td><td>0.989</td></tr><tr><td>Mono-ViFi (D-HRNet)</td><td>0.105</td><td>0.880</td><td>5.550</td><td>0.156</td><td>0.881</td><td>0.969</td><td>0.992</td></tr><tr><td colspan="8">Multi-frame Inference</td></tr><tr><td>ManyDepth (ResNet18) [58]</td><td>0.114</td><td>1.193</td><td>6.223</td><td>0.170</td><td>0.875</td><td>0.967</td><td>0.989</td></tr><tr><td>DynamicDepth [14]</td><td>0.104</td><td>1.011</td><td>5.987</td><td>0.159</td><td>0.889</td><td>0.972</td><td>0.991</td></tr><tr><td>TC-Depth [49]</td><td>0.110</td><td>0.958</td><td>5.820</td><td>-</td><td>0.867</td><td>-</td><td>0.991</td></tr><tr><td>Mono-ViFi (ResNet18)</td><td>0.107</td><td>0.934</td><td>5.688</td><td>0.159</td><td>0.880</td><td>0.969</td><td>0.990</td></tr><tr><td>Mono-ViFi (Lite-Mono)</td><td>0.105</td><td>0.904</td><td>5.646</td><td>0.157</td><td>0.880</td><td>0.970</td><td>0.991</td></tr><tr><td>Mono-ViFi (D-HRNet)</td><td>0.102</td><td>0.836</td><td>5.395</td><td>0.153</td><td>0.885</td><td>0.971</td><td>0.992</td></tr></table>
|
| 260 |
+
|
| 261 |
+
Table 4: Evaluation results on dynamic objects and static regions of Cityscapes [10] dataset under $416\times 128$ resolution. Note that DynamicDepth relies on depth prior and segmentation masks of dynamic objects when reasoning.
|
| 262 |
+
|
| 263 |
+
<table><tr><td>Method</td><td>Test Area</td><td>Abs Rel</td><td>Sq Rel</td><td>RMSE</td><td>RMSE log</td><td>δ1</td><td>δ2</td><td>δ3</td></tr><tr><td colspan="9">Multi-frame Inference</td></tr><tr><td rowspan="3">ManyDepth (ResNet18) [58]DynamicDepth [14]Mono-ViFI (ResNet18)</td><td rowspan="3">Dynamic Objects</td><td>0.172</td><td>2.206</td><td>5.924</td><td>0.216</td><td>0.792</td><td>0.917</td><td>0.967</td></tr><tr><td>0.144</td><td>1.554</td><td>5.238</td><td>0.185</td><td>0.831</td><td>0.950</td><td>0.982</td></tr><tr><td>0.162</td><td>1.453</td><td>5.114</td><td>0.199</td><td>0.794</td><td>0.924</td><td>0.977</td></tr><tr><td rowspan="3">ManyDepth (ResNet18) [58]DynamicDepth [14]Mono-ViFI (ResNet18)</td><td rowspan="3">Static Regions</td><td>0.107</td><td>1.090</td><td>6.314</td><td>0.160</td><td>0.882</td><td>0.972</td><td>0.991</td></tr><tr><td>0.100</td><td>0.977</td><td>6.155</td><td>0.155</td><td>0.892</td><td>0.974</td><td>0.992</td></tr><tr><td>0.099</td><td>0.907</td><td>5.835</td><td>0.149</td><td>0.890</td><td>0.975</td><td>0.992</td></tr></table>
|
| 264 |
+
|
| 265 |
+
we choose D-HRNet as the baseline depth model. It can be seen that depth maps look qualitatively similar among all four methods. However, the error maps reveal mistakes of D-HRNet baseline and RA-Depth in some ambiguous regions like bushes (see the red boxes). In contrast, our Mono-ViFi have dramatically lower error in these regions, especially under multi-frame reasoning.
|
| 266 |
+
|
| 267 |
+
Cityscapes Results. Results of Cityscapes in Table 3 demonstrate the same trend as KITTI. Moreover, proposed multi-frame model with ResNet18 can beat ManyDepth on all metrics, while our Mono-ViFI is better than DynamicDepth only on partial indexes. It is because Cityscapes with more dynamic scenes is more challenging than KITTI, and DynamicDepth leverages additional information to cope with dynamic objects. For fine-grained analysis, we separately evaluate on the dynamic objects and static regions, using the segmentation masks of dynamic objects provided in [14], as listed in Table 4. On dynamic objects, Mono-ViFI with ResNet18 surpasses ManyDepth by a large margin, demonstrating that our multi-frame inference paradigm can sidestep the issue of dynamic objects to a greater extent compared to ManyDepth. DynamicDepth utilizes depth prior and segmentation masks of dynamic objects to tackle this problem for multi-frame inference. Even though, our method still excels it on Sq Rel and
|
| 268 |
+
|
| 269 |
+

|
| 270 |
+
Fig. 4: Qualitative generalization results of a YouTube video example. All depth maps are predicted under $640 \times 360$ resolution using KITTI models trained under $640 \times 192$ .
|
| 271 |
+
|
| 272 |
+
Table 5: Ablation results of each proposed component on Cityscapes [10]. SA is spatial augmentation. TA is temporal augmentation. SVDC denotes standard view depth consistency. SADC denotes scale-aware depth consistency. FA is feature alignment. MAFA is motion-aware feature alignment. OAFF means occlusion-alleviated feature fusion.
|
| 273 |
+
|
| 274 |
+
<table><tr><td>Method</td><td>SA</td><td>TA</td><td>SVDC</td><td>SADC</td><td>FA</td><td>MAFA</td><td>OAFF</td><td>Abs Rel</td><td>Sq Rel</td><td>RMSE</td><td>δ1</td></tr><tr><td colspan="12">Single-frame Inference</td></tr><tr><td>Baseline (Monodepth2 [18])</td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td>0.129</td><td>1.569</td><td>6.876</td><td>0.849</td></tr><tr><td>Baseline + TA</td><td></td><td>✓</td><td></td><td></td><td></td><td></td><td></td><td>0.119</td><td>1.254</td><td>6.523</td><td>0.857</td></tr><tr><td>Baseline + SVDC</td><td></td><td></td><td>✓</td><td></td><td></td><td></td><td></td><td>0.117</td><td>1.186</td><td>6.492</td><td>0.862</td></tr><tr><td>Baseline + SA</td><td>✓</td><td></td><td></td><td></td><td></td><td></td><td></td><td>0.122</td><td>1.386</td><td>6.557</td><td>0.856</td></tr><tr><td>Baseline + SA + SADC</td><td>✓</td><td></td><td></td><td>✓</td><td></td><td></td><td></td><td>0.116</td><td>1.079</td><td>6.189</td><td>0.868</td></tr><tr><td>Mono-ViFI (Single-frame)</td><td>✓</td><td>✓</td><td>✓</td><td>✓</td><td></td><td></td><td></td><td>0.112</td><td>1.003</td><td>5.899</td><td>0.872</td></tr><tr><td colspan="12">Multi-frame Inference</td></tr><tr><td>Baseline (direct feature fusion)</td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td>0.120</td><td>1.360</td><td>6.491</td><td>0.869</td></tr><tr><td>FA (not motion-aware)</td><td></td><td></td><td></td><td></td><td>✓</td><td></td><td></td><td>0.116</td><td>1.217</td><td>6.387</td><td>0.871</td></tr><tr><td>MAFA</td><td></td><td></td><td></td><td></td><td></td><td>✓</td><td></td><td>0.114</td><td>1.150</td><td>6.115</td><td>0.874</td></tr><tr><td>MAFA + OAFF</td><td></td><td></td><td></td><td></td><td></td><td>✓</td><td>✓</td><td>0.111</td><td>1.107</td><td>6.012</td><td>0.876</td></tr><tr><td>MAFA + OAFF + TA</td><td></td><td>✓</td><td></td><td></td><td></td><td>✓</td><td>✓</td><td>0.109</td><td>1.023</td><td>5.892</td><td>0.877</td></tr><tr><td>Mono-ViFI (Multi-frame)</td><td>✓</td><td>✓</td><td>✓</td><td>✓</td><td></td><td>✓</td><td>✓</td><td>0.107</td><td>0.934</td><td>5.688</td><td>0.880</td></tr></table>
|
| 275 |
+
|
| 276 |
+
RMSE. When evaluating on static regions, proposed Mono-ViFI shows absolute superiority over ManyDepth and DynamicDepth.
|
| 277 |
+
|
| 278 |
+
Generalization Results. We evaluate the KITTI models on Cityscapes and Make3D without fine-tuning, results of which are put in the supplementary. For qualitative comparison, we give a visualization example in Fig. 4, where KITTI models are used to predict the multi-frame depth of a YouTube video frame downloaded from the Wind Walk Travel Videos channel. This indicates that our Mono-ViFI has better generalization ability than existing methods.
|
| 279 |
+
|
| 280 |
+
# 5.3 Ablation Study
|
| 281 |
+
|
| 282 |
+
In this part, we explore the influence of each component in Mono-ViFi. Our ablations are conducted on Cityscapes or KITTI with ResNet18 backbone.
|
| 283 |
+
|
| 284 |
+
Table 6: Ablation of different flow-based VFI models on KITTI [16].
|
| 285 |
+
|
| 286 |
+
<table><tr><td>VFI Model</td><td>Abs Rel</td><td>Sq Rel</td><td>RMSE</td><td>RMSE log</td><td>δ1</td><td>δ2</td><td>δ3</td></tr><tr><td colspan="8">Multi-frame Inference</td></tr><tr><td>IFRNet [29] (Ours)</td><td>0.099</td><td>0.661</td><td>4.321</td><td>0.174</td><td>0.898</td><td>0.966</td><td>0.985</td></tr><tr><td>RIFE [26]</td><td>0.100</td><td>0.672</td><td>4.352</td><td>0.175</td><td>0.896</td><td>0.965</td><td>0.985</td></tr><tr><td>AMT [34]</td><td>0.099</td><td>0.653</td><td>4.335</td><td>0.174</td><td>0.897</td><td>0.966</td><td>0.985</td></tr></table>
|
| 287 |
+
|
| 288 |
+
Single-frame Ablation. We first ablate on the single-frame model. As listed in the top half part of Table 5, the standard view depth consistency and spatial augmentation can bring gains individually. Note that scale-aware depth consistency $L_{sa}$ should be based on spatial augmentation. When adding $L_{sa}$ to spatial augmentation, the accuracy can be further improved. And this leads to the most significant enhancement among all components, since $L_{sa}$ can enforce the model to learn scale relationship from scene variations. In orthogonal dimension, our proposed temporal augmentation can also reduce the prediction error by providing more guidance from newly synthesized camera views. When integrating all above elements, our single-frame model can achieve the best performance.
|
| 289 |
+
|
| 290 |
+
Multi-frame Ablation. To probe the effect of each component in our VFI-assisted multi-frame fusion module, we conduct ablation experiments under multi-frame inference, whose results are summarized in the bottom half part of Table 5. The baseline is stacking the raw feature of three frames directly, i.e., direct feature fusion. Then, we gradually apply components of the proposed fusion module, including motion-aware feature alignment and occlusion-alleviated feature fusion. The results demonstrate that these strategies can bring cumulative improvements. Furthermore, when incorporating spatial augmentation, temporal augmentation and relevant depth consistency losses, our multi-frame model can be further promoted like the single-frame setting.
|
| 291 |
+
|
| 292 |
+
Ablation of different VFI models. Apart from IFRNet [29], we additionally experiment with another two flow-based VFI models, RIFE [26] and AMT [34]. The results are depicted in Table 6. It can be observed that there is only a negligible performance difference among the three VFI networks, which reveals that our method is robust to different VFI models.
|
| 293 |
+
|
| 294 |
+
# 6 Conclusion
|
| 295 |
+
|
| 296 |
+
In conclusion, this paper presents a novel self-supervised learning framework, named Mono-ViFI, to bilaterally connect single- and multi-frame depth estimation in a unified manner. First, a flow-based VFI model is employed for temporal augmentation. Then, we propose a VFI-assisted multi-frame fusion module for more accurate multi-frame depth. Besides, resizing-cropping is extended to affine transformation as spatial augmentation for enriching data diversity. Finally, we develop a new triplet depth consistency loss for better regularization. Experiments show that our approach can unleash the potential of existing depth models to a greater extent, but without increasing inference complexity.
|
| 297 |
+
|
| 298 |
+
# References
|
| 299 |
+
|
| 300 |
+
1. Bae, J., Moon, S., Im, S.: Deep digging into the generalization of self-supervised monocular depth estimation. In: AAAI. pp. 187-196 (2023)
|
| 301 |
+
2. Bangunharcana, A., Magd, A., Kim, K.S.: Dualrefine: Self-supervised depth and pose estimation through iterative epipolar sampling and refinement toward equilibrium. In: CVPR. pp. 726-738 (2023)
|
| 302 |
+
3. Bello, J.L.G., Kim, M.: Forget about the lidar: Self-supervised depth estimators with med probability volumes. In: NeurIPS. pp. 12626-12637 (2020)
|
| 303 |
+
4. Bello, J.L.G., Moon, J., Kim, M.: Positional information is all you need: A novel pipeline for self-supervised svde from videos. arXiv:2205.08851 (2022)
|
| 304 |
+
5. Bian, J., Li, Z., Wang, N., Zhan, H., Shen, C., Cheng, M.M., Reid, I.: Unsupervised scale-consistent depth and ego-motion learning from monocular video. In: NeurIPS. pp. 35-45 (2019)
|
| 305 |
+
6. Chan, K.C., Wang, X., Yu, K., Dong, C., Loy, C.C.: Basicvsr: The search for essential components in video super-resolution and beyond. In: CVPR. pp. 4945-4954 (2021)
|
| 306 |
+
7. Chan, K.C., Zhou, S., Xu, X., Loy, C.C.: Basicvsr++: Improving video superresolution with enhanced propagation and alignment. In: CVPR. pp. 5962-5971 (2022)
|
| 307 |
+
8. Chen, Y., Liu, S., Wang, X.: Learning continuous image representation with local implicit image function. In: CVPR. pp. 8624-8634 (2021)
|
| 308 |
+
9. Chen, Y., Schmid, C., Sminchisescu, C.: Self-supervised learning with geometric constraints in monocular video: Connecting flow, depth, and camera. In: ICCV. pp. 7063-7072 (2019)
|
| 309 |
+
10. Cordts, M., Omran, M., Ramos, S., Rehfeld, T., Enzweiler, M., Benenson, R., Franke, U., Roth, S., Schiele, B.: The cityscapes dataset for semantic urban scene understanding. In: CVPR. pp. 3213-3223 (2016)
|
| 310 |
+
1. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: CVPR. pp. 248-255 (2009)
|
| 311 |
+
2. Eigen, D., Fergus, R.: Predicting depth, surface normals and semantic labels with a common multi-scale convolutional architecture. In: ICCV. pp. 2650-2658 (2015)
|
| 312 |
+
3. Eigen, D., Puhrsch, C., Fergus, R.: Depth map prediction from a single image using a multi-scale deep network. In: NeurIPS. pp. 2366-2374 (2014)
|
| 313 |
+
4. Feng, Z., Yang, L., Jing, L., Wang, H., Tian, Y., Li, B.: Disentangling object motion and occlusion for unsupervised multi-frame monocular depth. In: ECCV. pp. 228-244 (2022)
|
| 314 |
+
5. Garg, R., Kumar, B.G.V., Carneiro, G., Reid, I.: Unsupervised cnn for single view depth estimation: Geometry to the rescue. In: ECCV. pp. 740-756 (2016)
|
| 315 |
+
6. Geiger, A., Lenz, P., Stiller, C., Urtasun, R.: Vision meets robotics: The kitti dataset. *IJRR* **32**(11), 1231–1237 (2013)
|
| 316 |
+
7. Godard, C., Aodha, O.M., Brostow, G.J.: Unsupervised monocular depth estimation with left-right consistency. In: CVPR. pp. 6602-6611 (2017)
|
| 317 |
+
8. Godard, C., Aodha, O.M., Firman, M., Brostow, G.J.: Digging into self-supervised monocular depth estimation. In: CVPR. pp. 3828-3838 (2019)
|
| 318 |
+
9. Gordon, A., Li, H., Jonschkowski, R., Angelova, A.: Depth from videos in the wild: Unsupervised monocular depth learning from unknown cameras. In: ICCV. pp. 8976-8985 (2019)
|
| 319 |
+
20. Guizilini, V., Ambrus, R., Chen, D., Zakharov, S., Gaidon, A.: Multi-frame self-supervised depth with transformers. In: CVPR. pp. 160-170 (2022)
|
| 320 |
+
|
| 321 |
+
21. Guizilini, V., Ambrus, R., Pillai, S., Gaidon, A.: 3d packing for self-supervised monocular depth estimation. In: CVPR. pp. 2482-2491 (2020)
|
| 322 |
+
22. Han, W., Yin, J., Jin, X., Dai, X., Shen, J.: Brnet: Exploring comprehensive features for monocular depth estimation. In: ECCV. pp. 586-602 (2022)
|
| 323 |
+
23. Han, W., Yin, J., Shen, J.: Self-supervised monocular depth estimation by direction-aware cumulative convolution network. In: ICCV (2023)
|
| 324 |
+
24. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: CVPR. pp. 770-778 (2016)
|
| 325 |
+
25. He, M., Hui, L., Bian, Y., Ren, J., Xie, J., Yang, J.: Ra-depth: Resolution adaptive self-supervised monocular depth estimation. In: ECCV. pp. 565-581 (2022)
|
| 326 |
+
26. Huang, Z., Zhang, T., Heng, W., Shi, B., Zhou, S.: Real-time intermediate flow estimation for video frame interpolation. In: ECCV. pp. 624-642 (2022)
|
| 327 |
+
27. Jung, H., Park, E., Yoo, S.: Fine-grained semantics-aware representation enhancement for self-supervised monocular depth estimation. In: ICCV. pp. 12622-12632 (2021)
|
| 328 |
+
28. Klingner, M., Termöhlen, J.A., Mikolajczyk, J., Fingscheidt, T.: Self-supervised monocular depth estimation: Solving the dynamic object problem by semantic guidance. In: ECCV. pp. 582-600 (2020)
|
| 329 |
+
29. Kong, L., Jiang, B., Luo, D., Chu, W., Huang, X., Tai, Y., Wang, C., Yang, J.: Ifrnet: Intermediate feature refine network for efficient frame interpolation. In: CVPR. pp. 1959-1968 (2022)
|
| 330 |
+
30. Kuznietsov, Y., Proesmans, M., Gool, L.V.: Comoda: Continuous monocular depth adaptation using past experiences. In: WACV. pp. 2906-2916 (2021)
|
| 331 |
+
31. Lee, S., Im, S., Lin, S., Kweon, I.S.: Learning monocular depth in dynamic scenes via instance-aware projection consistency. In: AAAI. pp. 1863-1872 (2021)
|
| 332 |
+
32. Lee, S., Rameau, F., Pan, F., Kweon, I.S.: Attentive and contrastive learning for joint depth and motion field estimation. In: ICCV. pp. 4862-4871 (2021)
|
| 333 |
+
33. Li, H., Gordon, A., Zhao, H., Casser, V., Angelova, A.: Unsupervised monocular depth learning in dynamic scenes. In: CoRL. pp. 1908-1917 (2020)
|
| 334 |
+
34. Li, Z., Zhu, Z.L., Han, L.H., Hou, Q., Guo, C.L., Cheng, M.M.: Amt: All-pairs multi-field transforms for efficient frame interpolation. In: CVPR. pp. 9801-9810 (2023)
|
| 335 |
+
35. Liu, J., Kong, L., Yang, J.: Designing and searching for lightweight monocular depth network. In: ICONIP. pp. 477-488 (2021)
|
| 336 |
+
36. Liu, J., Kong, L., Yang, J.: Atca: An arc trajectory based model with curvature attention for video frame interpolation. In: ICIP. pp. 1486-1490 (2022)
|
| 337 |
+
37. Liu, J., Kong, L., Yang, J., Liu, W.: Towards better data exploitation in self-supervised monocular depth estimation. IEEE Robotics and Automation Letters 9(1), 763-770 (2024)
|
| 338 |
+
38. Luo, X., Huang, J.B., Szeliski, R., Matzen, K., Kopf, J.: Consistent video depth estimation. TOG 39(71), 1-13 (2020)
|
| 339 |
+
39. Lyu, X., Liu, L., Wang, M., Kong, X., Liu, L., Liu, Y., Chen, X., Yuan, Y.: Hrdepth: High resolution self-supervised monocular depth estimation. In: AAAI. pp. 2294-2301 (2021)
|
| 340 |
+
40. Ma, J., Lei, X., Liu, N., Zhao, X., Pu, S.: Towards comprehensive representation enhancement in semantics-guided self-supervised monocular depth estimation. In: ECCV. pp. 304-321 (2022)
|
| 341 |
+
41. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. In: ECCV. pp. 405-421 (2020)
|
| 342 |
+
|
| 343 |
+
42. Patil, V., Gansbeke, W.V., Dai, D., Gool, L.V.: Don't forget the past: Recurrent depth estimation from monocular video. IEEE Robotics and Automation Letters 5(4), 6813-6820 (2020)
|
| 344 |
+
43. Peng, R., Wang, R., Lai, Y., Tang, L., Cai, Y.: Excavating the potential capacity of selfsupervised monocular depth estimation. In: ICCV. pp. 15560-15569 (2021)
|
| 345 |
+
44. Petrovai, A., Nedevschi, S.: Exploiting pseudo labels in a self-supervised learning framework for improved monocular depth estimation. In: CVPR. pp. 1568-1578 (2022)
|
| 346 |
+
45. Pilzer, A., Lathuilière, S., Sebe, N., Ricci, E.: Refine and distill: Exploiting cycle-inconsistency and knowledge distillation for unsupervised monocular depth estimation. In: CVPR. pp. 9760-9769 (2019)
|
| 347 |
+
46. Poggi, M., Aleotti, F., Tosi, F., Mattoccia, S.: On the uncertainty of self-supervised monocular depth estimation. In: CVPR. pp. 3224-3234 (2020)
|
| 348 |
+
47. Ranjan, A., Jampani, V., Balles, L., Kim, K., Sun, D., Wulff, J., Black, M.J.: Competitive collaboration: Joint unsupervised learning of depth, camera motion, optical flow and motion segmentation. In: CVPR. pp. 12240-12249 (2019)
|
| 349 |
+
48. Ren, W., Wang, L., Piao, Y., Zhang, M., Lu, H., Liu, T.: Adaptive co-teaching for unsupervised monocular depth estimation. In: ECCV. pp. 89-105 (2022)
|
| 350 |
+
49. Ruhkamp, P., Gao, D., Chen, H., Navab, N., Busam, B.: Attention meets geometry: Geometry guided spatial-temporal attention for consistent self-supervised monococular depth estimation. In: 3DV. pp. 837-847 (2021)
|
| 351 |
+
50. Saxena, A., Sun, M., Ng, A.Y.: Make3d: Learning 3d scene structure from a single still image. PAMI 31(5), 824-840 (2009)
|
| 352 |
+
51. Tancik, M., Srinivasan, P.P., Mildenhall, B., Fridovich-Keil, S., Raghavan, N., Singhal, U., Ramamoorthi, R., Barron, J.T., Ng, R.: Fourier features let networks learn high frequency functions in low dimensional domains. In: NeurIPS. pp. 7537-7547 (2020)
|
| 353 |
+
52. Uhrig, J., Schneider, N., Schneider, L., Franke, U., Brox, T., Geiger, A.: Sparsity invariant cnns. In: 3DV. pp. 11-20 (2017)
|
| 354 |
+
53. Wang, C., Buenaposada, J.M., Zhu, R., Lucey, S.: Learning depth from monocular videos using direct methods. In: CVPR. pp. 2022-2030 (2018)
|
| 355 |
+
54. Wang, R., Pizer, S.M., Frahm, J.M.: Recurrent neural network for (un-)supervised learning of monocular video visual odometry and depth. In: CVPR. pp. 5555-5564 (2019)
|
| 356 |
+
55. Wang, R., Yu, Z., Gao, S.: Planedepth: Self-supervised depth estimation via orthogonal planes. In: CVPR. pp. 21425-21434 (2023)
|
| 357 |
+
56. Wang, X., Zhu, Z., Huang, G., Chi, X., Ye, Y., Chen, Z., Wang, X.: Drafting monocular cues and velocity guidance for self-supervised multi-frame depth learning. In: AAAI. pp. 2689-2697 (2023)
|
| 358 |
+
57. Wang, Z., Bovik, A.C., Sheikh, H.R., Simoncelli, E.P.: Image quality assessment: From error visibility to structural similarity. IEEE TIP 13(4), 600-612 (2004)
|
| 359 |
+
58. Watson, J., Aodha, O.M., Prisacariu, V.A., Brostow, G.J., Firman, M.: The temporal opportunist: Self-supervised multi-frame monocular depth. In: CVPR. pp. 1164-1174 (2021)
|
| 360 |
+
59. Watson, J., Firman, M., Brostow, G.J., Turmukhambetov, D.: Self-supervised monocular depth hints. In: ICCV. pp. 2162-2171 (2019)
|
| 361 |
+
60. Yin, Z., Shi, J.: Geonet: Unsupervised learning of dense depth, optical flow and camera pose. In: CVPR. pp. 1983-1992 (2018)
|
| 362 |
+
61. Zhang, H., Li, Y., Cao, Y., Liu, Y., Shen, C., Yan, Y.: Exploiting temporal consistency for real-time video depth estimation. In: ICCV. pp. 1725-1734 (2019)
|
| 363 |
+
|
| 364 |
+
62. Zhang, N., Nex, F., Vosselman, G., Kerle, N.: Lite-mono: A lightweight cnn and transformer architecture for self-supervised monocular depth estimation. In: CVPR. pp. 18537-18546 (2023)
|
| 365 |
+
63. Zhao, C., Poggi, M., Tosi, F., Zhou, L., Sun, Q., Tanga, Y., Mattoccia, S.: Gasmono: Geometry-aided self-supervised monocular depth estimation for indoor scenes. In: ICCV. pp. 16163-16174 (2023)
|
| 366 |
+
64. Zhao, C., Zhang, Y., Poggi, M., Tosi, F., Guo, X., Zhu, Z., Huang, G., Tang, Y., Mattoccia, S.: Monovit: Self-supervised monocular depth estimation with a vision transformer. In: 3DV. pp. 668-678 (2022)
|
| 367 |
+
65. Zhou, H., Greenwood, D., Taylor, S.: Self-supervised monocular depth estimation with internal feature fusion. In: BMVC (2021)
|
| 368 |
+
66. Zhou, T., Brown, M., Snavely, N., Lowe, D.G.: Unsupervised learning of depth and ego-motion from video. In: CVPR. pp. 6612-6619 (2017)
|
| 369 |
+
67. Zhou, Z., Fan, X., Shi, P., Xin, Y.: R-msfm: Recurrent multi-scale feature modulation for monocular depth estimating. In: ICCV. pp. 12757-12766 (2021)
|
2024/Mono-ViFI_ A Unified Learning Framework for Self-supervised Single- and Multi-frame Monocular Depth Estimation/images.zip
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:a5bcf89f9c911ff4e9152b1aeb8b6f359e6b9122a2091d4bcbc9b9716565442c
|
| 3 |
+
size 730527
|
2024/Mono-ViFI_ A Unified Learning Framework for Self-supervised Single- and Multi-frame Monocular Depth Estimation/layout.json
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
2024/MonoTTA_ Fully Test-Time Adaptation for Monocular 3D Object Detection/cfa6dde5-bde3-4186-b95b-8bbf6c77c339_content_list.json
ADDED
|
@@ -0,0 +1,1590 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
{
|
| 3 |
+
"type": "text",
|
| 4 |
+
"text": "MonoTTA: Fully Test-Time Adaptation for Monocular 3D Object Detection",
|
| 5 |
+
"text_level": 1,
|
| 6 |
+
"bbox": [
|
| 7 |
+
256,
|
| 8 |
+
140,
|
| 9 |
+
750,
|
| 10 |
+
186
|
| 11 |
+
],
|
| 12 |
+
"page_idx": 0
|
| 13 |
+
},
|
| 14 |
+
{
|
| 15 |
+
"type": "text",
|
| 16 |
+
"text": "Hongbin Lin $^{1,2*}$ , Yifan Zhang $^{3,4*}$ , Shuaicheng Niu $^{5*}$ , Shuguang Cui $^{2,1}$ , Zhen Li $^{2,1\\dagger}$",
|
| 17 |
+
"bbox": [
|
| 18 |
+
292,
|
| 19 |
+
210,
|
| 20 |
+
709,
|
| 21 |
+
243
|
| 22 |
+
],
|
| 23 |
+
"page_idx": 0
|
| 24 |
+
},
|
| 25 |
+
{
|
| 26 |
+
"type": "text",
|
| 27 |
+
"text": "$^{1}$ FNii-Shenzhen, $^{2}$ SSE, CUHK-Shenzhen, $^{3}$ NUS, $^{4}$ Skywork AI, $^{5}$ NTU {hongbinlin@link.,shuguangcui@,lizhen@}cuhk.edu.cn, yifan.zhang@u.nus.edu, shuaicheng.niu@ntu.edu.sg",
|
| 28 |
+
"bbox": [
|
| 29 |
+
259,
|
| 30 |
+
253,
|
| 31 |
+
743,
|
| 32 |
+
297
|
| 33 |
+
],
|
| 34 |
+
"page_idx": 0
|
| 35 |
+
},
|
| 36 |
+
{
|
| 37 |
+
"type": "text",
|
| 38 |
+
"text": "Abstract. Monocular 3D object detection (Mono 3Det) aims to identify 3D objects from a single RGB image. However, existing methods often assume training and test data follow the same distribution, which may not hold in real-world test scenarios. To address the out-of-distribution (OOD) problems, we explore a new adaptation paradigm for Mono 3Det, termed Fully Test-time Adaptation which aims to adapt a well-trained model to unlabeled test data by handling potential data distribution shifts at test time. However, applying this paradigm in Mono 3Det poses significant challenges due to OOD test data causing a remarkable decline in object detection scores. This decline conflicts with the pre-defined score thresholds of existing detection methods, leading to severe object omissions (i.e., rare positive detections and many false negatives). Consequently, the limited positive detection and plenty of noisy predictions cause test-time adaptation to fail in Mono 3Det. To handle this problem, we propose a novel Monocular Test-Time Adaptation (MonoTTA) method, based on two new strategies. 1) Reliability-driven adaptation: we empirically find that high-score objects are still reliable and the optimization of high-score objects can enhance confidence across all detections. Thus, we devise a self-adaptive strategy to identify reliable objects for model adaptation, which discovers potential objects and alleviates omissions. 2) Noise-guard adaptation: since high-score objects may be scarce, we develop a negative regularization term to exploit the numerous low-score objects via negative learning, preventing overfitting to noise and trivial solutions. Experimental results show that MonoTTA brings significant performance gains for Mono 3Det models in OOD test scenarios, approximately $190\\%$ gains by average on KITTI and $198\\%$ gains on nuScenes. The source code is now available at Hongbin98/MonoTTA.",
|
| 39 |
+
"bbox": [
|
| 40 |
+
261,
|
| 41 |
+
323,
|
| 42 |
+
743,
|
| 43 |
+
698
|
| 44 |
+
],
|
| 45 |
+
"page_idx": 0
|
| 46 |
+
},
|
| 47 |
+
{
|
| 48 |
+
"type": "text",
|
| 49 |
+
"text": "Keywords: Test-time Adaptation $\\cdot$ Monocular 3D Object Detection",
|
| 50 |
+
"bbox": [
|
| 51 |
+
261,
|
| 52 |
+
710,
|
| 53 |
+
723,
|
| 54 |
+
724
|
| 55 |
+
],
|
| 56 |
+
"page_idx": 0
|
| 57 |
+
},
|
| 58 |
+
{
|
| 59 |
+
"type": "text",
|
| 60 |
+
"text": "1 Introduction",
|
| 61 |
+
"text_level": 1,
|
| 62 |
+
"bbox": [
|
| 63 |
+
215,
|
| 64 |
+
746,
|
| 65 |
+
375,
|
| 66 |
+
762
|
| 67 |
+
],
|
| 68 |
+
"page_idx": 0
|
| 69 |
+
},
|
| 70 |
+
{
|
| 71 |
+
"type": "text",
|
| 72 |
+
"text": "Three-dimensional (3D) Object Detection is a significant computer vision task, with the objective of identifying objects and determining their spatial and di",
|
| 73 |
+
"bbox": [
|
| 74 |
+
212,
|
| 75 |
+
773,
|
| 76 |
+
785,
|
| 77 |
+
805
|
| 78 |
+
],
|
| 79 |
+
"page_idx": 0
|
| 80 |
+
},
|
| 81 |
+
{
|
| 82 |
+
"type": "page_footnote",
|
| 83 |
+
"text": "* Authors contributed equally.",
|
| 84 |
+
"bbox": [
|
| 85 |
+
217,
|
| 86 |
+
810,
|
| 87 |
+
426,
|
| 88 |
+
825
|
| 89 |
+
],
|
| 90 |
+
"page_idx": 0
|
| 91 |
+
},
|
| 92 |
+
{
|
| 93 |
+
"type": "page_footnote",
|
| 94 |
+
"text": "† Corresponding authors.",
|
| 95 |
+
"bbox": [
|
| 96 |
+
222,
|
| 97 |
+
825,
|
| 98 |
+
392,
|
| 99 |
+
839
|
| 100 |
+
],
|
| 101 |
+
"page_idx": 0
|
| 102 |
+
},
|
| 103 |
+
{
|
| 104 |
+
"type": "image",
|
| 105 |
+
"img_path": "images/8eb188ab9bc574c6820cd170f345d2d622008005abc8b11024dcb6d4b96ad52a.jpg",
|
| 106 |
+
"image_caption": [
|
| 107 |
+
"Fig. 1: An illustration of the generalizability issue of Mono 3Det models. Compared with in-distribution (In-dis) scenarios (e.g., sunny), the detection scores within out-of-distribution (OOD) test data suffer severe degradation when the well-trained model (MonoFlex [48]) is directly applied to test scenarios affected by common natural disruptions, like weather changes (e.g., snow and fog). Since existing Mono 3Det methods mainly adopt a pre-defined score threshold (e.g., 0.2) for object detection, it leads to severe omissions and unreliable detections, thereby suffering serious performance degradation. Note that test images are the same but under different weather conditions."
|
| 108 |
+
],
|
| 109 |
+
"image_footnote": [],
|
| 110 |
+
"bbox": [
|
| 111 |
+
222,
|
| 112 |
+
143,
|
| 113 |
+
782,
|
| 114 |
+
296
|
| 115 |
+
],
|
| 116 |
+
"page_idx": 1
|
| 117 |
+
},
|
| 118 |
+
{
|
| 119 |
+
"type": "text",
|
| 120 |
+
"text": "mensional attributes through diverse sensor inputs [3,5,16,37,38]. To reduce the cost of sensors, there is an increasing trend towards implementing autonomous driving systems via Monocular 3D Object Detection (Mono 3Det) [2,42], where only one single RGB image and the camera calibration information are given. Even if this practical task is challenging, Mono 3Det methods have achieved promising results across various tasks and datasets [4,22,29,40,48]. Behind the success, a common presupposition is assuming that test images have the same distribution as the training images. However, this assumption could be possibly invalidated in many real-world scenarios due to prevalent natural corruptions such as weather changes, diminished sharpness, and other factors that introduce noise and contribute to uncalibrated cameras. In such circumstances, the well-trained model often suffers substantial performance degradation as a consequence of the data distributional shifts between the training images and the unlabeled test images. As shown in Fig. 1, the model performance degrades from $46.2\\mathrm{mAP}$ in in-distribution data to $0.3\\mathrm{mAP}$ in Snow and $7.2\\mathrm{mAP}$ in Fog. Considering the widespread application of Mono 3Det in autonomous driving, the severe performance degradation due to out-of-distribution (OOD) test data may lead to unexpected traffic accidents and pose serious safety risks. Therefore, it is crucial to deal with the OOD generalization problem for Mono 3Det.",
|
| 121 |
+
"bbox": [
|
| 122 |
+
212,
|
| 123 |
+
443,
|
| 124 |
+
787,
|
| 125 |
+
731
|
| 126 |
+
],
|
| 127 |
+
"page_idx": 1
|
| 128 |
+
},
|
| 129 |
+
{
|
| 130 |
+
"type": "text",
|
| 131 |
+
"text": "In addressing the OOD challenges specifically in test scenarios, one paradigm that has emerged as highly promising and gaining traction is Test-Time Adaptation (TTA), which seeks to tackle data distribution shifts by adapting a well-trained model to unlabeled test images in real time. Test-time training (TTT) [34] represents an initial approach of TTA in classification tasks, by adjusting the well-trained model to predict rotations through additional model training, while its computation demands at the adaptation stage are prohibitive",
|
| 132 |
+
"bbox": [
|
| 133 |
+
212,
|
| 134 |
+
734,
|
| 135 |
+
787,
|
| 136 |
+
840
|
| 137 |
+
],
|
| 138 |
+
"page_idx": 1
|
| 139 |
+
},
|
| 140 |
+
{
|
| 141 |
+
"type": "page_number",
|
| 142 |
+
"text": "2",
|
| 143 |
+
"bbox": [
|
| 144 |
+
217,
|
| 145 |
+
114,
|
| 146 |
+
228,
|
| 147 |
+
126
|
| 148 |
+
],
|
| 149 |
+
"page_idx": 1
|
| 150 |
+
},
|
| 151 |
+
{
|
| 152 |
+
"type": "header",
|
| 153 |
+
"text": "Hongbin Lin, Yifan Zhang and et al.",
|
| 154 |
+
"bbox": [
|
| 155 |
+
271,
|
| 156 |
+
114,
|
| 157 |
+
514,
|
| 158 |
+
128
|
| 159 |
+
],
|
| 160 |
+
"page_idx": 1
|
| 161 |
+
},
|
| 162 |
+
{
|
| 163 |
+
"type": "text",
|
| 164 |
+
"text": "in Mono 3Det applications, particularly in autonomous driving. To enhance efficiency, Tent [36] and EATA [26] have been developed for Fully Test-Time Adaptation (Fully TTA) where only unlabeled test images and a well-trained model are provided. Besides, Ev-TTA [12] and SOD [35] devise TTA methods to handle the event-based object recognition and weakly supervised salient object detection, respectively. Considering the constraints on time of Mono 3Det, we explore the fully TTA paradigm which seeks to deal with OOD test data in real time.",
|
| 165 |
+
"bbox": [
|
| 166 |
+
212,
|
| 167 |
+
146,
|
| 168 |
+
787,
|
| 169 |
+
253
|
| 170 |
+
],
|
| 171 |
+
"page_idx": 2
|
| 172 |
+
},
|
| 173 |
+
{
|
| 174 |
+
"type": "text",
|
| 175 |
+
"text": "To investigate this paradigm for Mono 3Det, we dig into the detection outcomes for objects within test scenarios with variations or corruptions that are commonly caused by weather or cameras. Specifically, we directly apply the well-trained model to the validation set of KITTI which has been artificially injected with four distinct types of weather-related corruptions, namely Brightness, Frost, Fog, and Snow. Subsequently, we plot their distributions of detection scores (c.f. Fig. 1). It is observed that the detection scores of test objects tend to markedly decline as well as the high-score objects are scarce in the extreme scenario (Snow) when the well-trained model is directly applied to the scenarios with corruptions. This phenomenon indicates that: 1) The pre-trained Mono 3Det model struggles to discriminate between objects and the background within OOD test data, presenting as quantities of omissions and unconfident detections. 2) Directly applying existing fully TTA methods to Mono 3Det could only get suboptimal performance since they struggle to optimize the model without enough high-score (positive) detections, especially in certain extreme scenarios.",
|
| 176 |
+
"bbox": [
|
| 177 |
+
212,
|
| 178 |
+
258,
|
| 179 |
+
789,
|
| 180 |
+
484
|
| 181 |
+
],
|
| 182 |
+
"page_idx": 2
|
| 183 |
+
},
|
| 184 |
+
{
|
| 185 |
+
"type": "text",
|
| 186 |
+
"text": "To handle it in Mono 3Det, we propose a Monocular Test-Time Adaptation (MonoTTA) method, consisting of the reliability-driven adaptation and noise-guard adaptation strategies: 1) Reliability-driven adaptation. Specifically, data distribution shifts lead to omissions and noisy detections while our empirical analysis suggests that high detection score objects are still reliable (c.f. Fig. 3 (a)). Moreover, even if we only optimize the model via high-score objects ( $e.g., \\geq 0.5$ ), both the numbers of low-score and high-score objects increase (c.f. Fig. 3 (b)). These investigations motivate us that exploiting high-score objects rather than all objects for model adaptation would be a more reliable way to alleviate data distribution shifts and discover potential objects. Hence, we develop a self-adaptive strategy for the identification of reliable high-score objects in test images and devise the adaptive optimization loss $\\mathcal{L}_{AO}$ to exploit the reliable subset for model adaptation, alleviating the detection score decline issue of OOD test data and digging out more potential objects. 2) Noise-guard adaptation. In addition, data distribution shifts may also result in a scarcity of high-score objects, i.e., the majority of objects presenting low scores as the 'Snow' scenario in Fig. 3 (a). To this end, we develop a negative regularization term to make rational use of the numerous low-scoring objects in the Negative Learning manner [13]. On the one hand, the negative regularization term $\\mathcal{L}_{Nreg}$ allows the model to conduct adaptation via numerous noisy low-scoring objects. Thus, the model can achieve more high-score objects after alleviating distribution shifts. On the other hand, this term also prevents the model from overfitting to noise and trivial solutions, i.e., assigning all classes of one object with high scores.",
|
| 187 |
+
"bbox": [
|
| 188 |
+
212,
|
| 189 |
+
492,
|
| 190 |
+
789,
|
| 191 |
+
843
|
| 192 |
+
],
|
| 193 |
+
"page_idx": 2
|
| 194 |
+
},
|
| 195 |
+
{
|
| 196 |
+
"type": "header",
|
| 197 |
+
"text": "MonoTTA",
|
| 198 |
+
"bbox": [
|
| 199 |
+
658,
|
| 200 |
+
114,
|
| 201 |
+
730,
|
| 202 |
+
127
|
| 203 |
+
],
|
| 204 |
+
"page_idx": 2
|
| 205 |
+
},
|
| 206 |
+
{
|
| 207 |
+
"type": "page_number",
|
| 208 |
+
"text": "3",
|
| 209 |
+
"bbox": [
|
| 210 |
+
774,
|
| 211 |
+
116,
|
| 212 |
+
785,
|
| 213 |
+
126
|
| 214 |
+
],
|
| 215 |
+
"page_idx": 2
|
| 216 |
+
},
|
| 217 |
+
{
|
| 218 |
+
"type": "text",
|
| 219 |
+
"text": "We summarize the main contributions as follows:",
|
| 220 |
+
"bbox": [
|
| 221 |
+
240,
|
| 222 |
+
146,
|
| 223 |
+
593,
|
| 224 |
+
161
|
| 225 |
+
],
|
| 226 |
+
"page_idx": 3
|
| 227 |
+
},
|
| 228 |
+
{
|
| 229 |
+
"type": "list",
|
| 230 |
+
"sub_type": "text",
|
| 231 |
+
"list_items": [
|
| 232 |
+
"- To the best of our knowledge, we are the first to explore Fully Test-Time Adaptation to address OOD generalization problems for Mono 3Det. We show that the explored novel paradigm can bring significant improvements to Mono 3Det models in OOD test scenarios, e.g., $137\\%$ and $244\\%$ average performance gains across 13 types of OOD shifts on KITTI.",
|
| 233 |
+
"- Our empirical investigation reveals an important insight that high-score objects maintain their reliability amidst various corruptions, while optimizing these high-score objects significantly boosts model confidence across all detections. This motivates the first effective test-time adaptation method (i.e., our MonoTTA) in Mono 3Det.",
|
| 234 |
+
"- Extensive experiments on 13 types of corruptions of KITTI and 2 real scenarios (daytime $\\leftrightarrow$ night) of nuScenes demonstrate the effectiveness of our MonoTTA in boosting existing Mono 3Det methods [29,48] to handle test-time OOD problems. Even for instance-level methods [40], MonoTTA also maintains sufficient improvement, which further confirms its applicability."
|
| 235 |
+
],
|
| 236 |
+
"bbox": [
|
| 237 |
+
225,
|
| 238 |
+
169,
|
| 239 |
+
782,
|
| 240 |
+
393
|
| 241 |
+
],
|
| 242 |
+
"page_idx": 3
|
| 243 |
+
},
|
| 244 |
+
{
|
| 245 |
+
"type": "text",
|
| 246 |
+
"text": "2 Related Work",
|
| 247 |
+
"text_level": 1,
|
| 248 |
+
"bbox": [
|
| 249 |
+
215,
|
| 250 |
+
416,
|
| 251 |
+
387,
|
| 252 |
+
431
|
| 253 |
+
],
|
| 254 |
+
"page_idx": 3
|
| 255 |
+
},
|
| 256 |
+
{
|
| 257 |
+
"type": "text",
|
| 258 |
+
"text": "We first review the literature on Monocular 3D Object Detection, and then discuss Source-free Domain Adaptation and Test-Time Adaptation methods. More discussions on Unsupervised Domain Adaptation [18,41,47] are in Appendix A. Monocular 3D Object Detection aims to perceive 3D objects from a single 2D image. Existing Mono 3Det methods could be divided into two groups according to the use of extra information. On the one hand, some existing methods leverage extra pre-trained depth estimation modules [6,39,49] to solve one of the most difficult problems in Mono 3Det, i.e., depth estimation from a single image. Other methods utilize LiDAR information, e.g., generating pseudo-LiDAR [23,31,37]. It is worth noting that Monoground [29] proposes to introduce the ground plane as prior information, and MonoNeRD [40] proposes to utilize scene geometric clues to enhance the detector's performance in the implicit reconstruction manner. On the other hand, some Mono 3Det methods try to detect 3D objects without extra data. For example, SMOKE [20] proposes to detect 3D objects as the key points estimation task. Then, Monoflex [48] improves this idea by providing a flexible definition of object centers, which unifies the centers of regular and truncated objects. GrooMeD-NMS [15] proposes a grouped mathematically differentiable Non-Maximal Suppression for Mono 3Det.",
|
| 259 |
+
"bbox": [
|
| 260 |
+
212,
|
| 261 |
+
446,
|
| 262 |
+
784,
|
| 263 |
+
715
|
| 264 |
+
],
|
| 265 |
+
"page_idx": 3
|
| 266 |
+
},
|
| 267 |
+
{
|
| 268 |
+
"type": "text",
|
| 269 |
+
"text": "Source-free Domain Adaptation (SFDA) aims to adapt the pre-trained source model to an unlabeled target domain without using the source data due to privacy issues [17, 30]. SF-UDA $^3$ D [32] first explores the SFDA framework to adapt the PointRCNN 3D detector to target domains, which consists of pseudolabeling, reversible scale-transformations and motion coherency. Recently, the authors [10] seek to exploit the source model more reliably and propose an uncertainty-aware teacher-student framework to filter incorrect pseudo labels during model adaptation, alleviating the negative impact of label noise.",
|
| 270 |
+
"bbox": [
|
| 271 |
+
212,
|
| 272 |
+
719,
|
| 273 |
+
782,
|
| 274 |
+
839
|
| 275 |
+
],
|
| 276 |
+
"page_idx": 3
|
| 277 |
+
},
|
| 278 |
+
{
|
| 279 |
+
"type": "page_number",
|
| 280 |
+
"text": "4",
|
| 281 |
+
"bbox": [
|
| 282 |
+
217,
|
| 283 |
+
114,
|
| 284 |
+
228,
|
| 285 |
+
126
|
| 286 |
+
],
|
| 287 |
+
"page_idx": 3
|
| 288 |
+
},
|
| 289 |
+
{
|
| 290 |
+
"type": "header",
|
| 291 |
+
"text": "Hongbin Lin, Yifan Zhang and et al.",
|
| 292 |
+
"bbox": [
|
| 293 |
+
271,
|
| 294 |
+
114,
|
| 295 |
+
514,
|
| 296 |
+
128
|
| 297 |
+
],
|
| 298 |
+
"page_idx": 3
|
| 299 |
+
},
|
| 300 |
+
{
|
| 301 |
+
"type": "text",
|
| 302 |
+
"text": "Nonetheless, SFDA assumes all target data to be known in advance and makes predictions after multiple epochs of optimization, which may not be viable for real-time applications due to computational or time constraints.",
|
| 303 |
+
"bbox": [
|
| 304 |
+
212,
|
| 305 |
+
146,
|
| 306 |
+
782,
|
| 307 |
+
191
|
| 308 |
+
],
|
| 309 |
+
"page_idx": 4
|
| 310 |
+
},
|
| 311 |
+
{
|
| 312 |
+
"type": "text",
|
| 313 |
+
"text": "Test-Time Adaptation (TTA) seeks to improve model performance on test data via model adaptation through test samples even if data shifts exist. Early TTA methods [19, 34] endeavor to conduct additional model optimization on training data by self-supervised objectives, and then adapt the well-trained model to the test data via self-supervised objectives. However, in Mono 3Det applications like autonomous driving, the computation demands of such methods are prohibitive. To solve this, Fully Test-Time Adaptation methods are developed to adapt the well-trained model, where only unlabeled test images are available. Specifically, certain methods [25, 27, 33] tackle data distribution shifts by adapting the batch normalization layer statistics, while others alleviate this issue either by the entropy minimization of test data [8, 36] or maximizing the prediction consistency of different augmentations [43-46]. As for object detection tasks, Ev-TTA [12] and SOD [35] try to handle the event-based object recognition and weakly supervised salient object detection offline, respectively.",
|
| 314 |
+
"bbox": [
|
| 315 |
+
212,
|
| 316 |
+
191,
|
| 317 |
+
784,
|
| 318 |
+
402
|
| 319 |
+
],
|
| 320 |
+
"page_idx": 4
|
| 321 |
+
},
|
| 322 |
+
{
|
| 323 |
+
"type": "text",
|
| 324 |
+
"text": "However, existing fully TTA methods struggle to optimize the model and solve distribution shifts in Mono 3Det due to numerous false negative detections. To the best of our knowledge, our MonoTTA stands as the first fully TTA method that handles distribution shifts for Mono 3Det models in real time.",
|
| 325 |
+
"bbox": [
|
| 326 |
+
212,
|
| 327 |
+
404,
|
| 328 |
+
784,
|
| 329 |
+
464
|
| 330 |
+
],
|
| 331 |
+
"page_idx": 4
|
| 332 |
+
},
|
| 333 |
+
{
|
| 334 |
+
"type": "text",
|
| 335 |
+
"text": "3 Monocular Test-Time Adaptation",
|
| 336 |
+
"text_level": 1,
|
| 337 |
+
"bbox": [
|
| 338 |
+
214,
|
| 339 |
+
489,
|
| 340 |
+
578,
|
| 341 |
+
508
|
| 342 |
+
],
|
| 343 |
+
"page_idx": 4
|
| 344 |
+
},
|
| 345 |
+
{
|
| 346 |
+
"type": "text",
|
| 347 |
+
"text": "3.1 Problem Statement",
|
| 348 |
+
"text_level": 1,
|
| 349 |
+
"bbox": [
|
| 350 |
+
214,
|
| 351 |
+
523,
|
| 352 |
+
423,
|
| 353 |
+
539
|
| 354 |
+
],
|
| 355 |
+
"page_idx": 4
|
| 356 |
+
},
|
| 357 |
+
{
|
| 358 |
+
"type": "text",
|
| 359 |
+
"text": "Without loss of generality, we denote the pre-trained (or well-trained) model as $f_{\\Theta_0}(\\mathbf{x})$ , which is achieved via training on labeled training images $\\{(\\mathbf{x}_i^s,\\mathbf{y}_i^s)\\}_{i = 1}^N$ . The training images follow the training distribution $P(\\mathbf{x})$ (i.e., $\\mathbf{x}^s\\sim P(\\mathbf{x})$ ). Here, $\\Theta_0$ represents the parameters of the pre-trained model and $N$ is the number of training data. During the training stage, the model is optimized to fit (or overfit) the training data. Then, at the test stage, the model will be able to perform well if the unlabeled test images $\\mathcal{D}_t = \\{\\mathbf{x}_i\\}_{i = 1}^{N_t}$ follows the identical data distribution, i.e., $\\mathbf{x}\\sim P(\\mathbf{x})$ where $N_{t}$ is the total number of test images. However, in real applications, it is possible for the pre-trained model to encounter Out-Of-Distribution (OOD) test samples due to prevalent natural corruptions, namely distribution shifts, i.e., $\\mathbf{x}\\sim Q(\\mathbf{x})$ and $P(\\mathbf{x})\\neq Q(\\mathbf{x})$ .",
|
| 360 |
+
"bbox": [
|
| 361 |
+
212,
|
| 362 |
+
551,
|
| 363 |
+
784,
|
| 364 |
+
718
|
| 365 |
+
],
|
| 366 |
+
"page_idx": 4
|
| 367 |
+
},
|
| 368 |
+
{
|
| 369 |
+
"type": "text",
|
| 370 |
+
"text": "To address this issue, fully test-time adaptation [36] seeks to tackle distribution shifts by adapting the pre-trained model $f_{\\Theta_0}(\\mathbf{x})$ to unlabeled test images $\\{\\mathbf{x}_i\\}_{i=1}^{N_t}$ in real time. To achieve this goal, existing methods typically endeavor to update the model through the minimization of unsupervised objectives defined on test samples by $\\min_{\\hat{\\Theta}} \\mathcal{L}(\\mathbf{x}; \\Theta)$ , where $\\mathbf{x} \\sim Q(\\mathbf{x})$ and $\\hat{\\Theta} \\subseteq \\Theta$ . Here, $\\hat{\\Theta}$ denotes the subset of model parameters that should be updated (i.e., batch normalization layers following existing methods [26,36]). Most existing fully TTA methods focus on classification tasks, heavily relying on sufficient positive predictions for",
|
| 371 |
+
"bbox": [
|
| 372 |
+
212,
|
| 373 |
+
719,
|
| 374 |
+
785,
|
| 375 |
+
839
|
| 376 |
+
],
|
| 377 |
+
"page_idx": 4
|
| 378 |
+
},
|
| 379 |
+
{
|
| 380 |
+
"type": "header",
|
| 381 |
+
"text": "MonoTTA",
|
| 382 |
+
"bbox": [
|
| 383 |
+
658,
|
| 384 |
+
114,
|
| 385 |
+
730,
|
| 386 |
+
127
|
| 387 |
+
],
|
| 388 |
+
"page_idx": 4
|
| 389 |
+
},
|
| 390 |
+
{
|
| 391 |
+
"type": "page_number",
|
| 392 |
+
"text": "5",
|
| 393 |
+
"bbox": [
|
| 394 |
+
774,
|
| 395 |
+
116,
|
| 396 |
+
784,
|
| 397 |
+
126
|
| 398 |
+
],
|
| 399 |
+
"page_idx": 4
|
| 400 |
+
},
|
| 401 |
+
{
|
| 402 |
+
"type": "image",
|
| 403 |
+
"img_path": "images/7ba916103de92e85b6fa56879c52dc92c2b51841c794c47ecc76865a941152c8.jpg",
|
| 404 |
+
"image_caption": [
|
| 405 |
+
"Fig. 2: An illustration of our MonoTTA. During the test phase, only the pre-trained model $f_{\\Theta_0}(\\mathbf{x})$ and unlabeled test images $\\{\\mathbf{x}_i\\}_{i=1}^{N_t}$ are given. To conduct model adaptation, we initialize the model $f_{\\Theta}(\\mathbf{x})$ by $\\Theta_0$ and only update the parameters of batch normalization layers. When a batch of test images arrives, we first compute test object scores and refine the adaptive threshold $\\alpha$ to select the reliable high-score objects, thereby optimizing $\\Theta$ via the adaptive optimization loss $\\mathcal{L}_{AO}$ . Meanwhile, we devise a negative regularization term $\\mathcal{L}_{Nreg}$ to facilitate the model to avoid overfitting to noise and trivial solutions. Here, Ped. and Cyc. represent Pedestrian and Cyclist in KITTI."
|
| 406 |
+
],
|
| 407 |
+
"image_footnote": [],
|
| 408 |
+
"bbox": [
|
| 409 |
+
236,
|
| 410 |
+
143,
|
| 411 |
+
774,
|
| 412 |
+
373
|
| 413 |
+
],
|
| 414 |
+
"page_idx": 5
|
| 415 |
+
},
|
| 416 |
+
{
|
| 417 |
+
"type": "text",
|
| 418 |
+
"text": "model adaptation. Nevertheless, there is a significant difference between conventional classification tasks and Mono 3Det. As previously indicated, the detection scores of test images $\\mathbf{x}$ derived from $f_{\\Theta_0}(\\mathbf{x})$ are prone to markedly decrease in the presence of corruptions as shown in Fig. 1, leading to severe omissions (numerous false negatives) in Mono 3Det. In such circumstances, the scarcity of positive detections presents a significant challenge for model adaptation to test distributions while adapting the model with unreliable low-score detections may significantly introduce the noise. Therefore, existing fully TTA methods tend to fail in the OOD generalization problems of Mono 3Det.",
|
| 419 |
+
"bbox": [
|
| 420 |
+
212,
|
| 421 |
+
517,
|
| 422 |
+
787,
|
| 423 |
+
654
|
| 424 |
+
],
|
| 425 |
+
"page_idx": 5
|
| 426 |
+
},
|
| 427 |
+
{
|
| 428 |
+
"type": "text",
|
| 429 |
+
"text": "3.2 Overall Scheme",
|
| 430 |
+
"text_level": 1,
|
| 431 |
+
"bbox": [
|
| 432 |
+
215,
|
| 433 |
+
676,
|
| 434 |
+
390,
|
| 435 |
+
691
|
| 436 |
+
],
|
| 437 |
+
"page_idx": 5
|
| 438 |
+
},
|
| 439 |
+
{
|
| 440 |
+
"type": "text",
|
| 441 |
+
"text": "After thoroughly examining the characteristics and challenges of Mono 3Det, we introduce a Monocular Test-Time Adaptation (MonoTTA) method to address the OOD problems for Mono 3Det models, which seeks to solve the object score declining issue within unlabeled OOD test data. As shown in Fig. 2, MonoTTA consists of two strategies: 1) Reliability-driven adaptation and 2) Noise-guard adaptation. We first briefly introduce the two strategies below.",
|
| 442 |
+
"bbox": [
|
| 443 |
+
212,
|
| 444 |
+
703,
|
| 445 |
+
787,
|
| 446 |
+
794
|
| 447 |
+
],
|
| 448 |
+
"page_idx": 5
|
| 449 |
+
},
|
| 450 |
+
{
|
| 451 |
+
"type": "text",
|
| 452 |
+
"text": "First, we develop a reliability-driven adaptation strategy (c.f. Section 3.3) to conduct reliable model adaptation for OOD test data based on dependable test objects. Our empirical investigations inspire us to exploit those relatively reliable",
|
| 453 |
+
"bbox": [
|
| 454 |
+
212,
|
| 455 |
+
795,
|
| 456 |
+
787,
|
| 457 |
+
840
|
| 458 |
+
],
|
| 459 |
+
"page_idx": 5
|
| 460 |
+
},
|
| 461 |
+
{
|
| 462 |
+
"type": "page_number",
|
| 463 |
+
"text": "6",
|
| 464 |
+
"bbox": [
|
| 465 |
+
217,
|
| 466 |
+
114,
|
| 467 |
+
228,
|
| 468 |
+
126
|
| 469 |
+
],
|
| 470 |
+
"page_idx": 5
|
| 471 |
+
},
|
| 472 |
+
{
|
| 473 |
+
"type": "header",
|
| 474 |
+
"text": "Hongbin Lin, Yifan Zhang and et al.",
|
| 475 |
+
"bbox": [
|
| 476 |
+
271,
|
| 477 |
+
114,
|
| 478 |
+
514,
|
| 479 |
+
128
|
| 480 |
+
],
|
| 481 |
+
"page_idx": 5
|
| 482 |
+
},
|
| 483 |
+
{
|
| 484 |
+
"type": "image",
|
| 485 |
+
"img_path": "images/3acbb989c5da1a458cc9645f9fc38d02e6b56bbe28819edb4829bbe02c59e89f.jpg",
|
| 486 |
+
"image_caption": [
|
| 487 |
+
"(a)"
|
| 488 |
+
],
|
| 489 |
+
"image_footnote": [],
|
| 490 |
+
"bbox": [
|
| 491 |
+
256,
|
| 492 |
+
143,
|
| 493 |
+
478,
|
| 494 |
+
277
|
| 495 |
+
],
|
| 496 |
+
"page_idx": 6
|
| 497 |
+
},
|
| 498 |
+
{
|
| 499 |
+
"type": "image",
|
| 500 |
+
"img_path": "images/b71877ed35d254d833d907c9acdf1fabec7900d73d66895bba8a8d1b23eeaa23.jpg",
|
| 501 |
+
"image_caption": [
|
| 502 |
+
"(b)",
|
| 503 |
+
"Fig. 3: Based on MonoGround [29], we conduct two empirical studies (Car, KITTI), with the 3D IoU threshold of 0.5. (a) We visualize the accuracy of the objects across varied scoring ranges, which shows that the accuracy of objects with high scores remains relatively stable even in the presence of diverse corruptions (Ideal means in-distribution scenarios). (b) We visualize the number of low & high-score objects before and after optimization. Although only high-score objects are optimized, the model treats low-score objects with more confidence."
|
| 504 |
+
],
|
| 505 |
+
"image_footnote": [],
|
| 506 |
+
"bbox": [
|
| 507 |
+
486,
|
| 508 |
+
148,
|
| 509 |
+
738,
|
| 510 |
+
277
|
| 511 |
+
],
|
| 512 |
+
"page_idx": 6
|
| 513 |
+
},
|
| 514 |
+
{
|
| 515 |
+
"type": "text",
|
| 516 |
+
"text": "test objects for alleviating distribution shifts, thereby discovering more potential objects. To this end, MonoTTA excludes unreliable test objects involving an adaptive threshold $\\alpha$ for any unlabeled test data. Subsequently, the model is optimized by the adaptive optimization loss $\\mathcal{L}_{AO}$ via the selected reliable objects.",
|
| 517 |
+
"bbox": [
|
| 518 |
+
212,
|
| 519 |
+
411,
|
| 520 |
+
784,
|
| 521 |
+
470
|
| 522 |
+
],
|
| 523 |
+
"page_idx": 6
|
| 524 |
+
},
|
| 525 |
+
{
|
| 526 |
+
"type": "text",
|
| 527 |
+
"text": "Second, we tend to utilize plenty of low-score objects to adapt the model in an indirect manner instead of directly optimizing the model since low-score objects are noisy. Hence, we devise a noise-guard adaptation strategy (c.f. Section 3.4) to prevent the model from overfitting to noisy predictions and falling into trivial solutions. Specifically, we randomly choose one of the negative classes of low-score objects and minimize the scores (e.g., score 0.03 of [0.10, 0.12, 0.03] c.f. Fig. 2) after simply filtering out extremely low-score detections. Even though the positive class is noisy (i.e., score 0.12), this term is capable of optimizing the model indirectly, i.e., learn to deny the negative category of the object.",
|
| 528 |
+
"bbox": [
|
| 529 |
+
212,
|
| 530 |
+
472,
|
| 531 |
+
787,
|
| 532 |
+
607
|
| 533 |
+
],
|
| 534 |
+
"page_idx": 6
|
| 535 |
+
},
|
| 536 |
+
{
|
| 537 |
+
"type": "text",
|
| 538 |
+
"text": "Overall, the training scheme of MonoTTA is as follows:",
|
| 539 |
+
"bbox": [
|
| 540 |
+
238,
|
| 541 |
+
608,
|
| 542 |
+
640,
|
| 543 |
+
623
|
| 544 |
+
],
|
| 545 |
+
"page_idx": 6
|
| 546 |
+
},
|
| 547 |
+
{
|
| 548 |
+
"type": "equation",
|
| 549 |
+
"text": "\n$$\n\\min _ {\\hat {\\Theta}} \\mathcal {L} _ {A O} (\\hat {\\Theta}) + \\lambda \\mathcal {L} _ {N r e g} (\\hat {\\Theta}), \\tag {1}\n$$\n",
|
| 550 |
+
"text_format": "latex",
|
| 551 |
+
"bbox": [
|
| 552 |
+
401,
|
| 553 |
+
633,
|
| 554 |
+
785,
|
| 555 |
+
659
|
| 556 |
+
],
|
| 557 |
+
"page_idx": 6
|
| 558 |
+
},
|
| 559 |
+
{
|
| 560 |
+
"type": "text",
|
| 561 |
+
"text": "where $\\lambda$ is the balance hyper-parameter. The pseudo-code of MonoTTA is summarized in Algorithm 1.",
|
| 562 |
+
"bbox": [
|
| 563 |
+
212,
|
| 564 |
+
670,
|
| 565 |
+
787,
|
| 566 |
+
702
|
| 567 |
+
],
|
| 568 |
+
"page_idx": 6
|
| 569 |
+
},
|
| 570 |
+
{
|
| 571 |
+
"type": "text",
|
| 572 |
+
"text": "3.3 Reliability-Driven Adaptation",
|
| 573 |
+
"text_level": 1,
|
| 574 |
+
"bbox": [
|
| 575 |
+
214,
|
| 576 |
+
723,
|
| 577 |
+
509,
|
| 578 |
+
739
|
| 579 |
+
],
|
| 580 |
+
"page_idx": 6
|
| 581 |
+
},
|
| 582 |
+
{
|
| 583 |
+
"type": "text",
|
| 584 |
+
"text": "To identify dependable test objects and conduct test-time model adaptation, we propose a reliability-driven adaptation strategy that consists of two components: 1) Reliable object identification and 2) Adaptive model optimization.",
|
| 585 |
+
"bbox": [
|
| 586 |
+
212,
|
| 587 |
+
750,
|
| 588 |
+
784,
|
| 589 |
+
794
|
| 590 |
+
],
|
| 591 |
+
"page_idx": 6
|
| 592 |
+
},
|
| 593 |
+
{
|
| 594 |
+
"type": "text",
|
| 595 |
+
"text": "Reliable object identification. When unlabeled test images arrive, it is difficult for the pre-trained model to get accurate detections on OOD test data due to the decline in detection scores. To resolve this, we dig into the detection",
|
| 596 |
+
"bbox": [
|
| 597 |
+
212,
|
| 598 |
+
794,
|
| 599 |
+
785,
|
| 600 |
+
839
|
| 601 |
+
],
|
| 602 |
+
"page_idx": 6
|
| 603 |
+
},
|
| 604 |
+
{
|
| 605 |
+
"type": "header",
|
| 606 |
+
"text": "MonoTTA",
|
| 607 |
+
"bbox": [
|
| 608 |
+
658,
|
| 609 |
+
114,
|
| 610 |
+
730,
|
| 611 |
+
126
|
| 612 |
+
],
|
| 613 |
+
"page_idx": 6
|
| 614 |
+
},
|
| 615 |
+
{
|
| 616 |
+
"type": "page_number",
|
| 617 |
+
"text": "7",
|
| 618 |
+
"bbox": [
|
| 619 |
+
774,
|
| 620 |
+
114,
|
| 621 |
+
784,
|
| 622 |
+
125
|
| 623 |
+
],
|
| 624 |
+
"page_idx": 6
|
| 625 |
+
},
|
| 626 |
+
{
|
| 627 |
+
"type": "text",
|
| 628 |
+
"text": "accuracy of the pre-trained model in test scenarios with various corruptions. Specifically, based on MonoGround [29], we visualize the detection precision for the Car category within the KITTI dataset across varied scoring ranges, with the 3D Intersection over Union (IoU) threshold of 0.5. As shown in Fig. 3 (a), we find that high-score objects are more reliable and relatively stable even in the presence of diverse corruptions. Following this, we propose to select reliable high-score objects to conduct model adaptation via an adaptive threshold $\\alpha$ . Specifically, we exploit the model $f_{\\Theta}(\\cdot)$ (initialized by $\\Theta_0$ ) to infer a batch of test images $\\{\\mathbf{x}_b\\}_{b=1}^B$ and obtain the heatmap $\\mathbf{h}_i$ of each image $\\mathbf{x}_i$ by $\\mathbf{h}_i = f_{\\Theta}(\\mathbf{x}_i)$ , where $i$ ranges from 0 to $B-1$ and $B$ denotes the batch size. Then, we could achieve object score maps $\\mathbf{s}_i \\in \\mathbb{R}^{B \\times N_m}$ like the peaks in $\\mathbf{h}_i$ [7] after normalization, where $K$ and $N_m$ denotes the number of classes and the maximum of detected objects, respectively. With the score map $\\mathbf{s}_i$ , we update the adaptive threshold $\\alpha_t$ at the iteration $t$ in the exponential moving average manner by:",
|
| 629 |
+
"bbox": [
|
| 630 |
+
212,
|
| 631 |
+
146,
|
| 632 |
+
787,
|
| 633 |
+
359
|
| 634 |
+
],
|
| 635 |
+
"page_idx": 7
|
| 636 |
+
},
|
| 637 |
+
{
|
| 638 |
+
"type": "equation",
|
| 639 |
+
"text": "\n$$\n\\alpha_ {t} = \\left\\{ \\begin{array}{l l} \\gamma , & \\text {i f} t = 1 \\\\ \\beta \\bar {m} _ {t} + (1 - \\beta) \\alpha_ {t - 1}, & \\text {i f} t > 1 \\end{array} \\right., \\tag {2}\n$$\n",
|
| 640 |
+
"text_format": "latex",
|
| 641 |
+
"bbox": [
|
| 642 |
+
367,
|
| 643 |
+
376,
|
| 644 |
+
785,
|
| 645 |
+
411
|
| 646 |
+
],
|
| 647 |
+
"page_idx": 7
|
| 648 |
+
},
|
| 649 |
+
{
|
| 650 |
+
"type": "equation",
|
| 651 |
+
"text": "\n$$\n\\bar {m} _ {t} = \\frac {1}{B} \\sum_ {i = 1} ^ {B} \\frac {\\sum_ {j = 1} ^ {N _ {m}} s _ {i j} \\cdot \\mathbb {I} \\left(s _ {i j} \\geq \\gamma\\right)}{\\sum_ {j = 1} ^ {N _ {m}} \\mathbb {I} \\left(s _ {i j} \\geq \\gamma\\right)}. \\tag {3}\n$$\n",
|
| 652 |
+
"text_format": "latex",
|
| 653 |
+
"bbox": [
|
| 654 |
+
377,
|
| 655 |
+
435,
|
| 656 |
+
785,
|
| 657 |
+
479
|
| 658 |
+
],
|
| 659 |
+
"page_idx": 7
|
| 660 |
+
},
|
| 661 |
+
{
|
| 662 |
+
"type": "text",
|
| 663 |
+
"text": "In Eqn. (2), $\\bar{m}_t$ denotes the average score of all detected objects in a single batch of $B$ test samples at iteration $t$ , while $\\beta \\in [0,1]$ is a decay coefficient. As for Eqn. (3), $s_{ij} \\in (0,1)$ denotes the score of the $j$ -th object in the $i$ -th test image while $\\mathbb{I}(\\cdot)$ is the indicator function. Note that $\\gamma \\in \\mathbb{R}^1$ is a pre-defined object detection threshold adopted from existing methods at their original inference stage as well as $N_m$ .",
|
| 664 |
+
"bbox": [
|
| 665 |
+
212,
|
| 666 |
+
491,
|
| 667 |
+
787,
|
| 668 |
+
582
|
| 669 |
+
],
|
| 670 |
+
"page_idx": 7
|
| 671 |
+
},
|
| 672 |
+
{
|
| 673 |
+
"type": "text",
|
| 674 |
+
"text": "Adaptive model optimization. As shown in Fig. 3 (b), the optimization of high-score objects can also enhance the confidence of the model for relatively low-score objects. It motivates us that exploiting high-score objects rather than all objects for model adaptation would be a more reliable way to learn from OOD test data. Therefore, with the adaptive threshold $\\alpha_{t}$ , we select the reliable subset of high scores from $\\mathbf{s}_i$ and calculate the adaptive optimization loss $\\mathcal{L}_{AO}$ to adapt the model by:",
|
| 675 |
+
"bbox": [
|
| 676 |
+
212,
|
| 677 |
+
584,
|
| 678 |
+
787,
|
| 679 |
+
690
|
| 680 |
+
],
|
| 681 |
+
"page_idx": 7
|
| 682 |
+
},
|
| 683 |
+
{
|
| 684 |
+
"type": "equation",
|
| 685 |
+
"text": "\n$$\n\\mathcal {L} _ {A O} = - \\frac {1}{B} \\sum_ {i = 1} ^ {B} \\sum_ {j = 1} ^ {N _ {m}} \\log \\left(s _ {i j} \\cdot \\mathbb {I} \\left(s _ {i j} \\geq \\alpha_ {t}\\right)\\right), \\tag {4}\n$$\n",
|
| 686 |
+
"text_format": "latex",
|
| 687 |
+
"bbox": [
|
| 688 |
+
356,
|
| 689 |
+
705,
|
| 690 |
+
785,
|
| 691 |
+
750
|
| 692 |
+
],
|
| 693 |
+
"page_idx": 7
|
| 694 |
+
},
|
| 695 |
+
{
|
| 696 |
+
"type": "text",
|
| 697 |
+
"text": "The adaptive optimization loss $\\mathcal{L}_{AO}$ alleviates the potential data distribution shifts in OOD test data by allowing the model to confidently identify high-score test objects. As we mentioned, it solves the score decline issue of OOD test data in a more reliable way rather than directly optimizing all test objects, thereby avoiding overfitting to noise and discovering more potential test objects.",
|
| 698 |
+
"bbox": [
|
| 699 |
+
212,
|
| 700 |
+
763,
|
| 701 |
+
787,
|
| 702 |
+
840
|
| 703 |
+
],
|
| 704 |
+
"page_idx": 7
|
| 705 |
+
},
|
| 706 |
+
{
|
| 707 |
+
"type": "page_number",
|
| 708 |
+
"text": "8",
|
| 709 |
+
"bbox": [
|
| 710 |
+
217,
|
| 711 |
+
114,
|
| 712 |
+
228,
|
| 713 |
+
126
|
| 714 |
+
],
|
| 715 |
+
"page_idx": 7
|
| 716 |
+
},
|
| 717 |
+
{
|
| 718 |
+
"type": "header",
|
| 719 |
+
"text": "Hongbin Lin, Yifan Zhang and et al.",
|
| 720 |
+
"bbox": [
|
| 721 |
+
271,
|
| 722 |
+
114,
|
| 723 |
+
514,
|
| 724 |
+
128
|
| 725 |
+
],
|
| 726 |
+
"page_idx": 7
|
| 727 |
+
},
|
| 728 |
+
{
|
| 729 |
+
"type": "text",
|
| 730 |
+
"text": "Algorithm 1 The pipeline of the proposed MonoTTA",
|
| 731 |
+
"text_level": 1,
|
| 732 |
+
"bbox": [
|
| 733 |
+
217,
|
| 734 |
+
146,
|
| 735 |
+
609,
|
| 736 |
+
161
|
| 737 |
+
],
|
| 738 |
+
"page_idx": 8
|
| 739 |
+
},
|
| 740 |
+
{
|
| 741 |
+
"type": "text",
|
| 742 |
+
"text": "Require: Unlabeled test data $\\mathcal{D}_t = \\{\\mathbf{x}_i\\}_{i=1}^{N_t}$ ; Pre-trained model $f_{\\Theta_0}(\\mathbf{x})$ ; Batch size $B$ ; Parameters $\\lambda, \\beta, \\eta$ .",
|
| 743 |
+
"bbox": [
|
| 744 |
+
215,
|
| 745 |
+
166,
|
| 746 |
+
785,
|
| 747 |
+
195
|
| 748 |
+
],
|
| 749 |
+
"page_idx": 8
|
| 750 |
+
},
|
| 751 |
+
{
|
| 752 |
+
"type": "list",
|
| 753 |
+
"sub_type": "text",
|
| 754 |
+
"list_items": [
|
| 755 |
+
"1: for a batch images $\\{\\mathbf{x}_b\\}_{b = 1}^B$ in $\\mathcal{D}_t$ do",
|
| 756 |
+
"2: Update the adaptive threshold $\\alpha$ based on Eqn. (2);",
|
| 757 |
+
"3: Calculate adaptive optimization loss $\\mathcal{L}_{AO}$ based on Eqn. (4);",
|
| 758 |
+
"4: Calculate negative regularization term $\\mathcal{L}_{Nreg}$ based on Eqn. (6);",
|
| 759 |
+
"5: Update $\\hat{\\Theta}$ by optimizing Eqn. (1)",
|
| 760 |
+
"6: end for",
|
| 761 |
+
"7: return Detection Results for all $\\mathbf{x} \\in \\mathcal{D}_t$ ."
|
| 762 |
+
],
|
| 763 |
+
"bbox": [
|
| 764 |
+
222,
|
| 765 |
+
195,
|
| 766 |
+
686,
|
| 767 |
+
292
|
| 768 |
+
],
|
| 769 |
+
"page_idx": 8
|
| 770 |
+
},
|
| 771 |
+
{
|
| 772 |
+
"type": "text",
|
| 773 |
+
"text": "3.4 Noise-Guard Adaptation",
|
| 774 |
+
"text_level": 1,
|
| 775 |
+
"bbox": [
|
| 776 |
+
215,
|
| 777 |
+
325,
|
| 778 |
+
465,
|
| 779 |
+
339
|
| 780 |
+
],
|
| 781 |
+
"page_idx": 8
|
| 782 |
+
},
|
| 783 |
+
{
|
| 784 |
+
"type": "text",
|
| 785 |
+
"text": "Through the optimization of adaptive optimization loss $\\mathcal{L}_{AO}$ , the model is refined to yield more confident detection outcomes. Nevertheless, high-scoring objects may be scarce due to the distribution shifts, making the adaptation procedure difficult. Meanwhile, exclusive reliance on $\\mathcal{L}_{AO}$ for adaptation may result in trivial solutions, whereby the model indiscriminately assigns high scores to all categories. Previous studies [13, 14] indicate that deep neural networks could learn from noisy pseudo labels in classification tasks through negative learning. Thus, MonoTTA proposes to learn from noisy low-score objects in a negative learning manner for Mono 3Det. Specifically, we denote $\\hat{\\mathbf{s}}_i\\in \\mathbb{R}^{B\\times N_m\\times K}$ as the multi-class score map for the test image $\\mathbf{x}_i$ , i.e. the multi-class score map $\\hat{\\mathbf{s}}_i$ contains not only the highest score of objects but also the scores for other classes. Here $\\mathbf{s}_i = \\arg \\max_k\\hat{\\mathbf{s}}_{ik}$ where $k$ is the class index. Employing a simple constant threshold $\\eta$ , we filter out extremely low-score objects and then randomly select a negative class $\\bar{k}$ for each object. Subsequently, we compute the regularization loss $e_k$ for each class $k$ with relatively low object scores $s_{ij}\\in [\\eta ,\\alpha_t)$ by:",
|
| 786 |
+
"bbox": [
|
| 787 |
+
212,
|
| 788 |
+
349,
|
| 789 |
+
787,
|
| 790 |
+
578
|
| 791 |
+
],
|
| 792 |
+
"page_idx": 8
|
| 793 |
+
},
|
| 794 |
+
{
|
| 795 |
+
"type": "equation",
|
| 796 |
+
"text": "\n$$\ne _ {k} = - \\sum_ {i = 1} ^ {B} \\sum_ {j = 1} ^ {N _ {m}} \\bar {y} _ {i j} \\log \\left(1 - s _ {i j \\bar {k}} \\cdot \\mathbb {I} (\\bar {k} = k)\\right), \\tag {5}\n$$\n",
|
| 797 |
+
"text_format": "latex",
|
| 798 |
+
"bbox": [
|
| 799 |
+
349,
|
| 800 |
+
587,
|
| 801 |
+
785,
|
| 802 |
+
631
|
| 803 |
+
],
|
| 804 |
+
"page_idx": 8
|
| 805 |
+
},
|
| 806 |
+
{
|
| 807 |
+
"type": "text",
|
| 808 |
+
"text": "where $\\bar{y}_{ij} = 1 - s_{ij\\bar{k}}$ is a constant weight. We further define $n_k$ as the frequency of negative scores corresponding to class $k$ in the test batch and balance $e_k$ by:",
|
| 809 |
+
"bbox": [
|
| 810 |
+
212,
|
| 811 |
+
641,
|
| 812 |
+
785,
|
| 813 |
+
672
|
| 814 |
+
],
|
| 815 |
+
"page_idx": 8
|
| 816 |
+
},
|
| 817 |
+
{
|
| 818 |
+
"type": "equation",
|
| 819 |
+
"text": "\n$$\n\\mathcal {L} _ {N r e g} = \\sum_ {k = 1} ^ {K} \\frac {e _ {k}}{n _ {k}}. \\tag {6}\n$$\n",
|
| 820 |
+
"text_format": "latex",
|
| 821 |
+
"bbox": [
|
| 822 |
+
439,
|
| 823 |
+
681,
|
| 824 |
+
785,
|
| 825 |
+
724
|
| 826 |
+
],
|
| 827 |
+
"page_idx": 8
|
| 828 |
+
},
|
| 829 |
+
{
|
| 830 |
+
"type": "text",
|
| 831 |
+
"text": "As we mentioned before, detections with low and intermediate scores tend to be more noisy (c.f. Fig. 3 (a)), which is unreliable for direct model adaptation. To this end, we introduce the regularization term $\\mathcal{L}_{Nreg}$ to leverage noisy low-score detections and improve the model for assimilating potentially accurate information (c.f. Term 2 in Fig. 2). Moreover, this term also prevents the model from trivial solutions, i.e., indiscriminately assigning high scores to all classes of a single object (e.g., TENT [36]). In addition, high-score objects may be absent",
|
| 832 |
+
"bbox": [
|
| 833 |
+
212,
|
| 834 |
+
734,
|
| 835 |
+
787,
|
| 836 |
+
842
|
| 837 |
+
],
|
| 838 |
+
"page_idx": 8
|
| 839 |
+
},
|
| 840 |
+
{
|
| 841 |
+
"type": "header",
|
| 842 |
+
"text": "MonoTTA",
|
| 843 |
+
"bbox": [
|
| 844 |
+
658,
|
| 845 |
+
114,
|
| 846 |
+
730,
|
| 847 |
+
126
|
| 848 |
+
],
|
| 849 |
+
"page_idx": 8
|
| 850 |
+
},
|
| 851 |
+
{
|
| 852 |
+
"type": "page_number",
|
| 853 |
+
"text": "9",
|
| 854 |
+
"bbox": [
|
| 855 |
+
774,
|
| 856 |
+
116,
|
| 857 |
+
785,
|
| 858 |
+
126
|
| 859 |
+
],
|
| 860 |
+
"page_idx": 8
|
| 861 |
+
},
|
| 862 |
+
{
|
| 863 |
+
"type": "table",
|
| 864 |
+
"img_path": "images/63d80a18755b299d0077d6e63c29c38fd292bb69547e45ed5e25d1416a9e0bdb.jpg",
|
| 865 |
+
"table_caption": [
|
| 866 |
+
"Table 1: Comparison with baselines on the KITTI-C validation set, severity level 1 regarding Mean $AP_{3D|R_{40}}$ . The bold number indicates the best result."
|
| 867 |
+
],
|
| 868 |
+
"table_footnote": [],
|
| 869 |
+
"table_body": "<table><tr><td colspan=\"15\">Car, IoU @ 0.7, 0.5, 0.5</td></tr><tr><td rowspan=\"2\">Method</td><td colspan=\"3\">Noise</td><td colspan=\"3\">Blur</td><td colspan=\"4\">Weather</td><td colspan=\"3\">Digital</td><td rowspan=\"2\">Avg.</td></tr><tr><td>Gauss.</td><td>Shot</td><td>Impul.</td><td>Defoc.</td><td>Glass</td><td>Motion</td><td>Snow</td><td>Frost</td><td>Fog</td><td>Brit.</td><td>Contr.</td><td>Pixel</td><td>Sat.</td></tr><tr><td>Monoflex [48]</td><td>3.84</td><td>7.48</td><td>5.31</td><td>2.59</td><td>3.73</td><td>11.05</td><td>0.23</td><td>7.77</td><td>7.57</td><td>24.87</td><td>6.92</td><td>28.16</td><td>31.46</td><td>10.84</td></tr><tr><td>• BN adaptation [33]</td><td>13.58</td><td>21.93</td><td>18.78</td><td>15.87</td><td>8.59</td><td>24.32</td><td>5.42</td><td>21.45</td><td>24.63</td><td>31.80</td><td>30.58</td><td>41.04</td><td>30.71</td><td>22.21</td></tr><tr><td>• TENT [36]</td><td>17.80</td><td>27.09</td><td>23.18</td><td>21.66</td><td>11.90</td><td>28.75</td><td>6.84</td><td>26.58</td><td>30.78</td><td>35.65</td><td>34.72</td><td>41.71</td><td>35.91</td><td>26.35</td></tr><tr><td>• EATA [26]</td><td>16.67</td><td>26.42</td><td>25.07</td><td>22.54</td><td>13.23</td><td>27.73</td><td>7.87</td><td>26.58</td><td>31.10</td><td>35.39</td><td>35.28</td><td>41.40</td><td>36.72</td><td>26.62</td></tr><tr><td>• ActMAD [24]</td><td>12.26</td><td>21.49</td><td>20.47</td><td>11.10</td><td>7.53</td><td>23.79</td><td>5.57</td><td>19.11</td><td>26.54</td><td>26.24</td><td>23.92</td><td>38.51</td><td>33.75</td><td>20.79</td></tr><tr><td>• MonoTTA (Ours)</td><td>21.15</td><td>28.65</td><td>26.64</td><td>25.91</td><td>19.26</td><td>31.48</td><td>12.43</td><td>30.24</td><td>33.75</td><td>36.84</td><td>36.83</td><td>41.97</td><td>38.13</td><td>29.48</td></tr><tr><td>MonoGround [29]</td><td>2.40</td><td>4.10</td><td>3.31</td><td>3.71</td><td>2.67</td><td>8.13</td><td>0.22</td><td>5.54</td><td>4.59</td><td>25.37</td><td>4.00</td><td>33.57</td><td>28.08</td><td>9.67</td></tr><tr><td>• BN adaptation [33]</td><td>13.49</td><td>23.52</td><td>19.69</td><td>16.33</td><td>7.61</td><td>23.99</td><td>7.98</td><td>20.71</td><td>24.00</td><td>31.34</td><td>29.03</td><td>43.06</td><td>32.99</td><td>22.60</td></tr><tr><td>• TENT [36]</td><td>17.94</td><td>29.60</td><td>19.90</td><td>23.45</td><td>13.90</td><td>29.39</td><td>10.32</td><td>26.65</td><td>33.35</td><td>35.96</td><td>36.39</td><td>43.35</td><td>37.79</td><td>27.54</td></tr><tr><td>• EATA [26]</td><td>16.03</td><td>26.08</td><td>18.08</td><td>20.28</td><td>12.40</td><td>27.37</td><td>9.22</td><td>23.79</td><td>29.49</td><td>33.65</td><td>32.58</td><td>43.61</td><td>36.00</td><td>25.28</td></tr><tr><td>• ActMAD [24]</td><td>13.65</td><td>22.39</td><td>20.56</td><td>14.99</td><td>8.17</td><td>21.09</td><td>9.13</td><td>18.77</td><td>19.52</td><td>28.34</td><td>25.32</td><td>42.63</td><td>31.58</td><td>21.24</td></tr><tr><td>• MonoTTA (Ours)</td><td>26.13</td><td>33.11</td><td>28.60</td><td>30.38</td><td>25.48</td><td>32.44</td><td>18.72</td><td>32.60</td><td>37.75</td><td>37.87</td><td>39.57</td><td>43.67</td><td>37.98</td><td>32.64</td></tr></table>",
|
| 870 |
+
"bbox": [
|
| 871 |
+
241,
|
| 872 |
+
179,
|
| 873 |
+
761,
|
| 874 |
+
287
|
| 875 |
+
],
|
| 876 |
+
"page_idx": 9
|
| 877 |
+
},
|
| 878 |
+
{
|
| 879 |
+
"type": "table",
|
| 880 |
+
"img_path": "images/4e086364b9fcd8756e6d9347724bdf5e142b88d84507bf9576eb6564ab9951ad.jpg",
|
| 881 |
+
"table_caption": [],
|
| 882 |
+
"table_footnote": [],
|
| 883 |
+
"table_body": "<table><tr><td colspan=\"14\">Pedestrian, IoU @ 0.5, 0.25</td><td></td></tr><tr><td rowspan=\"2\">Method</td><td colspan=\"3\">Noise</td><td colspan=\"3\">Blur</td><td colspan=\"4\">Weather</td><td colspan=\"3\">Digital</td><td rowspan=\"2\">Avg.</td></tr><tr><td>Gauss.</td><td>Shot</td><td>Impul.</td><td>Defoc.</td><td>Glass</td><td>Motion</td><td>Snow</td><td>Frost</td><td>Fog</td><td>Brit.</td><td>Contr.</td><td>Pixel</td><td>Sat.</td></tr><tr><td>Monoflex</td><td>0.19</td><td>1.62</td><td>0.32</td><td>3.72</td><td>8.47</td><td>6.22</td><td>0.00</td><td>4.27</td><td>2.25</td><td>9.19</td><td>2.08</td><td>1.83</td><td>9.11</td><td>3.79</td></tr><tr><td>• BN adaptation [33]</td><td>6.21</td><td>8.20</td><td>9.20</td><td>7.83</td><td>5.35</td><td>7.52</td><td>2.89</td><td>6.47</td><td>9.24</td><td>9.12</td><td>9.93</td><td>12.73</td><td>9.76</td><td>8.03</td></tr><tr><td>• TENT [36]</td><td>6.02</td><td>7.96</td><td>9.57</td><td>7.75</td><td>6.06</td><td>8.63</td><td>2.63</td><td>6.71</td><td>9.91</td><td>10.26</td><td>10.55</td><td>12.33</td><td>10.27</td><td>8.36</td></tr><tr><td>• EATA [26]</td><td>6.05</td><td>7.96</td><td>9.74</td><td>7.93</td><td>6.06</td><td>8.01</td><td>2.48</td><td>6.24</td><td>9.94</td><td>9.70</td><td>10.02</td><td>12.41</td><td>10.12</td><td>8.21</td></tr><tr><td>• ActMAD [24]</td><td>2.41</td><td>4.27</td><td>4.68</td><td>1.46</td><td>1.62</td><td>6.48</td><td>1.17</td><td>4.25</td><td>4.84</td><td>6.42</td><td>6.19</td><td>8.29</td><td>7.07</td><td>4.55</td></tr><tr><td>• MonotTA (Ours)</td><td>6.54</td><td>8.41</td><td>9.39</td><td>7.63</td><td>7.12</td><td>8.99</td><td>3.25</td><td>7.64</td><td>10.26</td><td>10.55</td><td>10.06</td><td>13.28</td><td>10.66</td><td>8.75</td></tr><tr><td>MonoGround</td><td>0.61</td><td>0.93</td><td>0.73</td><td>8.41</td><td>7.57</td><td>8.03</td><td>0.00</td><td>3.19</td><td>1.39</td><td>13.82</td><td>1.83</td><td>3.70</td><td>7.26</td><td>4.42</td></tr><tr><td>• BN adaptation [33]</td><td>5.09</td><td>7.08</td><td>7.77</td><td>7.63</td><td>6.63</td><td>9.45</td><td>2.24</td><td>6.72</td><td>8.40</td><td>9.72</td><td>11.06</td><td>16.70</td><td>12.43</td><td>8.53</td></tr><tr><td>• TENT [36]</td><td>7.27</td><td>10.10</td><td>10.02</td><td>8.53</td><td>8.30</td><td>11.03</td><td>3.54</td><td>8.73</td><td>9.11</td><td>11.74</td><td>12.12</td><td>17.70</td><td>15.03</td><td>10.25</td></tr><tr><td>• EATA [26]</td><td>5.92</td><td>8.05</td><td>8.58</td><td>8.12</td><td>7.59</td><td>10.95</td><td>3.31</td><td>7.91</td><td>10.10</td><td>11.02</td><td>10.95</td><td>17.30</td><td>14.34</td><td>9.55</td></tr><tr><td>• ActMAD [24]</td><td>3.96</td><td>5.75</td><td>6.48</td><td>5.61</td><td>5.73</td><td>7.14</td><td>1.58</td><td>5.37</td><td>5.74</td><td>7.27</td><td>8.10</td><td>13.12</td><td>9.27</td><td>6.55</td></tr><tr><td>• MonotTA (Ours)</td><td>8.58</td><td>11.18</td><td>11.79</td><td>9.22</td><td>9.40</td><td>13.20</td><td>4.83</td><td>9.95</td><td>14.46</td><td>12.85</td><td>13.25</td><td>17.13</td><td>14.85</td><td>11.59</td></tr></table>",
|
| 884 |
+
"bbox": [
|
| 885 |
+
241,
|
| 886 |
+
287,
|
| 887 |
+
761,
|
| 888 |
+
393
|
| 889 |
+
],
|
| 890 |
+
"page_idx": 9
|
| 891 |
+
},
|
| 892 |
+
{
|
| 893 |
+
"type": "table",
|
| 894 |
+
"img_path": "images/44412cd6db4e954e0783727b5a869eb8bf6953821f7a71c2f6828db9bd1d6114.jpg",
|
| 895 |
+
"table_caption": [],
|
| 896 |
+
"table_footnote": [],
|
| 897 |
+
"table_body": "<table><tr><td colspan=\"14\">Cyclist, IoU @ 0.5, 0.25, 0.25</td><td></td></tr><tr><td rowspan=\"2\">Method</td><td colspan=\"3\">Noise</td><td colspan=\"3\">Blur</td><td colspan=\"4\">Weather</td><td colspan=\"3\">Digital</td><td rowspan=\"2\">Avg.</td></tr><tr><td>Gauss.</td><td>Shot</td><td>Impul.</td><td>Defoc.</td><td>Glass</td><td>Motion</td><td>Snow</td><td>Frost</td><td>Fog</td><td>Brit.</td><td>Contr.</td><td>Pixel</td><td>Sat.</td></tr><tr><td>Monoflex</td><td>0.28</td><td>1.64</td><td>0.47</td><td>0.59</td><td>4.97</td><td>3.60</td><td>0.00</td><td>7.42</td><td>3.81</td><td>13.07</td><td>3.79</td><td>3.80</td><td>8.39</td><td>3.99</td></tr><tr><td>• BN adaptation [33]</td><td>2.39</td><td>6.26</td><td>4.36</td><td>5.78</td><td>6.76</td><td>9.09</td><td>1.70</td><td>8.53</td><td>9.16</td><td>12.91</td><td>11.26</td><td>10.55</td><td>11.02</td><td>7.67</td></tr><tr><td>• TENT [36]</td><td>2.72</td><td>7.94</td><td>5.63</td><td>6.27</td><td>7.20</td><td>9.49</td><td>1.07</td><td>8.94</td><td>10.96</td><td>12.75</td><td>12.72</td><td>9.64</td><td>11.28</td><td>8.20</td></tr><tr><td>• EATA [26]</td><td>2.33</td><td>7.46</td><td>5.46</td><td>7.19</td><td>7.23</td><td>7.51</td><td>1.24</td><td>8.60</td><td>10.18</td><td>12.86</td><td>11.02</td><td>10.89</td><td>10.30</td><td>7.87</td></tr><tr><td>• ActMAD [24]</td><td>2.22</td><td>5.32</td><td>4.36</td><td>2.29</td><td>1.57</td><td>5.64</td><td>0.48</td><td>5.68</td><td>6.46</td><td>10.56</td><td>9.31</td><td>9.97</td><td>7.80</td><td>5.51</td></tr><tr><td>• MonotTA (Ours)</td><td>3.01</td><td>7.24</td><td>5.98</td><td>7.00</td><td>6.09</td><td>9.51</td><td>1.45</td><td>10.63</td><td>11.22</td><td>12.85</td><td>11.85</td><td>11.59</td><td>10.81</td><td>8.40</td></tr><tr><td>MonoGround</td><td>0.12</td><td>0.46</td><td>0.48</td><td>0.33</td><td>0.72</td><td>2.03</td><td>0.00</td><td>0.56</td><td>0.35</td><td>5.55</td><td>0.52</td><td>2.08</td><td>5.24</td><td>1.42</td></tr><tr><td>• BN adaptation [33]</td><td>1.76</td><td>3.58</td><td>2.08</td><td>3.61</td><td>3.05</td><td>5.41</td><td>0.57</td><td>3.82</td><td>4.47</td><td>6.30</td><td>5.60</td><td>11.02</td><td>8.87</td><td>4.63</td></tr><tr><td>• TENT [36]</td><td>1.79</td><td>4.85</td><td>3.00</td><td>3.36</td><td>3.49</td><td>6.05</td><td>0.49</td><td>4.43</td><td>6.20</td><td>7.19</td><td>6.50</td><td>10.43</td><td>9.23</td><td>5.15</td></tr><tr><td>• EATA [26]</td><td>1.89</td><td>4.20</td><td>2.41</td><td>4.13</td><td>3.17</td><td>5.73</td><td>0.38</td><td>4.06</td><td>6.27</td><td>6.41</td><td>6.14</td><td>10.93</td><td>7.72</td><td>4.88</td></tr><tr><td>• ActMAD [24]</td><td>1.14</td><td>2.84</td><td>1.53</td><td>3.3</td><td>3.08</td><td>3.15</td><td>0.32</td><td>2.74</td><td>3.47</td><td>6.36</td><td>5.75</td><td>9.69</td><td>6.95</td><td>3.87</td></tr><tr><td>• MonotTA (Ours)</td><td>3.93</td><td>5.78</td><td>4.55</td><td>5.43</td><td>4.70</td><td>6.09</td><td>0.69</td><td>4.66</td><td>7.53</td><td>7.69</td><td>7.74</td><td>11.71</td><td>9.43</td><td>6.15</td></tr></table>",
|
| 898 |
+
"bbox": [
|
| 899 |
+
241,
|
| 900 |
+
393,
|
| 901 |
+
761,
|
| 902 |
+
498
|
| 903 |
+
],
|
| 904 |
+
"page_idx": 9
|
| 905 |
+
},
|
| 906 |
+
{
|
| 907 |
+
"type": "text",
|
| 908 |
+
"text": "in certain extreme scenarios. For instance, the Snow scenario (c.f. Fig. 3 (a)) lacks objects with scores exceeding 0.4 when the pre-trained model is directly applied. Under such a circumstance, $\\mathcal{L}_{Nreg}$ plays a more important role in model adaptation since it can alleviate distribution shifts even only low-score objects, i.e., deny the negative category. In other words, $\\mathcal{L}_{Nreg}$ enables the model to alleviate distribution shifts and achieve more relatively high-score objects, thereby laying a crucial foundation for $\\mathcal{L}_{AO}$ in extremely challenging scenarios.",
|
| 909 |
+
"bbox": [
|
| 910 |
+
212,
|
| 911 |
+
518,
|
| 912 |
+
787,
|
| 913 |
+
625
|
| 914 |
+
],
|
| 915 |
+
"page_idx": 9
|
| 916 |
+
},
|
| 917 |
+
{
|
| 918 |
+
"type": "text",
|
| 919 |
+
"text": "4 Experiments",
|
| 920 |
+
"text_level": 1,
|
| 921 |
+
"bbox": [
|
| 922 |
+
215,
|
| 923 |
+
652,
|
| 924 |
+
375,
|
| 925 |
+
670
|
| 926 |
+
],
|
| 927 |
+
"page_idx": 9
|
| 928 |
+
},
|
| 929 |
+
{
|
| 930 |
+
"type": "text",
|
| 931 |
+
"text": "We conduct experiments based on KITTI [9] and nuScenes [1]. The results presented in this manuscript represent the average value across three difficulty levels, i.e., Easy, Moderate, Hard. Note that we provide more results of higher severity levels of KITTI and more detailed results in Appendix $D$ .",
|
| 932 |
+
"bbox": [
|
| 933 |
+
212,
|
| 934 |
+
686,
|
| 935 |
+
785,
|
| 936 |
+
748
|
| 937 |
+
],
|
| 938 |
+
"page_idx": 9
|
| 939 |
+
},
|
| 940 |
+
{
|
| 941 |
+
"type": "text",
|
| 942 |
+
"text": "Datasets. For KITTI, we adopt the protocol established by Monoflex [48] to split the training images into the training set (3712) and validation set (3769) to perform model training and adaptation, respectively. Following the official KITTI evaluation criteria, we evaluate detection results on three levels of difficulty. Then, we construct the KITTI-C dataset through the incorporation of 13 distinct types of data corruptions [11] to the validation set and each corruption",
|
| 943 |
+
"bbox": [
|
| 944 |
+
212,
|
| 945 |
+
750,
|
| 946 |
+
787,
|
| 947 |
+
840
|
| 948 |
+
],
|
| 949 |
+
"page_idx": 9
|
| 950 |
+
},
|
| 951 |
+
{
|
| 952 |
+
"type": "page_number",
|
| 953 |
+
"text": "10",
|
| 954 |
+
"bbox": [
|
| 955 |
+
217,
|
| 956 |
+
114,
|
| 957 |
+
235,
|
| 958 |
+
126
|
| 959 |
+
],
|
| 960 |
+
"page_idx": 9
|
| 961 |
+
},
|
| 962 |
+
{
|
| 963 |
+
"type": "header",
|
| 964 |
+
"text": "Hongbin Lin, Yifan Zhang and et al.",
|
| 965 |
+
"bbox": [
|
| 966 |
+
271,
|
| 967 |
+
114,
|
| 968 |
+
514,
|
| 969 |
+
128
|
| 970 |
+
],
|
| 971 |
+
"page_idx": 9
|
| 972 |
+
},
|
| 973 |
+
{
|
| 974 |
+
"type": "image",
|
| 975 |
+
"img_path": "images/1016ad6fb21f14f8f303f665c45098a31a4d72442d89e7b833a4fc5e03d381d3.jpg",
|
| 976 |
+
"image_caption": [
|
| 977 |
+
"Average performance of 13 corruptions on KITTI-C Validation set, severity level 2",
|
| 978 |
+
"Fig. 4: We visualize the comparison with baselines on the KITTI-C validation set, severity level 2 regarding Mean $AP_{3D|R_{40}}$ . The bold number indicates the best result."
|
| 979 |
+
],
|
| 980 |
+
"image_footnote": [],
|
| 981 |
+
"bbox": [
|
| 982 |
+
272,
|
| 983 |
+
156,
|
| 984 |
+
725,
|
| 985 |
+
282
|
| 986 |
+
],
|
| 987 |
+
"page_idx": 10
|
| 988 |
+
},
|
| 989 |
+
{
|
| 990 |
+
"type": "text",
|
| 991 |
+
"text": "has 5 severity levels. The model is trained on the original training set and then tested on the KITTI-C validation set within one of the corruptions.",
|
| 992 |
+
"bbox": [
|
| 993 |
+
214,
|
| 994 |
+
347,
|
| 995 |
+
782,
|
| 996 |
+
378
|
| 997 |
+
],
|
| 998 |
+
"page_idx": 10
|
| 999 |
+
},
|
| 1000 |
+
{
|
| 1001 |
+
"type": "text",
|
| 1002 |
+
"text": "As for nuScenes, we adopt the front-view images and construct the Daytime and Night scenarios via their scene descriptions following [21]. Specifically, there exist $24.7\\mathrm{k} / 5.4\\mathrm{k}$ train/test images in Daytime while $3.3\\mathrm{k} / 0.6\\mathrm{k}$ train/test images in Night. Based on these splits, we construct two real-world adaptation tasks, i.e., Daytime $\\rightarrow$ Night and Night $\\rightarrow$ Daytime. For simplification, we transfer the nuScenes dataset into the KITTI format and only consider the Car category. More details of data construction are provided in Appendix $\\pmb{B}$ .",
|
| 1003 |
+
"bbox": [
|
| 1004 |
+
212,
|
| 1005 |
+
387,
|
| 1006 |
+
784,
|
| 1007 |
+
492
|
| 1008 |
+
],
|
| 1009 |
+
"page_idx": 10
|
| 1010 |
+
},
|
| 1011 |
+
{
|
| 1012 |
+
"type": "text",
|
| 1013 |
+
"text": "Implementation details. We implement our method and other baselines in PyTorch [28]. In MonoTTA, we conduct model adaptation based on the public pre-trained weights and the parameter settings provided by their authors [29, 40, 48]. Besides, we employ the Stochastic Gradient Descent (SGD) optimizer with a half learning rate of the initial rate used in base training over different methods, a momentum of 0.9 and a batch size of 16 for KITTI, 4 for nuScenes. Parameters $\\lambda$ , $\\beta$ , $\\eta$ are assigned default values of 1, 0.1, and 0.05, respectively. More training details of MonoTTA are provided in Appendix $C$ .",
|
| 1014 |
+
"bbox": [
|
| 1015 |
+
212,
|
| 1016 |
+
502,
|
| 1017 |
+
784,
|
| 1018 |
+
623
|
| 1019 |
+
],
|
| 1020 |
+
"page_idx": 10
|
| 1021 |
+
},
|
| 1022 |
+
{
|
| 1023 |
+
"type": "text",
|
| 1024 |
+
"text": "Compared methods. Based on three typical or state-of-the-art (SOTA) Mono 3Det methods [29, 40, 48], we fully compare MonoTTA with following methods: 1) source-only, i.e., directly apply the pre-trained model to the test data within corruptions; 2) BN adaptation [33] updates batch normalization statistics via target data; 3) TENT [36] minimizes the entropy loss of test data; 4) EATA [26] identifies reliable samples to update the model by entropy loss minimization. Here we compare MonoTTA with its variant Efficient Test-time Adaptation. 5) ActMAD [24] matches activation statistics for the distribution alignment.",
|
| 1025 |
+
"bbox": [
|
| 1026 |
+
214,
|
| 1027 |
+
633,
|
| 1028 |
+
784,
|
| 1029 |
+
753
|
| 1030 |
+
],
|
| 1031 |
+
"page_idx": 10
|
| 1032 |
+
},
|
| 1033 |
+
{
|
| 1034 |
+
"type": "text",
|
| 1035 |
+
"text": "Evaluation protocols. In order to fully evaluate the proposed method, we report our experimental results in the Average Precision (AP) for 3D bounding boxes, denoted as $AP_{3D|R_{40}}$ . The results present the mean values across three levels of difficulty and the Intersection over Union (IoU) thresholds are set to 0.7, 0.5, 0.5 for Cars and 0.5, 0.25, 0.25 for Pedestrians and Cyclists.",
|
| 1036 |
+
"bbox": [
|
| 1037 |
+
214,
|
| 1038 |
+
763,
|
| 1039 |
+
784,
|
| 1040 |
+
839
|
| 1041 |
+
],
|
| 1042 |
+
"page_idx": 10
|
| 1043 |
+
},
|
| 1044 |
+
{
|
| 1045 |
+
"type": "header",
|
| 1046 |
+
"text": "MonoTTA",
|
| 1047 |
+
"bbox": [
|
| 1048 |
+
658,
|
| 1049 |
+
114,
|
| 1050 |
+
730,
|
| 1051 |
+
126
|
| 1052 |
+
],
|
| 1053 |
+
"page_idx": 10
|
| 1054 |
+
},
|
| 1055 |
+
{
|
| 1056 |
+
"type": "page_number",
|
| 1057 |
+
"text": "11",
|
| 1058 |
+
"bbox": [
|
| 1059 |
+
767,
|
| 1060 |
+
114,
|
| 1061 |
+
782,
|
| 1062 |
+
126
|
| 1063 |
+
],
|
| 1064 |
+
"page_idx": 10
|
| 1065 |
+
},
|
| 1066 |
+
{
|
| 1067 |
+
"type": "image",
|
| 1068 |
+
"img_path": "images/21e234440e204e27ec0c18cc793bbd57c303929d334ce98b0f5e2193c398f2d8.jpg",
|
| 1069 |
+
"image_caption": [
|
| 1070 |
+
"Performance on Daytime to Night and Night to Daytime scenarios of nuScenes",
|
| 1071 |
+
"Fig. 5: Comparison with baselines on Daytime $\\rightarrow$ Night and Night $\\rightarrow$ Daytime of nuScenes, regarding Mean $AP_{3D|R_{40}}$ . The **bold** number indicates the best result."
|
| 1072 |
+
],
|
| 1073 |
+
"image_footnote": [],
|
| 1074 |
+
"bbox": [
|
| 1075 |
+
271,
|
| 1076 |
+
157,
|
| 1077 |
+
728,
|
| 1078 |
+
292
|
| 1079 |
+
],
|
| 1080 |
+
"page_idx": 11
|
| 1081 |
+
},
|
| 1082 |
+
{
|
| 1083 |
+
"type": "text",
|
| 1084 |
+
"text": "4.1 Comparisons with Previous Methods",
|
| 1085 |
+
"text_level": 1,
|
| 1086 |
+
"bbox": [
|
| 1087 |
+
214,
|
| 1088 |
+
354,
|
| 1089 |
+
568,
|
| 1090 |
+
371
|
| 1091 |
+
],
|
| 1092 |
+
"page_idx": 11
|
| 1093 |
+
},
|
| 1094 |
+
{
|
| 1095 |
+
"type": "text",
|
| 1096 |
+
"text": "We first compare our MonoTTA with previous methods in severity level 1 of KITTI-C. The results are reported in Table 1, which gives the following observations: 1) Due to distribution shifts, directly applying the pre-trained model to the test data (i.e., source-only) suffers severe performance degradation in all categories. 2) Existing TTA methods are able to mitigate the negative effect of distribution shifts for Mono 3Det to some degree. However, they only achieve suboptimal performance since they tend to increase the scores of all positive detections, containing severe noise. 3) MonoTTA consistently outperforms all compared methods over all categories within various base models in terms of mean $AP_{3D|R_{40}}$ . Specifically, MonoTTA achieves the best or comparable performance in all categories under all corruptions, attaining a large performance gain over TENT and EATA (e.g., improving an average $AP_{3D|R_{40}}$ about 5.1 and 7.4 of the Car category based on MonoGround).",
|
| 1097 |
+
"bbox": [
|
| 1098 |
+
212,
|
| 1099 |
+
380,
|
| 1100 |
+
787,
|
| 1101 |
+
579
|
| 1102 |
+
],
|
| 1103 |
+
"page_idx": 11
|
| 1104 |
+
},
|
| 1105 |
+
{
|
| 1106 |
+
"type": "text",
|
| 1107 |
+
"text": "4.2 More Severe Corruption and Real Scenario",
|
| 1108 |
+
"text_level": 1,
|
| 1109 |
+
"bbox": [
|
| 1110 |
+
214,
|
| 1111 |
+
601,
|
| 1112 |
+
619,
|
| 1113 |
+
617
|
| 1114 |
+
],
|
| 1115 |
+
"page_idx": 11
|
| 1116 |
+
},
|
| 1117 |
+
{
|
| 1118 |
+
"type": "text",
|
| 1119 |
+
"text": "On the one hand, to fully validate the effectiveness of our MonoTTA, we visualize experimental results under more severe corruption conditions at severity level 2 as shown in Fig. 4, which clearly gives additional observations: 1) With the escalation of severity level, the pre-trained models suffer a larger performance decline within various corruptions, enlarging the difficulty of TTA. 2) The performance improvements of existing TTA methods become relatively limited, particularly in Pedestrian and Cyclist classes. 3) Even if the tasks are more challenging, MonoTTA still stably obtains the best average performance within all corruptions since $\\mathcal{L}_{Nreg}$ plays an important role in alleviating distribution shifts for certain extreme scenarios, i.e., only low-score objects exist.",
|
| 1120 |
+
"bbox": [
|
| 1121 |
+
212,
|
| 1122 |
+
627,
|
| 1123 |
+
787,
|
| 1124 |
+
777
|
| 1125 |
+
],
|
| 1126 |
+
"page_idx": 11
|
| 1127 |
+
},
|
| 1128 |
+
{
|
| 1129 |
+
"type": "text",
|
| 1130 |
+
"text": "On the other hand, we further validate different methods within real scenarios as shown in Fig. 5. The experimental results also give the following observations: 1) Under real corruptions, the pre-trained model still suffers severe performance degradation due to the data distribution shifts. 2) TENT tends to increase the",
|
| 1131 |
+
"bbox": [
|
| 1132 |
+
214,
|
| 1133 |
+
779,
|
| 1134 |
+
787,
|
| 1135 |
+
840
|
| 1136 |
+
],
|
| 1137 |
+
"page_idx": 11
|
| 1138 |
+
},
|
| 1139 |
+
{
|
| 1140 |
+
"type": "page_number",
|
| 1141 |
+
"text": "12",
|
| 1142 |
+
"bbox": [
|
| 1143 |
+
217,
|
| 1144 |
+
114,
|
| 1145 |
+
235,
|
| 1146 |
+
126
|
| 1147 |
+
],
|
| 1148 |
+
"page_idx": 11
|
| 1149 |
+
},
|
| 1150 |
+
{
|
| 1151 |
+
"type": "header",
|
| 1152 |
+
"text": "Hongbin Lin, Yifan Zhang and et al.",
|
| 1153 |
+
"bbox": [
|
| 1154 |
+
271,
|
| 1155 |
+
114,
|
| 1156 |
+
514,
|
| 1157 |
+
128
|
| 1158 |
+
],
|
| 1159 |
+
"page_idx": 11
|
| 1160 |
+
},
|
| 1161 |
+
{
|
| 1162 |
+
"type": "table",
|
| 1163 |
+
"img_path": "images/bd5c73fe7e83b74d2fa3050e86c7f7fed709f5718e2f98020d5b4cc7b4c25e3a.jpg",
|
| 1164 |
+
"table_caption": [
|
| 1165 |
+
"Table 2: Comparison with baselines based on the instance-level method (i.e., batch size is 1) on the KITTI-C validation set, regarding Mean $AP_{3D|R_{40}}$ and the severity levels 1 and 2. The bold number indicates the best result."
|
| 1166 |
+
],
|
| 1167 |
+
"table_footnote": [],
|
| 1168 |
+
"table_body": "<table><tr><td rowspan=\"2\">Method</td><td colspan=\"4\">Level 1</td><td colspan=\"4\">Level 2</td></tr><tr><td>Car</td><td>Pedes.</td><td>Cyclist</td><td>Avg.</td><td>Car</td><td>Pedes.</td><td>Cyclist</td><td>Avg.</td></tr><tr><td>MonoNeRD</td><td>19.84</td><td>5.96</td><td>2.27</td><td>9.36</td><td>13.02</td><td>3.83</td><td>1.61</td><td>6.15</td></tr><tr><td>BN adaptation [33]</td><td>30.73</td><td>8.85</td><td>3.81</td><td>14.46</td><td>26.47</td><td>6.94</td><td>2.91</td><td>12.11</td></tr><tr><td>TENT [36]</td><td>35.72</td><td>9.99</td><td>4.75</td><td>16.82</td><td>31.85</td><td>7.81</td><td>3.40</td><td>14.35</td></tr><tr><td>EATA [26]</td><td>34.60</td><td>10.04</td><td>4.20</td><td>16.28</td><td>30.66</td><td>7.86</td><td>3.28</td><td>13.93</td></tr><tr><td>MonoTTA (Ours)</td><td>37.40</td><td>10.39</td><td>4.35</td><td>17.38</td><td>33.99</td><td>8.25</td><td>3.33</td><td>15.19</td></tr></table>",
|
| 1169 |
+
"bbox": [
|
| 1170 |
+
330,
|
| 1171 |
+
198,
|
| 1172 |
+
671,
|
| 1173 |
+
250
|
| 1174 |
+
],
|
| 1175 |
+
"page_idx": 12
|
| 1176 |
+
},
|
| 1177 |
+
{
|
| 1178 |
+
"type": "table",
|
| 1179 |
+
"img_path": "images/58fae23c2e680d7613aa885f32347b57be1fe215d277b1b17e8cd08a98225fa2.jpg",
|
| 1180 |
+
"table_caption": [
|
| 1181 |
+
"Table 3: Based on MonoGround [29], we conduct ablation studies of $\\mathcal{L}_{AO}$ and $\\mathcal{L}_{Nreg}$ on the KITTI-C validation set, regarding $AP_{3D|R_{40}}$ ."
|
| 1182 |
+
],
|
| 1183 |
+
"table_footnote": [],
|
| 1184 |
+
"table_body": "<table><tr><td rowspan=\"2\">Backbone</td><td rowspan=\"2\">\\( {\\mathcal{L}}_{AO} \\)</td><td rowspan=\"2\">\\( {\\mathcal{L}}_{Nreg} \\)</td><td colspan=\"4\">Level 1</td><td colspan=\"4\">Level 2</td></tr><tr><td>Car</td><td>Pedes.</td><td>Cyclist</td><td>Avg.</td><td>Car</td><td>Pedes.</td><td>Cyclist</td><td>Avg.</td></tr><tr><td>✓</td><td></td><td></td><td>9.67</td><td>4.42</td><td>1.42</td><td>5.17</td><td>5.72</td><td>2.43</td><td>0.95</td><td>3.03</td></tr><tr><td>✓</td><td>✓</td><td></td><td>27.94</td><td>10.17</td><td>5.10</td><td>14.40</td><td>21.27</td><td>7.79</td><td>3.82</td><td>10.96</td></tr><tr><td>✓</td><td></td><td>✓</td><td>23.26</td><td>8.60</td><td>4.60</td><td>12.15</td><td>18.93</td><td>6.81</td><td>3.55</td><td>9.77</td></tr><tr><td>✓</td><td>✓</td><td>✓</td><td>32.64</td><td>11.59</td><td>6.15</td><td>16.79</td><td>25.96</td><td>8.78</td><td>4.12</td><td>12.85</td></tr></table>",
|
| 1185 |
+
"bbox": [
|
| 1186 |
+
328,
|
| 1187 |
+
301,
|
| 1188 |
+
671,
|
| 1189 |
+
348
|
| 1190 |
+
],
|
| 1191 |
+
"page_idx": 12
|
| 1192 |
+
},
|
| 1193 |
+
{
|
| 1194 |
+
"type": "text",
|
| 1195 |
+
"text": "confidence of all positive detections and thus overfits to noise, i.e., failing to handle the extremely challenging task $\\mathrm{N} \\rightarrow \\mathrm{D}$ . 3) EATA still achieves sub-optimal performance while our MonoTTA brings sufficient average performance improvement on both MonoFlex (6.23 mAP) and MonoGround (8.26 mAP), maintaining the best performance in real scenarios. Detailed results are in Appendix $D$ .",
|
| 1196 |
+
"bbox": [
|
| 1197 |
+
212,
|
| 1198 |
+
366,
|
| 1199 |
+
787,
|
| 1200 |
+
444
|
| 1201 |
+
],
|
| 1202 |
+
"page_idx": 12
|
| 1203 |
+
},
|
| 1204 |
+
{
|
| 1205 |
+
"type": "text",
|
| 1206 |
+
"text": "4.3 Application to Instance-Level Inference Method",
|
| 1207 |
+
"text_level": 1,
|
| 1208 |
+
"bbox": [
|
| 1209 |
+
214,
|
| 1210 |
+
465,
|
| 1211 |
+
658,
|
| 1212 |
+
479
|
| 1213 |
+
],
|
| 1214 |
+
"page_idx": 12
|
| 1215 |
+
},
|
| 1216 |
+
{
|
| 1217 |
+
"type": "text",
|
| 1218 |
+
"text": "In this section, we seek to investigate whether the proposed MonoTTA can be used to effectively enhance Mono 3Det methods which only process a single image sequentially, i.e., $B = 1$ . To be specific, it may be crucial for Mono 3Det methods to make immediate decisions based on the most recent scene (image) in real-world scenarios like autonomous driving. To this end, it is essential for TTA methods devised for Mono 3Det to allow a single image as input and then conduct model adaptation. To validate it, we integrate MonoTTA into the SOTA Mono 3Det method namely MonoNeRD [40] which accepts one image each time at the test phase. As shown in Table 2, MonoTTA achieves the best or comparable performance across all categories at both severity levels 1 and 2, illustrating the applicability of our method to boost these approaches for handling OOD test data. Detailed results can be also found in Appendix $D$ .",
|
| 1219 |
+
"bbox": [
|
| 1220 |
+
212,
|
| 1221 |
+
489,
|
| 1222 |
+
787,
|
| 1223 |
+
672
|
| 1224 |
+
],
|
| 1225 |
+
"page_idx": 12
|
| 1226 |
+
},
|
| 1227 |
+
{
|
| 1228 |
+
"type": "text",
|
| 1229 |
+
"text": "4.4 Ablation Studies and Quantitative Results",
|
| 1230 |
+
"text_level": 1,
|
| 1231 |
+
"bbox": [
|
| 1232 |
+
214,
|
| 1233 |
+
693,
|
| 1234 |
+
612,
|
| 1235 |
+
709
|
| 1236 |
+
],
|
| 1237 |
+
"page_idx": 12
|
| 1238 |
+
},
|
| 1239 |
+
{
|
| 1240 |
+
"type": "text",
|
| 1241 |
+
"text": "Ablation studies. To examine the effectiveness of the losses in MonoTTA, we show the results of the models optimized by different losses. As shown in Table 3, introducing $\\mathcal{L}_{AO}$ or $\\mathcal{L}_{Nreg}$ enhances the model performance compared to directly applying the pre-trained model (i.e., source only). On the one hand, such a result verifies that our strategy is able to alleviate the score decline issue in test OOD data with high-score objects. On the other hand, introducing $\\mathcal{L}_{Nreg}$ could obtain a higher average $AP_{3D|R_{40}}$ value, which also verifies that the negative regularization term is able to enhance the pre-trained model even only",
|
| 1242 |
+
"bbox": [
|
| 1243 |
+
212,
|
| 1244 |
+
718,
|
| 1245 |
+
787,
|
| 1246 |
+
840
|
| 1247 |
+
],
|
| 1248 |
+
"page_idx": 12
|
| 1249 |
+
},
|
| 1250 |
+
{
|
| 1251 |
+
"type": "header",
|
| 1252 |
+
"text": "MonoTTA",
|
| 1253 |
+
"bbox": [
|
| 1254 |
+
658,
|
| 1255 |
+
114,
|
| 1256 |
+
730,
|
| 1257 |
+
127
|
| 1258 |
+
],
|
| 1259 |
+
"page_idx": 12
|
| 1260 |
+
},
|
| 1261 |
+
{
|
| 1262 |
+
"type": "page_number",
|
| 1263 |
+
"text": "13",
|
| 1264 |
+
"bbox": [
|
| 1265 |
+
767,
|
| 1266 |
+
114,
|
| 1267 |
+
785,
|
| 1268 |
+
126
|
| 1269 |
+
],
|
| 1270 |
+
"page_idx": 12
|
| 1271 |
+
},
|
| 1272 |
+
{
|
| 1273 |
+
"type": "image",
|
| 1274 |
+
"img_path": "images/7dee429f5322d7f05be1b976335d70ae3b10fcc09f1d687224b5877ee2f9e237.jpg",
|
| 1275 |
+
"image_caption": [
|
| 1276 |
+
"Fig.6: Qualitative results of baselines and the proposed MonoTTA based on MonoGround [29]. We visualize the results on KITTI-C validation set, where predicted cars, pedestrians and cyclists are in lime green, sky blue and yellow, respectively."
|
| 1277 |
+
],
|
| 1278 |
+
"image_footnote": [],
|
| 1279 |
+
"bbox": [
|
| 1280 |
+
272,
|
| 1281 |
+
143,
|
| 1282 |
+
730,
|
| 1283 |
+
359
|
| 1284 |
+
],
|
| 1285 |
+
"page_idx": 13
|
| 1286 |
+
},
|
| 1287 |
+
{
|
| 1288 |
+
"type": "text",
|
| 1289 |
+
"text": "with low-score objects. When combining the losses (i.e. $\\mathcal{L}_{AO}$ , $\\mathcal{L}_{Nreg}$ ) together, we obtain the best performance.",
|
| 1290 |
+
"bbox": [
|
| 1291 |
+
212,
|
| 1292 |
+
426,
|
| 1293 |
+
782,
|
| 1294 |
+
457
|
| 1295 |
+
],
|
| 1296 |
+
"page_idx": 13
|
| 1297 |
+
},
|
| 1298 |
+
{
|
| 1299 |
+
"type": "text",
|
| 1300 |
+
"text": "Quantitative Results. We provide visualizations within Snow, Fog and Frost based on Monoground as shown in Fig. 6. It is evident that the source-only setting suffers severe omissions, while BN adaptation, TENT and EATA alleviate distribution shifts to some degree and give more detections. As for our MonoTTA, it can produce superior detections even in severe conditions, including fewer omissions and accurate detections as highlighted by red circles.",
|
| 1301 |
+
"bbox": [
|
| 1302 |
+
212,
|
| 1303 |
+
458,
|
| 1304 |
+
784,
|
| 1305 |
+
547
|
| 1306 |
+
],
|
| 1307 |
+
"page_idx": 13
|
| 1308 |
+
},
|
| 1309 |
+
{
|
| 1310 |
+
"type": "text",
|
| 1311 |
+
"text": "5 Conclusion",
|
| 1312 |
+
"text_level": 1,
|
| 1313 |
+
"bbox": [
|
| 1314 |
+
215,
|
| 1315 |
+
569,
|
| 1316 |
+
359,
|
| 1317 |
+
584
|
| 1318 |
+
],
|
| 1319 |
+
"page_idx": 13
|
| 1320 |
+
},
|
| 1321 |
+
{
|
| 1322 |
+
"type": "text",
|
| 1323 |
+
"text": "In this paper, we propose a monocular test-time adaptation method to improve the pre-trained model on the shifted test data for monocular 3D object detection. Specifically, our method consists of two strategies: 1) Reliability-driven adaptation. To discover more potential objects, we devise a self-adaptive strategy to identify reliable objects for adaptive model adaptation. 2) Noise-guard adaptation. To avoid overfitting to noise and trivial solutions, we devise the negative regularization term to mitigate the negative effects of noisy detections and alleviate distribution shifts. Experiments on KITTI-C and nuScenes datasets demonstrate the effectiveness of MonoTTA in handling fully test-time adaptation for monocular 3D object detection.",
|
| 1324 |
+
"bbox": [
|
| 1325 |
+
212,
|
| 1326 |
+
598,
|
| 1327 |
+
787,
|
| 1328 |
+
748
|
| 1329 |
+
],
|
| 1330 |
+
"page_idx": 13
|
| 1331 |
+
},
|
| 1332 |
+
{
|
| 1333 |
+
"type": "text",
|
| 1334 |
+
"text": "Future directions. 1) Our work focuses on 2D images, while future studies could explore 3D information in handling distribution shifts. 2) This work explores TTA by assuming one OOD distribution at a time, where the forgetting issue is not severe as the source model weights are recoverable. Exploring scenarios with dynamically OOD distributions offers a compelling future direction, where the forgetting issue would become more severe.",
|
| 1335 |
+
"bbox": [
|
| 1336 |
+
212,
|
| 1337 |
+
750,
|
| 1338 |
+
787,
|
| 1339 |
+
839
|
| 1340 |
+
],
|
| 1341 |
+
"page_idx": 13
|
| 1342 |
+
},
|
| 1343 |
+
{
|
| 1344 |
+
"type": "page_number",
|
| 1345 |
+
"text": "14",
|
| 1346 |
+
"bbox": [
|
| 1347 |
+
217,
|
| 1348 |
+
114,
|
| 1349 |
+
235,
|
| 1350 |
+
126
|
| 1351 |
+
],
|
| 1352 |
+
"page_idx": 13
|
| 1353 |
+
},
|
| 1354 |
+
{
|
| 1355 |
+
"type": "header",
|
| 1356 |
+
"text": "Hongbin Lin, Yifan Zhang and et al.",
|
| 1357 |
+
"bbox": [
|
| 1358 |
+
271,
|
| 1359 |
+
114,
|
| 1360 |
+
514,
|
| 1361 |
+
128
|
| 1362 |
+
],
|
| 1363 |
+
"page_idx": 13
|
| 1364 |
+
},
|
| 1365 |
+
{
|
| 1366 |
+
"type": "text",
|
| 1367 |
+
"text": "Acknowledgements",
|
| 1368 |
+
"text_level": 1,
|
| 1369 |
+
"bbox": [
|
| 1370 |
+
217,
|
| 1371 |
+
143,
|
| 1372 |
+
401,
|
| 1373 |
+
162
|
| 1374 |
+
],
|
| 1375 |
+
"page_idx": 14
|
| 1376 |
+
},
|
| 1377 |
+
{
|
| 1378 |
+
"type": "text",
|
| 1379 |
+
"text": "This work was supported by NSFC with Grant No. 62293482, by the Basic Research Project No. HZQB-KCZYZ-2021067 of Hetao Shenzhen HK S&T Cooperation Zone, by Shenzhen General Program No. JCYJ20220530143600001, by Shenzhen-Hong Kong Joint Funding No. SGDX20211123112401002, by the Shenzhen Outstanding Talents Training Fund 202002, by Guangdong Research Project No. 2017ZT07X152 and No. 2019CX01X104, by the Guangdong Provincial Key Laboratory of Future Networks of Intelligence (Grant No. 2022B1212010001), by the Guangdong Provincial Key Laboratory of Big Data Computing, CHUK-Shenzhen, by the NSFC 61931024&12326610, by the Key Area R&D Program of Guangdong Province with grant No. 2018B030338001, by the Shenzhen Key Laboratory of Big Data and Artificial Intelligence (Grant No. ZDSYS201707251409055), and by Tencent & Huawei Open Fund.",
|
| 1380 |
+
"bbox": [
|
| 1381 |
+
212,
|
| 1382 |
+
176,
|
| 1383 |
+
789,
|
| 1384 |
+
357
|
| 1385 |
+
],
|
| 1386 |
+
"page_idx": 14
|
| 1387 |
+
},
|
| 1388 |
+
{
|
| 1389 |
+
"type": "text",
|
| 1390 |
+
"text": "References",
|
| 1391 |
+
"text_level": 1,
|
| 1392 |
+
"bbox": [
|
| 1393 |
+
217,
|
| 1394 |
+
380,
|
| 1395 |
+
321,
|
| 1396 |
+
396
|
| 1397 |
+
],
|
| 1398 |
+
"page_idx": 14
|
| 1399 |
+
},
|
| 1400 |
+
{
|
| 1401 |
+
"type": "list",
|
| 1402 |
+
"sub_type": "ref_text",
|
| 1403 |
+
"list_items": [
|
| 1404 |
+
"1. Caesar, H., Bankiti, V., Lang, A.H., Vora, S., Liong, V.E., Xu, Q., Krishnan, A., Pan, Y., Baldan, G., Beijbom, O.: nuscenes: A multimodal dataset for autonomous driving. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 11621-11631 (2020)",
|
| 1405 |
+
"2. Chen, X., Kundu, K., Zhang, Z., Ma, H., Fidler, S., Urtasun, R.: Monocular 3d object detection for autonomous driving. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2147-2156 (2016)",
|
| 1406 |
+
"3. Chen, X., Kundu, K., Zhu, Y., Berneshawi, A.G., Ma, H., Fidler, S., Urtasun, R.: 3d object proposals for accurate object class detection. Advances in neural information processing systems 28 (2015)",
|
| 1407 |
+
"4. Chen, Y., Tai, L., Sun, K., Li, M.: Monopair: Monocular 3d object detection using pairwise spatial relationships. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12093-12102 (2020)",
|
| 1408 |
+
"5. Chen, Y., Liu, J., Zhang, X., Qi, X., Jia, J.: Voxelnext: Fully sparse voxelnet for 3d object detection and tracking. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 21674-21683 (2023)",
|
| 1409 |
+
"6. Ding, M., Huo, Y., Yi, H., Wang, Z., Shi, J., Lu, Z., Luo, P.: Learning depth-guided convolutions for monocular 3d object detection. In: Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition workshops. pp. 1000-1001 (2020)",
|
| 1410 |
+
"7. Duan, K., Bai, S., Xie, L., Qi, H., Huang, Q., Tian, Q.: Centernet: Keypoint triplets for object detection. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 6569-6578 (2019)",
|
| 1411 |
+
"8. Fleuret, F., et al.: Test time adaptation through perturbation robustness. In: NeurIPS 2021 Workshop on Distribution Shifts: Connecting Methods and Applications (2021)",
|
| 1412 |
+
"9. Geiger, A., Lenz, P., Urtasun, R.: Are we ready for autonomous driving? the kitti vision benchmark suite. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3354-3361. IEEE (2012)",
|
| 1413 |
+
"0. Hegde, D., Kilic, V., Sindagi, V., Cooper, A.B., Foster, M., Patel, V.M.: Source-free unsupervised domain adaptation for 3d object detection in adverse weather."
|
| 1414 |
+
],
|
| 1415 |
+
"bbox": [
|
| 1416 |
+
225,
|
| 1417 |
+
410,
|
| 1418 |
+
785,
|
| 1419 |
+
839
|
| 1420 |
+
],
|
| 1421 |
+
"page_idx": 14
|
| 1422 |
+
},
|
| 1423 |
+
{
|
| 1424 |
+
"type": "header",
|
| 1425 |
+
"text": "MonoTTA",
|
| 1426 |
+
"bbox": [
|
| 1427 |
+
658,
|
| 1428 |
+
114,
|
| 1429 |
+
730,
|
| 1430 |
+
126
|
| 1431 |
+
],
|
| 1432 |
+
"page_idx": 14
|
| 1433 |
+
},
|
| 1434 |
+
{
|
| 1435 |
+
"type": "page_number",
|
| 1436 |
+
"text": "15",
|
| 1437 |
+
"bbox": [
|
| 1438 |
+
767,
|
| 1439 |
+
114,
|
| 1440 |
+
785,
|
| 1441 |
+
126
|
| 1442 |
+
],
|
| 1443 |
+
"page_idx": 14
|
| 1444 |
+
},
|
| 1445 |
+
{
|
| 1446 |
+
"type": "list",
|
| 1447 |
+
"sub_type": "ref_text",
|
| 1448 |
+
"list_items": [
|
| 1449 |
+
"In: 2023 IEEE International Conference on Robotics and Automation (ICRA). pp. 6973-6980. IEEE (2023)",
|
| 1450 |
+
"11. Hendrycks, D., Dietterich, T.: Benchmarking neural network robustness to common corruptions and perturbations. In: International Conference on Learning Representations (2018)",
|
| 1451 |
+
"12. Kim, J., Hwang, I., Kim, Y.M.: Ev-tta: Test-time adaptation for event-based object recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17745-17754 (2022)",
|
| 1452 |
+
"13. Kim, Y., Yim, J., Yun, J., Kim, J.: Nlnl: Negative learning for noisy labels. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 101-110 (2019)",
|
| 1453 |
+
"14. Kim, Y., Yun, J., Shon, H., Kim, J.: Joint negative and positive learning for noisy labels. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9442-9451 (2021)",
|
| 1454 |
+
"15. Kumar, A., Brazil, G., Liu, X.: Groomed-nms: Grouped mathematically differentiable nms for monocular 3d object detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 8973-8983 (2021)",
|
| 1455 |
+
"16. Li, P., Chen, X., Shen, S.: Stereo r-cnn based 3d object detection for autonomous driving. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7644-7652 (2019)",
|
| 1456 |
+
"17. Liang, J., Hu, D., Feng, J.: Do we really need to access the source data? source hypothesis transfer for unsupervised domain adaptation. In: International conference on machine learning. pp. 6028-6039. PMLR (2020)",
|
| 1457 |
+
"18. Lin, H., Zhang, Y., Qiu, Z., Niu, S., Gan, C., Liu, Y., Tan, M.: Prototype-guided continual adaptation for class-incremental unsupervised domain adaptation. In: European Conference on Computer Vision. pp. 351-368 (2022)",
|
| 1458 |
+
"19. Liu, Y., Kothari, P., Van Delft, B., Bellot-Gurlet, B., Mordan, T., Alahi, A.: Ttt++: When does self-supervised test-time training fail or thrive? In: Advances in Neural Information Processing Systems. vol. 34, pp. 21808-21820 (2021)",
|
| 1459 |
+
"20. Liu, Z., Wu, Z., Tóth, R.: Smoke: Single-stage monocular 3d object detection via keypoint estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops. pp. 996-997 (2020)",
|
| 1460 |
+
"21. Liu, Z., Tang, H., Amini, A., Yang, X., Mao, H., Rus, D.L., Han, S.: Bevfusion: Multi-task multi-sensor fusion with unified bird's-eye view representation. In: 2023 IEEE international conference on robotics and automation (ICRA). pp. 2774-2781. IEEE (2023)",
|
| 1461 |
+
"22. Luo, Y., Zheng, C., Yan, X., Kun, T., Zheng, C., Cui, S., Li, Z.: Latr: 3d lane detection from monocular images with transformer. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 7941-7952 (2023)",
|
| 1462 |
+
"23. Ma, X., Liu, S., Xia, Z., Zhang, H., Zeng, X., Ouyang, W.: Rethinking pseudo-lidar representation. In: European Conference on Computer Vision (2020)",
|
| 1463 |
+
"24. Mirza, M.J., Soneira, P.J., Lin, W., Kozinski, M., Possegger, H., Bischof, H.: Actmad: Activation matching to align distributions for test-time-training. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 24152-24161 (2023)",
|
| 1464 |
+
"25. Nado, Z., Padhy, S., Sculley, D., D'Amour, A., Lakshminarayanan, B., Snoek, J.: Evaluating prediction-time batch normalization for robustness under covariate shift. arXiv preprint arXiv:2006.10963 (2020)",
|
| 1465 |
+
"26. Niu, S., Wu, J., Zhang, Y., Chen, Y., Zheng, S., Zhao, P., Tan, M.: Efficient test-time model adaptation without forgetting. In: The International Conference on Machine Learning (2022)"
|
| 1466 |
+
],
|
| 1467 |
+
"bbox": [
|
| 1468 |
+
217,
|
| 1469 |
+
146,
|
| 1470 |
+
784,
|
| 1471 |
+
839
|
| 1472 |
+
],
|
| 1473 |
+
"page_idx": 15
|
| 1474 |
+
},
|
| 1475 |
+
{
|
| 1476 |
+
"type": "page_number",
|
| 1477 |
+
"text": "16",
|
| 1478 |
+
"bbox": [
|
| 1479 |
+
217,
|
| 1480 |
+
114,
|
| 1481 |
+
235,
|
| 1482 |
+
126
|
| 1483 |
+
],
|
| 1484 |
+
"page_idx": 15
|
| 1485 |
+
},
|
| 1486 |
+
{
|
| 1487 |
+
"type": "header",
|
| 1488 |
+
"text": "Hongbin Lin, Yifan Zhang and et al.",
|
| 1489 |
+
"bbox": [
|
| 1490 |
+
271,
|
| 1491 |
+
114,
|
| 1492 |
+
514,
|
| 1493 |
+
128
|
| 1494 |
+
],
|
| 1495 |
+
"page_idx": 15
|
| 1496 |
+
},
|
| 1497 |
+
{
|
| 1498 |
+
"type": "list",
|
| 1499 |
+
"sub_type": "ref_text",
|
| 1500 |
+
"list_items": [
|
| 1501 |
+
"27. Niu, S., Wu, J., Zhang, Y., Wen, Z., Chen, Y., Zhao, P., Tan, M.: Towards stable test-time adaptation in dynamic wild world. In: International Conference on Learning Representations (2023)",
|
| 1502 |
+
"28. Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J., Chanan, G., Killeen, T., Lin, Z., Gimelshein, N., Antiga, L., et al.: Pytorch: An imperative style, high-performance deep learning library. In: Advances in neural information processing systems. vol. 32 (2019)",
|
| 1503 |
+
"29. Qin, Z., Li, X.: Monoground: Detecting monocular 3d objects from the ground. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3793-3802 (2022)",
|
| 1504 |
+
"30. Qiu, Z., Zhang, Y., Lin, H., et al.: Source-free domain adaptation via avatar prototype generation and adaptation. In: International Joint Conference on Artificial Intelligence (2021)",
|
| 1505 |
+
"31. Reading, C., Harakeh, A., Chae, J., Waslander, S.L.: Categorical depth distribution network for monocular 3d object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8555-8564 (2021)",
|
| 1506 |
+
"32. Saltori, C., Lathuilière, S., Sebe, N., Ricci, E., Galasso, F.: Sf-uda 3d: Source-free unsupervised domain adaptation for lidar-based 3d object detection. In: 2020 International Conference on 3D Vision (3DV). pp. 771–780. IEEE (2020)",
|
| 1507 |
+
"33. Schneider, S., Rusak, E., Eck, L., Bringmann, O., Brendel, W., Bethge, M.: Improving robustness against common corruptions by covariate shift adaptation. In: Advances in neural information processing systems. pp. 11539-11551 (2020)",
|
| 1508 |
+
"34. Sun, Y., Wang, X., Liu, Z., Miller, J., Efros, A., Hardt, M.: Test-time training with self-supervision for generalization under distribution shifts. In: International conference on machine learning. pp. 9229-9248. PMLR (2020)",
|
| 1509 |
+
"35. Veksler, O.: Test time adaptation with regularized loss for weakly supervised salient object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7360-7369 (2023)",
|
| 1510 |
+
"36. Wang, D., Shelhamer, E., Liu, S., Olshausen, B., Darrell, T.: Tent: Fully test-time adaptation by entropy minimization. In: The International Conference on Machine Learning (2021)",
|
| 1511 |
+
"37. Wang, Y., Chao, W.L., Garg, D., Hariharan, B., Campbell, M., Weinberger, K.Q.: Pseudo-lidar from visual depth estimation: Bridging the gap in 3d object detection for autonomous driving. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8445-8453 (2019)",
|
| 1512 |
+
"38. Wu, H., Wen, C., Shi, S., Li, X., Wang, C.: Virtual sparse convolution for multimodal 3d object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 21653-21662 (2023)",
|
| 1513 |
+
"39. Xu, B., Chen, Z.: Multi-level fusion based 3d object detection from monocular images. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 2345-2353 (2018)",
|
| 1514 |
+
"40. Xu, J., Peng, L., Cheng, H., Li, H., Qian, W., Li, K., Wang, W., Cai, D.: Mononerd: Nerf-like representations for monocular 3d object detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 6814-6824 (2023)",
|
| 1515 |
+
"41. Yang, J., Shi, S., Wang, Z., Li, H., Qi, X.: St3d: Self-training for unsupervised domain adaptation on 3d object detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10368-10378 (2021)",
|
| 1516 |
+
"42. Ye, X., Shu, M., Li, H., Shi, Y., Li, Y., Wang, G., Tan, X., Ding, E.: Rope3d: The roadside perception dataset for autonomous driving and monocular 3d object detection task. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 21341-21350 (2022)"
|
| 1517 |
+
],
|
| 1518 |
+
"bbox": [
|
| 1519 |
+
212,
|
| 1520 |
+
146,
|
| 1521 |
+
787,
|
| 1522 |
+
839
|
| 1523 |
+
],
|
| 1524 |
+
"page_idx": 16
|
| 1525 |
+
},
|
| 1526 |
+
{
|
| 1527 |
+
"type": "header",
|
| 1528 |
+
"text": "MonoTTA",
|
| 1529 |
+
"bbox": [
|
| 1530 |
+
658,
|
| 1531 |
+
114,
|
| 1532 |
+
730,
|
| 1533 |
+
126
|
| 1534 |
+
],
|
| 1535 |
+
"page_idx": 16
|
| 1536 |
+
},
|
| 1537 |
+
{
|
| 1538 |
+
"type": "page_number",
|
| 1539 |
+
"text": "17",
|
| 1540 |
+
"bbox": [
|
| 1541 |
+
767,
|
| 1542 |
+
114,
|
| 1543 |
+
785,
|
| 1544 |
+
126
|
| 1545 |
+
],
|
| 1546 |
+
"page_idx": 16
|
| 1547 |
+
},
|
| 1548 |
+
{
|
| 1549 |
+
"type": "list",
|
| 1550 |
+
"sub_type": "ref_text",
|
| 1551 |
+
"list_items": [
|
| 1552 |
+
"43. Zhang, M., Levine, S., Finn, C.: Memo: Test time robustness via adaptation and augmentation. In: Advances in Neural Information Processing Systems. vol. 35, pp. 38629-38642 (2022)",
|
| 1553 |
+
"44. Zhang, Y., Hooi, B., Hong, L., Feng, J.: Self-supervised aggregation of diverse experts for test-agnostic long-tailed recognition. In: Advances in Neural Information Processing Systems. vol. 35, pp. 34077-34090 (2022)",
|
| 1554 |
+
"45. Zhang, Y., Hooi, B., Hu, D., Liang, J., Feng, J.: Unleashing the power of contrastive self-supervised visual models via contrast-regularized fine-tuning. In: Advances in Neural Information Processing Systems. vol. 34, pp. 29848-29860 (2021)",
|
| 1555 |
+
"46. Zhang, Y., Kang, B., Hooi, B., Yan, S., Feng, J.: Deep long-tailed learning: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence 45(9), 10795-10816 (2023)",
|
| 1556 |
+
"47. Zhang, Y., Wei, Y., Wu, Q., Zhao, P., Niu, S., Huang, J., Tan, M.: Collaborative unsupervised domain adaptation for medical image diagnosis. IEEE Transactions on Image Processing 29, 7834-7844 (2020)",
|
| 1557 |
+
"48. Zhang, Y., Lu, J., Zhou, J.: Objects are different: Flexible monocular 3d object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3289-3298 (2021)",
|
| 1558 |
+
"49. Zou, Z., Ye, X., Du, L., Cheng, X., Tan, X., Zhang, L., Feng, J., Xue, X., Ding, E.: The devil is in the task: Exploiting reciprocal appearance-localization features for monocular 3d object detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 2713-2722 (2021)"
|
| 1559 |
+
],
|
| 1560 |
+
"bbox": [
|
| 1561 |
+
215,
|
| 1562 |
+
146,
|
| 1563 |
+
785,
|
| 1564 |
+
452
|
| 1565 |
+
],
|
| 1566 |
+
"page_idx": 17
|
| 1567 |
+
},
|
| 1568 |
+
{
|
| 1569 |
+
"type": "page_number",
|
| 1570 |
+
"text": "18",
|
| 1571 |
+
"bbox": [
|
| 1572 |
+
217,
|
| 1573 |
+
114,
|
| 1574 |
+
235,
|
| 1575 |
+
126
|
| 1576 |
+
],
|
| 1577 |
+
"page_idx": 17
|
| 1578 |
+
},
|
| 1579 |
+
{
|
| 1580 |
+
"type": "header",
|
| 1581 |
+
"text": "Hongbin Lin, Yifan Zhang and et al.",
|
| 1582 |
+
"bbox": [
|
| 1583 |
+
271,
|
| 1584 |
+
114,
|
| 1585 |
+
514,
|
| 1586 |
+
128
|
| 1587 |
+
],
|
| 1588 |
+
"page_idx": 17
|
| 1589 |
+
}
|
| 1590 |
+
]
|
2024/MonoTTA_ Fully Test-Time Adaptation for Monocular 3D Object Detection/cfa6dde5-bde3-4186-b95b-8bbf6c77c339_model.json
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
2024/MonoTTA_ Fully Test-Time Adaptation for Monocular 3D Object Detection/cfa6dde5-bde3-4186-b95b-8bbf6c77c339_origin.pdf
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:dc94df8f5c91cea17eb3ea82484b77abb58e0226858df0903a392b3b0936e0c8
|
| 3 |
+
size 1579683
|
2024/MonoTTA_ Fully Test-Time Adaptation for Monocular 3D Object Detection/full.md
ADDED
|
@@ -0,0 +1,271 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# MonoTTA: Fully Test-Time Adaptation for Monocular 3D Object Detection
|
| 2 |
+
|
| 3 |
+
Hongbin Lin $^{1,2*}$ , Yifan Zhang $^{3,4*}$ , Shuaicheng Niu $^{5*}$ , Shuguang Cui $^{2,1}$ , Zhen Li $^{2,1\dagger}$
|
| 4 |
+
|
| 5 |
+
$^{1}$ FNii-Shenzhen, $^{2}$ SSE, CUHK-Shenzhen, $^{3}$ NUS, $^{4}$ Skywork AI, $^{5}$ NTU {hongbinlin@link.,shuguangcui@,lizhen@}cuhk.edu.cn, yifan.zhang@u.nus.edu, shuaicheng.niu@ntu.edu.sg
|
| 6 |
+
|
| 7 |
+
Abstract. Monocular 3D object detection (Mono 3Det) aims to identify 3D objects from a single RGB image. However, existing methods often assume training and test data follow the same distribution, which may not hold in real-world test scenarios. To address the out-of-distribution (OOD) problems, we explore a new adaptation paradigm for Mono 3Det, termed Fully Test-time Adaptation which aims to adapt a well-trained model to unlabeled test data by handling potential data distribution shifts at test time. However, applying this paradigm in Mono 3Det poses significant challenges due to OOD test data causing a remarkable decline in object detection scores. This decline conflicts with the pre-defined score thresholds of existing detection methods, leading to severe object omissions (i.e., rare positive detections and many false negatives). Consequently, the limited positive detection and plenty of noisy predictions cause test-time adaptation to fail in Mono 3Det. To handle this problem, we propose a novel Monocular Test-Time Adaptation (MonoTTA) method, based on two new strategies. 1) Reliability-driven adaptation: we empirically find that high-score objects are still reliable and the optimization of high-score objects can enhance confidence across all detections. Thus, we devise a self-adaptive strategy to identify reliable objects for model adaptation, which discovers potential objects and alleviates omissions. 2) Noise-guard adaptation: since high-score objects may be scarce, we develop a negative regularization term to exploit the numerous low-score objects via negative learning, preventing overfitting to noise and trivial solutions. Experimental results show that MonoTTA brings significant performance gains for Mono 3Det models in OOD test scenarios, approximately $190\%$ gains by average on KITTI and $198\%$ gains on nuScenes. The source code is now available at Hongbin98/MonoTTA.
|
| 8 |
+
|
| 9 |
+
Keywords: Test-time Adaptation $\cdot$ Monocular 3D Object Detection
|
| 10 |
+
|
| 11 |
+
# 1 Introduction
|
| 12 |
+
|
| 13 |
+
Three-dimensional (3D) Object Detection is a significant computer vision task, with the objective of identifying objects and determining their spatial and di
|
| 14 |
+
|
| 15 |
+

|
| 16 |
+
Fig. 1: An illustration of the generalizability issue of Mono 3Det models. Compared with in-distribution (In-dis) scenarios (e.g., sunny), the detection scores within out-of-distribution (OOD) test data suffer severe degradation when the well-trained model (MonoFlex [48]) is directly applied to test scenarios affected by common natural disruptions, like weather changes (e.g., snow and fog). Since existing Mono 3Det methods mainly adopt a pre-defined score threshold (e.g., 0.2) for object detection, it leads to severe omissions and unreliable detections, thereby suffering serious performance degradation. Note that test images are the same but under different weather conditions.
|
| 17 |
+
|
| 18 |
+
mensional attributes through diverse sensor inputs [3,5,16,37,38]. To reduce the cost of sensors, there is an increasing trend towards implementing autonomous driving systems via Monocular 3D Object Detection (Mono 3Det) [2,42], where only one single RGB image and the camera calibration information are given. Even if this practical task is challenging, Mono 3Det methods have achieved promising results across various tasks and datasets [4,22,29,40,48]. Behind the success, a common presupposition is assuming that test images have the same distribution as the training images. However, this assumption could be possibly invalidated in many real-world scenarios due to prevalent natural corruptions such as weather changes, diminished sharpness, and other factors that introduce noise and contribute to uncalibrated cameras. In such circumstances, the well-trained model often suffers substantial performance degradation as a consequence of the data distributional shifts between the training images and the unlabeled test images. As shown in Fig. 1, the model performance degrades from $46.2\mathrm{mAP}$ in in-distribution data to $0.3\mathrm{mAP}$ in Snow and $7.2\mathrm{mAP}$ in Fog. Considering the widespread application of Mono 3Det in autonomous driving, the severe performance degradation due to out-of-distribution (OOD) test data may lead to unexpected traffic accidents and pose serious safety risks. Therefore, it is crucial to deal with the OOD generalization problem for Mono 3Det.
|
| 19 |
+
|
| 20 |
+
In addressing the OOD challenges specifically in test scenarios, one paradigm that has emerged as highly promising and gaining traction is Test-Time Adaptation (TTA), which seeks to tackle data distribution shifts by adapting a well-trained model to unlabeled test images in real time. Test-time training (TTT) [34] represents an initial approach of TTA in classification tasks, by adjusting the well-trained model to predict rotations through additional model training, while its computation demands at the adaptation stage are prohibitive
|
| 21 |
+
|
| 22 |
+
in Mono 3Det applications, particularly in autonomous driving. To enhance efficiency, Tent [36] and EATA [26] have been developed for Fully Test-Time Adaptation (Fully TTA) where only unlabeled test images and a well-trained model are provided. Besides, Ev-TTA [12] and SOD [35] devise TTA methods to handle the event-based object recognition and weakly supervised salient object detection, respectively. Considering the constraints on time of Mono 3Det, we explore the fully TTA paradigm which seeks to deal with OOD test data in real time.
|
| 23 |
+
|
| 24 |
+
To investigate this paradigm for Mono 3Det, we dig into the detection outcomes for objects within test scenarios with variations or corruptions that are commonly caused by weather or cameras. Specifically, we directly apply the well-trained model to the validation set of KITTI which has been artificially injected with four distinct types of weather-related corruptions, namely Brightness, Frost, Fog, and Snow. Subsequently, we plot their distributions of detection scores (c.f. Fig. 1). It is observed that the detection scores of test objects tend to markedly decline as well as the high-score objects are scarce in the extreme scenario (Snow) when the well-trained model is directly applied to the scenarios with corruptions. This phenomenon indicates that: 1) The pre-trained Mono 3Det model struggles to discriminate between objects and the background within OOD test data, presenting as quantities of omissions and unconfident detections. 2) Directly applying existing fully TTA methods to Mono 3Det could only get suboptimal performance since they struggle to optimize the model without enough high-score (positive) detections, especially in certain extreme scenarios.
|
| 25 |
+
|
| 26 |
+
To handle it in Mono 3Det, we propose a Monocular Test-Time Adaptation (MonoTTA) method, consisting of the reliability-driven adaptation and noise-guard adaptation strategies: 1) Reliability-driven adaptation. Specifically, data distribution shifts lead to omissions and noisy detections while our empirical analysis suggests that high detection score objects are still reliable (c.f. Fig. 3 (a)). Moreover, even if we only optimize the model via high-score objects ( $e.g., \geq 0.5$ ), both the numbers of low-score and high-score objects increase (c.f. Fig. 3 (b)). These investigations motivate us that exploiting high-score objects rather than all objects for model adaptation would be a more reliable way to alleviate data distribution shifts and discover potential objects. Hence, we develop a self-adaptive strategy for the identification of reliable high-score objects in test images and devise the adaptive optimization loss $\mathcal{L}_{AO}$ to exploit the reliable subset for model adaptation, alleviating the detection score decline issue of OOD test data and digging out more potential objects. 2) Noise-guard adaptation. In addition, data distribution shifts may also result in a scarcity of high-score objects, i.e., the majority of objects presenting low scores as the 'Snow' scenario in Fig. 3 (a). To this end, we develop a negative regularization term to make rational use of the numerous low-scoring objects in the Negative Learning manner [13]. On the one hand, the negative regularization term $\mathcal{L}_{Nreg}$ allows the model to conduct adaptation via numerous noisy low-scoring objects. Thus, the model can achieve more high-score objects after alleviating distribution shifts. On the other hand, this term also prevents the model from overfitting to noise and trivial solutions, i.e., assigning all classes of one object with high scores.
|
| 27 |
+
|
| 28 |
+
We summarize the main contributions as follows:
|
| 29 |
+
|
| 30 |
+
- To the best of our knowledge, we are the first to explore Fully Test-Time Adaptation to address OOD generalization problems for Mono 3Det. We show that the explored novel paradigm can bring significant improvements to Mono 3Det models in OOD test scenarios, e.g., $137\%$ and $244\%$ average performance gains across 13 types of OOD shifts on KITTI.
|
| 31 |
+
- Our empirical investigation reveals an important insight that high-score objects maintain their reliability amidst various corruptions, while optimizing these high-score objects significantly boosts model confidence across all detections. This motivates the first effective test-time adaptation method (i.e., our MonoTTA) in Mono 3Det.
|
| 32 |
+
- Extensive experiments on 13 types of corruptions of KITTI and 2 real scenarios (daytime $\leftrightarrow$ night) of nuScenes demonstrate the effectiveness of our MonoTTA in boosting existing Mono 3Det methods [29,48] to handle test-time OOD problems. Even for instance-level methods [40], MonoTTA also maintains sufficient improvement, which further confirms its applicability.
|
| 33 |
+
|
| 34 |
+
# 2 Related Work
|
| 35 |
+
|
| 36 |
+
We first review the literature on Monocular 3D Object Detection, and then discuss Source-free Domain Adaptation and Test-Time Adaptation methods. More discussions on Unsupervised Domain Adaptation [18,41,47] are in Appendix A. Monocular 3D Object Detection aims to perceive 3D objects from a single 2D image. Existing Mono 3Det methods could be divided into two groups according to the use of extra information. On the one hand, some existing methods leverage extra pre-trained depth estimation modules [6,39,49] to solve one of the most difficult problems in Mono 3Det, i.e., depth estimation from a single image. Other methods utilize LiDAR information, e.g., generating pseudo-LiDAR [23,31,37]. It is worth noting that Monoground [29] proposes to introduce the ground plane as prior information, and MonoNeRD [40] proposes to utilize scene geometric clues to enhance the detector's performance in the implicit reconstruction manner. On the other hand, some Mono 3Det methods try to detect 3D objects without extra data. For example, SMOKE [20] proposes to detect 3D objects as the key points estimation task. Then, Monoflex [48] improves this idea by providing a flexible definition of object centers, which unifies the centers of regular and truncated objects. GrooMeD-NMS [15] proposes a grouped mathematically differentiable Non-Maximal Suppression for Mono 3Det.
|
| 37 |
+
|
| 38 |
+
Source-free Domain Adaptation (SFDA) aims to adapt the pre-trained source model to an unlabeled target domain without using the source data due to privacy issues [17, 30]. SF-UDA $^3$ D [32] first explores the SFDA framework to adapt the PointRCNN 3D detector to target domains, which consists of pseudolabeling, reversible scale-transformations and motion coherency. Recently, the authors [10] seek to exploit the source model more reliably and propose an uncertainty-aware teacher-student framework to filter incorrect pseudo labels during model adaptation, alleviating the negative impact of label noise.
|
| 39 |
+
|
| 40 |
+
Nonetheless, SFDA assumes all target data to be known in advance and makes predictions after multiple epochs of optimization, which may not be viable for real-time applications due to computational or time constraints.
|
| 41 |
+
|
| 42 |
+
Test-Time Adaptation (TTA) seeks to improve model performance on test data via model adaptation through test samples even if data shifts exist. Early TTA methods [19, 34] endeavor to conduct additional model optimization on training data by self-supervised objectives, and then adapt the well-trained model to the test data via self-supervised objectives. However, in Mono 3Det applications like autonomous driving, the computation demands of such methods are prohibitive. To solve this, Fully Test-Time Adaptation methods are developed to adapt the well-trained model, where only unlabeled test images are available. Specifically, certain methods [25, 27, 33] tackle data distribution shifts by adapting the batch normalization layer statistics, while others alleviate this issue either by the entropy minimization of test data [8, 36] or maximizing the prediction consistency of different augmentations [43-46]. As for object detection tasks, Ev-TTA [12] and SOD [35] try to handle the event-based object recognition and weakly supervised salient object detection offline, respectively.
|
| 43 |
+
|
| 44 |
+
However, existing fully TTA methods struggle to optimize the model and solve distribution shifts in Mono 3Det due to numerous false negative detections. To the best of our knowledge, our MonoTTA stands as the first fully TTA method that handles distribution shifts for Mono 3Det models in real time.
|
| 45 |
+
|
| 46 |
+
# 3 Monocular Test-Time Adaptation
|
| 47 |
+
|
| 48 |
+
# 3.1 Problem Statement
|
| 49 |
+
|
| 50 |
+
Without loss of generality, we denote the pre-trained (or well-trained) model as $f_{\Theta_0}(\mathbf{x})$ , which is achieved via training on labeled training images $\{(\mathbf{x}_i^s,\mathbf{y}_i^s)\}_{i = 1}^N$ . The training images follow the training distribution $P(\mathbf{x})$ (i.e., $\mathbf{x}^s\sim P(\mathbf{x})$ ). Here, $\Theta_0$ represents the parameters of the pre-trained model and $N$ is the number of training data. During the training stage, the model is optimized to fit (or overfit) the training data. Then, at the test stage, the model will be able to perform well if the unlabeled test images $\mathcal{D}_t = \{\mathbf{x}_i\}_{i = 1}^{N_t}$ follows the identical data distribution, i.e., $\mathbf{x}\sim P(\mathbf{x})$ where $N_{t}$ is the total number of test images. However, in real applications, it is possible for the pre-trained model to encounter Out-Of-Distribution (OOD) test samples due to prevalent natural corruptions, namely distribution shifts, i.e., $\mathbf{x}\sim Q(\mathbf{x})$ and $P(\mathbf{x})\neq Q(\mathbf{x})$ .
|
| 51 |
+
|
| 52 |
+
To address this issue, fully test-time adaptation [36] seeks to tackle distribution shifts by adapting the pre-trained model $f_{\Theta_0}(\mathbf{x})$ to unlabeled test images $\{\mathbf{x}_i\}_{i=1}^{N_t}$ in real time. To achieve this goal, existing methods typically endeavor to update the model through the minimization of unsupervised objectives defined on test samples by $\min_{\hat{\Theta}} \mathcal{L}(\mathbf{x}; \Theta)$ , where $\mathbf{x} \sim Q(\mathbf{x})$ and $\hat{\Theta} \subseteq \Theta$ . Here, $\hat{\Theta}$ denotes the subset of model parameters that should be updated (i.e., batch normalization layers following existing methods [26,36]). Most existing fully TTA methods focus on classification tasks, heavily relying on sufficient positive predictions for
|
| 53 |
+
|
| 54 |
+

|
| 55 |
+
Fig. 2: An illustration of our MonoTTA. During the test phase, only the pre-trained model $f_{\Theta_0}(\mathbf{x})$ and unlabeled test images $\{\mathbf{x}_i\}_{i=1}^{N_t}$ are given. To conduct model adaptation, we initialize the model $f_{\Theta}(\mathbf{x})$ by $\Theta_0$ and only update the parameters of batch normalization layers. When a batch of test images arrives, we first compute test object scores and refine the adaptive threshold $\alpha$ to select the reliable high-score objects, thereby optimizing $\Theta$ via the adaptive optimization loss $\mathcal{L}_{AO}$ . Meanwhile, we devise a negative regularization term $\mathcal{L}_{Nreg}$ to facilitate the model to avoid overfitting to noise and trivial solutions. Here, Ped. and Cyc. represent Pedestrian and Cyclist in KITTI.
|
| 56 |
+
|
| 57 |
+
model adaptation. Nevertheless, there is a significant difference between conventional classification tasks and Mono 3Det. As previously indicated, the detection scores of test images $\mathbf{x}$ derived from $f_{\Theta_0}(\mathbf{x})$ are prone to markedly decrease in the presence of corruptions as shown in Fig. 1, leading to severe omissions (numerous false negatives) in Mono 3Det. In such circumstances, the scarcity of positive detections presents a significant challenge for model adaptation to test distributions while adapting the model with unreliable low-score detections may significantly introduce the noise. Therefore, existing fully TTA methods tend to fail in the OOD generalization problems of Mono 3Det.
|
| 58 |
+
|
| 59 |
+
# 3.2 Overall Scheme
|
| 60 |
+
|
| 61 |
+
After thoroughly examining the characteristics and challenges of Mono 3Det, we introduce a Monocular Test-Time Adaptation (MonoTTA) method to address the OOD problems for Mono 3Det models, which seeks to solve the object score declining issue within unlabeled OOD test data. As shown in Fig. 2, MonoTTA consists of two strategies: 1) Reliability-driven adaptation and 2) Noise-guard adaptation. We first briefly introduce the two strategies below.
|
| 62 |
+
|
| 63 |
+
First, we develop a reliability-driven adaptation strategy (c.f. Section 3.3) to conduct reliable model adaptation for OOD test data based on dependable test objects. Our empirical investigations inspire us to exploit those relatively reliable
|
| 64 |
+
|
| 65 |
+

|
| 66 |
+
(a)
|
| 67 |
+
|
| 68 |
+

|
| 69 |
+
(b)
|
| 70 |
+
Fig. 3: Based on MonoGround [29], we conduct two empirical studies (Car, KITTI), with the 3D IoU threshold of 0.5. (a) We visualize the accuracy of the objects across varied scoring ranges, which shows that the accuracy of objects with high scores remains relatively stable even in the presence of diverse corruptions (Ideal means in-distribution scenarios). (b) We visualize the number of low & high-score objects before and after optimization. Although only high-score objects are optimized, the model treats low-score objects with more confidence.
|
| 71 |
+
|
| 72 |
+
test objects for alleviating distribution shifts, thereby discovering more potential objects. To this end, MonoTTA excludes unreliable test objects involving an adaptive threshold $\alpha$ for any unlabeled test data. Subsequently, the model is optimized by the adaptive optimization loss $\mathcal{L}_{AO}$ via the selected reliable objects.
|
| 73 |
+
|
| 74 |
+
Second, we tend to utilize plenty of low-score objects to adapt the model in an indirect manner instead of directly optimizing the model since low-score objects are noisy. Hence, we devise a noise-guard adaptation strategy (c.f. Section 3.4) to prevent the model from overfitting to noisy predictions and falling into trivial solutions. Specifically, we randomly choose one of the negative classes of low-score objects and minimize the scores (e.g., score 0.03 of [0.10, 0.12, 0.03] c.f. Fig. 2) after simply filtering out extremely low-score detections. Even though the positive class is noisy (i.e., score 0.12), this term is capable of optimizing the model indirectly, i.e., learn to deny the negative category of the object.
|
| 75 |
+
|
| 76 |
+
Overall, the training scheme of MonoTTA is as follows:
|
| 77 |
+
|
| 78 |
+
$$
|
| 79 |
+
\min _ {\hat {\Theta}} \mathcal {L} _ {A O} (\hat {\Theta}) + \lambda \mathcal {L} _ {N r e g} (\hat {\Theta}), \tag {1}
|
| 80 |
+
$$
|
| 81 |
+
|
| 82 |
+
where $\lambda$ is the balance hyper-parameter. The pseudo-code of MonoTTA is summarized in Algorithm 1.
|
| 83 |
+
|
| 84 |
+
# 3.3 Reliability-Driven Adaptation
|
| 85 |
+
|
| 86 |
+
To identify dependable test objects and conduct test-time model adaptation, we propose a reliability-driven adaptation strategy that consists of two components: 1) Reliable object identification and 2) Adaptive model optimization.
|
| 87 |
+
|
| 88 |
+
Reliable object identification. When unlabeled test images arrive, it is difficult for the pre-trained model to get accurate detections on OOD test data due to the decline in detection scores. To resolve this, we dig into the detection
|
| 89 |
+
|
| 90 |
+
accuracy of the pre-trained model in test scenarios with various corruptions. Specifically, based on MonoGround [29], we visualize the detection precision for the Car category within the KITTI dataset across varied scoring ranges, with the 3D Intersection over Union (IoU) threshold of 0.5. As shown in Fig. 3 (a), we find that high-score objects are more reliable and relatively stable even in the presence of diverse corruptions. Following this, we propose to select reliable high-score objects to conduct model adaptation via an adaptive threshold $\alpha$ . Specifically, we exploit the model $f_{\Theta}(\cdot)$ (initialized by $\Theta_0$ ) to infer a batch of test images $\{\mathbf{x}_b\}_{b=1}^B$ and obtain the heatmap $\mathbf{h}_i$ of each image $\mathbf{x}_i$ by $\mathbf{h}_i = f_{\Theta}(\mathbf{x}_i)$ , where $i$ ranges from 0 to $B-1$ and $B$ denotes the batch size. Then, we could achieve object score maps $\mathbf{s}_i \in \mathbb{R}^{B \times N_m}$ like the peaks in $\mathbf{h}_i$ [7] after normalization, where $K$ and $N_m$ denotes the number of classes and the maximum of detected objects, respectively. With the score map $\mathbf{s}_i$ , we update the adaptive threshold $\alpha_t$ at the iteration $t$ in the exponential moving average manner by:
|
| 91 |
+
|
| 92 |
+
$$
|
| 93 |
+
\alpha_ {t} = \left\{ \begin{array}{l l} \gamma , & \text {i f} t = 1 \\ \beta \bar {m} _ {t} + (1 - \beta) \alpha_ {t - 1}, & \text {i f} t > 1 \end{array} \right., \tag {2}
|
| 94 |
+
$$
|
| 95 |
+
|
| 96 |
+
$$
|
| 97 |
+
\bar {m} _ {t} = \frac {1}{B} \sum_ {i = 1} ^ {B} \frac {\sum_ {j = 1} ^ {N _ {m}} s _ {i j} \cdot \mathbb {I} \left(s _ {i j} \geq \gamma\right)}{\sum_ {j = 1} ^ {N _ {m}} \mathbb {I} \left(s _ {i j} \geq \gamma\right)}. \tag {3}
|
| 98 |
+
$$
|
| 99 |
+
|
| 100 |
+
In Eqn. (2), $\bar{m}_t$ denotes the average score of all detected objects in a single batch of $B$ test samples at iteration $t$ , while $\beta \in [0,1]$ is a decay coefficient. As for Eqn. (3), $s_{ij} \in (0,1)$ denotes the score of the $j$ -th object in the $i$ -th test image while $\mathbb{I}(\cdot)$ is the indicator function. Note that $\gamma \in \mathbb{R}^1$ is a pre-defined object detection threshold adopted from existing methods at their original inference stage as well as $N_m$ .
|
| 101 |
+
|
| 102 |
+
Adaptive model optimization. As shown in Fig. 3 (b), the optimization of high-score objects can also enhance the confidence of the model for relatively low-score objects. It motivates us that exploiting high-score objects rather than all objects for model adaptation would be a more reliable way to learn from OOD test data. Therefore, with the adaptive threshold $\alpha_{t}$ , we select the reliable subset of high scores from $\mathbf{s}_i$ and calculate the adaptive optimization loss $\mathcal{L}_{AO}$ to adapt the model by:
|
| 103 |
+
|
| 104 |
+
$$
|
| 105 |
+
\mathcal {L} _ {A O} = - \frac {1}{B} \sum_ {i = 1} ^ {B} \sum_ {j = 1} ^ {N _ {m}} \log \left(s _ {i j} \cdot \mathbb {I} \left(s _ {i j} \geq \alpha_ {t}\right)\right), \tag {4}
|
| 106 |
+
$$
|
| 107 |
+
|
| 108 |
+
The adaptive optimization loss $\mathcal{L}_{AO}$ alleviates the potential data distribution shifts in OOD test data by allowing the model to confidently identify high-score test objects. As we mentioned, it solves the score decline issue of OOD test data in a more reliable way rather than directly optimizing all test objects, thereby avoiding overfitting to noise and discovering more potential test objects.
|
| 109 |
+
|
| 110 |
+
# Algorithm 1 The pipeline of the proposed MonoTTA
|
| 111 |
+
|
| 112 |
+
Require: Unlabeled test data $\mathcal{D}_t = \{\mathbf{x}_i\}_{i=1}^{N_t}$ ; Pre-trained model $f_{\Theta_0}(\mathbf{x})$ ; Batch size $B$ ; Parameters $\lambda, \beta, \eta$ .
|
| 113 |
+
|
| 114 |
+
1: for a batch images $\{\mathbf{x}_b\}_{b = 1}^B$ in $\mathcal{D}_t$ do
|
| 115 |
+
2: Update the adaptive threshold $\alpha$ based on Eqn. (2);
|
| 116 |
+
3: Calculate adaptive optimization loss $\mathcal{L}_{AO}$ based on Eqn. (4);
|
| 117 |
+
4: Calculate negative regularization term $\mathcal{L}_{Nreg}$ based on Eqn. (6);
|
| 118 |
+
5: Update $\hat{\Theta}$ by optimizing Eqn. (1)
|
| 119 |
+
6: end for
|
| 120 |
+
7: return Detection Results for all $\mathbf{x} \in \mathcal{D}_t$ .
|
| 121 |
+
|
| 122 |
+
# 3.4 Noise-Guard Adaptation
|
| 123 |
+
|
| 124 |
+
Through the optimization of adaptive optimization loss $\mathcal{L}_{AO}$ , the model is refined to yield more confident detection outcomes. Nevertheless, high-scoring objects may be scarce due to the distribution shifts, making the adaptation procedure difficult. Meanwhile, exclusive reliance on $\mathcal{L}_{AO}$ for adaptation may result in trivial solutions, whereby the model indiscriminately assigns high scores to all categories. Previous studies [13, 14] indicate that deep neural networks could learn from noisy pseudo labels in classification tasks through negative learning. Thus, MonoTTA proposes to learn from noisy low-score objects in a negative learning manner for Mono 3Det. Specifically, we denote $\hat{\mathbf{s}}_i\in \mathbb{R}^{B\times N_m\times K}$ as the multi-class score map for the test image $\mathbf{x}_i$ , i.e. the multi-class score map $\hat{\mathbf{s}}_i$ contains not only the highest score of objects but also the scores for other classes. Here $\mathbf{s}_i = \arg \max_k\hat{\mathbf{s}}_{ik}$ where $k$ is the class index. Employing a simple constant threshold $\eta$ , we filter out extremely low-score objects and then randomly select a negative class $\bar{k}$ for each object. Subsequently, we compute the regularization loss $e_k$ for each class $k$ with relatively low object scores $s_{ij}\in [\eta ,\alpha_t)$ by:
|
| 125 |
+
|
| 126 |
+
$$
|
| 127 |
+
e _ {k} = - \sum_ {i = 1} ^ {B} \sum_ {j = 1} ^ {N _ {m}} \bar {y} _ {i j} \log \left(1 - s _ {i j \bar {k}} \cdot \mathbb {I} (\bar {k} = k)\right), \tag {5}
|
| 128 |
+
$$
|
| 129 |
+
|
| 130 |
+
where $\bar{y}_{ij} = 1 - s_{ij\bar{k}}$ is a constant weight. We further define $n_k$ as the frequency of negative scores corresponding to class $k$ in the test batch and balance $e_k$ by:
|
| 131 |
+
|
| 132 |
+
$$
|
| 133 |
+
\mathcal {L} _ {N r e g} = \sum_ {k = 1} ^ {K} \frac {e _ {k}}{n _ {k}}. \tag {6}
|
| 134 |
+
$$
|
| 135 |
+
|
| 136 |
+
As we mentioned before, detections with low and intermediate scores tend to be more noisy (c.f. Fig. 3 (a)), which is unreliable for direct model adaptation. To this end, we introduce the regularization term $\mathcal{L}_{Nreg}$ to leverage noisy low-score detections and improve the model for assimilating potentially accurate information (c.f. Term 2 in Fig. 2). Moreover, this term also prevents the model from trivial solutions, i.e., indiscriminately assigning high scores to all classes of a single object (e.g., TENT [36]). In addition, high-score objects may be absent
|
| 137 |
+
|
| 138 |
+
Table 1: Comparison with baselines on the KITTI-C validation set, severity level 1 regarding Mean $AP_{3D|R_{40}}$ . The bold number indicates the best result.
|
| 139 |
+
|
| 140 |
+
<table><tr><td colspan="15">Car, IoU @ 0.7, 0.5, 0.5</td></tr><tr><td rowspan="2">Method</td><td colspan="3">Noise</td><td colspan="3">Blur</td><td colspan="4">Weather</td><td colspan="3">Digital</td><td rowspan="2">Avg.</td></tr><tr><td>Gauss.</td><td>Shot</td><td>Impul.</td><td>Defoc.</td><td>Glass</td><td>Motion</td><td>Snow</td><td>Frost</td><td>Fog</td><td>Brit.</td><td>Contr.</td><td>Pixel</td><td>Sat.</td></tr><tr><td>Monoflex [48]</td><td>3.84</td><td>7.48</td><td>5.31</td><td>2.59</td><td>3.73</td><td>11.05</td><td>0.23</td><td>7.77</td><td>7.57</td><td>24.87</td><td>6.92</td><td>28.16</td><td>31.46</td><td>10.84</td></tr><tr><td>• BN adaptation [33]</td><td>13.58</td><td>21.93</td><td>18.78</td><td>15.87</td><td>8.59</td><td>24.32</td><td>5.42</td><td>21.45</td><td>24.63</td><td>31.80</td><td>30.58</td><td>41.04</td><td>30.71</td><td>22.21</td></tr><tr><td>• TENT [36]</td><td>17.80</td><td>27.09</td><td>23.18</td><td>21.66</td><td>11.90</td><td>28.75</td><td>6.84</td><td>26.58</td><td>30.78</td><td>35.65</td><td>34.72</td><td>41.71</td><td>35.91</td><td>26.35</td></tr><tr><td>• EATA [26]</td><td>16.67</td><td>26.42</td><td>25.07</td><td>22.54</td><td>13.23</td><td>27.73</td><td>7.87</td><td>26.58</td><td>31.10</td><td>35.39</td><td>35.28</td><td>41.40</td><td>36.72</td><td>26.62</td></tr><tr><td>• ActMAD [24]</td><td>12.26</td><td>21.49</td><td>20.47</td><td>11.10</td><td>7.53</td><td>23.79</td><td>5.57</td><td>19.11</td><td>26.54</td><td>26.24</td><td>23.92</td><td>38.51</td><td>33.75</td><td>20.79</td></tr><tr><td>• MonoTTA (Ours)</td><td>21.15</td><td>28.65</td><td>26.64</td><td>25.91</td><td>19.26</td><td>31.48</td><td>12.43</td><td>30.24</td><td>33.75</td><td>36.84</td><td>36.83</td><td>41.97</td><td>38.13</td><td>29.48</td></tr><tr><td>MonoGround [29]</td><td>2.40</td><td>4.10</td><td>3.31</td><td>3.71</td><td>2.67</td><td>8.13</td><td>0.22</td><td>5.54</td><td>4.59</td><td>25.37</td><td>4.00</td><td>33.57</td><td>28.08</td><td>9.67</td></tr><tr><td>• BN adaptation [33]</td><td>13.49</td><td>23.52</td><td>19.69</td><td>16.33</td><td>7.61</td><td>23.99</td><td>7.98</td><td>20.71</td><td>24.00</td><td>31.34</td><td>29.03</td><td>43.06</td><td>32.99</td><td>22.60</td></tr><tr><td>• TENT [36]</td><td>17.94</td><td>29.60</td><td>19.90</td><td>23.45</td><td>13.90</td><td>29.39</td><td>10.32</td><td>26.65</td><td>33.35</td><td>35.96</td><td>36.39</td><td>43.35</td><td>37.79</td><td>27.54</td></tr><tr><td>• EATA [26]</td><td>16.03</td><td>26.08</td><td>18.08</td><td>20.28</td><td>12.40</td><td>27.37</td><td>9.22</td><td>23.79</td><td>29.49</td><td>33.65</td><td>32.58</td><td>43.61</td><td>36.00</td><td>25.28</td></tr><tr><td>• ActMAD [24]</td><td>13.65</td><td>22.39</td><td>20.56</td><td>14.99</td><td>8.17</td><td>21.09</td><td>9.13</td><td>18.77</td><td>19.52</td><td>28.34</td><td>25.32</td><td>42.63</td><td>31.58</td><td>21.24</td></tr><tr><td>• MonoTTA (Ours)</td><td>26.13</td><td>33.11</td><td>28.60</td><td>30.38</td><td>25.48</td><td>32.44</td><td>18.72</td><td>32.60</td><td>37.75</td><td>37.87</td><td>39.57</td><td>43.67</td><td>37.98</td><td>32.64</td></tr></table>
|
| 141 |
+
|
| 142 |
+
<table><tr><td colspan="14">Pedestrian, IoU @ 0.5, 0.25</td><td></td></tr><tr><td rowspan="2">Method</td><td colspan="3">Noise</td><td colspan="3">Blur</td><td colspan="4">Weather</td><td colspan="3">Digital</td><td rowspan="2">Avg.</td></tr><tr><td>Gauss.</td><td>Shot</td><td>Impul.</td><td>Defoc.</td><td>Glass</td><td>Motion</td><td>Snow</td><td>Frost</td><td>Fog</td><td>Brit.</td><td>Contr.</td><td>Pixel</td><td>Sat.</td></tr><tr><td>Monoflex</td><td>0.19</td><td>1.62</td><td>0.32</td><td>3.72</td><td>8.47</td><td>6.22</td><td>0.00</td><td>4.27</td><td>2.25</td><td>9.19</td><td>2.08</td><td>1.83</td><td>9.11</td><td>3.79</td></tr><tr><td>• BN adaptation [33]</td><td>6.21</td><td>8.20</td><td>9.20</td><td>7.83</td><td>5.35</td><td>7.52</td><td>2.89</td><td>6.47</td><td>9.24</td><td>9.12</td><td>9.93</td><td>12.73</td><td>9.76</td><td>8.03</td></tr><tr><td>• TENT [36]</td><td>6.02</td><td>7.96</td><td>9.57</td><td>7.75</td><td>6.06</td><td>8.63</td><td>2.63</td><td>6.71</td><td>9.91</td><td>10.26</td><td>10.55</td><td>12.33</td><td>10.27</td><td>8.36</td></tr><tr><td>• EATA [26]</td><td>6.05</td><td>7.96</td><td>9.74</td><td>7.93</td><td>6.06</td><td>8.01</td><td>2.48</td><td>6.24</td><td>9.94</td><td>9.70</td><td>10.02</td><td>12.41</td><td>10.12</td><td>8.21</td></tr><tr><td>• ActMAD [24]</td><td>2.41</td><td>4.27</td><td>4.68</td><td>1.46</td><td>1.62</td><td>6.48</td><td>1.17</td><td>4.25</td><td>4.84</td><td>6.42</td><td>6.19</td><td>8.29</td><td>7.07</td><td>4.55</td></tr><tr><td>• MonotTA (Ours)</td><td>6.54</td><td>8.41</td><td>9.39</td><td>7.63</td><td>7.12</td><td>8.99</td><td>3.25</td><td>7.64</td><td>10.26</td><td>10.55</td><td>10.06</td><td>13.28</td><td>10.66</td><td>8.75</td></tr><tr><td>MonoGround</td><td>0.61</td><td>0.93</td><td>0.73</td><td>8.41</td><td>7.57</td><td>8.03</td><td>0.00</td><td>3.19</td><td>1.39</td><td>13.82</td><td>1.83</td><td>3.70</td><td>7.26</td><td>4.42</td></tr><tr><td>• BN adaptation [33]</td><td>5.09</td><td>7.08</td><td>7.77</td><td>7.63</td><td>6.63</td><td>9.45</td><td>2.24</td><td>6.72</td><td>8.40</td><td>9.72</td><td>11.06</td><td>16.70</td><td>12.43</td><td>8.53</td></tr><tr><td>• TENT [36]</td><td>7.27</td><td>10.10</td><td>10.02</td><td>8.53</td><td>8.30</td><td>11.03</td><td>3.54</td><td>8.73</td><td>9.11</td><td>11.74</td><td>12.12</td><td>17.70</td><td>15.03</td><td>10.25</td></tr><tr><td>• EATA [26]</td><td>5.92</td><td>8.05</td><td>8.58</td><td>8.12</td><td>7.59</td><td>10.95</td><td>3.31</td><td>7.91</td><td>10.10</td><td>11.02</td><td>10.95</td><td>17.30</td><td>14.34</td><td>9.55</td></tr><tr><td>• ActMAD [24]</td><td>3.96</td><td>5.75</td><td>6.48</td><td>5.61</td><td>5.73</td><td>7.14</td><td>1.58</td><td>5.37</td><td>5.74</td><td>7.27</td><td>8.10</td><td>13.12</td><td>9.27</td><td>6.55</td></tr><tr><td>• MonotTA (Ours)</td><td>8.58</td><td>11.18</td><td>11.79</td><td>9.22</td><td>9.40</td><td>13.20</td><td>4.83</td><td>9.95</td><td>14.46</td><td>12.85</td><td>13.25</td><td>17.13</td><td>14.85</td><td>11.59</td></tr></table>
|
| 143 |
+
|
| 144 |
+
<table><tr><td colspan="14">Cyclist, IoU @ 0.5, 0.25, 0.25</td><td></td></tr><tr><td rowspan="2">Method</td><td colspan="3">Noise</td><td colspan="3">Blur</td><td colspan="4">Weather</td><td colspan="3">Digital</td><td rowspan="2">Avg.</td></tr><tr><td>Gauss.</td><td>Shot</td><td>Impul.</td><td>Defoc.</td><td>Glass</td><td>Motion</td><td>Snow</td><td>Frost</td><td>Fog</td><td>Brit.</td><td>Contr.</td><td>Pixel</td><td>Sat.</td></tr><tr><td>Monoflex</td><td>0.28</td><td>1.64</td><td>0.47</td><td>0.59</td><td>4.97</td><td>3.60</td><td>0.00</td><td>7.42</td><td>3.81</td><td>13.07</td><td>3.79</td><td>3.80</td><td>8.39</td><td>3.99</td></tr><tr><td>• BN adaptation [33]</td><td>2.39</td><td>6.26</td><td>4.36</td><td>5.78</td><td>6.76</td><td>9.09</td><td>1.70</td><td>8.53</td><td>9.16</td><td>12.91</td><td>11.26</td><td>10.55</td><td>11.02</td><td>7.67</td></tr><tr><td>• TENT [36]</td><td>2.72</td><td>7.94</td><td>5.63</td><td>6.27</td><td>7.20</td><td>9.49</td><td>1.07</td><td>8.94</td><td>10.96</td><td>12.75</td><td>12.72</td><td>9.64</td><td>11.28</td><td>8.20</td></tr><tr><td>• EATA [26]</td><td>2.33</td><td>7.46</td><td>5.46</td><td>7.19</td><td>7.23</td><td>7.51</td><td>1.24</td><td>8.60</td><td>10.18</td><td>12.86</td><td>11.02</td><td>10.89</td><td>10.30</td><td>7.87</td></tr><tr><td>• ActMAD [24]</td><td>2.22</td><td>5.32</td><td>4.36</td><td>2.29</td><td>1.57</td><td>5.64</td><td>0.48</td><td>5.68</td><td>6.46</td><td>10.56</td><td>9.31</td><td>9.97</td><td>7.80</td><td>5.51</td></tr><tr><td>• MonotTA (Ours)</td><td>3.01</td><td>7.24</td><td>5.98</td><td>7.00</td><td>6.09</td><td>9.51</td><td>1.45</td><td>10.63</td><td>11.22</td><td>12.85</td><td>11.85</td><td>11.59</td><td>10.81</td><td>8.40</td></tr><tr><td>MonoGround</td><td>0.12</td><td>0.46</td><td>0.48</td><td>0.33</td><td>0.72</td><td>2.03</td><td>0.00</td><td>0.56</td><td>0.35</td><td>5.55</td><td>0.52</td><td>2.08</td><td>5.24</td><td>1.42</td></tr><tr><td>• BN adaptation [33]</td><td>1.76</td><td>3.58</td><td>2.08</td><td>3.61</td><td>3.05</td><td>5.41</td><td>0.57</td><td>3.82</td><td>4.47</td><td>6.30</td><td>5.60</td><td>11.02</td><td>8.87</td><td>4.63</td></tr><tr><td>• TENT [36]</td><td>1.79</td><td>4.85</td><td>3.00</td><td>3.36</td><td>3.49</td><td>6.05</td><td>0.49</td><td>4.43</td><td>6.20</td><td>7.19</td><td>6.50</td><td>10.43</td><td>9.23</td><td>5.15</td></tr><tr><td>• EATA [26]</td><td>1.89</td><td>4.20</td><td>2.41</td><td>4.13</td><td>3.17</td><td>5.73</td><td>0.38</td><td>4.06</td><td>6.27</td><td>6.41</td><td>6.14</td><td>10.93</td><td>7.72</td><td>4.88</td></tr><tr><td>• ActMAD [24]</td><td>1.14</td><td>2.84</td><td>1.53</td><td>3.3</td><td>3.08</td><td>3.15</td><td>0.32</td><td>2.74</td><td>3.47</td><td>6.36</td><td>5.75</td><td>9.69</td><td>6.95</td><td>3.87</td></tr><tr><td>• MonotTA (Ours)</td><td>3.93</td><td>5.78</td><td>4.55</td><td>5.43</td><td>4.70</td><td>6.09</td><td>0.69</td><td>4.66</td><td>7.53</td><td>7.69</td><td>7.74</td><td>11.71</td><td>9.43</td><td>6.15</td></tr></table>
|
| 145 |
+
|
| 146 |
+
in certain extreme scenarios. For instance, the Snow scenario (c.f. Fig. 3 (a)) lacks objects with scores exceeding 0.4 when the pre-trained model is directly applied. Under such a circumstance, $\mathcal{L}_{Nreg}$ plays a more important role in model adaptation since it can alleviate distribution shifts even only low-score objects, i.e., deny the negative category. In other words, $\mathcal{L}_{Nreg}$ enables the model to alleviate distribution shifts and achieve more relatively high-score objects, thereby laying a crucial foundation for $\mathcal{L}_{AO}$ in extremely challenging scenarios.
|
| 147 |
+
|
| 148 |
+
# 4 Experiments
|
| 149 |
+
|
| 150 |
+
We conduct experiments based on KITTI [9] and nuScenes [1]. The results presented in this manuscript represent the average value across three difficulty levels, i.e., Easy, Moderate, Hard. Note that we provide more results of higher severity levels of KITTI and more detailed results in Appendix $D$ .
|
| 151 |
+
|
| 152 |
+
Datasets. For KITTI, we adopt the protocol established by Monoflex [48] to split the training images into the training set (3712) and validation set (3769) to perform model training and adaptation, respectively. Following the official KITTI evaluation criteria, we evaluate detection results on three levels of difficulty. Then, we construct the KITTI-C dataset through the incorporation of 13 distinct types of data corruptions [11] to the validation set and each corruption
|
| 153 |
+
|
| 154 |
+

|
| 155 |
+
Average performance of 13 corruptions on KITTI-C Validation set, severity level 2
|
| 156 |
+
Fig. 4: We visualize the comparison with baselines on the KITTI-C validation set, severity level 2 regarding Mean $AP_{3D|R_{40}}$ . The bold number indicates the best result.
|
| 157 |
+
|
| 158 |
+
has 5 severity levels. The model is trained on the original training set and then tested on the KITTI-C validation set within one of the corruptions.
|
| 159 |
+
|
| 160 |
+
As for nuScenes, we adopt the front-view images and construct the Daytime and Night scenarios via their scene descriptions following [21]. Specifically, there exist $24.7\mathrm{k} / 5.4\mathrm{k}$ train/test images in Daytime while $3.3\mathrm{k} / 0.6\mathrm{k}$ train/test images in Night. Based on these splits, we construct two real-world adaptation tasks, i.e., Daytime $\rightarrow$ Night and Night $\rightarrow$ Daytime. For simplification, we transfer the nuScenes dataset into the KITTI format and only consider the Car category. More details of data construction are provided in Appendix $\pmb{B}$ .
|
| 161 |
+
|
| 162 |
+
Implementation details. We implement our method and other baselines in PyTorch [28]. In MonoTTA, we conduct model adaptation based on the public pre-trained weights and the parameter settings provided by their authors [29, 40, 48]. Besides, we employ the Stochastic Gradient Descent (SGD) optimizer with a half learning rate of the initial rate used in base training over different methods, a momentum of 0.9 and a batch size of 16 for KITTI, 4 for nuScenes. Parameters $\lambda$ , $\beta$ , $\eta$ are assigned default values of 1, 0.1, and 0.05, respectively. More training details of MonoTTA are provided in Appendix $C$ .
|
| 163 |
+
|
| 164 |
+
Compared methods. Based on three typical or state-of-the-art (SOTA) Mono 3Det methods [29, 40, 48], we fully compare MonoTTA with following methods: 1) source-only, i.e., directly apply the pre-trained model to the test data within corruptions; 2) BN adaptation [33] updates batch normalization statistics via target data; 3) TENT [36] minimizes the entropy loss of test data; 4) EATA [26] identifies reliable samples to update the model by entropy loss minimization. Here we compare MonoTTA with its variant Efficient Test-time Adaptation. 5) ActMAD [24] matches activation statistics for the distribution alignment.
|
| 165 |
+
|
| 166 |
+
Evaluation protocols. In order to fully evaluate the proposed method, we report our experimental results in the Average Precision (AP) for 3D bounding boxes, denoted as $AP_{3D|R_{40}}$ . The results present the mean values across three levels of difficulty and the Intersection over Union (IoU) thresholds are set to 0.7, 0.5, 0.5 for Cars and 0.5, 0.25, 0.25 for Pedestrians and Cyclists.
|
| 167 |
+
|
| 168 |
+

|
| 169 |
+
Performance on Daytime to Night and Night to Daytime scenarios of nuScenes
|
| 170 |
+
Fig. 5: Comparison with baselines on Daytime $\rightarrow$ Night and Night $\rightarrow$ Daytime of nuScenes, regarding Mean $AP_{3D|R_{40}}$ . The **bold** number indicates the best result.
|
| 171 |
+
|
| 172 |
+
# 4.1 Comparisons with Previous Methods
|
| 173 |
+
|
| 174 |
+
We first compare our MonoTTA with previous methods in severity level 1 of KITTI-C. The results are reported in Table 1, which gives the following observations: 1) Due to distribution shifts, directly applying the pre-trained model to the test data (i.e., source-only) suffers severe performance degradation in all categories. 2) Existing TTA methods are able to mitigate the negative effect of distribution shifts for Mono 3Det to some degree. However, they only achieve suboptimal performance since they tend to increase the scores of all positive detections, containing severe noise. 3) MonoTTA consistently outperforms all compared methods over all categories within various base models in terms of mean $AP_{3D|R_{40}}$ . Specifically, MonoTTA achieves the best or comparable performance in all categories under all corruptions, attaining a large performance gain over TENT and EATA (e.g., improving an average $AP_{3D|R_{40}}$ about 5.1 and 7.4 of the Car category based on MonoGround).
|
| 175 |
+
|
| 176 |
+
# 4.2 More Severe Corruption and Real Scenario
|
| 177 |
+
|
| 178 |
+
On the one hand, to fully validate the effectiveness of our MonoTTA, we visualize experimental results under more severe corruption conditions at severity level 2 as shown in Fig. 4, which clearly gives additional observations: 1) With the escalation of severity level, the pre-trained models suffer a larger performance decline within various corruptions, enlarging the difficulty of TTA. 2) The performance improvements of existing TTA methods become relatively limited, particularly in Pedestrian and Cyclist classes. 3) Even if the tasks are more challenging, MonoTTA still stably obtains the best average performance within all corruptions since $\mathcal{L}_{Nreg}$ plays an important role in alleviating distribution shifts for certain extreme scenarios, i.e., only low-score objects exist.
|
| 179 |
+
|
| 180 |
+
On the other hand, we further validate different methods within real scenarios as shown in Fig. 5. The experimental results also give the following observations: 1) Under real corruptions, the pre-trained model still suffers severe performance degradation due to the data distribution shifts. 2) TENT tends to increase the
|
| 181 |
+
|
| 182 |
+
Table 2: Comparison with baselines based on the instance-level method (i.e., batch size is 1) on the KITTI-C validation set, regarding Mean $AP_{3D|R_{40}}$ and the severity levels 1 and 2. The bold number indicates the best result.
|
| 183 |
+
|
| 184 |
+
<table><tr><td rowspan="2">Method</td><td colspan="4">Level 1</td><td colspan="4">Level 2</td></tr><tr><td>Car</td><td>Pedes.</td><td>Cyclist</td><td>Avg.</td><td>Car</td><td>Pedes.</td><td>Cyclist</td><td>Avg.</td></tr><tr><td>MonoNeRD</td><td>19.84</td><td>5.96</td><td>2.27</td><td>9.36</td><td>13.02</td><td>3.83</td><td>1.61</td><td>6.15</td></tr><tr><td>BN adaptation [33]</td><td>30.73</td><td>8.85</td><td>3.81</td><td>14.46</td><td>26.47</td><td>6.94</td><td>2.91</td><td>12.11</td></tr><tr><td>TENT [36]</td><td>35.72</td><td>9.99</td><td>4.75</td><td>16.82</td><td>31.85</td><td>7.81</td><td>3.40</td><td>14.35</td></tr><tr><td>EATA [26]</td><td>34.60</td><td>10.04</td><td>4.20</td><td>16.28</td><td>30.66</td><td>7.86</td><td>3.28</td><td>13.93</td></tr><tr><td>MonoTTA (Ours)</td><td>37.40</td><td>10.39</td><td>4.35</td><td>17.38</td><td>33.99</td><td>8.25</td><td>3.33</td><td>15.19</td></tr></table>
|
| 185 |
+
|
| 186 |
+
Table 3: Based on MonoGround [29], we conduct ablation studies of $\mathcal{L}_{AO}$ and $\mathcal{L}_{Nreg}$ on the KITTI-C validation set, regarding $AP_{3D|R_{40}}$ .
|
| 187 |
+
|
| 188 |
+
<table><tr><td rowspan="2">Backbone</td><td rowspan="2">\( {\mathcal{L}}_{AO} \)</td><td rowspan="2">\( {\mathcal{L}}_{Nreg} \)</td><td colspan="4">Level 1</td><td colspan="4">Level 2</td></tr><tr><td>Car</td><td>Pedes.</td><td>Cyclist</td><td>Avg.</td><td>Car</td><td>Pedes.</td><td>Cyclist</td><td>Avg.</td></tr><tr><td>✓</td><td></td><td></td><td>9.67</td><td>4.42</td><td>1.42</td><td>5.17</td><td>5.72</td><td>2.43</td><td>0.95</td><td>3.03</td></tr><tr><td>✓</td><td>✓</td><td></td><td>27.94</td><td>10.17</td><td>5.10</td><td>14.40</td><td>21.27</td><td>7.79</td><td>3.82</td><td>10.96</td></tr><tr><td>✓</td><td></td><td>✓</td><td>23.26</td><td>8.60</td><td>4.60</td><td>12.15</td><td>18.93</td><td>6.81</td><td>3.55</td><td>9.77</td></tr><tr><td>✓</td><td>✓</td><td>✓</td><td>32.64</td><td>11.59</td><td>6.15</td><td>16.79</td><td>25.96</td><td>8.78</td><td>4.12</td><td>12.85</td></tr></table>
|
| 189 |
+
|
| 190 |
+
confidence of all positive detections and thus overfits to noise, i.e., failing to handle the extremely challenging task $\mathrm{N} \rightarrow \mathrm{D}$ . 3) EATA still achieves sub-optimal performance while our MonoTTA brings sufficient average performance improvement on both MonoFlex (6.23 mAP) and MonoGround (8.26 mAP), maintaining the best performance in real scenarios. Detailed results are in Appendix $D$ .
|
| 191 |
+
|
| 192 |
+
# 4.3 Application to Instance-Level Inference Method
|
| 193 |
+
|
| 194 |
+
In this section, we seek to investigate whether the proposed MonoTTA can be used to effectively enhance Mono 3Det methods which only process a single image sequentially, i.e., $B = 1$ . To be specific, it may be crucial for Mono 3Det methods to make immediate decisions based on the most recent scene (image) in real-world scenarios like autonomous driving. To this end, it is essential for TTA methods devised for Mono 3Det to allow a single image as input and then conduct model adaptation. To validate it, we integrate MonoTTA into the SOTA Mono 3Det method namely MonoNeRD [40] which accepts one image each time at the test phase. As shown in Table 2, MonoTTA achieves the best or comparable performance across all categories at both severity levels 1 and 2, illustrating the applicability of our method to boost these approaches for handling OOD test data. Detailed results can be also found in Appendix $D$ .
|
| 195 |
+
|
| 196 |
+
# 4.4 Ablation Studies and Quantitative Results
|
| 197 |
+
|
| 198 |
+
Ablation studies. To examine the effectiveness of the losses in MonoTTA, we show the results of the models optimized by different losses. As shown in Table 3, introducing $\mathcal{L}_{AO}$ or $\mathcal{L}_{Nreg}$ enhances the model performance compared to directly applying the pre-trained model (i.e., source only). On the one hand, such a result verifies that our strategy is able to alleviate the score decline issue in test OOD data with high-score objects. On the other hand, introducing $\mathcal{L}_{Nreg}$ could obtain a higher average $AP_{3D|R_{40}}$ value, which also verifies that the negative regularization term is able to enhance the pre-trained model even only
|
| 199 |
+
|
| 200 |
+

|
| 201 |
+
Fig.6: Qualitative results of baselines and the proposed MonoTTA based on MonoGround [29]. We visualize the results on KITTI-C validation set, where predicted cars, pedestrians and cyclists are in lime green, sky blue and yellow, respectively.
|
| 202 |
+
|
| 203 |
+
with low-score objects. When combining the losses (i.e. $\mathcal{L}_{AO}$ , $\mathcal{L}_{Nreg}$ ) together, we obtain the best performance.
|
| 204 |
+
|
| 205 |
+
Quantitative Results. We provide visualizations within Snow, Fog and Frost based on Monoground as shown in Fig. 6. It is evident that the source-only setting suffers severe omissions, while BN adaptation, TENT and EATA alleviate distribution shifts to some degree and give more detections. As for our MonoTTA, it can produce superior detections even in severe conditions, including fewer omissions and accurate detections as highlighted by red circles.
|
| 206 |
+
|
| 207 |
+
# 5 Conclusion
|
| 208 |
+
|
| 209 |
+
In this paper, we propose a monocular test-time adaptation method to improve the pre-trained model on the shifted test data for monocular 3D object detection. Specifically, our method consists of two strategies: 1) Reliability-driven adaptation. To discover more potential objects, we devise a self-adaptive strategy to identify reliable objects for adaptive model adaptation. 2) Noise-guard adaptation. To avoid overfitting to noise and trivial solutions, we devise the negative regularization term to mitigate the negative effects of noisy detections and alleviate distribution shifts. Experiments on KITTI-C and nuScenes datasets demonstrate the effectiveness of MonoTTA in handling fully test-time adaptation for monocular 3D object detection.
|
| 210 |
+
|
| 211 |
+
Future directions. 1) Our work focuses on 2D images, while future studies could explore 3D information in handling distribution shifts. 2) This work explores TTA by assuming one OOD distribution at a time, where the forgetting issue is not severe as the source model weights are recoverable. Exploring scenarios with dynamically OOD distributions offers a compelling future direction, where the forgetting issue would become more severe.
|
| 212 |
+
|
| 213 |
+
# Acknowledgements
|
| 214 |
+
|
| 215 |
+
This work was supported by NSFC with Grant No. 62293482, by the Basic Research Project No. HZQB-KCZYZ-2021067 of Hetao Shenzhen HK S&T Cooperation Zone, by Shenzhen General Program No. JCYJ20220530143600001, by Shenzhen-Hong Kong Joint Funding No. SGDX20211123112401002, by the Shenzhen Outstanding Talents Training Fund 202002, by Guangdong Research Project No. 2017ZT07X152 and No. 2019CX01X104, by the Guangdong Provincial Key Laboratory of Future Networks of Intelligence (Grant No. 2022B1212010001), by the Guangdong Provincial Key Laboratory of Big Data Computing, CHUK-Shenzhen, by the NSFC 61931024&12326610, by the Key Area R&D Program of Guangdong Province with grant No. 2018B030338001, by the Shenzhen Key Laboratory of Big Data and Artificial Intelligence (Grant No. ZDSYS201707251409055), and by Tencent & Huawei Open Fund.
|
| 216 |
+
|
| 217 |
+
# References
|
| 218 |
+
|
| 219 |
+
1. Caesar, H., Bankiti, V., Lang, A.H., Vora, S., Liong, V.E., Xu, Q., Krishnan, A., Pan, Y., Baldan, G., Beijbom, O.: nuscenes: A multimodal dataset for autonomous driving. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 11621-11631 (2020)
|
| 220 |
+
2. Chen, X., Kundu, K., Zhang, Z., Ma, H., Fidler, S., Urtasun, R.: Monocular 3d object detection for autonomous driving. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2147-2156 (2016)
|
| 221 |
+
3. Chen, X., Kundu, K., Zhu, Y., Berneshawi, A.G., Ma, H., Fidler, S., Urtasun, R.: 3d object proposals for accurate object class detection. Advances in neural information processing systems 28 (2015)
|
| 222 |
+
4. Chen, Y., Tai, L., Sun, K., Li, M.: Monopair: Monocular 3d object detection using pairwise spatial relationships. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12093-12102 (2020)
|
| 223 |
+
5. Chen, Y., Liu, J., Zhang, X., Qi, X., Jia, J.: Voxelnext: Fully sparse voxelnet for 3d object detection and tracking. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 21674-21683 (2023)
|
| 224 |
+
6. Ding, M., Huo, Y., Yi, H., Wang, Z., Shi, J., Lu, Z., Luo, P.: Learning depth-guided convolutions for monocular 3d object detection. In: Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition workshops. pp. 1000-1001 (2020)
|
| 225 |
+
7. Duan, K., Bai, S., Xie, L., Qi, H., Huang, Q., Tian, Q.: Centernet: Keypoint triplets for object detection. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 6569-6578 (2019)
|
| 226 |
+
8. Fleuret, F., et al.: Test time adaptation through perturbation robustness. In: NeurIPS 2021 Workshop on Distribution Shifts: Connecting Methods and Applications (2021)
|
| 227 |
+
9. Geiger, A., Lenz, P., Urtasun, R.: Are we ready for autonomous driving? the kitti vision benchmark suite. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3354-3361. IEEE (2012)
|
| 228 |
+
0. Hegde, D., Kilic, V., Sindagi, V., Cooper, A.B., Foster, M., Patel, V.M.: Source-free unsupervised domain adaptation for 3d object detection in adverse weather.
|
| 229 |
+
|
| 230 |
+
In: 2023 IEEE International Conference on Robotics and Automation (ICRA). pp. 6973-6980. IEEE (2023)
|
| 231 |
+
11. Hendrycks, D., Dietterich, T.: Benchmarking neural network robustness to common corruptions and perturbations. In: International Conference on Learning Representations (2018)
|
| 232 |
+
12. Kim, J., Hwang, I., Kim, Y.M.: Ev-tta: Test-time adaptation for event-based object recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17745-17754 (2022)
|
| 233 |
+
13. Kim, Y., Yim, J., Yun, J., Kim, J.: Nlnl: Negative learning for noisy labels. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 101-110 (2019)
|
| 234 |
+
14. Kim, Y., Yun, J., Shon, H., Kim, J.: Joint negative and positive learning for noisy labels. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9442-9451 (2021)
|
| 235 |
+
15. Kumar, A., Brazil, G., Liu, X.: Groomed-nms: Grouped mathematically differentiable nms for monocular 3d object detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 8973-8983 (2021)
|
| 236 |
+
16. Li, P., Chen, X., Shen, S.: Stereo r-cnn based 3d object detection for autonomous driving. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7644-7652 (2019)
|
| 237 |
+
17. Liang, J., Hu, D., Feng, J.: Do we really need to access the source data? source hypothesis transfer for unsupervised domain adaptation. In: International conference on machine learning. pp. 6028-6039. PMLR (2020)
|
| 238 |
+
18. Lin, H., Zhang, Y., Qiu, Z., Niu, S., Gan, C., Liu, Y., Tan, M.: Prototype-guided continual adaptation for class-incremental unsupervised domain adaptation. In: European Conference on Computer Vision. pp. 351-368 (2022)
|
| 239 |
+
19. Liu, Y., Kothari, P., Van Delft, B., Bellot-Gurlet, B., Mordan, T., Alahi, A.: Ttt++: When does self-supervised test-time training fail or thrive? In: Advances in Neural Information Processing Systems. vol. 34, pp. 21808-21820 (2021)
|
| 240 |
+
20. Liu, Z., Wu, Z., Tóth, R.: Smoke: Single-stage monocular 3d object detection via keypoint estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops. pp. 996-997 (2020)
|
| 241 |
+
21. Liu, Z., Tang, H., Amini, A., Yang, X., Mao, H., Rus, D.L., Han, S.: Bevfusion: Multi-task multi-sensor fusion with unified bird's-eye view representation. In: 2023 IEEE international conference on robotics and automation (ICRA). pp. 2774-2781. IEEE (2023)
|
| 242 |
+
22. Luo, Y., Zheng, C., Yan, X., Kun, T., Zheng, C., Cui, S., Li, Z.: Latr: 3d lane detection from monocular images with transformer. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 7941-7952 (2023)
|
| 243 |
+
23. Ma, X., Liu, S., Xia, Z., Zhang, H., Zeng, X., Ouyang, W.: Rethinking pseudo-lidar representation. In: European Conference on Computer Vision (2020)
|
| 244 |
+
24. Mirza, M.J., Soneira, P.J., Lin, W., Kozinski, M., Possegger, H., Bischof, H.: Actmad: Activation matching to align distributions for test-time-training. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 24152-24161 (2023)
|
| 245 |
+
25. Nado, Z., Padhy, S., Sculley, D., D'Amour, A., Lakshminarayanan, B., Snoek, J.: Evaluating prediction-time batch normalization for robustness under covariate shift. arXiv preprint arXiv:2006.10963 (2020)
|
| 246 |
+
26. Niu, S., Wu, J., Zhang, Y., Chen, Y., Zheng, S., Zhao, P., Tan, M.: Efficient test-time model adaptation without forgetting. In: The International Conference on Machine Learning (2022)
|
| 247 |
+
|
| 248 |
+
27. Niu, S., Wu, J., Zhang, Y., Wen, Z., Chen, Y., Zhao, P., Tan, M.: Towards stable test-time adaptation in dynamic wild world. In: International Conference on Learning Representations (2023)
|
| 249 |
+
28. Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J., Chanan, G., Killeen, T., Lin, Z., Gimelshein, N., Antiga, L., et al.: Pytorch: An imperative style, high-performance deep learning library. In: Advances in neural information processing systems. vol. 32 (2019)
|
| 250 |
+
29. Qin, Z., Li, X.: Monoground: Detecting monocular 3d objects from the ground. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3793-3802 (2022)
|
| 251 |
+
30. Qiu, Z., Zhang, Y., Lin, H., et al.: Source-free domain adaptation via avatar prototype generation and adaptation. In: International Joint Conference on Artificial Intelligence (2021)
|
| 252 |
+
31. Reading, C., Harakeh, A., Chae, J., Waslander, S.L.: Categorical depth distribution network for monocular 3d object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8555-8564 (2021)
|
| 253 |
+
32. Saltori, C., Lathuilière, S., Sebe, N., Ricci, E., Galasso, F.: Sf-uda 3d: Source-free unsupervised domain adaptation for lidar-based 3d object detection. In: 2020 International Conference on 3D Vision (3DV). pp. 771–780. IEEE (2020)
|
| 254 |
+
33. Schneider, S., Rusak, E., Eck, L., Bringmann, O., Brendel, W., Bethge, M.: Improving robustness against common corruptions by covariate shift adaptation. In: Advances in neural information processing systems. pp. 11539-11551 (2020)
|
| 255 |
+
34. Sun, Y., Wang, X., Liu, Z., Miller, J., Efros, A., Hardt, M.: Test-time training with self-supervision for generalization under distribution shifts. In: International conference on machine learning. pp. 9229-9248. PMLR (2020)
|
| 256 |
+
35. Veksler, O.: Test time adaptation with regularized loss for weakly supervised salient object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7360-7369 (2023)
|
| 257 |
+
36. Wang, D., Shelhamer, E., Liu, S., Olshausen, B., Darrell, T.: Tent: Fully test-time adaptation by entropy minimization. In: The International Conference on Machine Learning (2021)
|
| 258 |
+
37. Wang, Y., Chao, W.L., Garg, D., Hariharan, B., Campbell, M., Weinberger, K.Q.: Pseudo-lidar from visual depth estimation: Bridging the gap in 3d object detection for autonomous driving. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8445-8453 (2019)
|
| 259 |
+
38. Wu, H., Wen, C., Shi, S., Li, X., Wang, C.: Virtual sparse convolution for multimodal 3d object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 21653-21662 (2023)
|
| 260 |
+
39. Xu, B., Chen, Z.: Multi-level fusion based 3d object detection from monocular images. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 2345-2353 (2018)
|
| 261 |
+
40. Xu, J., Peng, L., Cheng, H., Li, H., Qian, W., Li, K., Wang, W., Cai, D.: Mononerd: Nerf-like representations for monocular 3d object detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 6814-6824 (2023)
|
| 262 |
+
41. Yang, J., Shi, S., Wang, Z., Li, H., Qi, X.: St3d: Self-training for unsupervised domain adaptation on 3d object detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10368-10378 (2021)
|
| 263 |
+
42. Ye, X., Shu, M., Li, H., Shi, Y., Li, Y., Wang, G., Tan, X., Ding, E.: Rope3d: The roadside perception dataset for autonomous driving and monocular 3d object detection task. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 21341-21350 (2022)
|
| 264 |
+
|
| 265 |
+
43. Zhang, M., Levine, S., Finn, C.: Memo: Test time robustness via adaptation and augmentation. In: Advances in Neural Information Processing Systems. vol. 35, pp. 38629-38642 (2022)
|
| 266 |
+
44. Zhang, Y., Hooi, B., Hong, L., Feng, J.: Self-supervised aggregation of diverse experts for test-agnostic long-tailed recognition. In: Advances in Neural Information Processing Systems. vol. 35, pp. 34077-34090 (2022)
|
| 267 |
+
45. Zhang, Y., Hooi, B., Hu, D., Liang, J., Feng, J.: Unleashing the power of contrastive self-supervised visual models via contrast-regularized fine-tuning. In: Advances in Neural Information Processing Systems. vol. 34, pp. 29848-29860 (2021)
|
| 268 |
+
46. Zhang, Y., Kang, B., Hooi, B., Yan, S., Feng, J.: Deep long-tailed learning: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence 45(9), 10795-10816 (2023)
|
| 269 |
+
47. Zhang, Y., Wei, Y., Wu, Q., Zhao, P., Niu, S., Huang, J., Tan, M.: Collaborative unsupervised domain adaptation for medical image diagnosis. IEEE Transactions on Image Processing 29, 7834-7844 (2020)
|
| 270 |
+
48. Zhang, Y., Lu, J., Zhou, J.: Objects are different: Flexible monocular 3d object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3289-3298 (2021)
|
| 271 |
+
49. Zou, Z., Ye, X., Du, L., Cheng, X., Tan, X., Zhang, L., Feng, J., Xue, X., Ding, E.: The devil is in the task: Exploiting reciprocal appearance-localization features for monocular 3d object detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 2713-2722 (2021)
|
2024/MonoTTA_ Fully Test-Time Adaptation for Monocular 3D Object Detection/images.zip
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:852fa5ab2da8f6ead4b8949fd580dbd1767ff504f560e973a58ca04e21177342
|
| 3 |
+
size 564881
|
2024/MonoTTA_ Fully Test-Time Adaptation for Monocular 3D Object Detection/layout.json
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
2024/MonoWAD_ Weather-Adaptive Diffusion Model for Robust Monocular 3D Object Detection/4d73977a-6d19-47be-84e6-b3aec99a372e_content_list.json
ADDED
|
@@ -0,0 +1,1866 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
{
|
| 3 |
+
"type": "text",
|
| 4 |
+
"text": "MonoWAD: Weather-Adaptive Diffusion Model for Robust Monocular 3D Object Detection",
|
| 5 |
+
"text_level": 1,
|
| 6 |
+
"bbox": [
|
| 7 |
+
228,
|
| 8 |
+
140,
|
| 9 |
+
774,
|
| 10 |
+
186
|
| 11 |
+
],
|
| 12 |
+
"page_idx": 0
|
| 13 |
+
},
|
| 14 |
+
{
|
| 15 |
+
"type": "text",
|
| 16 |
+
"text": "Youngmin Oh<sup>1</sup>, Hyung-Il Kim<sup>2</sup>, Seong Tae Kim<sup>1†</sup>, and Jung Uk Kim<sup>1†</sup>",
|
| 17 |
+
"bbox": [
|
| 18 |
+
222,
|
| 19 |
+
210,
|
| 20 |
+
781,
|
| 21 |
+
228
|
| 22 |
+
],
|
| 23 |
+
"page_idx": 0
|
| 24 |
+
},
|
| 25 |
+
{
|
| 26 |
+
"type": "text",
|
| 27 |
+
"text": "1 Kyung Hee University, Yong-in, South Korea {oym9104, st.kim, ju.kim}@khu.ac.kr 2 ETRI, Daejeon, South Korea hikim@etri.re.kr",
|
| 28 |
+
"bbox": [
|
| 29 |
+
343,
|
| 30 |
+
238,
|
| 31 |
+
658,
|
| 32 |
+
295
|
| 33 |
+
],
|
| 34 |
+
"page_idx": 0
|
| 35 |
+
},
|
| 36 |
+
{
|
| 37 |
+
"type": "text",
|
| 38 |
+
"text": "Abstract. Monocular 3D object detection is an important challenging task in autonomous driving. Existing methods mainly focus on performing 3D detection in ideal weather conditions, characterized by scenarios with clear and optimal visibility. However, the challenge of autonomous driving requires the ability to handle changes in weather conditions, such as foggy weather, not just clear weather. We introduce MonoWAD, a novel weather-robust monocular 3D object detector with a weather-adaptive diffusion model. It contains two components: (1) the weather codebook to memorize the knowledge of the clear weather and generate a weather-reference feature for any input, and (2) the weather-adaptive diffusion model to enhance the feature representation of the input feature by incorporating a weather-reference feature. This serves as an attention role in indicating how much improvement is needed for the input feature according to the weather conditions. To achieve this goal, we introduce a weather-adaptive enhancement loss to enhance the feature representation under both clear and foggy weather conditions. Extensive experiments under various weather conditions demonstrate that MonoWAD achieves weather-robust monocular 3D object detection. The code and dataset are released at https://github.com/VisualAIKHU/MonoWAD.",
|
| 39 |
+
"bbox": [
|
| 40 |
+
261,
|
| 41 |
+
332,
|
| 42 |
+
743,
|
| 43 |
+
597
|
| 44 |
+
],
|
| 45 |
+
"page_idx": 0
|
| 46 |
+
},
|
| 47 |
+
{
|
| 48 |
+
"type": "text",
|
| 49 |
+
"text": "Keywords: Monocular 3D Object Detection $\\cdot$ Weather-Adaptive Diffusion $\\cdot$ Weather Codebook",
|
| 50 |
+
"bbox": [
|
| 51 |
+
259,
|
| 52 |
+
609,
|
| 53 |
+
740,
|
| 54 |
+
638
|
| 55 |
+
],
|
| 56 |
+
"page_idx": 0
|
| 57 |
+
},
|
| 58 |
+
{
|
| 59 |
+
"type": "text",
|
| 60 |
+
"text": "1 Introduction",
|
| 61 |
+
"text_level": 1,
|
| 62 |
+
"bbox": [
|
| 63 |
+
215,
|
| 64 |
+
662,
|
| 65 |
+
375,
|
| 66 |
+
679
|
| 67 |
+
],
|
| 68 |
+
"page_idx": 0
|
| 69 |
+
},
|
| 70 |
+
{
|
| 71 |
+
"type": "text",
|
| 72 |
+
"text": "Monocular 3D object detection aims to detect 3D objects using only a single camera [16, 18, 28, 46, 55, 57]. In contrast to LiDAR-based methods that rely on expensive LiDAR sensors for depth estimation [24, 44, 45, 56], and stereo-based methods that require synchronized stereo cameras, monocular 3D object detection only requires monocular images, offering the advantage of computational cost-effectiveness and requiring fewer resources. Due to this characteristic, the monocular 3D object detection is applied to a wide range of real-world applications, such as autonomous driving [12, 19, 53] and robotics [50].",
|
| 73 |
+
"bbox": [
|
| 74 |
+
212,
|
| 75 |
+
694,
|
| 76 |
+
787,
|
| 77 |
+
816
|
| 78 |
+
],
|
| 79 |
+
"page_idx": 0
|
| 80 |
+
},
|
| 81 |
+
{
|
| 82 |
+
"type": "page_footnote",
|
| 83 |
+
"text": "† Corresponding author",
|
| 84 |
+
"bbox": [
|
| 85 |
+
217,
|
| 86 |
+
824,
|
| 87 |
+
382,
|
| 88 |
+
840
|
| 89 |
+
],
|
| 90 |
+
"page_idx": 0
|
| 91 |
+
},
|
| 92 |
+
{
|
| 93 |
+
"type": "image",
|
| 94 |
+
"img_path": "images/4096afacdb8773efa89658e3f125d0abdc0e7e5a72b2515d361cce17f37609ca.jpg",
|
| 95 |
+
"image_caption": [
|
| 96 |
+
"Fig. 1: Conceptual diagram of the proposed method (foggy example). (a) In the training phase, weather codebook learns the clear knowledge to transfer it to weather-adaptive diffusion model to enhance content related to the weather conditions. (b) By doing so, even with input images under various weather conditions (e.g., foggy images), monocular 3D object detection becomes adaptable to various weather scenarios."
|
| 97 |
+
],
|
| 98 |
+
"image_footnote": [],
|
| 99 |
+
"bbox": [
|
| 100 |
+
210,
|
| 101 |
+
148,
|
| 102 |
+
792,
|
| 103 |
+
255
|
| 104 |
+
],
|
| 105 |
+
"page_idx": 1
|
| 106 |
+
},
|
| 107 |
+
{
|
| 108 |
+
"type": "text",
|
| 109 |
+
"text": "However, existing monocular 3D object detectors mainly focus on ideal autonomous driving environments (i.e., clear weather). There are challenges in applying them to real-world scenarios with adverse weather conditions, such as fog and rain. Among these, fog poses the most significant challenge compared to other weather [14, 17]. This is due to the dense and diffuse nature of fog, which strongly scatters and absorbs light, leading to difficulties in object detection [2]. Since monocular 3D object detection relies solely on visual information from a single image, unlike LiDAR, it is crucial to design detectors to achieve enhanced performance in challenging visibility scenarios.",
|
| 110 |
+
"bbox": [
|
| 111 |
+
212,
|
| 112 |
+
340,
|
| 113 |
+
787,
|
| 114 |
+
477
|
| 115 |
+
],
|
| 116 |
+
"page_idx": 1
|
| 117 |
+
},
|
| 118 |
+
{
|
| 119 |
+
"type": "text",
|
| 120 |
+
"text": "In this paper, we propose MonoWAD, a novel weather-robust monocular 3D object detector to address the aforementioned issues. As mentioned earlier, due to the inherent challenges posed by foggy weather among various adverse weather conditions, we focus primarily on clear and foggy weather (results for other weather conditions such as rainy and sunset are presented in Section 4). For weather-robust object detection, clear weather requires relatively modest improvements to enhance the visual representation of features. In contrast, significant enhancements in this feature representation are required for foggy weather. To address this, we consider two key aspects: (1) how to quantify the degree of improvement needed for the input image, and (2) how to guide the representation of the input image.",
|
| 121 |
+
"bbox": [
|
| 122 |
+
212,
|
| 123 |
+
478,
|
| 124 |
+
787,
|
| 125 |
+
643
|
| 126 |
+
],
|
| 127 |
+
"page_idx": 1
|
| 128 |
+
},
|
| 129 |
+
{
|
| 130 |
+
"type": "text",
|
| 131 |
+
"text": "To address the two key aspects, as shown in Fig. 1, our MonoWAD consists of a weather codebook and weather-adaptive diffusion model. First, we introduce the weather codebook to generate a weather-reference feature that contains knowledge about reference weather in a given scene. The reference weather acts as a guide, indicating the degree of weather improvement required. Since the clear weather contains a richer visual representation of objects, we adopt it as a reference weather. At this time, we devise a clear knowledge recalling (CKR) loss to guide the weather codebook to memorize information about clear weather and generate a weather-reference feature for any input (clear or foggy). As a result, our detector can understand where improvements are needed in the input features based on the weather-reference feature.",
|
| 132 |
+
"bbox": [
|
| 133 |
+
212,
|
| 134 |
+
643,
|
| 135 |
+
787,
|
| 136 |
+
808
|
| 137 |
+
],
|
| 138 |
+
"page_idx": 1
|
| 139 |
+
},
|
| 140 |
+
{
|
| 141 |
+
"type": "text",
|
| 142 |
+
"text": "Second, we propose a weather-adaptive diffusion model to effectively enhance feature representations in accordance with weather conditions. Given input fea-",
|
| 143 |
+
"bbox": [
|
| 144 |
+
212,
|
| 145 |
+
809,
|
| 146 |
+
787,
|
| 147 |
+
840
|
| 148 |
+
],
|
| 149 |
+
"page_idx": 1
|
| 150 |
+
},
|
| 151 |
+
{
|
| 152 |
+
"type": "page_number",
|
| 153 |
+
"text": "2",
|
| 154 |
+
"bbox": [
|
| 155 |
+
217,
|
| 156 |
+
114,
|
| 157 |
+
228,
|
| 158 |
+
126
|
| 159 |
+
],
|
| 160 |
+
"page_idx": 1
|
| 161 |
+
},
|
| 162 |
+
{
|
| 163 |
+
"type": "header",
|
| 164 |
+
"text": "Y. Oh et al.",
|
| 165 |
+
"bbox": [
|
| 166 |
+
271,
|
| 167 |
+
114,
|
| 168 |
+
354,
|
| 169 |
+
127
|
| 170 |
+
],
|
| 171 |
+
"page_idx": 1
|
| 172 |
+
},
|
| 173 |
+
{
|
| 174 |
+
"type": "text",
|
| 175 |
+
"text": "ture (clear or foggy), the weather-adaptive diffusion model dynamically enhances the representation of the input feature based on the weather-reference feature. The weather-reference feature plays a role of attention, determining the extent to which the input features need improvement. At this time, we define the difference between clear and foggy weather (i.e., weather changes) as the fog distribution to adopt it as the noise for our diffusion model. With fog distribution, our weather-adaptive diffusion model can enhance the feature representation according to the weather conditions through multiple steps of reverse processes. To achieve this goal, we introduce a weather-adaptive enhancement (WAE) loss. As a result, our MonoWAD performs weather-robust detection by adaptively improving feature representation according to the weather conditions.",
|
| 176 |
+
"bbox": [
|
| 177 |
+
212,
|
| 178 |
+
146,
|
| 179 |
+
782,
|
| 180 |
+
311
|
| 181 |
+
],
|
| 182 |
+
"page_idx": 2
|
| 183 |
+
},
|
| 184 |
+
{
|
| 185 |
+
"type": "text",
|
| 186 |
+
"text": "To adaptively enhance the feature representation through the difference between weather conditions, we generate a new foggy KITTI dataset based on the KITTI dataset [12]. Comprehensive experimental results on several datasets [11, 12] show that our MonoWAD outperforms the existing state-of-the-art monocular 3D object detectors [16,28,37,40,55] under foggy weather, which is the most challenging weather condition. While our method primarily focuses on foggy, experiments conducted under various weather conditions (e.g., foggy, rainy, and sunset) have demonstrated its applicability to other weather scenarios.",
|
| 187 |
+
"bbox": [
|
| 188 |
+
212,
|
| 189 |
+
313,
|
| 190 |
+
784,
|
| 191 |
+
431
|
| 192 |
+
],
|
| 193 |
+
"page_idx": 2
|
| 194 |
+
},
|
| 195 |
+
{
|
| 196 |
+
"type": "text",
|
| 197 |
+
"text": "The main contributions of our paper can be summarized as follows:",
|
| 198 |
+
"bbox": [
|
| 199 |
+
238,
|
| 200 |
+
433,
|
| 201 |
+
725,
|
| 202 |
+
448
|
| 203 |
+
],
|
| 204 |
+
"page_idx": 2
|
| 205 |
+
},
|
| 206 |
+
{
|
| 207 |
+
"type": "list",
|
| 208 |
+
"sub_type": "text",
|
| 209 |
+
"list_items": [
|
| 210 |
+
"- We introduce a new weather-robust monocular 3D object detector, called MonoWAD, that is robust to various weather conditions.",
|
| 211 |
+
"- We design a weather codebook with clear knowledge recalling loss for learning about clear weather, providing reference information for enhancement.",
|
| 212 |
+
"- We propose weather-adaptive diffusion model with weather-adaptive enhancement loss to dynamically enhance the feature representation of the input images according to the weather conditions."
|
| 213 |
+
],
|
| 214 |
+
"bbox": [
|
| 215 |
+
225,
|
| 216 |
+
458,
|
| 217 |
+
782,
|
| 218 |
+
561
|
| 219 |
+
],
|
| 220 |
+
"page_idx": 2
|
| 221 |
+
},
|
| 222 |
+
{
|
| 223 |
+
"type": "text",
|
| 224 |
+
"text": "2 Related Work",
|
| 225 |
+
"text_level": 1,
|
| 226 |
+
"bbox": [
|
| 227 |
+
215,
|
| 228 |
+
585,
|
| 229 |
+
387,
|
| 230 |
+
602
|
| 231 |
+
],
|
| 232 |
+
"page_idx": 2
|
| 233 |
+
},
|
| 234 |
+
{
|
| 235 |
+
"type": "text",
|
| 236 |
+
"text": "2.1 Monocular 3D Object Detection",
|
| 237 |
+
"text_level": 1,
|
| 238 |
+
"bbox": [
|
| 239 |
+
215,
|
| 240 |
+
618,
|
| 241 |
+
527,
|
| 242 |
+
633
|
| 243 |
+
],
|
| 244 |
+
"page_idx": 2
|
| 245 |
+
},
|
| 246 |
+
{
|
| 247 |
+
"type": "text",
|
| 248 |
+
"text": "Monocular 3D object detection task can be categorized into two directions according to the type of data used in the training phase: (1) using only a monocular image and (2) incorporating additional data, such as depth along with a monocular image. The first category relies on the geometric relationship between 2D and 3D [3,7,25,26,28,31,34,40,46,55,57]. For example, Deep3Dbox [34] utilizes the geometric information of 2D bounding boxes to predict 3D bounding boxes. In [3], M3D-RPN was proposed to understand the 3D scene from the depth-aware convolution. MonoRCNN [46] predicted 3D bounding boxes through geometry-based distance decomposition, and MonoCon [25] learns mono context for 3D object detection. MonoDETR [55] introduces a depth-guided transformer that utilizes geometric depth cues without requiring additional data.",
|
| 249 |
+
"bbox": [
|
| 250 |
+
212,
|
| 251 |
+
643,
|
| 252 |
+
784,
|
| 253 |
+
808
|
| 254 |
+
],
|
| 255 |
+
"page_idx": 2
|
| 256 |
+
},
|
| 257 |
+
{
|
| 258 |
+
"type": "text",
|
| 259 |
+
"text": "Moreover, since monocular image contains limited information for estimating 3D object cues, monocular 3D object detectors have adopted the additional",
|
| 260 |
+
"bbox": [
|
| 261 |
+
214,
|
| 262 |
+
810,
|
| 263 |
+
784,
|
| 264 |
+
839
|
| 265 |
+
],
|
| 266 |
+
"page_idx": 2
|
| 267 |
+
},
|
| 268 |
+
{
|
| 269 |
+
"type": "header",
|
| 270 |
+
"text": "MonoWAD",
|
| 271 |
+
"bbox": [
|
| 272 |
+
653,
|
| 273 |
+
114,
|
| 274 |
+
730,
|
| 275 |
+
126
|
| 276 |
+
],
|
| 277 |
+
"page_idx": 2
|
| 278 |
+
},
|
| 279 |
+
{
|
| 280 |
+
"type": "page_number",
|
| 281 |
+
"text": "3",
|
| 282 |
+
"bbox": [
|
| 283 |
+
774,
|
| 284 |
+
116,
|
| 285 |
+
784,
|
| 286 |
+
126
|
| 287 |
+
],
|
| 288 |
+
"page_idx": 2
|
| 289 |
+
},
|
| 290 |
+
{
|
| 291 |
+
"type": "text",
|
| 292 |
+
"text": "data for more robust detection [4,8,16,27,32,37,41,48]. For example, the depth-conditioned dynamic message propagation (DDMP) [48] was proposed to integrate prior depth information with the image context. CaDDN [41] was introduced to utilize the depth distribution of predicted categories for each pixel to project the context information onto 3D space, deriving 3D bounding boxes. MonoDTR [16] employs the transformer architecture to integrate depth features and context, thus estimating more accurate depth information.",
|
| 293 |
+
"bbox": [
|
| 294 |
+
212,
|
| 295 |
+
146,
|
| 296 |
+
787,
|
| 297 |
+
252
|
| 298 |
+
],
|
| 299 |
+
"page_idx": 3
|
| 300 |
+
},
|
| 301 |
+
{
|
| 302 |
+
"type": "text",
|
| 303 |
+
"text": "Despite recent progress, the existing monocular 3D object detectors mainly rely on the benchmark data collected under clear weather conditions. However, it is essential to account for challenging adverse weather conditions, such as fog, to more accurately reflect real-world scenarios. In this paper, we aim to introduce weather-robust 3D object detection by enhancing visual features with the proposed weather-adaptive diffusion model and weather codebook.",
|
| 304 |
+
"bbox": [
|
| 305 |
+
212,
|
| 306 |
+
252,
|
| 307 |
+
787,
|
| 308 |
+
343
|
| 309 |
+
],
|
| 310 |
+
"page_idx": 3
|
| 311 |
+
},
|
| 312 |
+
{
|
| 313 |
+
"type": "text",
|
| 314 |
+
"text": "2.2 Computer Vision Tasks for Foggy Weather",
|
| 315 |
+
"text_level": 1,
|
| 316 |
+
"bbox": [
|
| 317 |
+
214,
|
| 318 |
+
354,
|
| 319 |
+
614,
|
| 320 |
+
371
|
| 321 |
+
],
|
| 322 |
+
"page_idx": 3
|
| 323 |
+
},
|
| 324 |
+
{
|
| 325 |
+
"type": "text",
|
| 326 |
+
"text": "There have been a lot of studies on improving performance in various weather conditions for real-world application of computer vision technology [2, 13, 20, 21, 33, 38, 43, 51, 54]. In particular, fog is considered one of the most critical issues due to its significant degradation of visual information compared to other weather conditions. To deal with the foggy weather conditions, the authors in [43] generated synthetic fog images (so-called Foggy Cityscapes) from clear weather images, and utilized them for training semantic segmentation and 2D object detection. Similar to Foggy Cityscapes, Martin et al. [13] naturally synthesizes fog into LIDAR to enhance the performance of LiDAR-based 3D object detectors in foggy weather. Bijelic et al. [2] use real dataset including foggy weather, for 2D detection with the multimodal fusion networks. Mai et al. [33] synthesize fog for LiDAR and stereo images and perform fusion-based 3D object detection using the SLS-Fusion network. Xin et al. [51] focused on 2D detection in foggy weather by applying domain adaptation. In this context, we focus on monocular 3D object detection in scenarios that rely solely on visual information in challenging foggy environments. To address this challenge, we propose a weather-robust diffusion model that dynamically improves features based on reference feature.",
|
| 327 |
+
"bbox": [
|
| 328 |
+
212,
|
| 329 |
+
380,
|
| 330 |
+
787,
|
| 331 |
+
636
|
| 332 |
+
],
|
| 333 |
+
"page_idx": 3
|
| 334 |
+
},
|
| 335 |
+
{
|
| 336 |
+
"type": "text",
|
| 337 |
+
"text": "2.3 Diffusion Models",
|
| 338 |
+
"text_level": 1,
|
| 339 |
+
"bbox": [
|
| 340 |
+
214,
|
| 341 |
+
648,
|
| 342 |
+
401,
|
| 343 |
+
662
|
| 344 |
+
],
|
| 345 |
+
"page_idx": 3
|
| 346 |
+
},
|
| 347 |
+
{
|
| 348 |
+
"type": "text",
|
| 349 |
+
"text": "Recently, the diffusion model [15] has attracted considerable interest in computer vision due to its impressive progress in image generation [22, 29, 35, 42], as well as its potential application to other vision tasks such as segmentation [1] and image captioning [30]. Inspired by the remarkable generative ability, we design the diffusion model for robust monocular 3D object detection by considering a foggy effect (which is one of the challenging adverse weather conditions for monocular 3D object detection) as a form of noise in the model. That is, we propose a method in which visual features obscured by fog are progressively improved by training a diffusion model based on the forward/reverse diffusion process. In particular, we present an adaptive method that allows the diffusion model to control the degree of improvement by weather conditions.",
|
| 350 |
+
"bbox": [
|
| 351 |
+
212,
|
| 352 |
+
672,
|
| 353 |
+
787,
|
| 354 |
+
840
|
| 355 |
+
],
|
| 356 |
+
"page_idx": 3
|
| 357 |
+
},
|
| 358 |
+
{
|
| 359 |
+
"type": "page_number",
|
| 360 |
+
"text": "4",
|
| 361 |
+
"bbox": [
|
| 362 |
+
217,
|
| 363 |
+
114,
|
| 364 |
+
228,
|
| 365 |
+
126
|
| 366 |
+
],
|
| 367 |
+
"page_idx": 3
|
| 368 |
+
},
|
| 369 |
+
{
|
| 370 |
+
"type": "header",
|
| 371 |
+
"text": "Y. Oh et al.",
|
| 372 |
+
"bbox": [
|
| 373 |
+
271,
|
| 374 |
+
114,
|
| 375 |
+
354,
|
| 376 |
+
127
|
| 377 |
+
],
|
| 378 |
+
"page_idx": 3
|
| 379 |
+
},
|
| 380 |
+
{
|
| 381 |
+
"type": "image",
|
| 382 |
+
"img_path": "images/2105952e910218bb8044c128b12c704c3f65278f2fb9df10aeaedb76fc42a5ae.jpg",
|
| 383 |
+
"image_caption": [
|
| 384 |
+
"Fig. 2: Overview of our MonoWAD in the inference phase. It mainly contains three parts: weather codebook, weather-adaptive diffusion model, and detection block. Through the weather codebook and weather-adaptive diffusion model, our method can maintain robustness against various weather conditions (i.e., clear or foggy)."
|
| 385 |
+
],
|
| 386 |
+
"image_footnote": [],
|
| 387 |
+
"bbox": [
|
| 388 |
+
218,
|
| 389 |
+
146,
|
| 390 |
+
782,
|
| 391 |
+
268
|
| 392 |
+
],
|
| 393 |
+
"page_idx": 4
|
| 394 |
+
},
|
| 395 |
+
{
|
| 396 |
+
"type": "text",
|
| 397 |
+
"text": "3 Proposed Method",
|
| 398 |
+
"text_level": 1,
|
| 399 |
+
"bbox": [
|
| 400 |
+
215,
|
| 401 |
+
347,
|
| 402 |
+
426,
|
| 403 |
+
364
|
| 404 |
+
],
|
| 405 |
+
"page_idx": 4
|
| 406 |
+
},
|
| 407 |
+
{
|
| 408 |
+
"type": "text",
|
| 409 |
+
"text": "Fig. 2 shows the overall framework of the proposed MonoWAD in the inference phase. A backbone network receives an input image (clear image $I_{c}$ or foggy image $I_{f}$ ) to encode the corresponding input feature (input clear feature $x^{c}$ or input foggy feature $x^{f}$ ). It interacts with the weather codebook $\\mathcal{Z}$ to generate weather-reference feature $x^{r}$ , indicating the amount of enhancement for the given input feature. Subsequently, the weather-adaptive diffusion model attempts to enhance the input feature over $T$ timesteps to obtain an enhanced feature $\\tilde{x}^{c}$ or $\\tilde{x}^{f}$ . Finally, monocular 3D object detection is performed through the detection block. Note that, in the training phase, our MonoWAD use the clear feature $x^{c}$ to train our diffusion as well as the detection block.",
|
| 410 |
+
"bbox": [
|
| 411 |
+
212,
|
| 412 |
+
383,
|
| 413 |
+
787,
|
| 414 |
+
532
|
| 415 |
+
],
|
| 416 |
+
"page_idx": 4
|
| 417 |
+
},
|
| 418 |
+
{
|
| 419 |
+
"type": "text",
|
| 420 |
+
"text": "We address two key issues: (1) how to guide weather-reference feature $x^r$ to serve as a reference feature, and (2) how to guide the weather-adaptive diffusion model to effectively enhance feature representations based on weather conditions. Details are in the following subsections.",
|
| 421 |
+
"bbox": [
|
| 422 |
+
212,
|
| 423 |
+
535,
|
| 424 |
+
787,
|
| 425 |
+
595
|
| 426 |
+
],
|
| 427 |
+
"page_idx": 4
|
| 428 |
+
},
|
| 429 |
+
{
|
| 430 |
+
"type": "text",
|
| 431 |
+
"text": "3.1 Weather Codebook",
|
| 432 |
+
"text_level": 1,
|
| 433 |
+
"bbox": [
|
| 434 |
+
215,
|
| 435 |
+
613,
|
| 436 |
+
421,
|
| 437 |
+
627
|
| 438 |
+
],
|
| 439 |
+
"page_idx": 4
|
| 440 |
+
},
|
| 441 |
+
{
|
| 442 |
+
"type": "text",
|
| 443 |
+
"text": "In foggy weather conditions, the overall visual quality of the scene is generally poor, requiring a significant enhancement. Conversely, in clear weather, the amount of improvement is expected to be relatively minimal compared to foggy conditions. Therefore, inspired by [9,36], we devise a weather codebook $\\mathcal{Z}$ to provide the reference knowledge about the weather for appropriate enhancement based on the weather conditions. At this time, as the clear weather contains abundant visual representations, we use it as a reference weather knowledge.",
|
| 444 |
+
"bbox": [
|
| 445 |
+
212,
|
| 446 |
+
642,
|
| 447 |
+
787,
|
| 448 |
+
747
|
| 449 |
+
],
|
| 450 |
+
"page_idx": 4
|
| 451 |
+
},
|
| 452 |
+
{
|
| 453 |
+
"type": "text",
|
| 454 |
+
"text": "As shown in Fig. 3, the reference knowledge embedding procedure involves receiving paired clear-foggy features during the training phase. The weather codebook $\\mathcal{Z}$ consists of $K$ learnable slots, denoted as $\\mathcal{Z} = \\{z_k\\}_{k=1}^K (z_k \\in \\mathbb{R}^{1 \\times c})$ , where $c$ represents the dimensionality of each slot. The paired clear feature $x^c$ and foggy feature $x^f$ pass through a convolution layer to generate $\\hat{x}^c \\in \\mathbb{R}^{h \\times w \\times c}$ and $\\hat{x}^f \\in \\mathbb{R}^{h \\times w \\times c}$ ( $w$ denotes width and $h$ indicates height). Each element of the",
|
| 455 |
+
"bbox": [
|
| 456 |
+
212,
|
| 457 |
+
750,
|
| 458 |
+
787,
|
| 459 |
+
840
|
| 460 |
+
],
|
| 461 |
+
"page_idx": 4
|
| 462 |
+
},
|
| 463 |
+
{
|
| 464 |
+
"type": "header",
|
| 465 |
+
"text": "MonoWAD",
|
| 466 |
+
"bbox": [
|
| 467 |
+
653,
|
| 468 |
+
114,
|
| 469 |
+
730,
|
| 470 |
+
126
|
| 471 |
+
],
|
| 472 |
+
"page_idx": 4
|
| 473 |
+
},
|
| 474 |
+
{
|
| 475 |
+
"type": "page_number",
|
| 476 |
+
"text": "5",
|
| 477 |
+
"bbox": [
|
| 478 |
+
774,
|
| 479 |
+
116,
|
| 480 |
+
784,
|
| 481 |
+
126
|
| 482 |
+
],
|
| 483 |
+
"page_idx": 4
|
| 484 |
+
},
|
| 485 |
+
{
|
| 486 |
+
"type": "image",
|
| 487 |
+
"img_path": "images/a461609d2f407e75e3eec09b04ff65dc201e192f796c9785c68517f768250a5b.jpg",
|
| 488 |
+
"image_caption": [
|
| 489 |
+
"Fig. 3: Illustration of the proposed (a) weather-invariant guiding (WIG) loss and (b) clear knowledge embedding (CKE) loss. The clear knowledge recalling (CKR) loss, obtained from combining WIG and CKE, aims to memorize the knowledge of the clear weather and recall the same clear knowledge from the foggy weather."
|
| 490 |
+
],
|
| 491 |
+
"image_footnote": [],
|
| 492 |
+
"bbox": [
|
| 493 |
+
212,
|
| 494 |
+
146,
|
| 495 |
+
787,
|
| 496 |
+
342
|
| 497 |
+
],
|
| 498 |
+
"page_idx": 5
|
| 499 |
+
},
|
| 500 |
+
{
|
| 501 |
+
"type": "text",
|
| 502 |
+
"text": "feature denoted as $\\hat{x}_{ij}^{c}\\in \\mathbb{R}^{1\\times c}$ and $\\hat{x}_{ij}^{f}\\in \\mathbb{R}^{1\\times c}$ . Subsequently, we obtain weather-reference feature for clear weather, $x^{r(c)}\\in \\mathbb{R}^{h\\times w\\times c}$ , by conducting element-wise quantization process $\\mathbf{q}(\\cdot)$ , calculated as:",
|
| 503 |
+
"bbox": [
|
| 504 |
+
212,
|
| 505 |
+
412,
|
| 506 |
+
787,
|
| 507 |
+
462
|
| 508 |
+
],
|
| 509 |
+
"page_idx": 5
|
| 510 |
+
},
|
| 511 |
+
{
|
| 512 |
+
"type": "equation",
|
| 513 |
+
"text": "\n$$\nx ^ {r (c)} = \\mathbf {q} (\\hat {x} ^ {c}) := \\left(\\underset {z _ {k} \\in \\mathcal {Z}} {\\arg \\min } \\| \\hat {x} _ {i j} ^ {c} - z _ {k} \\|\\right). \\tag {1}\n$$\n",
|
| 514 |
+
"text_format": "latex",
|
| 515 |
+
"bbox": [
|
| 516 |
+
357,
|
| 517 |
+
472,
|
| 518 |
+
785,
|
| 519 |
+
506
|
| 520 |
+
],
|
| 521 |
+
"page_idx": 5
|
| 522 |
+
},
|
| 523 |
+
{
|
| 524 |
+
"type": "text",
|
| 525 |
+
"text": "Utilizing $x^{c}$ and $x^{r(c)}$ , we introduce a clear knowledge embedding (CKE) loss $\\mathcal{L}_{cke}$ to guide $x^{r(c)}$ to follow the representation of $x^{c}$ . To this end, we perform global average pooling (GAP) for $x^{c}$ and $x^{r(c)}$ with softmax, generating $s^{c}$ and $s^{r(c)}$ , respectively. Each element in the vector indicates the probability of the significance of each channel. With $s^{c}$ and $s^{r(c)}$ , we employ KL divergence $D_{KL}(\\cdot)$ for $\\mathcal{L}_{cke}$ to compare the probability distributions, formulated as:",
|
| 526 |
+
"bbox": [
|
| 527 |
+
214,
|
| 528 |
+
518,
|
| 529 |
+
787,
|
| 530 |
+
611
|
| 531 |
+
],
|
| 532 |
+
"page_idx": 5
|
| 533 |
+
},
|
| 534 |
+
{
|
| 535 |
+
"type": "equation",
|
| 536 |
+
"text": "\n$$\n\\mathcal {L} _ {c k e} = D _ {K L} \\left(s ^ {c} \\| s ^ {r (c)}\\right). \\tag {2}\n$$\n",
|
| 537 |
+
"text_format": "latex",
|
| 538 |
+
"bbox": [
|
| 539 |
+
416,
|
| 540 |
+
619,
|
| 541 |
+
785,
|
| 542 |
+
640
|
| 543 |
+
],
|
| 544 |
+
"page_idx": 5
|
| 545 |
+
},
|
| 546 |
+
{
|
| 547 |
+
"type": "text",
|
| 548 |
+
"text": "Through $\\mathcal{L}_{cke}$ , the weather codebook $\\mathcal{Z}$ can memorize the knowledge of the clear weather, allowing it to effectively reconstruct the clear weather knowledge.",
|
| 549 |
+
"bbox": [
|
| 550 |
+
212,
|
| 551 |
+
650,
|
| 552 |
+
785,
|
| 553 |
+
680
|
| 554 |
+
],
|
| 555 |
+
"page_idx": 5
|
| 556 |
+
},
|
| 557 |
+
{
|
| 558 |
+
"type": "text",
|
| 559 |
+
"text": "Additionally, as the paired clear-foggy images are identical except for weather conditions, the quantization process of the foggy feature $x^{f}$ with the weather codebook should generate an equivalent weather-reference feature. For obtaining weather-reference feature for foggy $x^{r(f)}$ , the element-wise quantization process is also conducted for $\\hat{x}^{f}$ and $\\mathcal{Z}$ :",
|
| 560 |
+
"bbox": [
|
| 561 |
+
212,
|
| 562 |
+
681,
|
| 563 |
+
787,
|
| 564 |
+
755
|
| 565 |
+
],
|
| 566 |
+
"page_idx": 5
|
| 567 |
+
},
|
| 568 |
+
{
|
| 569 |
+
"type": "equation",
|
| 570 |
+
"text": "\n$$\nx ^ {r (f)} = \\mathbf {q} (\\hat {x} ^ {f}) := \\left(\\underset {z _ {k} \\in \\mathcal {Z}} {\\arg \\min } \\| \\hat {x} _ {i j} ^ {f} - z _ {k} \\|\\right). \\tag {3}\n$$\n",
|
| 571 |
+
"text_format": "latex",
|
| 572 |
+
"bbox": [
|
| 573 |
+
356,
|
| 574 |
+
763,
|
| 575 |
+
785,
|
| 576 |
+
799
|
| 577 |
+
],
|
| 578 |
+
"page_idx": 5
|
| 579 |
+
},
|
| 580 |
+
{
|
| 581 |
+
"type": "text",
|
| 582 |
+
"text": "Next, we introduce the weather-invariant guiding (WIG) loss $\\mathcal{L}_{wig}$ to guide $\\mathcal{Z}$ that the weather codebook recalls the same clear knowledge of clear feature for",
|
| 583 |
+
"bbox": [
|
| 584 |
+
214,
|
| 585 |
+
809,
|
| 586 |
+
787,
|
| 587 |
+
840
|
| 588 |
+
],
|
| 589 |
+
"page_idx": 5
|
| 590 |
+
},
|
| 591 |
+
{
|
| 592 |
+
"type": "page_number",
|
| 593 |
+
"text": "6",
|
| 594 |
+
"bbox": [
|
| 595 |
+
217,
|
| 596 |
+
114,
|
| 597 |
+
228,
|
| 598 |
+
126
|
| 599 |
+
],
|
| 600 |
+
"page_idx": 5
|
| 601 |
+
},
|
| 602 |
+
{
|
| 603 |
+
"type": "header",
|
| 604 |
+
"text": "Y. Oh et al.",
|
| 605 |
+
"bbox": [
|
| 606 |
+
271,
|
| 607 |
+
114,
|
| 608 |
+
354,
|
| 609 |
+
126
|
| 610 |
+
],
|
| 611 |
+
"page_idx": 5
|
| 612 |
+
},
|
| 613 |
+
{
|
| 614 |
+
"type": "image",
|
| 615 |
+
"img_path": "images/30c2bd0b9fc9be3b440b6c3e119c54193994e8d55260f719d12000e097d5ad66.jpg",
|
| 616 |
+
"image_caption": [
|
| 617 |
+
"Fig. 4: Training process of the weather-adaptive diffusion model, which consists of two processes: (1) Adding fog variant $\\epsilon_{n}$ from input clear feature $x^{c}$ (forward process) and (2) enhancing representation with weather-reference feature $x^{r}$ (reverse process)."
|
| 618 |
+
],
|
| 619 |
+
"image_footnote": [],
|
| 620 |
+
"bbox": [
|
| 621 |
+
225,
|
| 622 |
+
146,
|
| 623 |
+
776,
|
| 624 |
+
311
|
| 625 |
+
],
|
| 626 |
+
"page_idx": 6
|
| 627 |
+
},
|
| 628 |
+
{
|
| 629 |
+
"type": "text",
|
| 630 |
+
"text": "the foggy feature, which can be represented as follows:",
|
| 631 |
+
"bbox": [
|
| 632 |
+
215,
|
| 633 |
+
371,
|
| 634 |
+
607,
|
| 635 |
+
385
|
| 636 |
+
],
|
| 637 |
+
"page_idx": 6
|
| 638 |
+
},
|
| 639 |
+
{
|
| 640 |
+
"type": "equation",
|
| 641 |
+
"text": "\n$$\n\\mathcal {L} _ {w i g} = \\left\\| x ^ {r (c)} - x ^ {r (f)} \\right\\| _ {2} ^ {2}. \\tag {4}\n$$\n",
|
| 642 |
+
"text_format": "latex",
|
| 643 |
+
"bbox": [
|
| 644 |
+
410,
|
| 645 |
+
397,
|
| 646 |
+
785,
|
| 647 |
+
426
|
| 648 |
+
],
|
| 649 |
+
"page_idx": 6
|
| 650 |
+
},
|
| 651 |
+
{
|
| 652 |
+
"type": "text",
|
| 653 |
+
"text": "Finally, the clear knowledge recalling (CKR) loss $\\mathcal{L}_{ckr}$ is obtained by adding $\\mathcal{L}_{cke}$ and $\\mathcal{L}_{wig}$ , which is defined as:",
|
| 654 |
+
"bbox": [
|
| 655 |
+
215,
|
| 656 |
+
436,
|
| 657 |
+
784,
|
| 658 |
+
467
|
| 659 |
+
],
|
| 660 |
+
"page_idx": 6
|
| 661 |
+
},
|
| 662 |
+
{
|
| 663 |
+
"type": "equation",
|
| 664 |
+
"text": "\n$$\n\\mathcal {L} _ {c k r} = \\mathcal {L} _ {c k e} + \\mathcal {L} _ {w i g}. \\tag {5}\n$$\n",
|
| 665 |
+
"text_format": "latex",
|
| 666 |
+
"bbox": [
|
| 667 |
+
426,
|
| 668 |
+
479,
|
| 669 |
+
784,
|
| 670 |
+
494
|
| 671 |
+
],
|
| 672 |
+
"page_idx": 6
|
| 673 |
+
},
|
| 674 |
+
{
|
| 675 |
+
"type": "text",
|
| 676 |
+
"text": "In the training phase, the weight parameters of embedding $K$ slots of weather codebook $\\mathcal{Z}$ are initialized randomly and they are updated through Eq. (5). In the inference phase, all parameters are fixed to recall clear weather, generating weather-reference features for any weather conditions.",
|
| 677 |
+
"bbox": [
|
| 678 |
+
212,
|
| 679 |
+
505,
|
| 680 |
+
784,
|
| 681 |
+
566
|
| 682 |
+
],
|
| 683 |
+
"page_idx": 6
|
| 684 |
+
},
|
| 685 |
+
{
|
| 686 |
+
"type": "text",
|
| 687 |
+
"text": "3.2 Weather-Adaptive Diffusion Model",
|
| 688 |
+
"text_level": 1,
|
| 689 |
+
"bbox": [
|
| 690 |
+
215,
|
| 691 |
+
588,
|
| 692 |
+
550,
|
| 693 |
+
603
|
| 694 |
+
],
|
| 695 |
+
"page_idx": 6
|
| 696 |
+
},
|
| 697 |
+
{
|
| 698 |
+
"type": "text",
|
| 699 |
+
"text": "Through Section 3.1, we now know that the weather codebook outputs the weather-reference feature $x^{r(c)}$ for the clear weather and $x^{r(f)}$ for the foggy weather through Eq. (5). From now on, since our method can receive any input images (clear or foggy), we denote the weather-reference feature as $x^r$ .",
|
| 700 |
+
"bbox": [
|
| 701 |
+
212,
|
| 702 |
+
613,
|
| 703 |
+
784,
|
| 704 |
+
672
|
| 705 |
+
],
|
| 706 |
+
"page_idx": 6
|
| 707 |
+
},
|
| 708 |
+
{
|
| 709 |
+
"type": "text",
|
| 710 |
+
"text": "Fig. 4 shows the training process of the proposed weather-adaptive diffusion model. The key idea of the diffusion model [15,42] is to gradually enhance $x_{T}$ to $x_{0}$ with fixed Markov Chain of $T$ timesteps. To this end, the forward and reverse processes are conducted in the training phase, and only the reverse process is used for the inference phase. Motivated by [15,42], we construct the weather-adaptive diffusion model to enhance representation related to the weather conditions. To this end, unlike traditional diffusion methods [15,42] that adopt the Gaussian noise to the image or latent space, we adopt $\\mathcal{F} = x^{f} - x^{c}$ , called fog distribution, to guide our diffusion model to be aware of the foggy weather. Ideally, the information contained within $\\mathcal{F}$ should include the information of fog, as it represents the difference between the foggy scene and the same scene with clear weather.",
|
| 711 |
+
"bbox": [
|
| 712 |
+
212,
|
| 713 |
+
674,
|
| 714 |
+
785,
|
| 715 |
+
839
|
| 716 |
+
],
|
| 717 |
+
"page_idx": 6
|
| 718 |
+
},
|
| 719 |
+
{
|
| 720 |
+
"type": "header",
|
| 721 |
+
"text": "MonoWAD",
|
| 722 |
+
"bbox": [
|
| 723 |
+
653,
|
| 724 |
+
114,
|
| 725 |
+
730,
|
| 726 |
+
126
|
| 727 |
+
],
|
| 728 |
+
"page_idx": 6
|
| 729 |
+
},
|
| 730 |
+
{
|
| 731 |
+
"type": "page_number",
|
| 732 |
+
"text": "7",
|
| 733 |
+
"bbox": [
|
| 734 |
+
774,
|
| 735 |
+
116,
|
| 736 |
+
785,
|
| 737 |
+
126
|
| 738 |
+
],
|
| 739 |
+
"page_idx": 6
|
| 740 |
+
},
|
| 741 |
+
{
|
| 742 |
+
"type": "text",
|
| 743 |
+
"text": "By doing so, our diffusion model learns the variation in weather by repeatedly adding and removing the fog. Note that, as we take clear feature $x^{c}$ for $x_{0}$ to make a reference input for our diffusion model, we newly denote $x_{0}$ as $x_{0}^{c}$ .",
|
| 744 |
+
"bbox": [
|
| 745 |
+
212,
|
| 746 |
+
146,
|
| 747 |
+
782,
|
| 748 |
+
191
|
| 749 |
+
],
|
| 750 |
+
"page_idx": 7
|
| 751 |
+
},
|
| 752 |
+
{
|
| 753 |
+
"type": "text",
|
| 754 |
+
"text": "For the forward process at the $t$ -th timestep, $q(x_{t}^{c}|x_{t - 1}^{c})$ takes the previous feature $x_{t - 1}^{c}$ and the noise related to the fog (i.e., $\\mathcal{F}$ ) as inputs to generate $x_{t}^{c}$ . This procedure is repeated over $T$ timesteps, which can be represented as:",
|
| 755 |
+
"bbox": [
|
| 756 |
+
212,
|
| 757 |
+
191,
|
| 758 |
+
784,
|
| 759 |
+
237
|
| 760 |
+
],
|
| 761 |
+
"page_idx": 7
|
| 762 |
+
},
|
| 763 |
+
{
|
| 764 |
+
"type": "equation",
|
| 765 |
+
"text": "\n$$\nq (x _ {t} ^ {c} | x _ {t - 1} ^ {c}) = \\mathcal {F} (x _ {t} ^ {c}; \\sqrt {1 - \\beta_ {t}} x _ {t - 1} ^ {c}, \\beta_ {t} \\mathbf {I}), \\tag {6}\n$$\n",
|
| 766 |
+
"text_format": "latex",
|
| 767 |
+
"bbox": [
|
| 768 |
+
366,
|
| 769 |
+
247,
|
| 770 |
+
785,
|
| 771 |
+
267
|
| 772 |
+
],
|
| 773 |
+
"page_idx": 7
|
| 774 |
+
},
|
| 775 |
+
{
|
| 776 |
+
"type": "equation",
|
| 777 |
+
"text": "\n$$\nq \\left(x _ {1} ^ {c}, \\dots , x _ {T} ^ {c} \\mid x ^ {c}\\right) = \\prod_ {t = 1} ^ {T} q \\left(x _ {t} ^ {c} \\mid x _ {t - 1} ^ {c}\\right), \\tag {7}\n$$\n",
|
| 778 |
+
"text_format": "latex",
|
| 779 |
+
"bbox": [
|
| 780 |
+
385,
|
| 781 |
+
277,
|
| 782 |
+
785,
|
| 783 |
+
316
|
| 784 |
+
],
|
| 785 |
+
"page_idx": 7
|
| 786 |
+
},
|
| 787 |
+
{
|
| 788 |
+
"type": "text",
|
| 789 |
+
"text": "where $\\beta_{t}$ denotes the variance schedule.",
|
| 790 |
+
"bbox": [
|
| 791 |
+
214,
|
| 792 |
+
323,
|
| 793 |
+
501,
|
| 794 |
+
335
|
| 795 |
+
],
|
| 796 |
+
"page_idx": 7
|
| 797 |
+
},
|
| 798 |
+
{
|
| 799 |
+
"type": "text",
|
| 800 |
+
"text": "Next, the reverse process at the $t$ -th timestep aims to estimate fog variant $\\epsilon_{n}$ using $x_{t}^{c}$ to enhance the feature representation of the foggy. To this end, we adopt conditional autoencoder $\\epsilon_{\\theta}(x_{t}^{c}, t, x^{r})$ that receives $x_{t}^{c}$ and reference feature $x^{r}$ from the weather codebook $\\mathcal{Z}$ . Specifically, $\\epsilon_{\\theta}$ estimates mean $\\mu_{\\theta}$ and variance $\\pmb{\\Sigma}_{\\theta}$ of the fog distribution at the $t$ -th timestep, denoted as $\\tilde{\\mathcal{F}}(\\cdot)$ . The reverse process is also repeated over $T$ timesteps, which can be represented as:",
|
| 801 |
+
"bbox": [
|
| 802 |
+
212,
|
| 803 |
+
338,
|
| 804 |
+
782,
|
| 805 |
+
428
|
| 806 |
+
],
|
| 807 |
+
"page_idx": 7
|
| 808 |
+
},
|
| 809 |
+
{
|
| 810 |
+
"type": "equation",
|
| 811 |
+
"text": "\n$$\np _ {\\theta} \\left(x _ {t - 1} ^ {c} \\mid x _ {t} ^ {c}, x ^ {r}\\right) = \\tilde {\\mathcal {F}} \\left(x _ {t - 1} ^ {c}; \\boldsymbol {\\mu} _ {\\theta} \\left(x _ {t} ^ {c}, t, x ^ {r}\\right), \\boldsymbol {\\Sigma} _ {\\theta} \\left(x _ {t} ^ {c}, t\\right)\\right), \\tag {8}\n$$\n",
|
| 812 |
+
"text_format": "latex",
|
| 813 |
+
"bbox": [
|
| 814 |
+
325,
|
| 815 |
+
440,
|
| 816 |
+
785,
|
| 817 |
+
458
|
| 818 |
+
],
|
| 819 |
+
"page_idx": 7
|
| 820 |
+
},
|
| 821 |
+
{
|
| 822 |
+
"type": "equation",
|
| 823 |
+
"text": "\n$$\np _ {\\theta} \\left(x ^ {c}, \\dots , x _ {T} ^ {c}\\right) = p \\left(x _ {T} ^ {c}\\right) \\prod_ {t = 1} ^ {T} p _ {\\theta} \\left(x _ {t - 1} ^ {c} \\mid x _ {t} ^ {c}, x ^ {r}\\right), \\tag {9}\n$$\n",
|
| 824 |
+
"text_format": "latex",
|
| 825 |
+
"bbox": [
|
| 826 |
+
356,
|
| 827 |
+
469,
|
| 828 |
+
785,
|
| 829 |
+
508
|
| 830 |
+
],
|
| 831 |
+
"page_idx": 7
|
| 832 |
+
},
|
| 833 |
+
{
|
| 834 |
+
"type": "text",
|
| 835 |
+
"text": "where $p_{\\theta}(x_{t - 1}^{c}|x_{t}^{c},x^{r})$ includes cross-attention layer in $\\epsilon_{\\theta}$",
|
| 836 |
+
"bbox": [
|
| 837 |
+
214,
|
| 838 |
+
515,
|
| 839 |
+
627,
|
| 840 |
+
530
|
| 841 |
+
],
|
| 842 |
+
"page_idx": 7
|
| 843 |
+
},
|
| 844 |
+
{
|
| 845 |
+
"type": "text",
|
| 846 |
+
"text": "The cross-attention layer receives the flattened clear feature $x_{t}^{c}$ and weather-reference feature $x^{r}$ , i.e., $\\bar{x}_{t}^{c}$ and $\\bar{x}^{r}$ , respectively. The similarity between $\\bar{x}^{r}$ and $\\bar{x}_{t}^{c}$ is calculated in the cross-attention layer to transfer the enhancement to $\\bar{x}_{t}^{c}$ , which can be formulated as:",
|
| 847 |
+
"bbox": [
|
| 848 |
+
212,
|
| 849 |
+
530,
|
| 850 |
+
782,
|
| 851 |
+
589
|
| 852 |
+
],
|
| 853 |
+
"page_idx": 7
|
| 854 |
+
},
|
| 855 |
+
{
|
| 856 |
+
"type": "equation",
|
| 857 |
+
"text": "\n$$\n\\operatorname {A t t e n t i o n} (Q, K, V) = \\operatorname {s o f t m a x} \\left(\\frac {Q K ^ {T}}{\\sqrt {d}}\\right) V, \\tag {10}\n$$\n",
|
| 858 |
+
"text_format": "latex",
|
| 859 |
+
"bbox": [
|
| 860 |
+
352,
|
| 861 |
+
598,
|
| 862 |
+
785,
|
| 863 |
+
633
|
| 864 |
+
],
|
| 865 |
+
"page_idx": 7
|
| 866 |
+
},
|
| 867 |
+
{
|
| 868 |
+
"type": "text",
|
| 869 |
+
"text": "where $Q = W_{i}^{q} \\cdot \\bar{x}_{t}^{c}, K = W_{i}^{k} \\cdot \\bar{x}^{r}, V = W_{i}^{v} \\cdot \\bar{x}^{r}$ with learnable parameters $W_{i}^{q}, W_{i}^{k}, W_{i}^{v}$ .",
|
| 870 |
+
"bbox": [
|
| 871 |
+
214,
|
| 872 |
+
642,
|
| 873 |
+
782,
|
| 874 |
+
672
|
| 875 |
+
],
|
| 876 |
+
"page_idx": 7
|
| 877 |
+
},
|
| 878 |
+
{
|
| 879 |
+
"type": "text",
|
| 880 |
+
"text": "To ensure that the estimated fog variant $\\epsilon_{\\theta}$ is similar to the fog variant $\\epsilon_{n}$ applied to make $x_{t}^{c}$ from the input clear feature $x^{c}$ , we propose a weather-adaptive enhancement loss $\\mathcal{L}_{wae}$ , which is formulated as:",
|
| 881 |
+
"bbox": [
|
| 882 |
+
212,
|
| 883 |
+
674,
|
| 884 |
+
782,
|
| 885 |
+
719
|
| 886 |
+
],
|
| 887 |
+
"page_idx": 7
|
| 888 |
+
},
|
| 889 |
+
{
|
| 890 |
+
"type": "equation",
|
| 891 |
+
"text": "\n$$\n\\mathcal {L} _ {w a e} = \\mathbb {E} _ {x ^ {c}, \\boldsymbol {\\epsilon} _ {n} \\sim \\mathcal {F}, t} \\left[ \\| \\boldsymbol {\\epsilon} _ {n} - \\boldsymbol {\\epsilon} _ {\\theta} \\left(x _ {t} ^ {c}, t, x ^ {r}\\right) \\| _ {2} ^ {2} \\right]. \\tag {11}\n$$\n",
|
| 892 |
+
"text_format": "latex",
|
| 893 |
+
"bbox": [
|
| 894 |
+
356,
|
| 895 |
+
729,
|
| 896 |
+
785,
|
| 897 |
+
755
|
| 898 |
+
],
|
| 899 |
+
"page_idx": 7
|
| 900 |
+
},
|
| 901 |
+
{
|
| 902 |
+
"type": "text",
|
| 903 |
+
"text": "With $\\mathcal{L}_{wae}$ , $\\epsilon_{\\theta}$ can estimate the fog variant by leveraging the fog distribution as noise for our diffusion model through multiple forward/reverse processes. Additionally, the cross-attention layer within $\\epsilon_{\\theta}$ dynamically enhances the feature representation based on combining knowledge of the input feature (foggy or clear) and the weather-reference feature. Since our diffusion model learns the",
|
| 904 |
+
"bbox": [
|
| 905 |
+
212,
|
| 906 |
+
763,
|
| 907 |
+
785,
|
| 908 |
+
839
|
| 909 |
+
],
|
| 910 |
+
"page_idx": 7
|
| 911 |
+
},
|
| 912 |
+
{
|
| 913 |
+
"type": "page_number",
|
| 914 |
+
"text": "8",
|
| 915 |
+
"bbox": [
|
| 916 |
+
217,
|
| 917 |
+
114,
|
| 918 |
+
228,
|
| 919 |
+
126
|
| 920 |
+
],
|
| 921 |
+
"page_idx": 7
|
| 922 |
+
},
|
| 923 |
+
{
|
| 924 |
+
"type": "header",
|
| 925 |
+
"text": "Y. Oh et al.",
|
| 926 |
+
"bbox": [
|
| 927 |
+
271,
|
| 928 |
+
114,
|
| 929 |
+
354,
|
| 930 |
+
127
|
| 931 |
+
],
|
| 932 |
+
"page_idx": 7
|
| 933 |
+
},
|
| 934 |
+
{
|
| 935 |
+
"type": "text",
|
| 936 |
+
"text": "degree of improvement needed corresponding to weather conditions, it can improve the representation of any input, whether clear or foggy in inference phase. This leads our monocular 3D object detector can handle both clear and foggy input images, resulting in weather-robust detection.",
|
| 937 |
+
"bbox": [
|
| 938 |
+
212,
|
| 939 |
+
146,
|
| 940 |
+
782,
|
| 941 |
+
205
|
| 942 |
+
],
|
| 943 |
+
"page_idx": 8
|
| 944 |
+
},
|
| 945 |
+
{
|
| 946 |
+
"type": "text",
|
| 947 |
+
"text": "3.3 Total Loss Function",
|
| 948 |
+
"text_level": 1,
|
| 949 |
+
"bbox": [
|
| 950 |
+
214,
|
| 951 |
+
218,
|
| 952 |
+
426,
|
| 953 |
+
232
|
| 954 |
+
],
|
| 955 |
+
"page_idx": 8
|
| 956 |
+
},
|
| 957 |
+
{
|
| 958 |
+
"type": "text",
|
| 959 |
+
"text": "The total loss function of our MonoWAD is represented as follows:",
|
| 960 |
+
"bbox": [
|
| 961 |
+
214,
|
| 962 |
+
243,
|
| 963 |
+
694,
|
| 964 |
+
258
|
| 965 |
+
],
|
| 966 |
+
"page_idx": 8
|
| 967 |
+
},
|
| 968 |
+
{
|
| 969 |
+
"type": "equation",
|
| 970 |
+
"text": "\n$$\n\\mathcal {L} _ {\\text {T o t a l}} = \\mathcal {L} _ {O D} + \\lambda_ {1} \\mathcal {L} _ {\\text {c k r}} + \\lambda_ {2} \\mathcal {L} _ {\\text {w a e}}, \\tag {12}\n$$\n",
|
| 971 |
+
"text_format": "latex",
|
| 972 |
+
"bbox": [
|
| 973 |
+
380,
|
| 974 |
+
272,
|
| 975 |
+
784,
|
| 976 |
+
287
|
| 977 |
+
],
|
| 978 |
+
"page_idx": 8
|
| 979 |
+
},
|
| 980 |
+
{
|
| 981 |
+
"type": "text",
|
| 982 |
+
"text": "where $\\lambda_{1}$ and $\\lambda_{2}$ denote balancing hyper-parameters. In our experiments, we set $\\lambda_{1} = \\lambda_{2} = 1$ . $\\mathcal{L}_{OD}$ is the detection loss for 3D object detection [16,55]. It includes loss functions of classification, regression, and depth loss, that are similar to prior works [8,16,31]. The overall weight parameters are updated through $\\mathcal{L}_{Total}$ .",
|
| 983 |
+
"bbox": [
|
| 984 |
+
214,
|
| 985 |
+
299,
|
| 986 |
+
782,
|
| 987 |
+
359
|
| 988 |
+
],
|
| 989 |
+
"page_idx": 8
|
| 990 |
+
},
|
| 991 |
+
{
|
| 992 |
+
"type": "text",
|
| 993 |
+
"text": "4 Experiments",
|
| 994 |
+
"text_level": 1,
|
| 995 |
+
"bbox": [
|
| 996 |
+
214,
|
| 997 |
+
382,
|
| 998 |
+
375,
|
| 999 |
+
400
|
| 1000 |
+
],
|
| 1001 |
+
"page_idx": 8
|
| 1002 |
+
},
|
| 1003 |
+
{
|
| 1004 |
+
"type": "text",
|
| 1005 |
+
"text": "4.1 Dataset and Evaluation Metrics",
|
| 1006 |
+
"text_level": 1,
|
| 1007 |
+
"bbox": [
|
| 1008 |
+
214,
|
| 1009 |
+
414,
|
| 1010 |
+
526,
|
| 1011 |
+
429
|
| 1012 |
+
],
|
| 1013 |
+
"page_idx": 8
|
| 1014 |
+
},
|
| 1015 |
+
{
|
| 1016 |
+
"type": "text",
|
| 1017 |
+
"text": "Datasets. We utilize the KITTI 3D object detection dataset [12], the most widely adopted for 3D object detection. It contains 7,481 training images and 7,518 test images under clear weather conditions. Due to unavailable ground-truth annotations for test images and limited evaluation on the test server, we follow [6] by splitting 3,712 training images and 3,769 validation images. In addition, as our work requires paired images to learn about weather changes, we generate foggy images from all images in the KITTI dataset, called foggy KITTI, that emulate the foggy scene. Following the protocols of [2,43], we generate foggy images based on object distances using depth maps estimated by DORN [10]. Please refer to the supplementary materials for details.",
|
| 1018 |
+
"bbox": [
|
| 1019 |
+
212,
|
| 1020 |
+
439,
|
| 1021 |
+
784,
|
| 1022 |
+
589
|
| 1023 |
+
],
|
| 1024 |
+
"page_idx": 8
|
| 1025 |
+
},
|
| 1026 |
+
{
|
| 1027 |
+
"type": "text",
|
| 1028 |
+
"text": "In addition, as our work is focused on robust monocular 3D object detection in various weather conditions, we further adopt the Virtual KITTI dataset [5,11], which contains photo-realistic synthetic images under various weather conditions (e.g., foggy, rainy, sunset). It is associated with the original real-world KITTI dataset, which has 3D annotations for each weather.",
|
| 1029 |
+
"bbox": [
|
| 1030 |
+
212,
|
| 1031 |
+
590,
|
| 1032 |
+
784,
|
| 1033 |
+
666
|
| 1034 |
+
],
|
| 1035 |
+
"page_idx": 8
|
| 1036 |
+
},
|
| 1037 |
+
{
|
| 1038 |
+
"type": "text",
|
| 1039 |
+
"text": "Evaluation Metrics. We adopt average precision in both 3D detection $(\\mathrm{AP}_{3D})$ and bird-eye view detection $(\\mathrm{AP}_{BEV})$ under three difficulty levels ('Easy', 'Moderate', 'Hard') according to size, occlusion, and truncation. Following [47], we use 40 recall position metric $\\mathrm{AP}_{40}$ and report scores for the car category under the IoU threshold 0.7 for the KITTI dataset and 0.5 for the Virtual KITTI dataset.",
|
| 1040 |
+
"bbox": [
|
| 1041 |
+
214,
|
| 1042 |
+
672,
|
| 1043 |
+
782,
|
| 1044 |
+
750
|
| 1045 |
+
],
|
| 1046 |
+
"page_idx": 8
|
| 1047 |
+
},
|
| 1048 |
+
{
|
| 1049 |
+
"type": "text",
|
| 1050 |
+
"text": "4.2 Implementation Details",
|
| 1051 |
+
"text_level": 1,
|
| 1052 |
+
"bbox": [
|
| 1053 |
+
214,
|
| 1054 |
+
768,
|
| 1055 |
+
455,
|
| 1056 |
+
784
|
| 1057 |
+
],
|
| 1058 |
+
"page_idx": 8
|
| 1059 |
+
},
|
| 1060 |
+
{
|
| 1061 |
+
"type": "text",
|
| 1062 |
+
"text": "We use DLA-102 [52] as our backbone and adopt the transformer architecture of [16] for the detection block. We train MonoWAD on a single RTX 4090 GPU with a batch size of 4 over 120 epochs using Adam optimizer [23] (initial learning",
|
| 1063 |
+
"bbox": [
|
| 1064 |
+
212,
|
| 1065 |
+
794,
|
| 1066 |
+
782,
|
| 1067 |
+
840
|
| 1068 |
+
],
|
| 1069 |
+
"page_idx": 8
|
| 1070 |
+
},
|
| 1071 |
+
{
|
| 1072 |
+
"type": "header",
|
| 1073 |
+
"text": "MonoWAD",
|
| 1074 |
+
"bbox": [
|
| 1075 |
+
653,
|
| 1076 |
+
114,
|
| 1077 |
+
730,
|
| 1078 |
+
126
|
| 1079 |
+
],
|
| 1080 |
+
"page_idx": 8
|
| 1081 |
+
},
|
| 1082 |
+
{
|
| 1083 |
+
"type": "page_number",
|
| 1084 |
+
"text": "9",
|
| 1085 |
+
"bbox": [
|
| 1086 |
+
774,
|
| 1087 |
+
116,
|
| 1088 |
+
785,
|
| 1089 |
+
126
|
| 1090 |
+
],
|
| 1091 |
+
"page_idx": 8
|
| 1092 |
+
},
|
| 1093 |
+
{
|
| 1094 |
+
"type": "table",
|
| 1095 |
+
"img_path": "images/43d987fd65abcd4539352a30f7ddc7a32010e6856d7d49d74baa9d01a9cba868.jpg",
|
| 1096 |
+
"table_caption": [
|
| 1097 |
+
"Table 1: Detection results of car category on KITTI validation set under foggy weather and clear weather conditions. The results of the state-of-the-art methods under foggy weather are obtained through our reproduction with the official source code. Bold/underlined fonts indicate the best/second-best results."
|
| 1098 |
+
],
|
| 1099 |
+
"table_footnote": [],
|
| 1100 |
+
"table_body": "<table><tr><td rowspan=\"2\">Method</td><td colspan=\"3\">Foggy (AP3D)</td><td colspan=\"3\">Clear (AP3D)</td><td colspan=\"3\">Average</td></tr><tr><td>Easy</td><td>Mod.</td><td>Hard</td><td>Easy</td><td>Mod.</td><td>Hard</td><td>Easy</td><td>Mod.</td><td>Hard</td></tr><tr><td>GUPNet [28] (ICCV'21)</td><td>2.74</td><td>2.19</td><td>2.16</td><td>22.76</td><td>16.46</td><td>13.72</td><td>12.75</td><td>9.33</td><td>7.94</td></tr><tr><td>DID-M3D [37] (ECCV'22)</td><td>1.15</td><td>0.61</td><td>0.64</td><td>22.98</td><td>16.12</td><td>14.03</td><td>12.07</td><td>8.37</td><td>7.34</td></tr><tr><td>MonoGround [40] (CVPR'22)</td><td>0.00</td><td>0.00</td><td>0.06</td><td>25.24</td><td>18.69</td><td>15.58</td><td>12.62</td><td>9.35</td><td>7.82</td></tr><tr><td>MonoDTR [16] (CVPR'22)</td><td>16.89</td><td>11.86</td><td>9.87</td><td>24.52</td><td>18.57</td><td>15.51</td><td>20.71</td><td>15.22</td><td>12.69</td></tr><tr><td>MonoDETR [55] (ICCV'23)</td><td>7.40</td><td>5.74</td><td>4.53</td><td>28.84</td><td>20.61</td><td>16.38</td><td>18.12</td><td>13.18</td><td>10.46</td></tr><tr><td>MonoWAD (Ours)</td><td>27.17</td><td>19.57</td><td>16.21</td><td>29.10</td><td>21.08</td><td>17.73</td><td>28.14</td><td>20.33</td><td>16.97</td></tr></table>",
|
| 1101 |
+
"bbox": [
|
| 1102 |
+
220,
|
| 1103 |
+
203,
|
| 1104 |
+
782,
|
| 1105 |
+
306
|
| 1106 |
+
],
|
| 1107 |
+
"page_idx": 9
|
| 1108 |
+
},
|
| 1109 |
+
{
|
| 1110 |
+
"type": "table",
|
| 1111 |
+
"img_path": "images/34ef2b7aba8cdc8c396a824d68a1df5f4a92af2375954e1aaeb4d4e5c784a916.jpg",
|
| 1112 |
+
"table_caption": [
|
| 1113 |
+
"Table 2: Detection results of car category on KITTI test set under foggy weather. Bold/underlined fonts indicate the best/second-best results."
|
| 1114 |
+
],
|
| 1115 |
+
"table_footnote": [],
|
| 1116 |
+
"table_body": "<table><tr><td rowspan=\"2\">Method</td><td colspan=\"3\">AP3D</td><td colspan=\"3\">APBEV</td></tr><tr><td>Easy</td><td>Mod.</td><td>Hard</td><td>Easy</td><td>Mod.</td><td>Hard</td></tr><tr><td>GUPNet [28] (ICCV'21)</td><td>3.01</td><td>2.42</td><td>1.13</td><td>4.90</td><td>3.02</td><td>2.91</td></tr><tr><td>DID-M3D [37] (ECCV'22)</td><td>3.10</td><td>2.39</td><td>2.19</td><td>5.34</td><td>3.01</td><td>2.91</td></tr><tr><td>MonoGround [40] (CVPR'22)</td><td>0.14</td><td>0.20</td><td>0.22</td><td>0.23</td><td>0.38</td><td>0.39</td></tr><tr><td>MonoDTR [16] (CVPR'22)</td><td>11.07</td><td>7.41</td><td>5.26</td><td>15.76</td><td>10.15</td><td>7.53</td></tr><tr><td>MonoDETR [55] (ICCV'23)</td><td>9.33</td><td>5.54</td><td>4.06</td><td>13.15</td><td>8.06</td><td>6.30</td></tr><tr><td>MonoWAD (Ours)</td><td>19.75</td><td>13.32</td><td>11.04</td><td>27.95</td><td>19.06</td><td>15.61</td></tr></table>",
|
| 1117 |
+
"bbox": [
|
| 1118 |
+
259,
|
| 1119 |
+
349,
|
| 1120 |
+
736,
|
| 1121 |
+
455
|
| 1122 |
+
],
|
| 1123 |
+
"page_idx": 9
|
| 1124 |
+
},
|
| 1125 |
+
{
|
| 1126 |
+
"type": "text",
|
| 1127 |
+
"text": "rate of $10^{-4}$ ). For the weather codebook, we utilize embedding slot $K = 4096$ and set the dimension of each slot $D = 256$ . We set 4 heads for cross-attention of the weather-adaptive diffusion model, and the timesteps for the forward and reverse process of diffusion default to 15. The number of channels for the diffusion output features is $C = 256$ . As recent monocular 3D object detectors [16,28,37,40,55] have not been explored under adverse weather conditions, we implemented them with available official source code to faithfully reproduce them.",
|
| 1128 |
+
"bbox": [
|
| 1129 |
+
212,
|
| 1130 |
+
469,
|
| 1131 |
+
787,
|
| 1132 |
+
575
|
| 1133 |
+
],
|
| 1134 |
+
"page_idx": 9
|
| 1135 |
+
},
|
| 1136 |
+
{
|
| 1137 |
+
"type": "text",
|
| 1138 |
+
"text": "4.3 Comparison",
|
| 1139 |
+
"text_level": 1,
|
| 1140 |
+
"bbox": [
|
| 1141 |
+
215,
|
| 1142 |
+
589,
|
| 1143 |
+
362,
|
| 1144 |
+
604
|
| 1145 |
+
],
|
| 1146 |
+
"page_idx": 9
|
| 1147 |
+
},
|
| 1148 |
+
{
|
| 1149 |
+
"type": "text",
|
| 1150 |
+
"text": "Results on KITTI 3D Dataset. We compared MonoWAD with state-of-the-art monocular 3D object detectors [16,28,37,40,55] that do not use additional data (e.g., depth maps or LiDAR) during inference on the KITTI and foggy KITTI validation sets. Table 1 shows the $AP_{3D}$ results, and $AP_{BEV}$ results are in the supplementary materials. While recent methods have shown improved performance in clear weather, their performance drops in foggy weather, limiting their applicability in real-world applications (e.g., autonomous driving and robotics). In contrast, MonoWAD showed stable 3D detection performance under both foggy and clear weather. Since our weather codebook learns knowledge about clear weather and the weather-adaptive diffusion model can enhance the feature representation of the input images under both clear and foggy weather, MonoWAD shows a more weather-robust detection performance than that of the existing methods. Also, to explore the weather robustness of MonoWAD, we conducted experiments with mixed foggy and clear weather conditions at various ratios (please see the supplementary materials).",
|
| 1151 |
+
"bbox": [
|
| 1152 |
+
212,
|
| 1153 |
+
612,
|
| 1154 |
+
787,
|
| 1155 |
+
840
|
| 1156 |
+
],
|
| 1157 |
+
"page_idx": 9
|
| 1158 |
+
},
|
| 1159 |
+
{
|
| 1160 |
+
"type": "page_number",
|
| 1161 |
+
"text": "10",
|
| 1162 |
+
"bbox": [
|
| 1163 |
+
217,
|
| 1164 |
+
114,
|
| 1165 |
+
235,
|
| 1166 |
+
126
|
| 1167 |
+
],
|
| 1168 |
+
"page_idx": 9
|
| 1169 |
+
},
|
| 1170 |
+
{
|
| 1171 |
+
"type": "header",
|
| 1172 |
+
"text": "Y. Oh et al.",
|
| 1173 |
+
"bbox": [
|
| 1174 |
+
271,
|
| 1175 |
+
114,
|
| 1176 |
+
354,
|
| 1177 |
+
127
|
| 1178 |
+
],
|
| 1179 |
+
"page_idx": 9
|
| 1180 |
+
},
|
| 1181 |
+
{
|
| 1182 |
+
"type": "table",
|
| 1183 |
+
"img_path": "images/8f1f5b8491bac96f0f5e3211761154bb7fe958c2abbea8ff8288f4ecb58b66f1.jpg",
|
| 1184 |
+
"table_caption": [
|
| 1185 |
+
"Table 3: Detection results of car category on Virtual KITTI under foggy, rainy, sunset conditions. Bold/underlined fonts indicate the best/second-best results."
|
| 1186 |
+
],
|
| 1187 |
+
"table_footnote": [],
|
| 1188 |
+
"table_body": "<table><tr><td rowspan=\"2\">Method</td><td colspan=\"3\">Foggy (AP3D)</td><td colspan=\"3\">Rainy (AP3D)</td><td colspan=\"3\">Sunset (AP3D)</td></tr><tr><td>Easy</td><td>Mod.</td><td>Hard</td><td>Easy</td><td>Mod.</td><td>Hard</td><td>Easy</td><td>Mod.</td><td>Hard</td></tr><tr><td>GUPNet [28] (ICCV'21)</td><td>1.76</td><td>1.57</td><td>1.57</td><td>2.34</td><td>1.24</td><td>1.21</td><td>2.77</td><td>1.64</td><td>1.65</td></tr><tr><td>DID-M3D [37] (ECCV'22)</td><td>0.91</td><td>0.39</td><td>0.39</td><td>0.40</td><td>0.13</td><td>0.13</td><td>0.34</td><td>0.10</td><td>0.10</td></tr><tr><td>MonoGround [40] (CVPR'22)</td><td>0.29</td><td>0.30</td><td>0.25</td><td>5.49</td><td>2.82</td><td>2.77</td><td>7.68</td><td>4.24</td><td>4.20</td></tr><tr><td>MonoDTR [16] (CVPR'22)</td><td>8.79</td><td>5.75</td><td>5.72</td><td>11.73</td><td>6.25</td><td>6.74</td><td>9.86</td><td>5.42</td><td>5.42</td></tr><tr><td>MonoDETR [55] (ICCV'23)</td><td>4.50</td><td>2.99</td><td>2.96</td><td>6.61</td><td>3.46</td><td>3.42</td><td>7.08</td><td>4.17</td><td>4.16</td></tr><tr><td>MonoWAD (Ours)</td><td>13.33</td><td>8.56</td><td>8.50</td><td>14.12</td><td>8.33</td><td>8.24</td><td>13.38</td><td>7.89</td><td>7.80</td></tr></table>",
|
| 1189 |
+
"bbox": [
|
| 1190 |
+
220,
|
| 1191 |
+
176,
|
| 1192 |
+
777,
|
| 1193 |
+
280
|
| 1194 |
+
],
|
| 1195 |
+
"page_idx": 10
|
| 1196 |
+
},
|
| 1197 |
+
{
|
| 1198 |
+
"type": "table",
|
| 1199 |
+
"img_path": "images/92ef456f169f950eafccc5b80a73a7c9bb1b71ae5ff26f4d5d6731618d513186.jpg",
|
| 1200 |
+
"table_caption": [
|
| 1201 |
+
"Table 4: Effect of the proposed method on KITTI validation set for car category. WC denotes our weather codebook, WAD indicates our weather-adaptive diffusion model."
|
| 1202 |
+
],
|
| 1203 |
+
"table_footnote": [],
|
| 1204 |
+
"table_body": "<table><tr><td rowspan=\"2\">Method</td><td rowspan=\"2\">WC</td><td rowspan=\"2\">WAD</td><td colspan=\"3\">Foggy (AP3D)</td><td colspan=\"3\">Clear (AP3D)</td></tr><tr><td>Easy</td><td>Mod.</td><td>Hard</td><td>Easy</td><td>Mod.</td><td>Hard</td></tr><tr><td>Baseline</td><td>-</td><td>-</td><td>13.75</td><td>9.61</td><td>8.10</td><td>22.63</td><td>17.16</td><td>14.28</td></tr><tr><td rowspan=\"2\">MonoWAD (Ours)</td><td>-</td><td>✓</td><td>25.62</td><td>18.66</td><td>15.56</td><td>26.34</td><td>19.17</td><td>16.15</td></tr><tr><td>✓</td><td>✓</td><td>27.17</td><td>19.57</td><td>16.21</td><td>29.10</td><td>21.08</td><td>17.73</td></tr></table>",
|
| 1205 |
+
"bbox": [
|
| 1206 |
+
279,
|
| 1207 |
+
313,
|
| 1208 |
+
717,
|
| 1209 |
+
385
|
| 1210 |
+
],
|
| 1211 |
+
"page_idx": 10
|
| 1212 |
+
},
|
| 1213 |
+
{
|
| 1214 |
+
"type": "table",
|
| 1215 |
+
"img_path": "images/7ac33ed1df4e17379d7d4fd068da7ffec2ff38f541fb5f02817366a9c2141fb6.jpg",
|
| 1216 |
+
"table_caption": [
|
| 1217 |
+
"Table 5: Detection results on KITTI validation set by changing diffusion timestep $T$ ."
|
| 1218 |
+
],
|
| 1219 |
+
"table_footnote": [],
|
| 1220 |
+
"table_body": "<table><tr><td rowspan=\"2\">Timestep T</td><td colspan=\"3\">Foggy (AP3D)</td><td colspan=\"3\">Clear (AP3D)</td></tr><tr><td>Easy</td><td>Mod.</td><td>Hard</td><td>Easy</td><td>Mod.</td><td>Hard</td></tr><tr><td>-</td><td>13.75</td><td>9.61</td><td>8.10</td><td>22.63</td><td>17.16</td><td>14.28</td></tr><tr><td>5</td><td>23.57</td><td>17.91</td><td>14.96</td><td>26.03</td><td>19.21</td><td>16.01</td></tr><tr><td>10</td><td>25.28</td><td>18.49</td><td>15.42</td><td>26.79</td><td>19.90</td><td>16.78</td></tr><tr><td>15</td><td>27.17</td><td>19.57</td><td>16.21</td><td>29.10</td><td>21.08</td><td>17.73</td></tr><tr><td>20</td><td>24.54</td><td>18.29</td><td>15.10</td><td>24.85</td><td>18.54</td><td>15.34</td></tr></table>",
|
| 1221 |
+
"bbox": [
|
| 1222 |
+
274,
|
| 1223 |
+
407,
|
| 1224 |
+
722,
|
| 1225 |
+
502
|
| 1226 |
+
],
|
| 1227 |
+
"page_idx": 10
|
| 1228 |
+
},
|
| 1229 |
+
{
|
| 1230 |
+
"type": "text",
|
| 1231 |
+
"text": "We further compared 3D detection performances on the foggy KITTI test set (Table 2). Similar to Table 1, existing monocular 3D object detectors show lower performance under foggy weather. Through the results, our method shows the weather-robust monocular 3D performance under foggy and clear weather.",
|
| 1232 |
+
"bbox": [
|
| 1233 |
+
215,
|
| 1234 |
+
513,
|
| 1235 |
+
782,
|
| 1236 |
+
574
|
| 1237 |
+
],
|
| 1238 |
+
"page_idx": 10
|
| 1239 |
+
},
|
| 1240 |
+
{
|
| 1241 |
+
"type": "text",
|
| 1242 |
+
"text": "Results on Virtual KITTI Dataset. We also conducted experiments on the Virtual KITTI dataset to validate the generalization ability of our method across various weather conditions (i.e., foggy, rainy, sunset). The results are shown in Table 3. MonoWAD also shows the highest performance in foggy weather. Moreover, it outperforms the existing method in the rainy and sunset conditions. The results demonstrate that MonoWAD is robust and insensitive to various weather conditions that can be faced in real-world scenarios, not just clear weather.",
|
| 1243 |
+
"bbox": [
|
| 1244 |
+
215,
|
| 1245 |
+
582,
|
| 1246 |
+
784,
|
| 1247 |
+
688
|
| 1248 |
+
],
|
| 1249 |
+
"page_idx": 10
|
| 1250 |
+
},
|
| 1251 |
+
{
|
| 1252 |
+
"type": "text",
|
| 1253 |
+
"text": "4.4 Ablation Studies",
|
| 1254 |
+
"text_level": 1,
|
| 1255 |
+
"bbox": [
|
| 1256 |
+
215,
|
| 1257 |
+
707,
|
| 1258 |
+
398,
|
| 1259 |
+
722
|
| 1260 |
+
],
|
| 1261 |
+
"page_idx": 10
|
| 1262 |
+
},
|
| 1263 |
+
{
|
| 1264 |
+
"type": "text",
|
| 1265 |
+
"text": "We conducted ablation studies to examine (1) the effect of each proposed component (i.e., weather codebook and weather-adaptive diffusion model) and (2) the effect of the weather-adaptive diffusion model by varying the timestep $T$ . These experiments were performed on the KITTI and foggy KITTI 3D validation sets.",
|
| 1266 |
+
"bbox": [
|
| 1267 |
+
215,
|
| 1268 |
+
729,
|
| 1269 |
+
782,
|
| 1270 |
+
789
|
| 1271 |
+
],
|
| 1272 |
+
"page_idx": 10
|
| 1273 |
+
},
|
| 1274 |
+
{
|
| 1275 |
+
"type": "text",
|
| 1276 |
+
"text": "Effect of the Proposed Modules. The results regarding the effectiveness of the proposed weather codebook and weather-adaptive diffusion model are pre-",
|
| 1277 |
+
"bbox": [
|
| 1278 |
+
215,
|
| 1279 |
+
809,
|
| 1280 |
+
784,
|
| 1281 |
+
839
|
| 1282 |
+
],
|
| 1283 |
+
"page_idx": 10
|
| 1284 |
+
},
|
| 1285 |
+
{
|
| 1286 |
+
"type": "header",
|
| 1287 |
+
"text": "MonoWAD",
|
| 1288 |
+
"bbox": [
|
| 1289 |
+
653,
|
| 1290 |
+
114,
|
| 1291 |
+
730,
|
| 1292 |
+
126
|
| 1293 |
+
],
|
| 1294 |
+
"page_idx": 10
|
| 1295 |
+
},
|
| 1296 |
+
{
|
| 1297 |
+
"type": "page_number",
|
| 1298 |
+
"text": "11",
|
| 1299 |
+
"bbox": [
|
| 1300 |
+
767,
|
| 1301 |
+
116,
|
| 1302 |
+
782,
|
| 1303 |
+
126
|
| 1304 |
+
],
|
| 1305 |
+
"page_idx": 10
|
| 1306 |
+
},
|
| 1307 |
+
{
|
| 1308 |
+
"type": "table",
|
| 1309 |
+
"img_path": "images/c1a042afafb71c984ae10b776ed7bb8d251369f7e429fb46a9e0001f466d4e66.jpg",
|
| 1310 |
+
"table_caption": [
|
| 1311 |
+
"Table 6: Performance comparison on KITTI validation set under foggy and clear weather conditions. We compare our MonoWAD with MonoDTR [16], which demonstrates superior performance in foggy weather using state-of-the-art dehazing methods."
|
| 1312 |
+
],
|
| 1313 |
+
"table_footnote": [],
|
| 1314 |
+
"table_body": "<table><tr><td rowspan=\"2\">Method</td><td rowspan=\"2\">Dehazing Method</td><td colspan=\"3\">Foggy (AP3D)</td><td colspan=\"3\">Clear (AP3D)</td><td colspan=\"3\">Average</td></tr><tr><td>Easy</td><td>Mod.</td><td>Hard</td><td>Easy</td><td>Mod.</td><td>Hard</td><td>Easy</td><td>Mod.</td><td>Hard</td></tr><tr><td>MonoDTR(B) [16] (CVPR'22)</td><td>-</td><td>16.89</td><td>11.86</td><td>9.87</td><td>24.52</td><td>18.57</td><td>15.51</td><td>20.71</td><td>15.22</td><td>12.69</td></tr><tr><td>B + RIDCP [49] (CVPR'23)</td><td>Image-level</td><td>17.23</td><td>12.41</td><td>10.44</td><td>24.02</td><td>17.89</td><td>14.78</td><td>20.63</td><td>15.15</td><td>12.61</td></tr><tr><td>B + DENet [39] (ACCV'22)</td><td>Feature-level</td><td>22.35</td><td>17.44</td><td>14.47</td><td>7.10</td><td>5.70</td><td>4.53</td><td>14.73</td><td>11.57</td><td>9.50</td></tr><tr><td>B + Yang et al. [51] (ACCV'22)</td><td>Feature-level</td><td>22.87</td><td>15.21</td><td>12.17</td><td>17.96</td><td>13.10</td><td>10.64</td><td>20.42</td><td>14.16</td><td>11.41</td></tr><tr><td>MonoWAD (Ours)</td><td>-</td><td>27.17</td><td>19.57</td><td>16.21</td><td>29.10</td><td>21.08</td><td>17.73</td><td>28.14</td><td>20.33</td><td>16.97</td></tr></table>",
|
| 1315 |
+
"bbox": [
|
| 1316 |
+
220,
|
| 1317 |
+
188,
|
| 1318 |
+
777,
|
| 1319 |
+
282
|
| 1320 |
+
],
|
| 1321 |
+
"page_idx": 11
|
| 1322 |
+
},
|
| 1323 |
+
{
|
| 1324 |
+
"type": "table",
|
| 1325 |
+
"img_path": "images/f612a81c020046afd3345f974015b5172f1ffaa62f69e2d8a3c7124aaae1bd75.jpg",
|
| 1326 |
+
"table_caption": [
|
| 1327 |
+
"Table 7: Comparison of diffusion models on KITTI validation set for car category: $\\mathcal{B}$ is baseline detection block (transformer encoder-decoder), CDM (Conditional Diffusion Model [42]), WC (weather codebook), and WAD (weather adaptive diffusion model)."
|
| 1328 |
+
],
|
| 1329 |
+
"table_footnote": [],
|
| 1330 |
+
"table_body": "<table><tr><td rowspan=\"2\">Method</td><td colspan=\"3\">Foggy (AP3D)</td><td colspan=\"3\">Clear (AP3D)</td></tr><tr><td>Easy</td><td>Mod.</td><td>Hard</td><td>Easy</td><td>Mod.</td><td>Hard</td></tr><tr><td>B+DDPM [15]</td><td>5.32</td><td>3.84</td><td>2.77</td><td>18.31</td><td>12.71</td><td>10.20</td></tr><tr><td>B+CDM [42]</td><td>2.74</td><td>2.10</td><td>2.00</td><td>20.50</td><td>14.51</td><td>11.63</td></tr><tr><td>B+WC+CDM [42]</td><td>17.51</td><td>12.74</td><td>10.40</td><td>21.05</td><td>15.11</td><td>12.54</td></tr><tr><td>B+WAD</td><td>25.62</td><td>18.66</td><td>15.56</td><td>26.34</td><td>19.17</td><td>16.15</td></tr><tr><td>MonoWAD(B+WC+WAD)</td><td>27.17</td><td>19.57</td><td>16.21</td><td>29.10</td><td>21.08</td><td>17.73</td></tr></table>",
|
| 1331 |
+
"bbox": [
|
| 1332 |
+
274,
|
| 1333 |
+
340,
|
| 1334 |
+
723,
|
| 1335 |
+
450
|
| 1336 |
+
],
|
| 1337 |
+
"page_idx": 11
|
| 1338 |
+
},
|
| 1339 |
+
{
|
| 1340 |
+
"type": "text",
|
| 1341 |
+
"text": "sented in Table 4. Since our weather-adaptive diffusion model is designed to enhance feature representation, it alone has shown significant performance improvement. When the weather codebook is additionally considered, the weather-adaptive diffusion model can leverage the knowledge of the weather-reference feature for clear weather, leading to enhanced performance. It makes MonoWAD can be robust to various weather conditions for monocular 3D object detection.",
|
| 1342 |
+
"bbox": [
|
| 1343 |
+
212,
|
| 1344 |
+
465,
|
| 1345 |
+
787,
|
| 1346 |
+
556
|
| 1347 |
+
],
|
| 1348 |
+
"page_idx": 11
|
| 1349 |
+
},
|
| 1350 |
+
{
|
| 1351 |
+
"type": "text",
|
| 1352 |
+
"text": "Effect of Timestep $T$ . We also conduct experiments by varying timestep $T$ of the weather-adaptive diffusion model. Timestep is the number of steps for the forward and reverse process of the diffusion model. Table 5 indicates that the highest monocular 3D detection performance is achieved when $T = 15$ under both clear and foggy weather conditions. Our MonoWAD consistently outperforms the baseline and other existing methods across all timesteps.",
|
| 1353 |
+
"bbox": [
|
| 1354 |
+
212,
|
| 1355 |
+
566,
|
| 1356 |
+
787,
|
| 1357 |
+
657
|
| 1358 |
+
],
|
| 1359 |
+
"page_idx": 11
|
| 1360 |
+
},
|
| 1361 |
+
{
|
| 1362 |
+
"type": "text",
|
| 1363 |
+
"text": "4.5 Discussions",
|
| 1364 |
+
"text_level": 1,
|
| 1365 |
+
"bbox": [
|
| 1366 |
+
215,
|
| 1367 |
+
672,
|
| 1368 |
+
356,
|
| 1369 |
+
686
|
| 1370 |
+
],
|
| 1371 |
+
"page_idx": 11
|
| 1372 |
+
},
|
| 1373 |
+
{
|
| 1374 |
+
"type": "text",
|
| 1375 |
+
"text": "Comparison with Dehazing Methods. We investigate the weather-robustness of our method for monocular 3D object detection compared to the other monocular detectors using dehazing methods. To this end, we compared MonoWAD with MonoDTR [16], which shows the best performance in foggy weather through the application of state-of-the-art image-level and feature-level dehazing methods [39,49,51]. The results are shown in Table 6. Recent dehazing methods are primarily focused on specific weather conditions, such as foggy weather. While they have shown some improvement under foggy conditions, they exhibit reduced performances under clear weather conditions. In contrast, our MonoWAD shows robust performance across both foggy and clear weather conditions.",
|
| 1376 |
+
"bbox": [
|
| 1377 |
+
212,
|
| 1378 |
+
696,
|
| 1379 |
+
792,
|
| 1380 |
+
849
|
| 1381 |
+
],
|
| 1382 |
+
"page_idx": 11
|
| 1383 |
+
},
|
| 1384 |
+
{
|
| 1385 |
+
"type": "page_number",
|
| 1386 |
+
"text": "12",
|
| 1387 |
+
"bbox": [
|
| 1388 |
+
217,
|
| 1389 |
+
114,
|
| 1390 |
+
235,
|
| 1391 |
+
126
|
| 1392 |
+
],
|
| 1393 |
+
"page_idx": 11
|
| 1394 |
+
},
|
| 1395 |
+
{
|
| 1396 |
+
"type": "header",
|
| 1397 |
+
"text": "Y. Oh et al.",
|
| 1398 |
+
"bbox": [
|
| 1399 |
+
271,
|
| 1400 |
+
114,
|
| 1401 |
+
354,
|
| 1402 |
+
127
|
| 1403 |
+
],
|
| 1404 |
+
"page_idx": 11
|
| 1405 |
+
},
|
| 1406 |
+
{
|
| 1407 |
+
"type": "image",
|
| 1408 |
+
"img_path": "images/1ce9cb7a8af6a016833dac592e5eb58cf51c6f754a3faf5c18414187a51fa69e.jpg",
|
| 1409 |
+
"image_caption": [
|
| 1410 |
+
"Fig. 5: Comparison of 3D detection examples (green: ground-truth, red: predicted 3D bounding-box) between our MonoWAD and two detectors, MonoDTR [16] and MonoDETR [55], that show the most improved performances among existing methods."
|
| 1411 |
+
],
|
| 1412 |
+
"image_footnote": [],
|
| 1413 |
+
"bbox": [
|
| 1414 |
+
223,
|
| 1415 |
+
146,
|
| 1416 |
+
781,
|
| 1417 |
+
290
|
| 1418 |
+
],
|
| 1419 |
+
"page_idx": 12
|
| 1420 |
+
},
|
| 1421 |
+
{
|
| 1422 |
+
"type": "image",
|
| 1423 |
+
"img_path": "images/4df94299132fc98cb15113316d4d2591056ee507d5585a7166dd4da1fed82b6a.jpg",
|
| 1424 |
+
"image_caption": [
|
| 1425 |
+
"Fig. 6: 3D detection results on real-world images of various weather conditions."
|
| 1426 |
+
],
|
| 1427 |
+
"image_footnote": [],
|
| 1428 |
+
"bbox": [
|
| 1429 |
+
212,
|
| 1430 |
+
345,
|
| 1431 |
+
789,
|
| 1432 |
+
594
|
| 1433 |
+
],
|
| 1434 |
+
"page_idx": 12
|
| 1435 |
+
},
|
| 1436 |
+
{
|
| 1437 |
+
"type": "text",
|
| 1438 |
+
"text": "Effect of Weather-Adaptive Diffusion Model. Table 7 shows the effectiveness of MonoWAD with existing diffusion models [15,42] on the KITTI and foggy KITTI validation sets. Existing methods adopt Gaussian noise for forward and reverse processes, but they can not fully understand about weather. In contrast, our weather-adaptive diffusion understands weather variances, allowing our MonoWAD to surpass existing methods in clear and foggy weather.",
|
| 1439 |
+
"bbox": [
|
| 1440 |
+
212,
|
| 1441 |
+
622,
|
| 1442 |
+
787,
|
| 1443 |
+
714
|
| 1444 |
+
],
|
| 1445 |
+
"page_idx": 12
|
| 1446 |
+
},
|
| 1447 |
+
{
|
| 1448 |
+
"type": "text",
|
| 1449 |
+
"text": "4.6 Visualization Results",
|
| 1450 |
+
"text_level": 1,
|
| 1451 |
+
"bbox": [
|
| 1452 |
+
215,
|
| 1453 |
+
737,
|
| 1454 |
+
434,
|
| 1455 |
+
752
|
| 1456 |
+
],
|
| 1457 |
+
"page_idx": 12
|
| 1458 |
+
},
|
| 1459 |
+
{
|
| 1460 |
+
"type": "text",
|
| 1461 |
+
"text": "Results on KITTI Dataset. We visualize several 3D detection results on the KITTI 3D dataset, comparing MonoWAD with MonoDTR and MonoDETR, which exhibit the highest performances among existing methods under foggy condition (Fig. 5). Existing methods struggle to detect objects obscured by fog, indicating limitations in detecting only fully visible objects. In contrast, even",
|
| 1462 |
+
"bbox": [
|
| 1463 |
+
212,
|
| 1464 |
+
763,
|
| 1465 |
+
787,
|
| 1466 |
+
840
|
| 1467 |
+
],
|
| 1468 |
+
"page_idx": 12
|
| 1469 |
+
},
|
| 1470 |
+
{
|
| 1471 |
+
"type": "header",
|
| 1472 |
+
"text": "MonoWAD",
|
| 1473 |
+
"bbox": [
|
| 1474 |
+
653,
|
| 1475 |
+
114,
|
| 1476 |
+
730,
|
| 1477 |
+
126
|
| 1478 |
+
],
|
| 1479 |
+
"page_idx": 12
|
| 1480 |
+
},
|
| 1481 |
+
{
|
| 1482 |
+
"type": "page_number",
|
| 1483 |
+
"text": "13",
|
| 1484 |
+
"bbox": [
|
| 1485 |
+
767,
|
| 1486 |
+
114,
|
| 1487 |
+
785,
|
| 1488 |
+
126
|
| 1489 |
+
],
|
| 1490 |
+
"page_idx": 12
|
| 1491 |
+
},
|
| 1492 |
+
{
|
| 1493 |
+
"type": "image",
|
| 1494 |
+
"img_path": "images/9a7aeccd9d723fbb3756503070f2e04bc5fb3e1ddc1cc2d526c39a3329162339.jpg",
|
| 1495 |
+
"image_caption": [
|
| 1496 |
+
"Fig. 7: t-SNE visualization results (Red: Clear, Blue: Foggy)."
|
| 1497 |
+
],
|
| 1498 |
+
"image_footnote": [],
|
| 1499 |
+
"bbox": [
|
| 1500 |
+
251,
|
| 1501 |
+
143,
|
| 1502 |
+
401,
|
| 1503 |
+
239
|
| 1504 |
+
],
|
| 1505 |
+
"page_idx": 13
|
| 1506 |
+
},
|
| 1507 |
+
{
|
| 1508 |
+
"type": "image",
|
| 1509 |
+
"img_path": "images/a32b11d56c1326e5fd5cc8acac38664ea01b1127ef3278cbfcfff7f6b61d76a0.jpg",
|
| 1510 |
+
"image_caption": [],
|
| 1511 |
+
"image_footnote": [],
|
| 1512 |
+
"bbox": [
|
| 1513 |
+
426,
|
| 1514 |
+
143,
|
| 1515 |
+
575,
|
| 1516 |
+
239
|
| 1517 |
+
],
|
| 1518 |
+
"page_idx": 13
|
| 1519 |
+
},
|
| 1520 |
+
{
|
| 1521 |
+
"type": "image",
|
| 1522 |
+
"img_path": "images/2590033f084d124308f014d93a14eb5c6c55e66ce963acf9541378c89e17472c.jpg",
|
| 1523 |
+
"image_caption": [],
|
| 1524 |
+
"image_footnote": [],
|
| 1525 |
+
"bbox": [
|
| 1526 |
+
602,
|
| 1527 |
+
143,
|
| 1528 |
+
751,
|
| 1529 |
+
239
|
| 1530 |
+
],
|
| 1531 |
+
"page_idx": 13
|
| 1532 |
+
},
|
| 1533 |
+
{
|
| 1534 |
+
"type": "text",
|
| 1535 |
+
"text": "in dense fog, MonoWAD effectively detects both close and fog-obscured objects with the aid of the weather codebook and weather-adaptive diffusion model.",
|
| 1536 |
+
"bbox": [
|
| 1537 |
+
212,
|
| 1538 |
+
266,
|
| 1539 |
+
784,
|
| 1540 |
+
297
|
| 1541 |
+
],
|
| 1542 |
+
"page_idx": 13
|
| 1543 |
+
},
|
| 1544 |
+
{
|
| 1545 |
+
"type": "text",
|
| 1546 |
+
"text": "Results on Real-World Images. We further visualize 3D detection results on real-world images from the Seeing Through Fog dataset [2] under various weather conditions (i.e., foggy, rainy, snowy). In Fig. 6, MonoWAD shows robust detection under diverse weather conditions. This demonstrates that our proposed method maintains weather-robustness even in real-world scenarios by dynamically enhancing the input scenarios.",
|
| 1547 |
+
"bbox": [
|
| 1548 |
+
212,
|
| 1549 |
+
304,
|
| 1550 |
+
782,
|
| 1551 |
+
396
|
| 1552 |
+
],
|
| 1553 |
+
"page_idx": 13
|
| 1554 |
+
},
|
| 1555 |
+
{
|
| 1556 |
+
"type": "text",
|
| 1557 |
+
"text": "t-SNE Visualization. We conducted t-SNE visualization to analyze feature representations of MonoDTR, MonoDETR, and our MonoWAD on the KITTI and foggy KITTI validation set. As depicted in Fig. 7, the existing methods exhibit distinct feature representations for foggy and clear weather conditions. In contrast, MonoWAD, leveraging weather-robust feature learning from the weather codebook and weather-adaptive diffusion model, demonstrates similar feature representations for both clear and foggy weather conditions.",
|
| 1558 |
+
"bbox": [
|
| 1559 |
+
212,
|
| 1560 |
+
402,
|
| 1561 |
+
784,
|
| 1562 |
+
508
|
| 1563 |
+
],
|
| 1564 |
+
"page_idx": 13
|
| 1565 |
+
},
|
| 1566 |
+
{
|
| 1567 |
+
"type": "text",
|
| 1568 |
+
"text": "4.7 Limitations",
|
| 1569 |
+
"text_level": 1,
|
| 1570 |
+
"bbox": [
|
| 1571 |
+
214,
|
| 1572 |
+
522,
|
| 1573 |
+
356,
|
| 1574 |
+
536
|
| 1575 |
+
],
|
| 1576 |
+
"page_idx": 13
|
| 1577 |
+
},
|
| 1578 |
+
{
|
| 1579 |
+
"type": "text",
|
| 1580 |
+
"text": "The experimental results show the weather-robustness of our method. However, due to the iterative nature of the diffusion model, our method shows $144\\mathrm{ms}/$ image at timestep $T = 15$ (110ms/image at $T = 5$ ), slower than the latest work, MonoDETR (38ms/image). Moreover, our method dynamically enhances the representation of the input feature based on the weather-reference feature and weather difference which needs paired images. Thus, exploring a method to achieve faster processing speeds while maintaining weather-robust performance without paired images could be an interesting direction for our future work.",
|
| 1581 |
+
"bbox": [
|
| 1582 |
+
212,
|
| 1583 |
+
542,
|
| 1584 |
+
784,
|
| 1585 |
+
662
|
| 1586 |
+
],
|
| 1587 |
+
"page_idx": 13
|
| 1588 |
+
},
|
| 1589 |
+
{
|
| 1590 |
+
"type": "text",
|
| 1591 |
+
"text": "5 Conclusion",
|
| 1592 |
+
"text_level": 1,
|
| 1593 |
+
"bbox": [
|
| 1594 |
+
214,
|
| 1595 |
+
676,
|
| 1596 |
+
359,
|
| 1597 |
+
691
|
| 1598 |
+
],
|
| 1599 |
+
"page_idx": 13
|
| 1600 |
+
},
|
| 1601 |
+
{
|
| 1602 |
+
"type": "text",
|
| 1603 |
+
"text": "We proposed MonoWAD, a novel weather-robust monocular 3D object detector to handle various weather conditions. Addressing challenges in applying existing monocular 3D object detectors to real-world scenarios with various weather, we design a weather codebook with clear knowledge recalling loss to memorize the knowledge of the clear weather and to generate a weather-reference feature from both clear and foggy features. Also, we design a weather-adaptive diffusion model with weather-adaptive enhancement loss to enhance feature representation according to the weather conditions. As a result, our MonoWAD can detect objects occluded by fog and perform well in clear weather.",
|
| 1604 |
+
"bbox": [
|
| 1605 |
+
212,
|
| 1606 |
+
703,
|
| 1607 |
+
785,
|
| 1608 |
+
839
|
| 1609 |
+
],
|
| 1610 |
+
"page_idx": 13
|
| 1611 |
+
},
|
| 1612 |
+
{
|
| 1613 |
+
"type": "page_number",
|
| 1614 |
+
"text": "14",
|
| 1615 |
+
"bbox": [
|
| 1616 |
+
217,
|
| 1617 |
+
114,
|
| 1618 |
+
235,
|
| 1619 |
+
126
|
| 1620 |
+
],
|
| 1621 |
+
"page_idx": 13
|
| 1622 |
+
},
|
| 1623 |
+
{
|
| 1624 |
+
"type": "header",
|
| 1625 |
+
"text": "Y. Oh et al.",
|
| 1626 |
+
"bbox": [
|
| 1627 |
+
271,
|
| 1628 |
+
114,
|
| 1629 |
+
354,
|
| 1630 |
+
126
|
| 1631 |
+
],
|
| 1632 |
+
"page_idx": 13
|
| 1633 |
+
},
|
| 1634 |
+
{
|
| 1635 |
+
"type": "text",
|
| 1636 |
+
"text": "Acknowledgements",
|
| 1637 |
+
"text_level": 1,
|
| 1638 |
+
"bbox": [
|
| 1639 |
+
217,
|
| 1640 |
+
143,
|
| 1641 |
+
401,
|
| 1642 |
+
162
|
| 1643 |
+
],
|
| 1644 |
+
"page_idx": 14
|
| 1645 |
+
},
|
| 1646 |
+
{
|
| 1647 |
+
"type": "text",
|
| 1648 |
+
"text": "This work was supported by the NRF grant funded by the Korea government (MSIT) (No. RS-2023-00252391), and by IITP grant funded by the Korea government (MSIT) (No. RS-2022-00155911: Artificial Intelligence Convergence Innovation Human Resources Development (Kyung Hee University), IITP-2023-RS-2023-00266615: Convergence Security Core Talent Training Business Support Program, No. 2022-0-00124: Development of Artificial Intelligence Technology for Self-Improving Competency-Aware Learning Capabilities).",
|
| 1649 |
+
"bbox": [
|
| 1650 |
+
212,
|
| 1651 |
+
176,
|
| 1652 |
+
787,
|
| 1653 |
+
285
|
| 1654 |
+
],
|
| 1655 |
+
"page_idx": 14
|
| 1656 |
+
},
|
| 1657 |
+
{
|
| 1658 |
+
"type": "text",
|
| 1659 |
+
"text": "References",
|
| 1660 |
+
"text_level": 1,
|
| 1661 |
+
"bbox": [
|
| 1662 |
+
217,
|
| 1663 |
+
306,
|
| 1664 |
+
321,
|
| 1665 |
+
321
|
| 1666 |
+
],
|
| 1667 |
+
"page_idx": 14
|
| 1668 |
+
},
|
| 1669 |
+
{
|
| 1670 |
+
"type": "list",
|
| 1671 |
+
"sub_type": "ref_text",
|
| 1672 |
+
"list_items": [
|
| 1673 |
+
"1. Baranchuk, D., Rubachev, I., Voynov, A., Khrulkov, V., Babenko, A.: Label-efficient semantic segmentation with diffusion models. In: Int. Conf. Learn. Represent. (2021)",
|
| 1674 |
+
"2. Bijelic, M., Gruber, T., Mannan, F., Kraus, F., Ritter, W., Dietmayer, K., Heide, F.: Seeing through fog without seeing fog: Deep multimodal sensor fusion in unseen adverse weather. In: IEEE Conf. Comput. Vis. Pattern Recog. (2020)",
|
| 1675 |
+
"3. Brazil, G., Liu, X.: M3d-rpn: Monocular 3d region proposal network for object detection. In: Int. Conf. Comput. Vis. (2019)",
|
| 1676 |
+
"4. Brazil, G., Pons-Moll, G., Liu, X., Schiele, B.: Kinematic 3d object detection in monocular video. In: Eur. Conf. Comput. Vis. Springer (2020)",
|
| 1677 |
+
"5. Cabon, Y., Murray, N., Humenberger, M.: Virtual kitti 2 (2020)",
|
| 1678 |
+
"6. Chen, X., Kundu, K., Zhang, Z., Ma, H., Fidler, S., Urtasun, R.: Monocular 3d object detection for autonomous driving. In: IEEE Conf. Comput. Vis. Pattern Recog. (2016)",
|
| 1679 |
+
"7. Chen, Y., Tai, L., Sun, K., Li, M.: Monopair: Monocular 3d object detection using pairwise spatial relationships. In: IEEE Conf. Comput. Vis. Pattern Recog. (2020)",
|
| 1680 |
+
"8. Ding, M., Huo, Y., Yi, H., Wang, Z., Shi, J., Lu, Z., Luo, P.: Learning depth-guided convolutions for monocular 3d object detection. In: IEEE Conf. Comput. Vis. Pattern Recog. Worksh. (2020)",
|
| 1681 |
+
"9. Esser, P., Rombach, R., Ommer, B.: Taming transformers for high-resolution image synthesis. In: IEEE Conf. Comput. Vis. Pattern Recog. (2021)",
|
| 1682 |
+
"10. Fu, H., Gong, M., Wang, C., Batmanghelich, K., Tao, D.: Deep ordinal regression network for monocular depth estimation. In: IEEE Conf. Comput. Vis. Pattern Recog. (2018)",
|
| 1683 |
+
"1. Gaidon, A., Wang, Q., Cabon, Y., Vig, E.: Virtual worlds as proxy for multi-object tracking analysis. In: IEEE Conf. Comput. Vis. Pattern Recog. (2016)",
|
| 1684 |
+
"2. Geiger, A., Lenz, P., Urtasun, R.: Are we ready for autonomous driving? the kitti vision benchmark suite. In: IEEE Conf. Comput. Vis. Pattern Recog. (2012). https://doi.org/10.1109/CVPR.2012.6248074",
|
| 1685 |
+
"3. Hahner, M., Sakaridis, C., Dai, D., Van Gool, L.: Fog simulation on real lidar point clouds for 3d object detection in adverse weather. In: Int. Conf. Comput. Vis. (2021)",
|
| 1686 |
+
"4. Hamilton, B., Tefft, B., Arnold, L., Grabowski, J.: Hidden highways: Fog and traffic crashes on america's roads (november 2014). Montana 40 (2006)",
|
| 1687 |
+
"5. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. In: Adv. Neural Inform. Process. Syst. vol. 33 (2020)"
|
| 1688 |
+
],
|
| 1689 |
+
"bbox": [
|
| 1690 |
+
225,
|
| 1691 |
+
339,
|
| 1692 |
+
784,
|
| 1693 |
+
839
|
| 1694 |
+
],
|
| 1695 |
+
"page_idx": 14
|
| 1696 |
+
},
|
| 1697 |
+
{
|
| 1698 |
+
"type": "header",
|
| 1699 |
+
"text": "MonoWAD",
|
| 1700 |
+
"bbox": [
|
| 1701 |
+
653,
|
| 1702 |
+
114,
|
| 1703 |
+
730,
|
| 1704 |
+
126
|
| 1705 |
+
],
|
| 1706 |
+
"page_idx": 14
|
| 1707 |
+
},
|
| 1708 |
+
{
|
| 1709 |
+
"type": "page_number",
|
| 1710 |
+
"text": "15",
|
| 1711 |
+
"bbox": [
|
| 1712 |
+
767,
|
| 1713 |
+
116,
|
| 1714 |
+
784,
|
| 1715 |
+
126
|
| 1716 |
+
],
|
| 1717 |
+
"page_idx": 14
|
| 1718 |
+
},
|
| 1719 |
+
{
|
| 1720 |
+
"type": "list",
|
| 1721 |
+
"sub_type": "ref_text",
|
| 1722 |
+
"list_items": [
|
| 1723 |
+
"16. Huang, K.C., Wu, T.H., Su, H.T., Hsu, W.H.: Monodtr: Monocular 3d object detection with depth-aware transformer. In: IEEE Conf. Comput. Vis. Pattern Recog. (2022)",
|
| 1724 |
+
"17. Juneja, A., Kumar, V., Singla, S.K.: A systematic review on foggy datasets: Applications and challenges. Archives of Computational Methods in Engineering 29(3) (2022)",
|
| 1725 |
+
"18. Kim, J.U., Kim, H.I., Ro, Y.M.: Stereoscopic vision recalling memory for monocular 3d object detection. IEEE Trans. Image Process. 32 (2023). https://doi.org/10.1109/TIP.2023.3274479",
|
| 1726 |
+
"19. Kim, J.U., Park, S., Ro, Y.M.: Robust small-scale pedestrian detection with cued recall via memory learning. In: Int. Conf. Comput. Vis. (2021)",
|
| 1727 |
+
"20. Kim, J.U., Park, S., Ro, Y.M.: Uncertainty-guided cross-modal learning for robust multispectral pedestrian detection. IEEE Trans. Circuit Syst. Video Technol. 32(3) (2021)",
|
| 1728 |
+
"21. Kim, J.U., Park, S., Ro, Y.M.: Towards versatile pedestrian detector with multisensory-matching and multispectral recalling memory. In: AAAI. vol. 36 (2022)",
|
| 1729 |
+
"22. Kim, S.W., Brown, B., Yin, K., Kreis, K., Schwarz, K., Li, D., Rombach, R., Torralba, A., Fidler, S.: Neuralfield-ldm: Scene generation with hierarchical latent diffusion models. In: IEEE Conf. Comput. Vis. Pattern Recog. (2023)",
|
| 1730 |
+
"23. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014)",
|
| 1731 |
+
"24. Lang, A.H., Vora, S., Caesar, H., Zhou, L., Yang, J., Beijbom, O.: Pointpillars: Fast encoders for object detection from point clouds. In: IEEE Conf. Comput. Vis. Pattern Recog. (2019)",
|
| 1732 |
+
"25. Liu, X., Xue, N., Wu, T.: Learning auxiliary monocular contexts helps monocular 3d object detection. In: AAAI. vol. 36 (2022)",
|
| 1733 |
+
"26. Liu, Z., Wu, Z., Toth, R.: Smoke: Single-stage monocular 3d object detection via keypoint estimation. In: IEEE Conf. Comput. Vis. Pattern Recog. Worksh. (2020)",
|
| 1734 |
+
"27. Liu, Z., Zhou, D., Lu, F., Fang, J., Zhang, L.: Autoshape: Real-time shape-aware monocular 3d object detection. In: Int. Conf. Comput. Vis. (2021)",
|
| 1735 |
+
"28. Lu, Y., Ma, X., Yang, L., Zhang, T., Liu, Y., Chu, Q., Yan, J., Ouyang, W.: Geometry uncertainty projection network for monocular 3d object detection. In: Int. Conf. Comput. Vis. (2021)",
|
| 1736 |
+
"29. Lugmayr, A., Danelljan, M., Romero, A., Yu, F., Timofte, R., Van Gool, L.: Repaint: Inpainting using denoising diffusion probabilistic models. In: IEEE Conf. Comput. Vis. Pattern Recog. (2022)",
|
| 1737 |
+
"30. Luo, J., Li, Y., Pan, Y., Yao, T., Feng, J., Chao, H., Mei, T.: Semantic-conditional diffusion networks for image captioning. In: IEEE Conf. Comput. Vis. Pattern Recog. (2023)",
|
| 1738 |
+
"31. Luo, S., Dai, H., Shao, L., Ding, Y.: M3dssd: Monocular 3d single stage object detector. In: IEEE Conf. Comput. Vis. Pattern Recog. (2021)",
|
| 1739 |
+
"32. Ma, X., Liu, S., Xia, Z., Zhang, H., Zeng, X., Ouyang, W.: Rethinking pseudo-lidar representation. In: Eur. Conf. Comput. Vis. Springer (2020)",
|
| 1740 |
+
"33. Mai, N.A.M., Duthon, P., Khoudour, L., Crouzil, A., Velastin, S.A.: 3d object detection with sls-fusion network in foggy weather conditions. Sensors 21(20) (2021)",
|
| 1741 |
+
"34. Mousavian, A., Anguelov, D., Flynn, J., Kosecka, J.: 3d bounding box estimation using deep learning and geometry. In: IEEE Conf. Comput. Vis. Pattern Recog. (2017)"
|
| 1742 |
+
],
|
| 1743 |
+
"bbox": [
|
| 1744 |
+
215,
|
| 1745 |
+
146,
|
| 1746 |
+
784,
|
| 1747 |
+
839
|
| 1748 |
+
],
|
| 1749 |
+
"page_idx": 15
|
| 1750 |
+
},
|
| 1751 |
+
{
|
| 1752 |
+
"type": "page_number",
|
| 1753 |
+
"text": "16",
|
| 1754 |
+
"bbox": [
|
| 1755 |
+
217,
|
| 1756 |
+
114,
|
| 1757 |
+
235,
|
| 1758 |
+
126
|
| 1759 |
+
],
|
| 1760 |
+
"page_idx": 15
|
| 1761 |
+
},
|
| 1762 |
+
{
|
| 1763 |
+
"type": "header",
|
| 1764 |
+
"text": "Y. Oh et al.",
|
| 1765 |
+
"bbox": [
|
| 1766 |
+
271,
|
| 1767 |
+
114,
|
| 1768 |
+
352,
|
| 1769 |
+
126
|
| 1770 |
+
],
|
| 1771 |
+
"page_idx": 15
|
| 1772 |
+
},
|
| 1773 |
+
{
|
| 1774 |
+
"type": "list",
|
| 1775 |
+
"sub_type": "ref_text",
|
| 1776 |
+
"list_items": [
|
| 1777 |
+
"35. Ni, H., Shi, C., Li, K., Huang, S.X., Min, M.R.: Conditional image-to-video generation with latent flow diffusion models. In: IEEE Conf. Comput. Vis. Pattern Recog. (2023)",
|
| 1778 |
+
"36. van den Oord, A., Vinyals, O., Kavukcuoglu, K.: Neural discrete representation learning (2018)",
|
| 1779 |
+
"37. Peng, L., Wu, X., Yang, Z., Liu, H., Cai, D.: Did-m3d: Decoupling instance depth for monocular 3d object detection. In: Eur. Conf. Comput. Vis. Springer (2022)",
|
| 1780 |
+
"38. Pfeuffer, A., Dietmayer, K.: Robust semantic segmentation in adverse weather conditions by means of sensor data fusion. In: 2019 22th International Conference on Information Fusion (FUSION) (2019). https://doi.org/10.23919/FUSION43075.2019.9011192",
|
| 1781 |
+
"39. Qin, Q., Chang, K., Huang, M., Li, G.: Denet: Detection-driven enhancement network for object detection under adverse weather conditions. In: Asian Conf. Comput. Vis. (2022)",
|
| 1782 |
+
"40. Qin, Z., Li, X.: Monoground: Detecting monocular 3d objects from the ground. In: IEEE Conf. Comput. Vis. Pattern Recog. (2022)",
|
| 1783 |
+
"41. Reading, C., Harakeh, A., Chae, J., Waslander, S.L.: Categorical depth distribution network for monocular 3d object detection. In: IEEE Conf. Comput. Vis. Pattern Recog. (2021)",
|
| 1784 |
+
"42. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: IEEE Conf. Comput. Vis. Pattern Recog. (2022)",
|
| 1785 |
+
"43. Sakaridis, C., Dai, D., Van Gool, L.: Semantic foggy scene understanding with synthetic data. Int. J. Comput. Vis. 126 (2018)",
|
| 1786 |
+
"44. Shi, S., Guo, C., Jiang, L., Wang, Z., Shi, J., Wang, X., Li, H.: Pv-rcnn: Point-voxel feature set abstraction for 3d object detection. In: IEEE Conf. Comput. Vis. Pattern Recog. (2020)",
|
| 1787 |
+
"45. Shi, S., Wang, X., Li, H.: Pointcnn: 3d object proposal generation and detection from point cloud. In: IEEE Conf. Comput. Vis. Pattern Recog. (2019)",
|
| 1788 |
+
"46. Shi, X., Ye, Q., Chen, X., Chen, C., Chen, Z., Kim, T.K.: Geometry-based distance decomposition for monocular 3d object detection. In: Int. Conf. Comput. Vis. (2021)",
|
| 1789 |
+
"47. Simonelli, A., Bulo, S.R., Porzi, L., Lopez-Antequera, M., Kontschieder, P.: Disentangling monocular 3d object detection. In: Int. Conf. Comput. Vis. (2019)",
|
| 1790 |
+
"48. Wang, L., Du, L., Ye, X., Fu, Y., Guo, G., Xue, X., Feng, J., Zhang, L.: Depth-conditioned dynamic message propagation for monocular 3d object detection. In: IEEE Conf. Comput. Vis. Pattern Recog. (2021)",
|
| 1791 |
+
"49. Wu, R.Q., Duan, Z.P., Guo, C.L., Chai, Z., Li, C.: Ridcp: Revitalizing real image dehazing via high-quality codebook priors. In: IEEE Conf. Comput. Vis. Pattern Recog. (2023)",
|
| 1792 |
+
"50. Yang, S., Scherer, S.: Cubeslam: Monocular 3-d object slam. IEEE Transactions on Robotics 35(4) (2019). https://doi.org/10.1109/TR0.2019.2909168",
|
| 1793 |
+
"51. Yang, X., Mi, M.B., Yuan, Y., Wang, X., Tan, R.T.: Object detection in foggy scenes by embedding depth and reconstruction into domain adaptation. In: Asian Conf. Comput. Vis. (2022)",
|
| 1794 |
+
"52. Yu, F., Wang, D., Shelhamer, E., Darrell, T.: Deep layer aggregation. In: IEEE Conf. Comput. Vis. Pattern Recog. (2018)",
|
| 1795 |
+
"53. Yurtsever, E., Lambert, J., Carballo, A., Takeda, K.: A survey of autonomous driving: Common practices and emerging technologies. IEEE Access 8 (2020). https://doi.org/10.1109/ACCESS.2020.2983149"
|
| 1796 |
+
],
|
| 1797 |
+
"bbox": [
|
| 1798 |
+
212,
|
| 1799 |
+
146,
|
| 1800 |
+
787,
|
| 1801 |
+
840
|
| 1802 |
+
],
|
| 1803 |
+
"page_idx": 16
|
| 1804 |
+
},
|
| 1805 |
+
{
|
| 1806 |
+
"type": "header",
|
| 1807 |
+
"text": "MonoWAD",
|
| 1808 |
+
"bbox": [
|
| 1809 |
+
653,
|
| 1810 |
+
114,
|
| 1811 |
+
730,
|
| 1812 |
+
126
|
| 1813 |
+
],
|
| 1814 |
+
"page_idx": 16
|
| 1815 |
+
},
|
| 1816 |
+
{
|
| 1817 |
+
"type": "page_number",
|
| 1818 |
+
"text": "17",
|
| 1819 |
+
"bbox": [
|
| 1820 |
+
767,
|
| 1821 |
+
114,
|
| 1822 |
+
785,
|
| 1823 |
+
126
|
| 1824 |
+
],
|
| 1825 |
+
"page_idx": 16
|
| 1826 |
+
},
|
| 1827 |
+
{
|
| 1828 |
+
"type": "list",
|
| 1829 |
+
"sub_type": "ref_text",
|
| 1830 |
+
"list_items": [
|
| 1831 |
+
"54. Zhang, C., Wang, H., Cai, Y., Chen, L., Li, Y., Sotelo, M.A., Li, Z.: Robust-fusionnet: Deep multimodal sensor fusion for 3-d object detection under severe weather conditions. IEEE Transactions on Instrumentation and Measurement 71 (2022). https://doi.org/10.1109/TIM.2022.3191724",
|
| 1832 |
+
"55. Zhang, R., Qiu, H., Wang, T., Guo, Z., Cui, Z., Qiao, Y., Li, H., Gao, P.: Monodetr: Depth-guided transformer for monocular 3d object detection. In: Int. Conf. Comput. Vis. (2023)",
|
| 1833 |
+
"56. Zhou, Y., Tuzel, O.: Voxelnet: End-to-end learning for point cloud based 3d object detection. In: IEEE Conf. Comput. Vis. Pattern Recog. (2018)",
|
| 1834 |
+
"57. Zhou, Y., He, Y., Zhu, H., Wang, C., Li, H., Jiang, Q.: Monocular 3d object detection: An extrinsic parameter free approach. In: IEEE Conf. Comput. Vis. Pattern Recog. (2021)"
|
| 1835 |
+
],
|
| 1836 |
+
"bbox": [
|
| 1837 |
+
215,
|
| 1838 |
+
146,
|
| 1839 |
+
785,
|
| 1840 |
+
314
|
| 1841 |
+
],
|
| 1842 |
+
"page_idx": 17
|
| 1843 |
+
},
|
| 1844 |
+
{
|
| 1845 |
+
"type": "page_number",
|
| 1846 |
+
"text": "18",
|
| 1847 |
+
"bbox": [
|
| 1848 |
+
217,
|
| 1849 |
+
114,
|
| 1850 |
+
235,
|
| 1851 |
+
126
|
| 1852 |
+
],
|
| 1853 |
+
"page_idx": 17
|
| 1854 |
+
},
|
| 1855 |
+
{
|
| 1856 |
+
"type": "header",
|
| 1857 |
+
"text": "Y. Oh et al.",
|
| 1858 |
+
"bbox": [
|
| 1859 |
+
271,
|
| 1860 |
+
114,
|
| 1861 |
+
352,
|
| 1862 |
+
126
|
| 1863 |
+
],
|
| 1864 |
+
"page_idx": 17
|
| 1865 |
+
}
|
| 1866 |
+
]
|
2024/MonoWAD_ Weather-Adaptive Diffusion Model for Robust Monocular 3D Object Detection/4d73977a-6d19-47be-84e6-b3aec99a372e_model.json
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
2024/MonoWAD_ Weather-Adaptive Diffusion Model for Robust Monocular 3D Object Detection/4d73977a-6d19-47be-84e6-b3aec99a372e_origin.pdf
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:10ee0c2027c15d1c0979b668e589c704a6c74cbf0d253e646523fc170187a059
|
| 3 |
+
size 3334913
|
2024/MonoWAD_ Weather-Adaptive Diffusion Model for Robust Monocular 3D Object Detection/full.md
ADDED
|
@@ -0,0 +1,337 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# MonoWAD: Weather-Adaptive Diffusion Model for Robust Monocular 3D Object Detection
|
| 2 |
+
|
| 3 |
+
Youngmin Oh<sup>1</sup>, Hyung-Il Kim<sup>2</sup>, Seong Tae Kim<sup>1†</sup>, and Jung Uk Kim<sup>1†</sup>
|
| 4 |
+
|
| 5 |
+
1 Kyung Hee University, Yong-in, South Korea {oym9104, st.kim, ju.kim}@khu.ac.kr 2 ETRI, Daejeon, South Korea hikim@etri.re.kr
|
| 6 |
+
|
| 7 |
+
Abstract. Monocular 3D object detection is an important challenging task in autonomous driving. Existing methods mainly focus on performing 3D detection in ideal weather conditions, characterized by scenarios with clear and optimal visibility. However, the challenge of autonomous driving requires the ability to handle changes in weather conditions, such as foggy weather, not just clear weather. We introduce MonoWAD, a novel weather-robust monocular 3D object detector with a weather-adaptive diffusion model. It contains two components: (1) the weather codebook to memorize the knowledge of the clear weather and generate a weather-reference feature for any input, and (2) the weather-adaptive diffusion model to enhance the feature representation of the input feature by incorporating a weather-reference feature. This serves as an attention role in indicating how much improvement is needed for the input feature according to the weather conditions. To achieve this goal, we introduce a weather-adaptive enhancement loss to enhance the feature representation under both clear and foggy weather conditions. Extensive experiments under various weather conditions demonstrate that MonoWAD achieves weather-robust monocular 3D object detection. The code and dataset are released at https://github.com/VisualAIKHU/MonoWAD.
|
| 8 |
+
|
| 9 |
+
Keywords: Monocular 3D Object Detection $\cdot$ Weather-Adaptive Diffusion $\cdot$ Weather Codebook
|
| 10 |
+
|
| 11 |
+
# 1 Introduction
|
| 12 |
+
|
| 13 |
+
Monocular 3D object detection aims to detect 3D objects using only a single camera [16, 18, 28, 46, 55, 57]. In contrast to LiDAR-based methods that rely on expensive LiDAR sensors for depth estimation [24, 44, 45, 56], and stereo-based methods that require synchronized stereo cameras, monocular 3D object detection only requires monocular images, offering the advantage of computational cost-effectiveness and requiring fewer resources. Due to this characteristic, the monocular 3D object detection is applied to a wide range of real-world applications, such as autonomous driving [12, 19, 53] and robotics [50].
|
| 14 |
+
|
| 15 |
+

|
| 16 |
+
Fig. 1: Conceptual diagram of the proposed method (foggy example). (a) In the training phase, weather codebook learns the clear knowledge to transfer it to weather-adaptive diffusion model to enhance content related to the weather conditions. (b) By doing so, even with input images under various weather conditions (e.g., foggy images), monocular 3D object detection becomes adaptable to various weather scenarios.
|
| 17 |
+
|
| 18 |
+
However, existing monocular 3D object detectors mainly focus on ideal autonomous driving environments (i.e., clear weather). There are challenges in applying them to real-world scenarios with adverse weather conditions, such as fog and rain. Among these, fog poses the most significant challenge compared to other weather [14, 17]. This is due to the dense and diffuse nature of fog, which strongly scatters and absorbs light, leading to difficulties in object detection [2]. Since monocular 3D object detection relies solely on visual information from a single image, unlike LiDAR, it is crucial to design detectors to achieve enhanced performance in challenging visibility scenarios.
|
| 19 |
+
|
| 20 |
+
In this paper, we propose MonoWAD, a novel weather-robust monocular 3D object detector to address the aforementioned issues. As mentioned earlier, due to the inherent challenges posed by foggy weather among various adverse weather conditions, we focus primarily on clear and foggy weather (results for other weather conditions such as rainy and sunset are presented in Section 4). For weather-robust object detection, clear weather requires relatively modest improvements to enhance the visual representation of features. In contrast, significant enhancements in this feature representation are required for foggy weather. To address this, we consider two key aspects: (1) how to quantify the degree of improvement needed for the input image, and (2) how to guide the representation of the input image.
|
| 21 |
+
|
| 22 |
+
To address the two key aspects, as shown in Fig. 1, our MonoWAD consists of a weather codebook and weather-adaptive diffusion model. First, we introduce the weather codebook to generate a weather-reference feature that contains knowledge about reference weather in a given scene. The reference weather acts as a guide, indicating the degree of weather improvement required. Since the clear weather contains a richer visual representation of objects, we adopt it as a reference weather. At this time, we devise a clear knowledge recalling (CKR) loss to guide the weather codebook to memorize information about clear weather and generate a weather-reference feature for any input (clear or foggy). As a result, our detector can understand where improvements are needed in the input features based on the weather-reference feature.
|
| 23 |
+
|
| 24 |
+
Second, we propose a weather-adaptive diffusion model to effectively enhance feature representations in accordance with weather conditions. Given input fea-
|
| 25 |
+
|
| 26 |
+
ture (clear or foggy), the weather-adaptive diffusion model dynamically enhances the representation of the input feature based on the weather-reference feature. The weather-reference feature plays a role of attention, determining the extent to which the input features need improvement. At this time, we define the difference between clear and foggy weather (i.e., weather changes) as the fog distribution to adopt it as the noise for our diffusion model. With fog distribution, our weather-adaptive diffusion model can enhance the feature representation according to the weather conditions through multiple steps of reverse processes. To achieve this goal, we introduce a weather-adaptive enhancement (WAE) loss. As a result, our MonoWAD performs weather-robust detection by adaptively improving feature representation according to the weather conditions.
|
| 27 |
+
|
| 28 |
+
To adaptively enhance the feature representation through the difference between weather conditions, we generate a new foggy KITTI dataset based on the KITTI dataset [12]. Comprehensive experimental results on several datasets [11, 12] show that our MonoWAD outperforms the existing state-of-the-art monocular 3D object detectors [16,28,37,40,55] under foggy weather, which is the most challenging weather condition. While our method primarily focuses on foggy, experiments conducted under various weather conditions (e.g., foggy, rainy, and sunset) have demonstrated its applicability to other weather scenarios.
|
| 29 |
+
|
| 30 |
+
The main contributions of our paper can be summarized as follows:
|
| 31 |
+
|
| 32 |
+
- We introduce a new weather-robust monocular 3D object detector, called MonoWAD, that is robust to various weather conditions.
|
| 33 |
+
- We design a weather codebook with clear knowledge recalling loss for learning about clear weather, providing reference information for enhancement.
|
| 34 |
+
- We propose weather-adaptive diffusion model with weather-adaptive enhancement loss to dynamically enhance the feature representation of the input images according to the weather conditions.
|
| 35 |
+
|
| 36 |
+
# 2 Related Work
|
| 37 |
+
|
| 38 |
+
# 2.1 Monocular 3D Object Detection
|
| 39 |
+
|
| 40 |
+
Monocular 3D object detection task can be categorized into two directions according to the type of data used in the training phase: (1) using only a monocular image and (2) incorporating additional data, such as depth along with a monocular image. The first category relies on the geometric relationship between 2D and 3D [3,7,25,26,28,31,34,40,46,55,57]. For example, Deep3Dbox [34] utilizes the geometric information of 2D bounding boxes to predict 3D bounding boxes. In [3], M3D-RPN was proposed to understand the 3D scene from the depth-aware convolution. MonoRCNN [46] predicted 3D bounding boxes through geometry-based distance decomposition, and MonoCon [25] learns mono context for 3D object detection. MonoDETR [55] introduces a depth-guided transformer that utilizes geometric depth cues without requiring additional data.
|
| 41 |
+
|
| 42 |
+
Moreover, since monocular image contains limited information for estimating 3D object cues, monocular 3D object detectors have adopted the additional
|
| 43 |
+
|
| 44 |
+
data for more robust detection [4,8,16,27,32,37,41,48]. For example, the depth-conditioned dynamic message propagation (DDMP) [48] was proposed to integrate prior depth information with the image context. CaDDN [41] was introduced to utilize the depth distribution of predicted categories for each pixel to project the context information onto 3D space, deriving 3D bounding boxes. MonoDTR [16] employs the transformer architecture to integrate depth features and context, thus estimating more accurate depth information.
|
| 45 |
+
|
| 46 |
+
Despite recent progress, the existing monocular 3D object detectors mainly rely on the benchmark data collected under clear weather conditions. However, it is essential to account for challenging adverse weather conditions, such as fog, to more accurately reflect real-world scenarios. In this paper, we aim to introduce weather-robust 3D object detection by enhancing visual features with the proposed weather-adaptive diffusion model and weather codebook.
|
| 47 |
+
|
| 48 |
+
# 2.2 Computer Vision Tasks for Foggy Weather
|
| 49 |
+
|
| 50 |
+
There have been a lot of studies on improving performance in various weather conditions for real-world application of computer vision technology [2, 13, 20, 21, 33, 38, 43, 51, 54]. In particular, fog is considered one of the most critical issues due to its significant degradation of visual information compared to other weather conditions. To deal with the foggy weather conditions, the authors in [43] generated synthetic fog images (so-called Foggy Cityscapes) from clear weather images, and utilized them for training semantic segmentation and 2D object detection. Similar to Foggy Cityscapes, Martin et al. [13] naturally synthesizes fog into LIDAR to enhance the performance of LiDAR-based 3D object detectors in foggy weather. Bijelic et al. [2] use real dataset including foggy weather, for 2D detection with the multimodal fusion networks. Mai et al. [33] synthesize fog for LiDAR and stereo images and perform fusion-based 3D object detection using the SLS-Fusion network. Xin et al. [51] focused on 2D detection in foggy weather by applying domain adaptation. In this context, we focus on monocular 3D object detection in scenarios that rely solely on visual information in challenging foggy environments. To address this challenge, we propose a weather-robust diffusion model that dynamically improves features based on reference feature.
|
| 51 |
+
|
| 52 |
+
# 2.3 Diffusion Models
|
| 53 |
+
|
| 54 |
+
Recently, the diffusion model [15] has attracted considerable interest in computer vision due to its impressive progress in image generation [22, 29, 35, 42], as well as its potential application to other vision tasks such as segmentation [1] and image captioning [30]. Inspired by the remarkable generative ability, we design the diffusion model for robust monocular 3D object detection by considering a foggy effect (which is one of the challenging adverse weather conditions for monocular 3D object detection) as a form of noise in the model. That is, we propose a method in which visual features obscured by fog are progressively improved by training a diffusion model based on the forward/reverse diffusion process. In particular, we present an adaptive method that allows the diffusion model to control the degree of improvement by weather conditions.
|
| 55 |
+
|
| 56 |
+

|
| 57 |
+
Fig. 2: Overview of our MonoWAD in the inference phase. It mainly contains three parts: weather codebook, weather-adaptive diffusion model, and detection block. Through the weather codebook and weather-adaptive diffusion model, our method can maintain robustness against various weather conditions (i.e., clear or foggy).
|
| 58 |
+
|
| 59 |
+
# 3 Proposed Method
|
| 60 |
+
|
| 61 |
+
Fig. 2 shows the overall framework of the proposed MonoWAD in the inference phase. A backbone network receives an input image (clear image $I_{c}$ or foggy image $I_{f}$ ) to encode the corresponding input feature (input clear feature $x^{c}$ or input foggy feature $x^{f}$ ). It interacts with the weather codebook $\mathcal{Z}$ to generate weather-reference feature $x^{r}$ , indicating the amount of enhancement for the given input feature. Subsequently, the weather-adaptive diffusion model attempts to enhance the input feature over $T$ timesteps to obtain an enhanced feature $\tilde{x}^{c}$ or $\tilde{x}^{f}$ . Finally, monocular 3D object detection is performed through the detection block. Note that, in the training phase, our MonoWAD use the clear feature $x^{c}$ to train our diffusion as well as the detection block.
|
| 62 |
+
|
| 63 |
+
We address two key issues: (1) how to guide weather-reference feature $x^r$ to serve as a reference feature, and (2) how to guide the weather-adaptive diffusion model to effectively enhance feature representations based on weather conditions. Details are in the following subsections.
|
| 64 |
+
|
| 65 |
+
# 3.1 Weather Codebook
|
| 66 |
+
|
| 67 |
+
In foggy weather conditions, the overall visual quality of the scene is generally poor, requiring a significant enhancement. Conversely, in clear weather, the amount of improvement is expected to be relatively minimal compared to foggy conditions. Therefore, inspired by [9,36], we devise a weather codebook $\mathcal{Z}$ to provide the reference knowledge about the weather for appropriate enhancement based on the weather conditions. At this time, as the clear weather contains abundant visual representations, we use it as a reference weather knowledge.
|
| 68 |
+
|
| 69 |
+
As shown in Fig. 3, the reference knowledge embedding procedure involves receiving paired clear-foggy features during the training phase. The weather codebook $\mathcal{Z}$ consists of $K$ learnable slots, denoted as $\mathcal{Z} = \{z_k\}_{k=1}^K (z_k \in \mathbb{R}^{1 \times c})$ , where $c$ represents the dimensionality of each slot. The paired clear feature $x^c$ and foggy feature $x^f$ pass through a convolution layer to generate $\hat{x}^c \in \mathbb{R}^{h \times w \times c}$ and $\hat{x}^f \in \mathbb{R}^{h \times w \times c}$ ( $w$ denotes width and $h$ indicates height). Each element of the
|
| 70 |
+
|
| 71 |
+

|
| 72 |
+
Fig. 3: Illustration of the proposed (a) weather-invariant guiding (WIG) loss and (b) clear knowledge embedding (CKE) loss. The clear knowledge recalling (CKR) loss, obtained from combining WIG and CKE, aims to memorize the knowledge of the clear weather and recall the same clear knowledge from the foggy weather.
|
| 73 |
+
|
| 74 |
+
feature denoted as $\hat{x}_{ij}^{c}\in \mathbb{R}^{1\times c}$ and $\hat{x}_{ij}^{f}\in \mathbb{R}^{1\times c}$ . Subsequently, we obtain weather-reference feature for clear weather, $x^{r(c)}\in \mathbb{R}^{h\times w\times c}$ , by conducting element-wise quantization process $\mathbf{q}(\cdot)$ , calculated as:
|
| 75 |
+
|
| 76 |
+
$$
|
| 77 |
+
x ^ {r (c)} = \mathbf {q} (\hat {x} ^ {c}) := \left(\underset {z _ {k} \in \mathcal {Z}} {\arg \min } \| \hat {x} _ {i j} ^ {c} - z _ {k} \|\right). \tag {1}
|
| 78 |
+
$$
|
| 79 |
+
|
| 80 |
+
Utilizing $x^{c}$ and $x^{r(c)}$ , we introduce a clear knowledge embedding (CKE) loss $\mathcal{L}_{cke}$ to guide $x^{r(c)}$ to follow the representation of $x^{c}$ . To this end, we perform global average pooling (GAP) for $x^{c}$ and $x^{r(c)}$ with softmax, generating $s^{c}$ and $s^{r(c)}$ , respectively. Each element in the vector indicates the probability of the significance of each channel. With $s^{c}$ and $s^{r(c)}$ , we employ KL divergence $D_{KL}(\cdot)$ for $\mathcal{L}_{cke}$ to compare the probability distributions, formulated as:
|
| 81 |
+
|
| 82 |
+
$$
|
| 83 |
+
\mathcal {L} _ {c k e} = D _ {K L} \left(s ^ {c} \| s ^ {r (c)}\right). \tag {2}
|
| 84 |
+
$$
|
| 85 |
+
|
| 86 |
+
Through $\mathcal{L}_{cke}$ , the weather codebook $\mathcal{Z}$ can memorize the knowledge of the clear weather, allowing it to effectively reconstruct the clear weather knowledge.
|
| 87 |
+
|
| 88 |
+
Additionally, as the paired clear-foggy images are identical except for weather conditions, the quantization process of the foggy feature $x^{f}$ with the weather codebook should generate an equivalent weather-reference feature. For obtaining weather-reference feature for foggy $x^{r(f)}$ , the element-wise quantization process is also conducted for $\hat{x}^{f}$ and $\mathcal{Z}$ :
|
| 89 |
+
|
| 90 |
+
$$
|
| 91 |
+
x ^ {r (f)} = \mathbf {q} (\hat {x} ^ {f}) := \left(\underset {z _ {k} \in \mathcal {Z}} {\arg \min } \| \hat {x} _ {i j} ^ {f} - z _ {k} \|\right). \tag {3}
|
| 92 |
+
$$
|
| 93 |
+
|
| 94 |
+
Next, we introduce the weather-invariant guiding (WIG) loss $\mathcal{L}_{wig}$ to guide $\mathcal{Z}$ that the weather codebook recalls the same clear knowledge of clear feature for
|
| 95 |
+
|
| 96 |
+

|
| 97 |
+
Fig. 4: Training process of the weather-adaptive diffusion model, which consists of two processes: (1) Adding fog variant $\epsilon_{n}$ from input clear feature $x^{c}$ (forward process) and (2) enhancing representation with weather-reference feature $x^{r}$ (reverse process).
|
| 98 |
+
|
| 99 |
+
the foggy feature, which can be represented as follows:
|
| 100 |
+
|
| 101 |
+
$$
|
| 102 |
+
\mathcal {L} _ {w i g} = \left\| x ^ {r (c)} - x ^ {r (f)} \right\| _ {2} ^ {2}. \tag {4}
|
| 103 |
+
$$
|
| 104 |
+
|
| 105 |
+
Finally, the clear knowledge recalling (CKR) loss $\mathcal{L}_{ckr}$ is obtained by adding $\mathcal{L}_{cke}$ and $\mathcal{L}_{wig}$ , which is defined as:
|
| 106 |
+
|
| 107 |
+
$$
|
| 108 |
+
\mathcal {L} _ {c k r} = \mathcal {L} _ {c k e} + \mathcal {L} _ {w i g}. \tag {5}
|
| 109 |
+
$$
|
| 110 |
+
|
| 111 |
+
In the training phase, the weight parameters of embedding $K$ slots of weather codebook $\mathcal{Z}$ are initialized randomly and they are updated through Eq. (5). In the inference phase, all parameters are fixed to recall clear weather, generating weather-reference features for any weather conditions.
|
| 112 |
+
|
| 113 |
+
# 3.2 Weather-Adaptive Diffusion Model
|
| 114 |
+
|
| 115 |
+
Through Section 3.1, we now know that the weather codebook outputs the weather-reference feature $x^{r(c)}$ for the clear weather and $x^{r(f)}$ for the foggy weather through Eq. (5). From now on, since our method can receive any input images (clear or foggy), we denote the weather-reference feature as $x^r$ .
|
| 116 |
+
|
| 117 |
+
Fig. 4 shows the training process of the proposed weather-adaptive diffusion model. The key idea of the diffusion model [15,42] is to gradually enhance $x_{T}$ to $x_{0}$ with fixed Markov Chain of $T$ timesteps. To this end, the forward and reverse processes are conducted in the training phase, and only the reverse process is used for the inference phase. Motivated by [15,42], we construct the weather-adaptive diffusion model to enhance representation related to the weather conditions. To this end, unlike traditional diffusion methods [15,42] that adopt the Gaussian noise to the image or latent space, we adopt $\mathcal{F} = x^{f} - x^{c}$ , called fog distribution, to guide our diffusion model to be aware of the foggy weather. Ideally, the information contained within $\mathcal{F}$ should include the information of fog, as it represents the difference between the foggy scene and the same scene with clear weather.
|
| 118 |
+
|
| 119 |
+
By doing so, our diffusion model learns the variation in weather by repeatedly adding and removing the fog. Note that, as we take clear feature $x^{c}$ for $x_{0}$ to make a reference input for our diffusion model, we newly denote $x_{0}$ as $x_{0}^{c}$ .
|
| 120 |
+
|
| 121 |
+
For the forward process at the $t$ -th timestep, $q(x_{t}^{c}|x_{t - 1}^{c})$ takes the previous feature $x_{t - 1}^{c}$ and the noise related to the fog (i.e., $\mathcal{F}$ ) as inputs to generate $x_{t}^{c}$ . This procedure is repeated over $T$ timesteps, which can be represented as:
|
| 122 |
+
|
| 123 |
+
$$
|
| 124 |
+
q (x _ {t} ^ {c} | x _ {t - 1} ^ {c}) = \mathcal {F} (x _ {t} ^ {c}; \sqrt {1 - \beta_ {t}} x _ {t - 1} ^ {c}, \beta_ {t} \mathbf {I}), \tag {6}
|
| 125 |
+
$$
|
| 126 |
+
|
| 127 |
+
$$
|
| 128 |
+
q \left(x _ {1} ^ {c}, \dots , x _ {T} ^ {c} \mid x ^ {c}\right) = \prod_ {t = 1} ^ {T} q \left(x _ {t} ^ {c} \mid x _ {t - 1} ^ {c}\right), \tag {7}
|
| 129 |
+
$$
|
| 130 |
+
|
| 131 |
+
where $\beta_{t}$ denotes the variance schedule.
|
| 132 |
+
|
| 133 |
+
Next, the reverse process at the $t$ -th timestep aims to estimate fog variant $\epsilon_{n}$ using $x_{t}^{c}$ to enhance the feature representation of the foggy. To this end, we adopt conditional autoencoder $\epsilon_{\theta}(x_{t}^{c}, t, x^{r})$ that receives $x_{t}^{c}$ and reference feature $x^{r}$ from the weather codebook $\mathcal{Z}$ . Specifically, $\epsilon_{\theta}$ estimates mean $\mu_{\theta}$ and variance $\pmb{\Sigma}_{\theta}$ of the fog distribution at the $t$ -th timestep, denoted as $\tilde{\mathcal{F}}(\cdot)$ . The reverse process is also repeated over $T$ timesteps, which can be represented as:
|
| 134 |
+
|
| 135 |
+
$$
|
| 136 |
+
p _ {\theta} \left(x _ {t - 1} ^ {c} \mid x _ {t} ^ {c}, x ^ {r}\right) = \tilde {\mathcal {F}} \left(x _ {t - 1} ^ {c}; \boldsymbol {\mu} _ {\theta} \left(x _ {t} ^ {c}, t, x ^ {r}\right), \boldsymbol {\Sigma} _ {\theta} \left(x _ {t} ^ {c}, t\right)\right), \tag {8}
|
| 137 |
+
$$
|
| 138 |
+
|
| 139 |
+
$$
|
| 140 |
+
p _ {\theta} \left(x ^ {c}, \dots , x _ {T} ^ {c}\right) = p \left(x _ {T} ^ {c}\right) \prod_ {t = 1} ^ {T} p _ {\theta} \left(x _ {t - 1} ^ {c} \mid x _ {t} ^ {c}, x ^ {r}\right), \tag {9}
|
| 141 |
+
$$
|
| 142 |
+
|
| 143 |
+
where $p_{\theta}(x_{t - 1}^{c}|x_{t}^{c},x^{r})$ includes cross-attention layer in $\epsilon_{\theta}$
|
| 144 |
+
|
| 145 |
+
The cross-attention layer receives the flattened clear feature $x_{t}^{c}$ and weather-reference feature $x^{r}$ , i.e., $\bar{x}_{t}^{c}$ and $\bar{x}^{r}$ , respectively. The similarity between $\bar{x}^{r}$ and $\bar{x}_{t}^{c}$ is calculated in the cross-attention layer to transfer the enhancement to $\bar{x}_{t}^{c}$ , which can be formulated as:
|
| 146 |
+
|
| 147 |
+
$$
|
| 148 |
+
\operatorname {A t t e n t i o n} (Q, K, V) = \operatorname {s o f t m a x} \left(\frac {Q K ^ {T}}{\sqrt {d}}\right) V, \tag {10}
|
| 149 |
+
$$
|
| 150 |
+
|
| 151 |
+
where $Q = W_{i}^{q} \cdot \bar{x}_{t}^{c}, K = W_{i}^{k} \cdot \bar{x}^{r}, V = W_{i}^{v} \cdot \bar{x}^{r}$ with learnable parameters $W_{i}^{q}, W_{i}^{k}, W_{i}^{v}$ .
|
| 152 |
+
|
| 153 |
+
To ensure that the estimated fog variant $\epsilon_{\theta}$ is similar to the fog variant $\epsilon_{n}$ applied to make $x_{t}^{c}$ from the input clear feature $x^{c}$ , we propose a weather-adaptive enhancement loss $\mathcal{L}_{wae}$ , which is formulated as:
|
| 154 |
+
|
| 155 |
+
$$
|
| 156 |
+
\mathcal {L} _ {w a e} = \mathbb {E} _ {x ^ {c}, \boldsymbol {\epsilon} _ {n} \sim \mathcal {F}, t} \left[ \| \boldsymbol {\epsilon} _ {n} - \boldsymbol {\epsilon} _ {\theta} \left(x _ {t} ^ {c}, t, x ^ {r}\right) \| _ {2} ^ {2} \right]. \tag {11}
|
| 157 |
+
$$
|
| 158 |
+
|
| 159 |
+
With $\mathcal{L}_{wae}$ , $\epsilon_{\theta}$ can estimate the fog variant by leveraging the fog distribution as noise for our diffusion model through multiple forward/reverse processes. Additionally, the cross-attention layer within $\epsilon_{\theta}$ dynamically enhances the feature representation based on combining knowledge of the input feature (foggy or clear) and the weather-reference feature. Since our diffusion model learns the
|
| 160 |
+
|
| 161 |
+
degree of improvement needed corresponding to weather conditions, it can improve the representation of any input, whether clear or foggy in inference phase. This leads our monocular 3D object detector can handle both clear and foggy input images, resulting in weather-robust detection.
|
| 162 |
+
|
| 163 |
+
# 3.3 Total Loss Function
|
| 164 |
+
|
| 165 |
+
The total loss function of our MonoWAD is represented as follows:
|
| 166 |
+
|
| 167 |
+
$$
|
| 168 |
+
\mathcal {L} _ {\text {T o t a l}} = \mathcal {L} _ {O D} + \lambda_ {1} \mathcal {L} _ {\text {c k r}} + \lambda_ {2} \mathcal {L} _ {\text {w a e}}, \tag {12}
|
| 169 |
+
$$
|
| 170 |
+
|
| 171 |
+
where $\lambda_{1}$ and $\lambda_{2}$ denote balancing hyper-parameters. In our experiments, we set $\lambda_{1} = \lambda_{2} = 1$ . $\mathcal{L}_{OD}$ is the detection loss for 3D object detection [16,55]. It includes loss functions of classification, regression, and depth loss, that are similar to prior works [8,16,31]. The overall weight parameters are updated through $\mathcal{L}_{Total}$ .
|
| 172 |
+
|
| 173 |
+
# 4 Experiments
|
| 174 |
+
|
| 175 |
+
# 4.1 Dataset and Evaluation Metrics
|
| 176 |
+
|
| 177 |
+
Datasets. We utilize the KITTI 3D object detection dataset [12], the most widely adopted for 3D object detection. It contains 7,481 training images and 7,518 test images under clear weather conditions. Due to unavailable ground-truth annotations for test images and limited evaluation on the test server, we follow [6] by splitting 3,712 training images and 3,769 validation images. In addition, as our work requires paired images to learn about weather changes, we generate foggy images from all images in the KITTI dataset, called foggy KITTI, that emulate the foggy scene. Following the protocols of [2,43], we generate foggy images based on object distances using depth maps estimated by DORN [10]. Please refer to the supplementary materials for details.
|
| 178 |
+
|
| 179 |
+
In addition, as our work is focused on robust monocular 3D object detection in various weather conditions, we further adopt the Virtual KITTI dataset [5,11], which contains photo-realistic synthetic images under various weather conditions (e.g., foggy, rainy, sunset). It is associated with the original real-world KITTI dataset, which has 3D annotations for each weather.
|
| 180 |
+
|
| 181 |
+
Evaluation Metrics. We adopt average precision in both 3D detection $(\mathrm{AP}_{3D})$ and bird-eye view detection $(\mathrm{AP}_{BEV})$ under three difficulty levels ('Easy', 'Moderate', 'Hard') according to size, occlusion, and truncation. Following [47], we use 40 recall position metric $\mathrm{AP}_{40}$ and report scores for the car category under the IoU threshold 0.7 for the KITTI dataset and 0.5 for the Virtual KITTI dataset.
|
| 182 |
+
|
| 183 |
+
# 4.2 Implementation Details
|
| 184 |
+
|
| 185 |
+
We use DLA-102 [52] as our backbone and adopt the transformer architecture of [16] for the detection block. We train MonoWAD on a single RTX 4090 GPU with a batch size of 4 over 120 epochs using Adam optimizer [23] (initial learning
|
| 186 |
+
|
| 187 |
+
Table 1: Detection results of car category on KITTI validation set under foggy weather and clear weather conditions. The results of the state-of-the-art methods under foggy weather are obtained through our reproduction with the official source code. Bold/underlined fonts indicate the best/second-best results.
|
| 188 |
+
|
| 189 |
+
<table><tr><td rowspan="2">Method</td><td colspan="3">Foggy (AP3D)</td><td colspan="3">Clear (AP3D)</td><td colspan="3">Average</td></tr><tr><td>Easy</td><td>Mod.</td><td>Hard</td><td>Easy</td><td>Mod.</td><td>Hard</td><td>Easy</td><td>Mod.</td><td>Hard</td></tr><tr><td>GUPNet [28] (ICCV'21)</td><td>2.74</td><td>2.19</td><td>2.16</td><td>22.76</td><td>16.46</td><td>13.72</td><td>12.75</td><td>9.33</td><td>7.94</td></tr><tr><td>DID-M3D [37] (ECCV'22)</td><td>1.15</td><td>0.61</td><td>0.64</td><td>22.98</td><td>16.12</td><td>14.03</td><td>12.07</td><td>8.37</td><td>7.34</td></tr><tr><td>MonoGround [40] (CVPR'22)</td><td>0.00</td><td>0.00</td><td>0.06</td><td>25.24</td><td>18.69</td><td>15.58</td><td>12.62</td><td>9.35</td><td>7.82</td></tr><tr><td>MonoDTR [16] (CVPR'22)</td><td>16.89</td><td>11.86</td><td>9.87</td><td>24.52</td><td>18.57</td><td>15.51</td><td>20.71</td><td>15.22</td><td>12.69</td></tr><tr><td>MonoDETR [55] (ICCV'23)</td><td>7.40</td><td>5.74</td><td>4.53</td><td>28.84</td><td>20.61</td><td>16.38</td><td>18.12</td><td>13.18</td><td>10.46</td></tr><tr><td>MonoWAD (Ours)</td><td>27.17</td><td>19.57</td><td>16.21</td><td>29.10</td><td>21.08</td><td>17.73</td><td>28.14</td><td>20.33</td><td>16.97</td></tr></table>
|
| 190 |
+
|
| 191 |
+
Table 2: Detection results of car category on KITTI test set under foggy weather. Bold/underlined fonts indicate the best/second-best results.
|
| 192 |
+
|
| 193 |
+
<table><tr><td rowspan="2">Method</td><td colspan="3">AP3D</td><td colspan="3">APBEV</td></tr><tr><td>Easy</td><td>Mod.</td><td>Hard</td><td>Easy</td><td>Mod.</td><td>Hard</td></tr><tr><td>GUPNet [28] (ICCV'21)</td><td>3.01</td><td>2.42</td><td>1.13</td><td>4.90</td><td>3.02</td><td>2.91</td></tr><tr><td>DID-M3D [37] (ECCV'22)</td><td>3.10</td><td>2.39</td><td>2.19</td><td>5.34</td><td>3.01</td><td>2.91</td></tr><tr><td>MonoGround [40] (CVPR'22)</td><td>0.14</td><td>0.20</td><td>0.22</td><td>0.23</td><td>0.38</td><td>0.39</td></tr><tr><td>MonoDTR [16] (CVPR'22)</td><td>11.07</td><td>7.41</td><td>5.26</td><td>15.76</td><td>10.15</td><td>7.53</td></tr><tr><td>MonoDETR [55] (ICCV'23)</td><td>9.33</td><td>5.54</td><td>4.06</td><td>13.15</td><td>8.06</td><td>6.30</td></tr><tr><td>MonoWAD (Ours)</td><td>19.75</td><td>13.32</td><td>11.04</td><td>27.95</td><td>19.06</td><td>15.61</td></tr></table>
|
| 194 |
+
|
| 195 |
+
rate of $10^{-4}$ ). For the weather codebook, we utilize embedding slot $K = 4096$ and set the dimension of each slot $D = 256$ . We set 4 heads for cross-attention of the weather-adaptive diffusion model, and the timesteps for the forward and reverse process of diffusion default to 15. The number of channels for the diffusion output features is $C = 256$ . As recent monocular 3D object detectors [16,28,37,40,55] have not been explored under adverse weather conditions, we implemented them with available official source code to faithfully reproduce them.
|
| 196 |
+
|
| 197 |
+
# 4.3 Comparison
|
| 198 |
+
|
| 199 |
+
Results on KITTI 3D Dataset. We compared MonoWAD with state-of-the-art monocular 3D object detectors [16,28,37,40,55] that do not use additional data (e.g., depth maps or LiDAR) during inference on the KITTI and foggy KITTI validation sets. Table 1 shows the $AP_{3D}$ results, and $AP_{BEV}$ results are in the supplementary materials. While recent methods have shown improved performance in clear weather, their performance drops in foggy weather, limiting their applicability in real-world applications (e.g., autonomous driving and robotics). In contrast, MonoWAD showed stable 3D detection performance under both foggy and clear weather. Since our weather codebook learns knowledge about clear weather and the weather-adaptive diffusion model can enhance the feature representation of the input images under both clear and foggy weather, MonoWAD shows a more weather-robust detection performance than that of the existing methods. Also, to explore the weather robustness of MonoWAD, we conducted experiments with mixed foggy and clear weather conditions at various ratios (please see the supplementary materials).
|
| 200 |
+
|
| 201 |
+
Table 3: Detection results of car category on Virtual KITTI under foggy, rainy, sunset conditions. Bold/underlined fonts indicate the best/second-best results.
|
| 202 |
+
|
| 203 |
+
<table><tr><td rowspan="2">Method</td><td colspan="3">Foggy (AP3D)</td><td colspan="3">Rainy (AP3D)</td><td colspan="3">Sunset (AP3D)</td></tr><tr><td>Easy</td><td>Mod.</td><td>Hard</td><td>Easy</td><td>Mod.</td><td>Hard</td><td>Easy</td><td>Mod.</td><td>Hard</td></tr><tr><td>GUPNet [28] (ICCV'21)</td><td>1.76</td><td>1.57</td><td>1.57</td><td>2.34</td><td>1.24</td><td>1.21</td><td>2.77</td><td>1.64</td><td>1.65</td></tr><tr><td>DID-M3D [37] (ECCV'22)</td><td>0.91</td><td>0.39</td><td>0.39</td><td>0.40</td><td>0.13</td><td>0.13</td><td>0.34</td><td>0.10</td><td>0.10</td></tr><tr><td>MonoGround [40] (CVPR'22)</td><td>0.29</td><td>0.30</td><td>0.25</td><td>5.49</td><td>2.82</td><td>2.77</td><td>7.68</td><td>4.24</td><td>4.20</td></tr><tr><td>MonoDTR [16] (CVPR'22)</td><td>8.79</td><td>5.75</td><td>5.72</td><td>11.73</td><td>6.25</td><td>6.74</td><td>9.86</td><td>5.42</td><td>5.42</td></tr><tr><td>MonoDETR [55] (ICCV'23)</td><td>4.50</td><td>2.99</td><td>2.96</td><td>6.61</td><td>3.46</td><td>3.42</td><td>7.08</td><td>4.17</td><td>4.16</td></tr><tr><td>MonoWAD (Ours)</td><td>13.33</td><td>8.56</td><td>8.50</td><td>14.12</td><td>8.33</td><td>8.24</td><td>13.38</td><td>7.89</td><td>7.80</td></tr></table>
|
| 204 |
+
|
| 205 |
+
Table 4: Effect of the proposed method on KITTI validation set for car category. WC denotes our weather codebook, WAD indicates our weather-adaptive diffusion model.
|
| 206 |
+
|
| 207 |
+
<table><tr><td rowspan="2">Method</td><td rowspan="2">WC</td><td rowspan="2">WAD</td><td colspan="3">Foggy (AP3D)</td><td colspan="3">Clear (AP3D)</td></tr><tr><td>Easy</td><td>Mod.</td><td>Hard</td><td>Easy</td><td>Mod.</td><td>Hard</td></tr><tr><td>Baseline</td><td>-</td><td>-</td><td>13.75</td><td>9.61</td><td>8.10</td><td>22.63</td><td>17.16</td><td>14.28</td></tr><tr><td rowspan="2">MonoWAD (Ours)</td><td>-</td><td>✓</td><td>25.62</td><td>18.66</td><td>15.56</td><td>26.34</td><td>19.17</td><td>16.15</td></tr><tr><td>✓</td><td>✓</td><td>27.17</td><td>19.57</td><td>16.21</td><td>29.10</td><td>21.08</td><td>17.73</td></tr></table>
|
| 208 |
+
|
| 209 |
+
Table 5: Detection results on KITTI validation set by changing diffusion timestep $T$ .
|
| 210 |
+
|
| 211 |
+
<table><tr><td rowspan="2">Timestep T</td><td colspan="3">Foggy (AP3D)</td><td colspan="3">Clear (AP3D)</td></tr><tr><td>Easy</td><td>Mod.</td><td>Hard</td><td>Easy</td><td>Mod.</td><td>Hard</td></tr><tr><td>-</td><td>13.75</td><td>9.61</td><td>8.10</td><td>22.63</td><td>17.16</td><td>14.28</td></tr><tr><td>5</td><td>23.57</td><td>17.91</td><td>14.96</td><td>26.03</td><td>19.21</td><td>16.01</td></tr><tr><td>10</td><td>25.28</td><td>18.49</td><td>15.42</td><td>26.79</td><td>19.90</td><td>16.78</td></tr><tr><td>15</td><td>27.17</td><td>19.57</td><td>16.21</td><td>29.10</td><td>21.08</td><td>17.73</td></tr><tr><td>20</td><td>24.54</td><td>18.29</td><td>15.10</td><td>24.85</td><td>18.54</td><td>15.34</td></tr></table>
|
| 212 |
+
|
| 213 |
+
We further compared 3D detection performances on the foggy KITTI test set (Table 2). Similar to Table 1, existing monocular 3D object detectors show lower performance under foggy weather. Through the results, our method shows the weather-robust monocular 3D performance under foggy and clear weather.
|
| 214 |
+
|
| 215 |
+
Results on Virtual KITTI Dataset. We also conducted experiments on the Virtual KITTI dataset to validate the generalization ability of our method across various weather conditions (i.e., foggy, rainy, sunset). The results are shown in Table 3. MonoWAD also shows the highest performance in foggy weather. Moreover, it outperforms the existing method in the rainy and sunset conditions. The results demonstrate that MonoWAD is robust and insensitive to various weather conditions that can be faced in real-world scenarios, not just clear weather.
|
| 216 |
+
|
| 217 |
+
# 4.4 Ablation Studies
|
| 218 |
+
|
| 219 |
+
We conducted ablation studies to examine (1) the effect of each proposed component (i.e., weather codebook and weather-adaptive diffusion model) and (2) the effect of the weather-adaptive diffusion model by varying the timestep $T$ . These experiments were performed on the KITTI and foggy KITTI 3D validation sets.
|
| 220 |
+
|
| 221 |
+
Effect of the Proposed Modules. The results regarding the effectiveness of the proposed weather codebook and weather-adaptive diffusion model are pre-
|
| 222 |
+
|
| 223 |
+
Table 6: Performance comparison on KITTI validation set under foggy and clear weather conditions. We compare our MonoWAD with MonoDTR [16], which demonstrates superior performance in foggy weather using state-of-the-art dehazing methods.
|
| 224 |
+
|
| 225 |
+
<table><tr><td rowspan="2">Method</td><td rowspan="2">Dehazing Method</td><td colspan="3">Foggy (AP3D)</td><td colspan="3">Clear (AP3D)</td><td colspan="3">Average</td></tr><tr><td>Easy</td><td>Mod.</td><td>Hard</td><td>Easy</td><td>Mod.</td><td>Hard</td><td>Easy</td><td>Mod.</td><td>Hard</td></tr><tr><td>MonoDTR(B) [16] (CVPR'22)</td><td>-</td><td>16.89</td><td>11.86</td><td>9.87</td><td>24.52</td><td>18.57</td><td>15.51</td><td>20.71</td><td>15.22</td><td>12.69</td></tr><tr><td>B + RIDCP [49] (CVPR'23)</td><td>Image-level</td><td>17.23</td><td>12.41</td><td>10.44</td><td>24.02</td><td>17.89</td><td>14.78</td><td>20.63</td><td>15.15</td><td>12.61</td></tr><tr><td>B + DENet [39] (ACCV'22)</td><td>Feature-level</td><td>22.35</td><td>17.44</td><td>14.47</td><td>7.10</td><td>5.70</td><td>4.53</td><td>14.73</td><td>11.57</td><td>9.50</td></tr><tr><td>B + Yang et al. [51] (ACCV'22)</td><td>Feature-level</td><td>22.87</td><td>15.21</td><td>12.17</td><td>17.96</td><td>13.10</td><td>10.64</td><td>20.42</td><td>14.16</td><td>11.41</td></tr><tr><td>MonoWAD (Ours)</td><td>-</td><td>27.17</td><td>19.57</td><td>16.21</td><td>29.10</td><td>21.08</td><td>17.73</td><td>28.14</td><td>20.33</td><td>16.97</td></tr></table>
|
| 226 |
+
|
| 227 |
+
Table 7: Comparison of diffusion models on KITTI validation set for car category: $\mathcal{B}$ is baseline detection block (transformer encoder-decoder), CDM (Conditional Diffusion Model [42]), WC (weather codebook), and WAD (weather adaptive diffusion model).
|
| 228 |
+
|
| 229 |
+
<table><tr><td rowspan="2">Method</td><td colspan="3">Foggy (AP3D)</td><td colspan="3">Clear (AP3D)</td></tr><tr><td>Easy</td><td>Mod.</td><td>Hard</td><td>Easy</td><td>Mod.</td><td>Hard</td></tr><tr><td>B+DDPM [15]</td><td>5.32</td><td>3.84</td><td>2.77</td><td>18.31</td><td>12.71</td><td>10.20</td></tr><tr><td>B+CDM [42]</td><td>2.74</td><td>2.10</td><td>2.00</td><td>20.50</td><td>14.51</td><td>11.63</td></tr><tr><td>B+WC+CDM [42]</td><td>17.51</td><td>12.74</td><td>10.40</td><td>21.05</td><td>15.11</td><td>12.54</td></tr><tr><td>B+WAD</td><td>25.62</td><td>18.66</td><td>15.56</td><td>26.34</td><td>19.17</td><td>16.15</td></tr><tr><td>MonoWAD(B+WC+WAD)</td><td>27.17</td><td>19.57</td><td>16.21</td><td>29.10</td><td>21.08</td><td>17.73</td></tr></table>
|
| 230 |
+
|
| 231 |
+
sented in Table 4. Since our weather-adaptive diffusion model is designed to enhance feature representation, it alone has shown significant performance improvement. When the weather codebook is additionally considered, the weather-adaptive diffusion model can leverage the knowledge of the weather-reference feature for clear weather, leading to enhanced performance. It makes MonoWAD can be robust to various weather conditions for monocular 3D object detection.
|
| 232 |
+
|
| 233 |
+
Effect of Timestep $T$ . We also conduct experiments by varying timestep $T$ of the weather-adaptive diffusion model. Timestep is the number of steps for the forward and reverse process of the diffusion model. Table 5 indicates that the highest monocular 3D detection performance is achieved when $T = 15$ under both clear and foggy weather conditions. Our MonoWAD consistently outperforms the baseline and other existing methods across all timesteps.
|
| 234 |
+
|
| 235 |
+
# 4.5 Discussions
|
| 236 |
+
|
| 237 |
+
Comparison with Dehazing Methods. We investigate the weather-robustness of our method for monocular 3D object detection compared to the other monocular detectors using dehazing methods. To this end, we compared MonoWAD with MonoDTR [16], which shows the best performance in foggy weather through the application of state-of-the-art image-level and feature-level dehazing methods [39,49,51]. The results are shown in Table 6. Recent dehazing methods are primarily focused on specific weather conditions, such as foggy weather. While they have shown some improvement under foggy conditions, they exhibit reduced performances under clear weather conditions. In contrast, our MonoWAD shows robust performance across both foggy and clear weather conditions.
|
| 238 |
+
|
| 239 |
+

|
| 240 |
+
Fig. 5: Comparison of 3D detection examples (green: ground-truth, red: predicted 3D bounding-box) between our MonoWAD and two detectors, MonoDTR [16] and MonoDETR [55], that show the most improved performances among existing methods.
|
| 241 |
+
|
| 242 |
+

|
| 243 |
+
Fig. 6: 3D detection results on real-world images of various weather conditions.
|
| 244 |
+
|
| 245 |
+
Effect of Weather-Adaptive Diffusion Model. Table 7 shows the effectiveness of MonoWAD with existing diffusion models [15,42] on the KITTI and foggy KITTI validation sets. Existing methods adopt Gaussian noise for forward and reverse processes, but they can not fully understand about weather. In contrast, our weather-adaptive diffusion understands weather variances, allowing our MonoWAD to surpass existing methods in clear and foggy weather.
|
| 246 |
+
|
| 247 |
+
# 4.6 Visualization Results
|
| 248 |
+
|
| 249 |
+
Results on KITTI Dataset. We visualize several 3D detection results on the KITTI 3D dataset, comparing MonoWAD with MonoDTR and MonoDETR, which exhibit the highest performances among existing methods under foggy condition (Fig. 5). Existing methods struggle to detect objects obscured by fog, indicating limitations in detecting only fully visible objects. In contrast, even
|
| 250 |
+
|
| 251 |
+

|
| 252 |
+
Fig. 7: t-SNE visualization results (Red: Clear, Blue: Foggy).
|
| 253 |
+
|
| 254 |
+

|
| 255 |
+
|
| 256 |
+

|
| 257 |
+
|
| 258 |
+
in dense fog, MonoWAD effectively detects both close and fog-obscured objects with the aid of the weather codebook and weather-adaptive diffusion model.
|
| 259 |
+
|
| 260 |
+
Results on Real-World Images. We further visualize 3D detection results on real-world images from the Seeing Through Fog dataset [2] under various weather conditions (i.e., foggy, rainy, snowy). In Fig. 6, MonoWAD shows robust detection under diverse weather conditions. This demonstrates that our proposed method maintains weather-robustness even in real-world scenarios by dynamically enhancing the input scenarios.
|
| 261 |
+
|
| 262 |
+
t-SNE Visualization. We conducted t-SNE visualization to analyze feature representations of MonoDTR, MonoDETR, and our MonoWAD on the KITTI and foggy KITTI validation set. As depicted in Fig. 7, the existing methods exhibit distinct feature representations for foggy and clear weather conditions. In contrast, MonoWAD, leveraging weather-robust feature learning from the weather codebook and weather-adaptive diffusion model, demonstrates similar feature representations for both clear and foggy weather conditions.
|
| 263 |
+
|
| 264 |
+
# 4.7 Limitations
|
| 265 |
+
|
| 266 |
+
The experimental results show the weather-robustness of our method. However, due to the iterative nature of the diffusion model, our method shows $144\mathrm{ms}/$ image at timestep $T = 15$ (110ms/image at $T = 5$ ), slower than the latest work, MonoDETR (38ms/image). Moreover, our method dynamically enhances the representation of the input feature based on the weather-reference feature and weather difference which needs paired images. Thus, exploring a method to achieve faster processing speeds while maintaining weather-robust performance without paired images could be an interesting direction for our future work.
|
| 267 |
+
|
| 268 |
+
# 5 Conclusion
|
| 269 |
+
|
| 270 |
+
We proposed MonoWAD, a novel weather-robust monocular 3D object detector to handle various weather conditions. Addressing challenges in applying existing monocular 3D object detectors to real-world scenarios with various weather, we design a weather codebook with clear knowledge recalling loss to memorize the knowledge of the clear weather and to generate a weather-reference feature from both clear and foggy features. Also, we design a weather-adaptive diffusion model with weather-adaptive enhancement loss to enhance feature representation according to the weather conditions. As a result, our MonoWAD can detect objects occluded by fog and perform well in clear weather.
|
| 271 |
+
|
| 272 |
+
# Acknowledgements
|
| 273 |
+
|
| 274 |
+
This work was supported by the NRF grant funded by the Korea government (MSIT) (No. RS-2023-00252391), and by IITP grant funded by the Korea government (MSIT) (No. RS-2022-00155911: Artificial Intelligence Convergence Innovation Human Resources Development (Kyung Hee University), IITP-2023-RS-2023-00266615: Convergence Security Core Talent Training Business Support Program, No. 2022-0-00124: Development of Artificial Intelligence Technology for Self-Improving Competency-Aware Learning Capabilities).
|
| 275 |
+
|
| 276 |
+
# References
|
| 277 |
+
|
| 278 |
+
1. Baranchuk, D., Rubachev, I., Voynov, A., Khrulkov, V., Babenko, A.: Label-efficient semantic segmentation with diffusion models. In: Int. Conf. Learn. Represent. (2021)
|
| 279 |
+
2. Bijelic, M., Gruber, T., Mannan, F., Kraus, F., Ritter, W., Dietmayer, K., Heide, F.: Seeing through fog without seeing fog: Deep multimodal sensor fusion in unseen adverse weather. In: IEEE Conf. Comput. Vis. Pattern Recog. (2020)
|
| 280 |
+
3. Brazil, G., Liu, X.: M3d-rpn: Monocular 3d region proposal network for object detection. In: Int. Conf. Comput. Vis. (2019)
|
| 281 |
+
4. Brazil, G., Pons-Moll, G., Liu, X., Schiele, B.: Kinematic 3d object detection in monocular video. In: Eur. Conf. Comput. Vis. Springer (2020)
|
| 282 |
+
5. Cabon, Y., Murray, N., Humenberger, M.: Virtual kitti 2 (2020)
|
| 283 |
+
6. Chen, X., Kundu, K., Zhang, Z., Ma, H., Fidler, S., Urtasun, R.: Monocular 3d object detection for autonomous driving. In: IEEE Conf. Comput. Vis. Pattern Recog. (2016)
|
| 284 |
+
7. Chen, Y., Tai, L., Sun, K., Li, M.: Monopair: Monocular 3d object detection using pairwise spatial relationships. In: IEEE Conf. Comput. Vis. Pattern Recog. (2020)
|
| 285 |
+
8. Ding, M., Huo, Y., Yi, H., Wang, Z., Shi, J., Lu, Z., Luo, P.: Learning depth-guided convolutions for monocular 3d object detection. In: IEEE Conf. Comput. Vis. Pattern Recog. Worksh. (2020)
|
| 286 |
+
9. Esser, P., Rombach, R., Ommer, B.: Taming transformers for high-resolution image synthesis. In: IEEE Conf. Comput. Vis. Pattern Recog. (2021)
|
| 287 |
+
10. Fu, H., Gong, M., Wang, C., Batmanghelich, K., Tao, D.: Deep ordinal regression network for monocular depth estimation. In: IEEE Conf. Comput. Vis. Pattern Recog. (2018)
|
| 288 |
+
1. Gaidon, A., Wang, Q., Cabon, Y., Vig, E.: Virtual worlds as proxy for multi-object tracking analysis. In: IEEE Conf. Comput. Vis. Pattern Recog. (2016)
|
| 289 |
+
2. Geiger, A., Lenz, P., Urtasun, R.: Are we ready for autonomous driving? the kitti vision benchmark suite. In: IEEE Conf. Comput. Vis. Pattern Recog. (2012). https://doi.org/10.1109/CVPR.2012.6248074
|
| 290 |
+
3. Hahner, M., Sakaridis, C., Dai, D., Van Gool, L.: Fog simulation on real lidar point clouds for 3d object detection in adverse weather. In: Int. Conf. Comput. Vis. (2021)
|
| 291 |
+
4. Hamilton, B., Tefft, B., Arnold, L., Grabowski, J.: Hidden highways: Fog and traffic crashes on america's roads (november 2014). Montana 40 (2006)
|
| 292 |
+
5. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. In: Adv. Neural Inform. Process. Syst. vol. 33 (2020)
|
| 293 |
+
|
| 294 |
+
16. Huang, K.C., Wu, T.H., Su, H.T., Hsu, W.H.: Monodtr: Monocular 3d object detection with depth-aware transformer. In: IEEE Conf. Comput. Vis. Pattern Recog. (2022)
|
| 295 |
+
17. Juneja, A., Kumar, V., Singla, S.K.: A systematic review on foggy datasets: Applications and challenges. Archives of Computational Methods in Engineering 29(3) (2022)
|
| 296 |
+
18. Kim, J.U., Kim, H.I., Ro, Y.M.: Stereoscopic vision recalling memory for monocular 3d object detection. IEEE Trans. Image Process. 32 (2023). https://doi.org/10.1109/TIP.2023.3274479
|
| 297 |
+
19. Kim, J.U., Park, S., Ro, Y.M.: Robust small-scale pedestrian detection with cued recall via memory learning. In: Int. Conf. Comput. Vis. (2021)
|
| 298 |
+
20. Kim, J.U., Park, S., Ro, Y.M.: Uncertainty-guided cross-modal learning for robust multispectral pedestrian detection. IEEE Trans. Circuit Syst. Video Technol. 32(3) (2021)
|
| 299 |
+
21. Kim, J.U., Park, S., Ro, Y.M.: Towards versatile pedestrian detector with multisensory-matching and multispectral recalling memory. In: AAAI. vol. 36 (2022)
|
| 300 |
+
22. Kim, S.W., Brown, B., Yin, K., Kreis, K., Schwarz, K., Li, D., Rombach, R., Torralba, A., Fidler, S.: Neuralfield-ldm: Scene generation with hierarchical latent diffusion models. In: IEEE Conf. Comput. Vis. Pattern Recog. (2023)
|
| 301 |
+
23. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014)
|
| 302 |
+
24. Lang, A.H., Vora, S., Caesar, H., Zhou, L., Yang, J., Beijbom, O.: Pointpillars: Fast encoders for object detection from point clouds. In: IEEE Conf. Comput. Vis. Pattern Recog. (2019)
|
| 303 |
+
25. Liu, X., Xue, N., Wu, T.: Learning auxiliary monocular contexts helps monocular 3d object detection. In: AAAI. vol. 36 (2022)
|
| 304 |
+
26. Liu, Z., Wu, Z., Toth, R.: Smoke: Single-stage monocular 3d object detection via keypoint estimation. In: IEEE Conf. Comput. Vis. Pattern Recog. Worksh. (2020)
|
| 305 |
+
27. Liu, Z., Zhou, D., Lu, F., Fang, J., Zhang, L.: Autoshape: Real-time shape-aware monocular 3d object detection. In: Int. Conf. Comput. Vis. (2021)
|
| 306 |
+
28. Lu, Y., Ma, X., Yang, L., Zhang, T., Liu, Y., Chu, Q., Yan, J., Ouyang, W.: Geometry uncertainty projection network for monocular 3d object detection. In: Int. Conf. Comput. Vis. (2021)
|
| 307 |
+
29. Lugmayr, A., Danelljan, M., Romero, A., Yu, F., Timofte, R., Van Gool, L.: Repaint: Inpainting using denoising diffusion probabilistic models. In: IEEE Conf. Comput. Vis. Pattern Recog. (2022)
|
| 308 |
+
30. Luo, J., Li, Y., Pan, Y., Yao, T., Feng, J., Chao, H., Mei, T.: Semantic-conditional diffusion networks for image captioning. In: IEEE Conf. Comput. Vis. Pattern Recog. (2023)
|
| 309 |
+
31. Luo, S., Dai, H., Shao, L., Ding, Y.: M3dssd: Monocular 3d single stage object detector. In: IEEE Conf. Comput. Vis. Pattern Recog. (2021)
|
| 310 |
+
32. Ma, X., Liu, S., Xia, Z., Zhang, H., Zeng, X., Ouyang, W.: Rethinking pseudo-lidar representation. In: Eur. Conf. Comput. Vis. Springer (2020)
|
| 311 |
+
33. Mai, N.A.M., Duthon, P., Khoudour, L., Crouzil, A., Velastin, S.A.: 3d object detection with sls-fusion network in foggy weather conditions. Sensors 21(20) (2021)
|
| 312 |
+
34. Mousavian, A., Anguelov, D., Flynn, J., Kosecka, J.: 3d bounding box estimation using deep learning and geometry. In: IEEE Conf. Comput. Vis. Pattern Recog. (2017)
|
| 313 |
+
|
| 314 |
+
35. Ni, H., Shi, C., Li, K., Huang, S.X., Min, M.R.: Conditional image-to-video generation with latent flow diffusion models. In: IEEE Conf. Comput. Vis. Pattern Recog. (2023)
|
| 315 |
+
36. van den Oord, A., Vinyals, O., Kavukcuoglu, K.: Neural discrete representation learning (2018)
|
| 316 |
+
37. Peng, L., Wu, X., Yang, Z., Liu, H., Cai, D.: Did-m3d: Decoupling instance depth for monocular 3d object detection. In: Eur. Conf. Comput. Vis. Springer (2022)
|
| 317 |
+
38. Pfeuffer, A., Dietmayer, K.: Robust semantic segmentation in adverse weather conditions by means of sensor data fusion. In: 2019 22th International Conference on Information Fusion (FUSION) (2019). https://doi.org/10.23919/FUSION43075.2019.9011192
|
| 318 |
+
39. Qin, Q., Chang, K., Huang, M., Li, G.: Denet: Detection-driven enhancement network for object detection under adverse weather conditions. In: Asian Conf. Comput. Vis. (2022)
|
| 319 |
+
40. Qin, Z., Li, X.: Monoground: Detecting monocular 3d objects from the ground. In: IEEE Conf. Comput. Vis. Pattern Recog. (2022)
|
| 320 |
+
41. Reading, C., Harakeh, A., Chae, J., Waslander, S.L.: Categorical depth distribution network for monocular 3d object detection. In: IEEE Conf. Comput. Vis. Pattern Recog. (2021)
|
| 321 |
+
42. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: IEEE Conf. Comput. Vis. Pattern Recog. (2022)
|
| 322 |
+
43. Sakaridis, C., Dai, D., Van Gool, L.: Semantic foggy scene understanding with synthetic data. Int. J. Comput. Vis. 126 (2018)
|
| 323 |
+
44. Shi, S., Guo, C., Jiang, L., Wang, Z., Shi, J., Wang, X., Li, H.: Pv-rcnn: Point-voxel feature set abstraction for 3d object detection. In: IEEE Conf. Comput. Vis. Pattern Recog. (2020)
|
| 324 |
+
45. Shi, S., Wang, X., Li, H.: Pointcnn: 3d object proposal generation and detection from point cloud. In: IEEE Conf. Comput. Vis. Pattern Recog. (2019)
|
| 325 |
+
46. Shi, X., Ye, Q., Chen, X., Chen, C., Chen, Z., Kim, T.K.: Geometry-based distance decomposition for monocular 3d object detection. In: Int. Conf. Comput. Vis. (2021)
|
| 326 |
+
47. Simonelli, A., Bulo, S.R., Porzi, L., Lopez-Antequera, M., Kontschieder, P.: Disentangling monocular 3d object detection. In: Int. Conf. Comput. Vis. (2019)
|
| 327 |
+
48. Wang, L., Du, L., Ye, X., Fu, Y., Guo, G., Xue, X., Feng, J., Zhang, L.: Depth-conditioned dynamic message propagation for monocular 3d object detection. In: IEEE Conf. Comput. Vis. Pattern Recog. (2021)
|
| 328 |
+
49. Wu, R.Q., Duan, Z.P., Guo, C.L., Chai, Z., Li, C.: Ridcp: Revitalizing real image dehazing via high-quality codebook priors. In: IEEE Conf. Comput. Vis. Pattern Recog. (2023)
|
| 329 |
+
50. Yang, S., Scherer, S.: Cubeslam: Monocular 3-d object slam. IEEE Transactions on Robotics 35(4) (2019). https://doi.org/10.1109/TR0.2019.2909168
|
| 330 |
+
51. Yang, X., Mi, M.B., Yuan, Y., Wang, X., Tan, R.T.: Object detection in foggy scenes by embedding depth and reconstruction into domain adaptation. In: Asian Conf. Comput. Vis. (2022)
|
| 331 |
+
52. Yu, F., Wang, D., Shelhamer, E., Darrell, T.: Deep layer aggregation. In: IEEE Conf. Comput. Vis. Pattern Recog. (2018)
|
| 332 |
+
53. Yurtsever, E., Lambert, J., Carballo, A., Takeda, K.: A survey of autonomous driving: Common practices and emerging technologies. IEEE Access 8 (2020). https://doi.org/10.1109/ACCESS.2020.2983149
|
| 333 |
+
|
| 334 |
+
54. Zhang, C., Wang, H., Cai, Y., Chen, L., Li, Y., Sotelo, M.A., Li, Z.: Robust-fusionnet: Deep multimodal sensor fusion for 3-d object detection under severe weather conditions. IEEE Transactions on Instrumentation and Measurement 71 (2022). https://doi.org/10.1109/TIM.2022.3191724
|
| 335 |
+
55. Zhang, R., Qiu, H., Wang, T., Guo, Z., Cui, Z., Qiao, Y., Li, H., Gao, P.: Monodetr: Depth-guided transformer for monocular 3d object detection. In: Int. Conf. Comput. Vis. (2023)
|
| 336 |
+
56. Zhou, Y., Tuzel, O.: Voxelnet: End-to-end learning for point cloud based 3d object detection. In: IEEE Conf. Comput. Vis. Pattern Recog. (2018)
|
| 337 |
+
57. Zhou, Y., He, Y., Zhu, H., Wang, C., Li, H., Jiang, Q.: Monocular 3d object detection: An extrinsic parameter free approach. In: IEEE Conf. Comput. Vis. Pattern Recog. (2021)
|
2024/MonoWAD_ Weather-Adaptive Diffusion Model for Robust Monocular 3D Object Detection/images.zip
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:3ba630e9abcd9904fa8bba5a36e2b70df6967262edb91ecd6ebb1f6556d9b4fd
|
| 3 |
+
size 713153
|
2024/MonoWAD_ Weather-Adaptive Diffusion Model for Robust Monocular 3D Object Detection/layout.json
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
2024/Monocular Occupancy Prediction for Scalable Indoor Scenes/325d434b-3564-4d20-a82e-4ec4e345eb58_content_list.json
ADDED
|
@@ -0,0 +1,1824 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
{
|
| 3 |
+
"type": "text",
|
| 4 |
+
"text": "Monocular Occupancy Prediction for Scalable Indoor Scenes",
|
| 5 |
+
"text_level": 1,
|
| 6 |
+
"bbox": [
|
| 7 |
+
238,
|
| 8 |
+
140,
|
| 9 |
+
763,
|
| 10 |
+
183
|
| 11 |
+
],
|
| 12 |
+
"page_idx": 0
|
| 13 |
+
},
|
| 14 |
+
{
|
| 15 |
+
"type": "text",
|
| 16 |
+
"text": "Hongxiao $\\mathrm{Yu}^{1,2}$ Yuqi Wang $^{1,2}$ , Yuntao Chen $^{3}$ , and Zhaoxiang Zhang $^{1,2,3}$",
|
| 17 |
+
"bbox": [
|
| 18 |
+
215,
|
| 19 |
+
210,
|
| 20 |
+
785,
|
| 21 |
+
228
|
| 22 |
+
],
|
| 23 |
+
"page_idx": 0
|
| 24 |
+
},
|
| 25 |
+
{
|
| 26 |
+
"type": "text",
|
| 27 |
+
"text": "<sup>1</sup> School of Artificial Intelligence, University of Chinese Academy of Sciences (UCAS)",
|
| 28 |
+
"bbox": [
|
| 29 |
+
215,
|
| 30 |
+
238,
|
| 31 |
+
782,
|
| 32 |
+
252
|
| 33 |
+
],
|
| 34 |
+
"page_idx": 0
|
| 35 |
+
},
|
| 36 |
+
{
|
| 37 |
+
"type": "text",
|
| 38 |
+
"text": "$^{2}$ NLPR, MAIS, Institute of Automation, Chinese Academy of Sciences (CASIA)",
|
| 39 |
+
"bbox": [
|
| 40 |
+
228,
|
| 41 |
+
253,
|
| 42 |
+
769,
|
| 43 |
+
268
|
| 44 |
+
],
|
| 45 |
+
"page_idx": 0
|
| 46 |
+
},
|
| 47 |
+
{
|
| 48 |
+
"type": "text",
|
| 49 |
+
"text": "<sup>3</sup> Centre for Artificial Intelligence and Robotics (HKISI_CAS) {yuhongxiao2023, wangyuqi2020, zhaoxiang.zhang}@ia.ac.cn, chenyuntao08@gmail.com",
|
| 50 |
+
"bbox": [
|
| 51 |
+
281,
|
| 52 |
+
268,
|
| 53 |
+
720,
|
| 54 |
+
309
|
| 55 |
+
],
|
| 56 |
+
"page_idx": 0
|
| 57 |
+
},
|
| 58 |
+
{
|
| 59 |
+
"type": "text",
|
| 60 |
+
"text": "Abstract. Camera-based 3D occupancy prediction has recently garnered increasing attention in outdoor driving scenes. However, research in indoor scenes remains relatively unexplored. The core differences in indoor scenes lie in the complexity of scene scale and the variance in object size. In this paper, we propose a novel method, named ISO, for predicting indoor scene occupancy using monocular images. ISO harnesses the advantages of a pretrained depth model to achieve accurate depth predictions. Furthermore, we introduce the Dual Feature Line of Sight Projection (D-FLoSP) module within ISO, which enhances the learning of 3D voxel features. To foster further research in this domain, we introduce Occ-ScanNet, a large-scale occupancy benchmark for indoor scenes. With a dataset size 40 times larger than the NYUv2 dataset, it facilitates future scalable research in indoor scene analysis. Experimental results on both NYUv2 and Occ-ScanNet demonstrate that our method achieves state-of-the-art performance. The dataset and code are made publicly at https://github.com/hongxiaoy/ISO.git.",
|
| 61 |
+
"bbox": [
|
| 62 |
+
259,
|
| 63 |
+
342,
|
| 64 |
+
738,
|
| 65 |
+
564
|
| 66 |
+
],
|
| 67 |
+
"page_idx": 0
|
| 68 |
+
},
|
| 69 |
+
{
|
| 70 |
+
"type": "text",
|
| 71 |
+
"text": "Keywords: 3D occupancy prediction $\\cdot$ Camera-based scene understanding $\\cdot$ Semantic scene completion",
|
| 72 |
+
"bbox": [
|
| 73 |
+
259,
|
| 74 |
+
575,
|
| 75 |
+
740,
|
| 76 |
+
604
|
| 77 |
+
],
|
| 78 |
+
"page_idx": 0
|
| 79 |
+
},
|
| 80 |
+
{
|
| 81 |
+
"type": "text",
|
| 82 |
+
"text": "1 Introduction",
|
| 83 |
+
"text_level": 1,
|
| 84 |
+
"bbox": [
|
| 85 |
+
215,
|
| 86 |
+
628,
|
| 87 |
+
375,
|
| 88 |
+
643
|
| 89 |
+
],
|
| 90 |
+
"page_idx": 0
|
| 91 |
+
},
|
| 92 |
+
{
|
| 93 |
+
"type": "text",
|
| 94 |
+
"text": "3D scene understanding is a crucial task in computer vision, becoming increasingly important for applications such as robotic navigation [10], augmented reality [2], and autonomous driving [47]. While humans possess a natural ability to comprehend 3D environments through vision, enabling this ability in computers poses a significant challenge due to limitations like restricted fields of view, sparse sensing capabilities, and measurement noise. However, with the rapid development of deep learning and the availability of large-scale 3D driving datasets, camera-based 3D object detection [14, 15, 20, 21, 35, 37, 39, 43] and occupancy prediction [5, 22, 33, 38, 40] have developed rapidly, achieving significant improvements in performance.",
|
| 95 |
+
"bbox": [
|
| 96 |
+
212,
|
| 97 |
+
657,
|
| 98 |
+
787,
|
| 99 |
+
809
|
| 100 |
+
],
|
| 101 |
+
"page_idx": 0
|
| 102 |
+
},
|
| 103 |
+
{
|
| 104 |
+
"type": "text",
|
| 105 |
+
"text": "Occupancy prediction has gained popularity recently due to its effectiveness in representing both background and foreground objects within a scene using a",
|
| 106 |
+
"bbox": [
|
| 107 |
+
212,
|
| 108 |
+
809,
|
| 109 |
+
785,
|
| 110 |
+
839
|
| 111 |
+
],
|
| 112 |
+
"page_idx": 0
|
| 113 |
+
},
|
| 114 |
+
{
|
| 115 |
+
"type": "text",
|
| 116 |
+
"text": "unified representation. Although significant progress has been made in outdoor driving scenarios, research in indoor scenarios remains limited. Indoor scenes differ from outdoor driving scenes in two key aspects: (1) scene-scale complexity: Indoor rooms often exhibit a more diverse range of sizes compared to outdoor environments, where driving scenarios typically focus on a fixed 3D space for perception. This diversity, ranging from spacious living rooms to narrow kitchens, poses higher precision requirements for depth prediction. (2) object complexity: Indoor scenes feature a higher density and greater variety of objects. Unlike outdoor objects such as vehicles and pedestrians, which typically have consistent sizes within their respective categories and are relatively well-separated, indoor furniture and other objects often exhibit significant variations in scale and are closely positioned to each other. This increased complexity necessitates more sophisticated 3D perception techniques to accurately capture and understand the intricate geometry and relationships among objects within indoor spaces. Furthermore, existing works [5,45] focusing on indoor scenes primarily utilize the NYUv2 dataset [29], which lacks a more scalable and generalizable benchmark for comprehensive evaluation.",
|
| 117 |
+
"bbox": [
|
| 118 |
+
212,
|
| 119 |
+
146,
|
| 120 |
+
787,
|
| 121 |
+
402
|
| 122 |
+
],
|
| 123 |
+
"page_idx": 1
|
| 124 |
+
},
|
| 125 |
+
{
|
| 126 |
+
"type": "text",
|
| 127 |
+
"text": "To address the above issues, we firstly introduce a more scalable benchmark, Occ-ScanNet, for 3D occupancy prediction in indoor scenes. This benchmark builds upon the large-scale ScanNet [8] dataset, offering a 40 times more samples compared to the NYUv2 [29] dataset, thus greatly expanding the scope and diversity of indoor scene study. To effectively address the unique complexities of indoor scenes, we propose a novel method named ISO (Indoor Scene Occupancy). ISO leverages a powerful pretrained depth model and integrates a D-FLoSP (Dual Feature Line of Sight Projection) module. This module enables precise depth estimation and facilitates learning voxel features for accurate predictions. To accommodate the varying sizes of indoor objects, we introduce a multi-scale feature fusion module. This module enhances the learning of object features across different scales. Leveraging a powerful depth model, our method is adept at handling diverse indoor scenes, offering a robust and versatile solution.",
|
| 128 |
+
"bbox": [
|
| 129 |
+
212,
|
| 130 |
+
405,
|
| 131 |
+
787,
|
| 132 |
+
602
|
| 133 |
+
],
|
| 134 |
+
"page_idx": 1
|
| 135 |
+
},
|
| 136 |
+
{
|
| 137 |
+
"type": "text",
|
| 138 |
+
"text": "Our main contributions can be summarized as follows.",
|
| 139 |
+
"bbox": [
|
| 140 |
+
238,
|
| 141 |
+
604,
|
| 142 |
+
632,
|
| 143 |
+
618
|
| 144 |
+
],
|
| 145 |
+
"page_idx": 1
|
| 146 |
+
},
|
| 147 |
+
{
|
| 148 |
+
"type": "list",
|
| 149 |
+
"sub_type": "text",
|
| 150 |
+
"list_items": [
|
| 151 |
+
"- We introduce a new benchmark called Occ-ScanNet for monocular 3D occupancy prediction in indoor scenes. With a dataset size 40 times larger than the NYUv2 dataset, it significantly enhances the potential for scalable research in indoor scene analysis.",
|
| 152 |
+
"- We propose a novel approach called $ISO$ , which primarily comprises the DFLoSP (Dual Feature Line of Sight Projection) module and a multi-scale feature fusion module. Together, these components effectively address the challenges posed by variations in scene and object sizes, enabling more accurate and robust 3D occupancy prediction.",
|
| 153 |
+
"- Experiments on the Occ-ScanNet and NYUv2 datasets demonstrate that our approach achieves state-of-the-art performance in monocular 3D occupancy prediction. Furthermore, our method exhibits significant scalability potential."
|
| 154 |
+
],
|
| 155 |
+
"bbox": [
|
| 156 |
+
223,
|
| 157 |
+
638,
|
| 158 |
+
784,
|
| 159 |
+
838
|
| 160 |
+
],
|
| 161 |
+
"page_idx": 1
|
| 162 |
+
},
|
| 163 |
+
{
|
| 164 |
+
"type": "page_number",
|
| 165 |
+
"text": "2",
|
| 166 |
+
"bbox": [
|
| 167 |
+
217,
|
| 168 |
+
114,
|
| 169 |
+
228,
|
| 170 |
+
126
|
| 171 |
+
],
|
| 172 |
+
"page_idx": 1
|
| 173 |
+
},
|
| 174 |
+
{
|
| 175 |
+
"type": "header",
|
| 176 |
+
"text": "H. Yu et al.",
|
| 177 |
+
"bbox": [
|
| 178 |
+
271,
|
| 179 |
+
114,
|
| 180 |
+
351,
|
| 181 |
+
127
|
| 182 |
+
],
|
| 183 |
+
"page_idx": 1
|
| 184 |
+
},
|
| 185 |
+
{
|
| 186 |
+
"type": "text",
|
| 187 |
+
"text": "2 Related works",
|
| 188 |
+
"text_level": 1,
|
| 189 |
+
"bbox": [
|
| 190 |
+
215,
|
| 191 |
+
143,
|
| 192 |
+
392,
|
| 193 |
+
160
|
| 194 |
+
],
|
| 195 |
+
"page_idx": 2
|
| 196 |
+
},
|
| 197 |
+
{
|
| 198 |
+
"type": "text",
|
| 199 |
+
"text": "2.1 Monocular 3D Semantic Scene Completion",
|
| 200 |
+
"text_level": 1,
|
| 201 |
+
"bbox": [
|
| 202 |
+
215,
|
| 203 |
+
176,
|
| 204 |
+
616,
|
| 205 |
+
191
|
| 206 |
+
],
|
| 207 |
+
"page_idx": 2
|
| 208 |
+
},
|
| 209 |
+
{
|
| 210 |
+
"type": "text",
|
| 211 |
+
"text": "Monocular 3D Semantic Scene Completion (SSC) aims to infer the complete 3D structure and corresponding semantics from a single image. This monocular setting was first introduced by Monoscene [5], advancing upon prior SSC methods [18,23,30,48,51] by relying solely on vision, without additional 3D inputs. Monoscene [5] introduces the Features Line of Sight Projection (FLoSP), inspired by optics, to obtain 3D features through ray projection. However, the shared 2D features lifted to 3D rays via FLoSP results in depth ambiguity. Consequently, subsequent efforts have placed greater emphasis on leveraging depth information. VoxFormer [19] proposed a novel query proposal network based on 2D convolutions, generating sparse queries from image depth, which showed impressive performance in driving scenes. Meanwhile, NDC-Scene [45] further devised a Depth-Adaptive Dual Decoder to concurrently upsample and merge the 2D and 3D feature maps, thereby enhancing overall performance.",
|
| 212 |
+
"bbox": [
|
| 213 |
+
212,
|
| 214 |
+
200,
|
| 215 |
+
787,
|
| 216 |
+
398
|
| 217 |
+
],
|
| 218 |
+
"page_idx": 2
|
| 219 |
+
},
|
| 220 |
+
{
|
| 221 |
+
"type": "text",
|
| 222 |
+
"text": "2.2 Multiview 3D Occupancy Prediction",
|
| 223 |
+
"text_level": 1,
|
| 224 |
+
"bbox": [
|
| 225 |
+
215,
|
| 226 |
+
420,
|
| 227 |
+
565,
|
| 228 |
+
435
|
| 229 |
+
],
|
| 230 |
+
"page_idx": 2
|
| 231 |
+
},
|
| 232 |
+
{
|
| 233 |
+
"type": "text",
|
| 234 |
+
"text": "Multiview 3D Occupancy Prediction predicts the semantic occupancy of the surrounding 3D scene given multiview images, has recently garnered significant attention in the field of autonomous driving. TPVFormer [17] pioneers exploration in driving scenes, employing sparse LiDAR labels for supervision. It introduces a tri-perspective view (TPV) representation that accompanies BEV with two additional perpendicular planes. However, due to the sparse supervision provided by LiDAR, subsequent works [33, 34, 36, 40] have focused on providing more dense occupancy benchmarks. FB-OCC [22] integrates the lifting of 2D to 3D and querying of 3D to 2D, achieving a more efficient 3D feature transformation. Meanwhile, PanoOcc [38] unifies detection and semantic occupancy tasks, enabling comprehensive panoramic scene understanding. Recent methods also explore improving model efficiency [46] and utilizing weaker forms of supervision [16, 25].",
|
| 235 |
+
"bbox": [
|
| 236 |
+
212,
|
| 237 |
+
445,
|
| 238 |
+
787,
|
| 239 |
+
642
|
| 240 |
+
],
|
| 241 |
+
"page_idx": 2
|
| 242 |
+
},
|
| 243 |
+
{
|
| 244 |
+
"type": "text",
|
| 245 |
+
"text": "2.3 3D Reconstruction from Image",
|
| 246 |
+
"text_level": 1,
|
| 247 |
+
"bbox": [
|
| 248 |
+
215,
|
| 249 |
+
664,
|
| 250 |
+
517,
|
| 251 |
+
679
|
| 252 |
+
],
|
| 253 |
+
"page_idx": 2
|
| 254 |
+
},
|
| 255 |
+
{
|
| 256 |
+
"type": "text",
|
| 257 |
+
"text": "3D reconstruction is a technique that involves recovering a 3D representation of objects [1, 11, 26], scenes [9, 41, 49], or even human bodies [12, 13, 50] from camera images. The task of 3D reconstruction can be classified into two categories: monocular reconstruction and multi-view reconstruction, depending on the number of images utilized. In the context of indoor scenes, 3D scene reconstruction aims to determine the surface geometry of the entire scene, often without incorporating semantic information. Atlas [24] proposes an end-to-end 3D reconstruction framework using TSDF regression from RGB images, bypassing traditional depth map estimation for efficient semantic segmentation. Panoptic-Reconstruction [7] unifies geometric reconstruction, 3D semantic, and",
|
| 258 |
+
"bbox": [
|
| 259 |
+
212,
|
| 260 |
+
689,
|
| 261 |
+
787,
|
| 262 |
+
840
|
| 263 |
+
],
|
| 264 |
+
"page_idx": 2
|
| 265 |
+
},
|
| 266 |
+
{
|
| 267 |
+
"type": "header",
|
| 268 |
+
"text": "Monocular Occupancy Prediction for Scalable Indoor Scenes",
|
| 269 |
+
"bbox": [
|
| 270 |
+
328,
|
| 271 |
+
114,
|
| 272 |
+
730,
|
| 273 |
+
128
|
| 274 |
+
],
|
| 275 |
+
"page_idx": 2
|
| 276 |
+
},
|
| 277 |
+
{
|
| 278 |
+
"type": "page_number",
|
| 279 |
+
"text": "3",
|
| 280 |
+
"bbox": [
|
| 281 |
+
774,
|
| 282 |
+
116,
|
| 283 |
+
784,
|
| 284 |
+
126
|
| 285 |
+
],
|
| 286 |
+
"page_idx": 2
|
| 287 |
+
},
|
| 288 |
+
{
|
| 289 |
+
"type": "image",
|
| 290 |
+
"img_path": "images/cb27d372c27774b713bc3f76a42488bc37ffc0f9440739cdc29e72df35393e5f.jpg",
|
| 291 |
+
"image_caption": [
|
| 292 |
+
"Fig. 1: The core design of ISO centers around the transformation of features from 2D to 3D spaces, encompassing the Depth Branch and the D-FLoSP module. A depth branch is initially integrated, it leverages a pre-trained depth model to estimate a pixel-wise depth map which is processed by the DepthNet to generate the final depth distribution. An element-wise multiplication between the voxel depth and features followed by summation are subsequently performed to derive the initial 3D voxel feature. The 3D feature is further processed to predict the 3D scene occupancy."
|
| 293 |
+
],
|
| 294 |
+
"image_footnote": [],
|
| 295 |
+
"bbox": [
|
| 296 |
+
215,
|
| 297 |
+
150,
|
| 298 |
+
785,
|
| 299 |
+
287
|
| 300 |
+
],
|
| 301 |
+
"page_idx": 3
|
| 302 |
+
},
|
| 303 |
+
{
|
| 304 |
+
"type": "text",
|
| 305 |
+
"text": "instance segmentation tasks. It predicts complete geometric reconstruction, semantic, and instance segmentations from a single RGB image's camera frustum. SCFusion [42], a real-time scene reconstruction framework, integrates continuous depth data using neural architecture for occupancy maps. It efficiently combines semantic completion with voxel states for simultaneous scene reconstruction and semantic understanding in real-time 3D environments.",
|
| 306 |
+
"bbox": [
|
| 307 |
+
212,
|
| 308 |
+
429,
|
| 309 |
+
787,
|
| 310 |
+
521
|
| 311 |
+
],
|
| 312 |
+
"page_idx": 3
|
| 313 |
+
},
|
| 314 |
+
{
|
| 315 |
+
"type": "text",
|
| 316 |
+
"text": "3 Method",
|
| 317 |
+
"text_level": 1,
|
| 318 |
+
"bbox": [
|
| 319 |
+
215,
|
| 320 |
+
547,
|
| 321 |
+
330,
|
| 322 |
+
565
|
| 323 |
+
],
|
| 324 |
+
"page_idx": 3
|
| 325 |
+
},
|
| 326 |
+
{
|
| 327 |
+
"type": "text",
|
| 328 |
+
"text": "Problem definition. We focus on the problem of monocular 3D Occupancy Prediction. Specifically, this task takes a single RGB image $\\mathbf{I}^{RGB}$ as input and output a voxel-wise occupancy along with semantic categories $\\mathbf{Y}^{X\\times Y\\times Z\\times C}$ . $X$ , $Y$ , and $Z$ represent the dimensions of the predicted 3D scene, while $C$ denotes the total number of semantic categories.",
|
| 329 |
+
"bbox": [
|
| 330 |
+
212,
|
| 331 |
+
585,
|
| 332 |
+
787,
|
| 333 |
+
662
|
| 334 |
+
],
|
| 335 |
+
"page_idx": 3
|
| 336 |
+
},
|
| 337 |
+
{
|
| 338 |
+
"type": "text",
|
| 339 |
+
"text": "Model overview. In this section, we introduce the overall architecture of ISO, as shown in Fig. 1. Following Monoscene [5], we utilize 2D Unet and 3D Unet architectures to handle 2D and 3D features. Our core design focuses on transforming features from 2D to 3D, incorporating the Depth Branch and D-FLoSP module. Specifically, we incorporate a depth branch to estimate the pixel-wise depth map using a pre-trained depth model. Processed by the DepthNet, the model outputs a refined depth distribution. Then, voxel depth and features are multiplied element-wise and summed to obtain the initial 3D voxel feature volume $\\mathbf{X}^{3\\mathbf{d}} \\in \\mathbb{R}^{X \\times Y \\times Z \\times C}$ . Finally, after processing the 3D voxel features, the model outputs the 3D scene occupancy.",
|
| 340 |
+
"bbox": [
|
| 341 |
+
212,
|
| 342 |
+
688,
|
| 343 |
+
787,
|
| 344 |
+
840
|
| 345 |
+
],
|
| 346 |
+
"page_idx": 3
|
| 347 |
+
},
|
| 348 |
+
{
|
| 349 |
+
"type": "page_number",
|
| 350 |
+
"text": "4",
|
| 351 |
+
"bbox": [
|
| 352 |
+
217,
|
| 353 |
+
114,
|
| 354 |
+
228,
|
| 355 |
+
126
|
| 356 |
+
],
|
| 357 |
+
"page_idx": 3
|
| 358 |
+
},
|
| 359 |
+
{
|
| 360 |
+
"type": "header",
|
| 361 |
+
"text": "H. Yu et al.",
|
| 362 |
+
"bbox": [
|
| 363 |
+
271,
|
| 364 |
+
114,
|
| 365 |
+
352,
|
| 366 |
+
126
|
| 367 |
+
],
|
| 368 |
+
"page_idx": 3
|
| 369 |
+
},
|
| 370 |
+
{
|
| 371 |
+
"type": "text",
|
| 372 |
+
"text": "3.1 Depth Branch",
|
| 373 |
+
"text_level": 1,
|
| 374 |
+
"bbox": [
|
| 375 |
+
215,
|
| 376 |
+
146,
|
| 377 |
+
380,
|
| 378 |
+
161
|
| 379 |
+
],
|
| 380 |
+
"page_idx": 4
|
| 381 |
+
},
|
| 382 |
+
{
|
| 383 |
+
"type": "text",
|
| 384 |
+
"text": "In this section, we introduce how to effectively estimate depth information from a single image.",
|
| 385 |
+
"bbox": [
|
| 386 |
+
212,
|
| 387 |
+
167,
|
| 388 |
+
782,
|
| 389 |
+
198
|
| 390 |
+
],
|
| 391 |
+
"page_idx": 4
|
| 392 |
+
},
|
| 393 |
+
{
|
| 394 |
+
"type": "text",
|
| 395 |
+
"text": "Coarse depth estimation. Learning depth from scratch can be quite challenging. However, thanks to the recent rapid advancements in depth estimation [3,4,27,44], we can leverage pre-trained depth models to initially estimate a coarse depth map.",
|
| 396 |
+
"bbox": [
|
| 397 |
+
212,
|
| 398 |
+
217,
|
| 399 |
+
782,
|
| 400 |
+
276
|
| 401 |
+
],
|
| 402 |
+
"page_idx": 4
|
| 403 |
+
},
|
| 404 |
+
{
|
| 405 |
+
"type": "text",
|
| 406 |
+
"text": "Compared to past models that could only estimate relative depth, we opted for Depth-Anything [44] as the pre-trained depth model, because it excels in predicting metric depth, as the Eq. (1) shows:",
|
| 407 |
+
"bbox": [
|
| 408 |
+
212,
|
| 409 |
+
277,
|
| 410 |
+
784,
|
| 411 |
+
324
|
| 412 |
+
],
|
| 413 |
+
"page_idx": 4
|
| 414 |
+
},
|
| 415 |
+
{
|
| 416 |
+
"type": "equation",
|
| 417 |
+
"text": "\n$$\n\\mathbf {D} ^ {\\mathrm {m e t r i c}} = \\mathbf {N} _ {\\mathrm {d e p t h}} (\\mathbf {I} ^ {\\mathrm {r g b}}) \\in \\mathbb {R} ^ {1 \\times H \\times W}, \\tag {1}\n$$\n",
|
| 418 |
+
"text_format": "latex",
|
| 419 |
+
"bbox": [
|
| 420 |
+
372,
|
| 421 |
+
328,
|
| 422 |
+
785,
|
| 423 |
+
348
|
| 424 |
+
],
|
| 425 |
+
"page_idx": 4
|
| 426 |
+
},
|
| 427 |
+
{
|
| 428 |
+
"type": "text",
|
| 429 |
+
"text": "where $\\mathbf{N}_{\\mathrm{depth}}$ denotes the pre-trained depth model [3, 44], and $\\mathbf{I}^{\\mathrm{rgb}}$ represents the input image. $H$ and $W$ represents the height and width of the input image.",
|
| 430 |
+
"bbox": [
|
| 431 |
+
214,
|
| 432 |
+
354,
|
| 433 |
+
782,
|
| 434 |
+
386
|
| 435 |
+
],
|
| 436 |
+
"page_idx": 4
|
| 437 |
+
},
|
| 438 |
+
{
|
| 439 |
+
"type": "text",
|
| 440 |
+
"text": "Depth refinement. The coarse depth estimation is subsequently refined through model learning. The depth from pre-trained depth model is not precise enough for a higher mIoU, so we design a fine-tuning strategy. Specifically, the predicted metric depth $\\mathbf{D}^{\\mathrm{metric}}$ is then concatenated with the image feature $\\mathbf{X}$ that has the same spatial scale. The augmented feature is continue processed by a following DepthNet to get a refined depth distribution $\\mathbf{D}_{s=1}^{\\mathrm{dist}}$ , as Eq. (2) shows:",
|
| 441 |
+
"bbox": [
|
| 442 |
+
214,
|
| 443 |
+
404,
|
| 444 |
+
789,
|
| 445 |
+
496
|
| 446 |
+
],
|
| 447 |
+
"page_idx": 4
|
| 448 |
+
},
|
| 449 |
+
{
|
| 450 |
+
"type": "equation",
|
| 451 |
+
"text": "\n$$\n\\mathbf {D} _ {\\mathrm {s} = 1} ^ {\\text {d i s t}} = \\mathbf {F} _ {\\text {d e p t h}} \\left(\\operatorname {C o n c a t} \\left(\\mathbf {D} ^ {\\text {m e t r i c}}, \\mathbf {X} _ {\\mathrm {s} = 1} ^ {2 \\mathrm {d}}\\right)\\right), \\quad \\mathbf {D} _ {\\mathrm {s} = 1} ^ {\\text {d i s t}} \\in \\mathbb {R} ^ {N _ {\\text {b i n s}} \\times H \\times W}, \\tag {2}\n$$\n",
|
| 452 |
+
"text_format": "latex",
|
| 453 |
+
"bbox": [
|
| 454 |
+
267,
|
| 455 |
+
513,
|
| 456 |
+
784,
|
| 457 |
+
547
|
| 458 |
+
],
|
| 459 |
+
"page_idx": 4
|
| 460 |
+
},
|
| 461 |
+
{
|
| 462 |
+
"type": "text",
|
| 463 |
+
"text": "where $\\mathbf{F}_{\\mathrm{depth}}$ is the DepthNet, $\\mathbf{X}_{s=1}^{2d}$ is the 2D image feature, and $N_{\\mathrm{bins}}$ is the number of discrete depth bins. $s$ denotes the level of feature scale, where $s = 1$ implies that the feature is at a 1:1 original image ratio.",
|
| 464 |
+
"bbox": [
|
| 465 |
+
212,
|
| 466 |
+
554,
|
| 467 |
+
782,
|
| 468 |
+
599
|
| 469 |
+
],
|
| 470 |
+
"page_idx": 4
|
| 471 |
+
},
|
| 472 |
+
{
|
| 473 |
+
"type": "text",
|
| 474 |
+
"text": "We down-samples the scale $\\frac{1}{1}$ distribution to get other smaller-scale 2D image feature maps' depth distribution. We represent the procedure as Eq. (3):",
|
| 475 |
+
"bbox": [
|
| 476 |
+
212,
|
| 477 |
+
599,
|
| 478 |
+
782,
|
| 479 |
+
631
|
| 480 |
+
],
|
| 481 |
+
"page_idx": 4
|
| 482 |
+
},
|
| 483 |
+
{
|
| 484 |
+
"type": "equation",
|
| 485 |
+
"text": "\n$$\n\\mathbf {D} _ {\\mathrm {s} = k} ^ {\\text {d i s t}} = \\operatorname {D o w n S a m p l e} \\left(\\mathbf {D} _ {\\mathrm {s} = 1} ^ {\\text {d i s t}}\\right), k = 2, 4, 8. \\tag {3}\n$$\n",
|
| 486 |
+
"text_format": "latex",
|
| 487 |
+
"bbox": [
|
| 488 |
+
346,
|
| 489 |
+
636,
|
| 490 |
+
784,
|
| 491 |
+
655
|
| 492 |
+
],
|
| 493 |
+
"page_idx": 4
|
| 494 |
+
},
|
| 495 |
+
{
|
| 496 |
+
"type": "text",
|
| 497 |
+
"text": "The refined depth is supervised by the ground truth $\\mathbf{D}^{\\mathrm{GT}}$ . The ground truth has single depth value at each pixel, so we convert the single value to a one-hot vector of length $N_{\\mathrm{bins}}$ . The refined depth prediction can be optimized using the BCE loss function as Eq. (4) shows.",
|
| 498 |
+
"bbox": [
|
| 499 |
+
212,
|
| 500 |
+
662,
|
| 501 |
+
782,
|
| 502 |
+
724
|
| 503 |
+
],
|
| 504 |
+
"page_idx": 4
|
| 505 |
+
},
|
| 506 |
+
{
|
| 507 |
+
"type": "equation",
|
| 508 |
+
"text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {\\mathrm {d e p t h}} = - \\frac {1}{N _ {\\mathrm {b i n s}} \\times H \\times W} \\\\ \\sum_ {d = 1} ^ {N _ {\\mathrm {b i n s}}} \\sum_ {h = 1} ^ {H} \\sum_ {w = 1} ^ {W} \\left[ \\mathbf {D} _ {N _ {\\mathrm {b i n s}}, d, h, w} ^ {\\mathrm {G T}} \\cdot \\log \\left(\\mathbf {D} _ {N _ {\\mathrm {b i n s}}, d, h, w} ^ {\\mathrm {d i s t}}\\right) \\right. \\\\ \\left. + \\left(1 - \\mathbf {D} _ {N _ {\\mathrm {b i n s}}, d, h, w} ^ {\\mathrm {G T}}\\right) \\cdot \\log \\left(1 - \\mathbf {D} _ {N _ {\\mathrm {b i n s}}, d, h, w} ^ {\\mathrm {d i s t}}\\right) \\right] \\quad (4) \\\\ \\end{array}\n$$\n",
|
| 509 |
+
"text_format": "latex",
|
| 510 |
+
"bbox": [
|
| 511 |
+
228,
|
| 512 |
+
744,
|
| 513 |
+
785,
|
| 514 |
+
843
|
| 515 |
+
],
|
| 516 |
+
"page_idx": 4
|
| 517 |
+
},
|
| 518 |
+
{
|
| 519 |
+
"type": "header",
|
| 520 |
+
"text": "Monocular Occupancy Prediction for Scalable Indoor Scenes",
|
| 521 |
+
"bbox": [
|
| 522 |
+
328,
|
| 523 |
+
114,
|
| 524 |
+
732,
|
| 525 |
+
128
|
| 526 |
+
],
|
| 527 |
+
"page_idx": 4
|
| 528 |
+
},
|
| 529 |
+
{
|
| 530 |
+
"type": "page_number",
|
| 531 |
+
"text": "5",
|
| 532 |
+
"bbox": [
|
| 533 |
+
774,
|
| 534 |
+
116,
|
| 535 |
+
784,
|
| 536 |
+
126
|
| 537 |
+
],
|
| 538 |
+
"page_idx": 4
|
| 539 |
+
},
|
| 540 |
+
{
|
| 541 |
+
"type": "text",
|
| 542 |
+
"text": "3.2 Dual Feature Line of Sight Projection (D-FLoSP)",
|
| 543 |
+
"text_level": 1,
|
| 544 |
+
"bbox": [
|
| 545 |
+
215,
|
| 546 |
+
146,
|
| 547 |
+
671,
|
| 548 |
+
162
|
| 549 |
+
],
|
| 550 |
+
"page_idx": 5
|
| 551 |
+
},
|
| 552 |
+
{
|
| 553 |
+
"type": "text",
|
| 554 |
+
"text": "Compared to the FLoSP module in MonoScene [5], which projects 3D voxels to 2D, and 3D feature vectors are the corresponding pixel feature vectors. It only considers the projection of 3D rays and overlooks depth information, our proposed D-FLoSP module can more effectively integrate depth information.",
|
| 555 |
+
"bbox": [
|
| 556 |
+
212,
|
| 557 |
+
180,
|
| 558 |
+
782,
|
| 559 |
+
241
|
| 560 |
+
],
|
| 561 |
+
"page_idx": 5
|
| 562 |
+
},
|
| 563 |
+
{
|
| 564 |
+
"type": "text",
|
| 565 |
+
"text": "Specifically, after getting depth distribution in the camera frame, we implement the FLoSP module to depth distribution. Each 3D voxel centroid position $x^{c}$ in the world frame can be projected to the camera frame and the pixel frame using camera pose and intrinsic matrix. We assign the projected depth values to discrete depth bins index by Eqs. (5) and (6), according to [32],",
|
| 566 |
+
"bbox": [
|
| 567 |
+
212,
|
| 568 |
+
243,
|
| 569 |
+
784,
|
| 570 |
+
318
|
| 571 |
+
],
|
| 572 |
+
"page_idx": 5
|
| 573 |
+
},
|
| 574 |
+
{
|
| 575 |
+
"type": "equation",
|
| 576 |
+
"text": "\n$$\n\\delta = \\frac {2 \\left(d _ {\\max } - d _ {\\min }\\right)}{N _ {b i n s} \\left(1 + N _ {b i n s}\\right)}, \\tag {5}\n$$\n",
|
| 577 |
+
"text_format": "latex",
|
| 578 |
+
"bbox": [
|
| 579 |
+
382,
|
| 580 |
+
333,
|
| 581 |
+
784,
|
| 582 |
+
364
|
| 583 |
+
],
|
| 584 |
+
"page_idx": 5
|
| 585 |
+
},
|
| 586 |
+
{
|
| 587 |
+
"type": "equation",
|
| 588 |
+
"text": "\n$$\nl = - 0. 5 + 0. 5 \\sqrt {1 + \\frac {8 \\left(d - d _ {\\operatorname* {m i n}}\\right)}{\\delta}}, \\tag {6}\n$$\n",
|
| 589 |
+
"text_format": "latex",
|
| 590 |
+
"bbox": [
|
| 591 |
+
385,
|
| 592 |
+
369,
|
| 593 |
+
784,
|
| 594 |
+
401
|
| 595 |
+
],
|
| 596 |
+
"page_idx": 5
|
| 597 |
+
},
|
| 598 |
+
{
|
| 599 |
+
"type": "text",
|
| 600 |
+
"text": "and the depth distribution probability it lies in serves as the depth probability of that voxel. The process is illustrated by Eq. (7):",
|
| 601 |
+
"bbox": [
|
| 602 |
+
212,
|
| 603 |
+
415,
|
| 604 |
+
782,
|
| 605 |
+
445
|
| 606 |
+
],
|
| 607 |
+
"page_idx": 5
|
| 608 |
+
},
|
| 609 |
+
{
|
| 610 |
+
"type": "equation",
|
| 611 |
+
"text": "\n$$\n\\mathbf {D} _ {\\mathrm {s} = k} ^ {\\mathrm {3 d}} = \\Phi_ {\\rho \\left(x ^ {c}\\right)} ^ {\\mathrm {3 d}} \\left(\\mathbf {D} _ {\\mathrm {s} = k} ^ {\\mathrm {d i s t}}\\right) \\in \\mathbb {R} ^ {X \\times Y \\times Z \\times C}, k = 1, 2, 4, 8, \\tag {7}\n$$\n",
|
| 612 |
+
"text_format": "latex",
|
| 613 |
+
"bbox": [
|
| 614 |
+
323,
|
| 615 |
+
459,
|
| 616 |
+
784,
|
| 617 |
+
479
|
| 618 |
+
],
|
| 619 |
+
"page_idx": 5
|
| 620 |
+
},
|
| 621 |
+
{
|
| 622 |
+
"type": "text",
|
| 623 |
+
"text": "where $\\Phi_{a}^{3\\mathrm{d}}(b)$ is the 3D sampling of b at coordinates a, and $\\rho (\\cdot)$ is the perspective projection. Thus, the depth distribution is lifted from 2D to 3D. The bin-based representation with quantization is used for mapping continuous depth value to a discrete form, each bin will be assigned a different depth distribution score to make voxel aware of depth. The original depth precision is a continuous value, and hard to integrate it with the discrete form of voxel.",
|
| 624 |
+
"bbox": [
|
| 625 |
+
212,
|
| 626 |
+
494,
|
| 627 |
+
784,
|
| 628 |
+
584
|
| 629 |
+
],
|
| 630 |
+
"page_idx": 5
|
| 631 |
+
},
|
| 632 |
+
{
|
| 633 |
+
"type": "text",
|
| 634 |
+
"text": "Following [5], we also employ the FLoSP module to generate a 3D voxel feature from a 2D feature map. Similar to what we did with the voxel centroids, the projected voxel can sample corresponding $1:k$ scaled feature maps $F_{1:k}$ from the 2D UNet decoder. This process is illustrated by Equation Eq. (8):",
|
| 635 |
+
"bbox": [
|
| 636 |
+
212,
|
| 637 |
+
585,
|
| 638 |
+
784,
|
| 639 |
+
648
|
| 640 |
+
],
|
| 641 |
+
"page_idx": 5
|
| 642 |
+
},
|
| 643 |
+
{
|
| 644 |
+
"type": "equation",
|
| 645 |
+
"text": "\n$$\n\\mathbf {X} _ {\\mathrm {s} = k} ^ {3 \\mathrm {d}} = \\Phi_ {\\rho \\left(x ^ {c}\\right)} ^ {2 \\mathrm {d}} \\left(\\mathbf {X} _ {\\mathrm {s} = k} ^ {2 \\mathrm {d}}\\right) \\in \\mathbb {R} ^ {X \\times Y \\times Z \\times C}, k = 1, 2, 4, 8, \\tag {8}\n$$\n",
|
| 646 |
+
"text_format": "latex",
|
| 647 |
+
"bbox": [
|
| 648 |
+
323,
|
| 649 |
+
662,
|
| 650 |
+
784,
|
| 651 |
+
681
|
| 652 |
+
],
|
| 653 |
+
"page_idx": 5
|
| 654 |
+
},
|
| 655 |
+
{
|
| 656 |
+
"type": "text",
|
| 657 |
+
"text": "where $\\Phi_{a}^{2\\mathrm{d}}(b)$ is the 2D sampling of b at coordinates a, and $\\rho (\\cdot)$ is the perspective projection.",
|
| 658 |
+
"bbox": [
|
| 659 |
+
212,
|
| 660 |
+
696,
|
| 661 |
+
782,
|
| 662 |
+
727
|
| 663 |
+
],
|
| 664 |
+
"page_idx": 5
|
| 665 |
+
},
|
| 666 |
+
{
|
| 667 |
+
"type": "text",
|
| 668 |
+
"text": "The final 3D feature map $\\mathbf{X}^{3\\mathrm{d}}$ are summed in Eq. (9):",
|
| 669 |
+
"bbox": [
|
| 670 |
+
238,
|
| 671 |
+
729,
|
| 672 |
+
630,
|
| 673 |
+
744
|
| 674 |
+
],
|
| 675 |
+
"page_idx": 5
|
| 676 |
+
},
|
| 677 |
+
{
|
| 678 |
+
"type": "equation",
|
| 679 |
+
"text": "\n$$\n\\mathbf {X} ^ {\\mathrm {3 d}} = \\sum_ {\\mathrm {s} \\in \\{1, 2, 4, 8 \\}} \\mathbf {X} _ {\\mathrm {s} = k} ^ {\\mathrm {3 d}} \\odot \\mathbf {D} _ {\\mathrm {s} = k} ^ {\\mathrm {3 d}}, \\tag {9}\n$$\n",
|
| 680 |
+
"text_format": "latex",
|
| 681 |
+
"bbox": [
|
| 682 |
+
392,
|
| 683 |
+
760,
|
| 684 |
+
784,
|
| 685 |
+
792
|
| 686 |
+
],
|
| 687 |
+
"page_idx": 5
|
| 688 |
+
},
|
| 689 |
+
{
|
| 690 |
+
"type": "text",
|
| 691 |
+
"text": "where $\\odot$ is the element-wise multiplication. The output map $\\mathbf{X}^{3\\mathrm{d}}$ is then served as the input of 3D UNet.",
|
| 692 |
+
"bbox": [
|
| 693 |
+
212,
|
| 694 |
+
809,
|
| 695 |
+
782,
|
| 696 |
+
839
|
| 697 |
+
],
|
| 698 |
+
"page_idx": 5
|
| 699 |
+
},
|
| 700 |
+
{
|
| 701 |
+
"type": "page_number",
|
| 702 |
+
"text": "6",
|
| 703 |
+
"bbox": [
|
| 704 |
+
217,
|
| 705 |
+
114,
|
| 706 |
+
228,
|
| 707 |
+
126
|
| 708 |
+
],
|
| 709 |
+
"page_idx": 5
|
| 710 |
+
},
|
| 711 |
+
{
|
| 712 |
+
"type": "header",
|
| 713 |
+
"text": "H. Yu et al.",
|
| 714 |
+
"bbox": [
|
| 715 |
+
271,
|
| 716 |
+
114,
|
| 717 |
+
351,
|
| 718 |
+
127
|
| 719 |
+
],
|
| 720 |
+
"page_idx": 5
|
| 721 |
+
},
|
| 722 |
+
{
|
| 723 |
+
"type": "image",
|
| 724 |
+
"img_path": "images/4e4b51667dfd71e9661d016ed42f74358aa31bb99bf8b59fd60904e8f7f647cb.jpg",
|
| 725 |
+
"image_caption": [
|
| 726 |
+
"(a) Depth ranges in NYUv2 and our Occ-ScanNet Benchmark"
|
| 727 |
+
],
|
| 728 |
+
"image_footnote": [],
|
| 729 |
+
"bbox": [
|
| 730 |
+
271,
|
| 731 |
+
162,
|
| 732 |
+
429,
|
| 733 |
+
287
|
| 734 |
+
],
|
| 735 |
+
"page_idx": 6
|
| 736 |
+
},
|
| 737 |
+
{
|
| 738 |
+
"type": "image",
|
| 739 |
+
"img_path": "images/225cd6aab82aef724a36995945316cee74bdd5373a4c636b37ae5faa67dde27d.jpg",
|
| 740 |
+
"image_caption": [
|
| 741 |
+
"(b) Dataset attributes of NYUv2 and our OccScanNet Benchmark",
|
| 742 |
+
"Fig. 2: Comparison of NYUv2 and Occ-ScanNet Benchmark. In (a), the depth ranges of NYUv2 and Occ-ScanNet are distinguished by dark and light green, respectively, with the horizontal axis indicating the minimum depth and the vertical axis showing the maximum depth of scenes. (b) quantitatively demonstrates that Occ-ScanNet possesses a significantly larger data scale compared to the NYUv2 dataset."
|
| 743 |
+
],
|
| 744 |
+
"image_footnote": [],
|
| 745 |
+
"bbox": [
|
| 746 |
+
558,
|
| 747 |
+
162,
|
| 748 |
+
717,
|
| 749 |
+
286
|
| 750 |
+
],
|
| 751 |
+
"page_idx": 6
|
| 752 |
+
},
|
| 753 |
+
{
|
| 754 |
+
"type": "text",
|
| 755 |
+
"text": "4 Occ-ScanNet Benchmark",
|
| 756 |
+
"text_level": 1,
|
| 757 |
+
"bbox": [
|
| 758 |
+
215,
|
| 759 |
+
426,
|
| 760 |
+
493,
|
| 761 |
+
441
|
| 762 |
+
],
|
| 763 |
+
"page_idx": 6
|
| 764 |
+
},
|
| 765 |
+
{
|
| 766 |
+
"type": "text",
|
| 767 |
+
"text": "Compared to the previous widely used indoor scene benchmark NYUv2 [29], which contains only 795 / 654 for train / test samples, our benchmark boasts 45,755 / 19,764 samples. Our benchmark significantly surpasses NYUv2 [29] in both data quantity and richness of scene depth, as illustrated in Fig. 2. This dataset is available at https://huggingface.co/datasets/hongxiaoy/OccScanNet.",
|
| 768 |
+
"bbox": [
|
| 769 |
+
212,
|
| 770 |
+
455,
|
| 771 |
+
787,
|
| 772 |
+
546
|
| 773 |
+
],
|
| 774 |
+
"page_idx": 6
|
| 775 |
+
},
|
| 776 |
+
{
|
| 777 |
+
"type": "text",
|
| 778 |
+
"text": "4.1 Overview",
|
| 779 |
+
"text_level": 1,
|
| 780 |
+
"bbox": [
|
| 781 |
+
215,
|
| 782 |
+
565,
|
| 783 |
+
339,
|
| 784 |
+
579
|
| 785 |
+
],
|
| 786 |
+
"page_idx": 6
|
| 787 |
+
},
|
| 788 |
+
{
|
| 789 |
+
"type": "text",
|
| 790 |
+
"text": "Occ-ScanNet benchmark features a train/Validation split of 45,755 / 19,764 samples. As shown in Fig. 3, Occ-ScanNet exhibits rich diversity in scenes and viewpoints. This diverse dataset not only challenges the task of predicting scene occupancy but also fosters future research endeavors towards developing larger-scale and more versatile occupancy models.",
|
| 791 |
+
"bbox": [
|
| 792 |
+
212,
|
| 793 |
+
587,
|
| 794 |
+
785,
|
| 795 |
+
662
|
| 796 |
+
],
|
| 797 |
+
"page_idx": 6
|
| 798 |
+
},
|
| 799 |
+
{
|
| 800 |
+
"type": "text",
|
| 801 |
+
"text": "4.2 Occupancy Label Generation",
|
| 802 |
+
"text_level": 1,
|
| 803 |
+
"bbox": [
|
| 804 |
+
215,
|
| 805 |
+
681,
|
| 806 |
+
503,
|
| 807 |
+
698
|
| 808 |
+
],
|
| 809 |
+
"page_idx": 6
|
| 810 |
+
},
|
| 811 |
+
{
|
| 812 |
+
"type": "text",
|
| 813 |
+
"text": "We adhere to the data formulation used in NYUv2 [29] dataset. The pipeline of occupancy generation process is showed in Fig. 4. For generating 3D voxel labels, we initially followed the generation process in CompleteScanNet and then employed manual inspection.",
|
| 814 |
+
"bbox": [
|
| 815 |
+
212,
|
| 816 |
+
704,
|
| 817 |
+
785,
|
| 818 |
+
763
|
| 819 |
+
],
|
| 820 |
+
"page_idx": 6
|
| 821 |
+
},
|
| 822 |
+
{
|
| 823 |
+
"type": "text",
|
| 824 |
+
"text": "More specifically, from the official ScanNet [8] scenes, we extracted various data components, including color images, depth images, camera intrinsic matrices, and camera poses. A total of 100 frames were sampled from each scene and were randomly divided into training and validation sets with a $7/3$ split ratio per scene.",
|
| 825 |
+
"bbox": [
|
| 826 |
+
212,
|
| 827 |
+
763,
|
| 828 |
+
785,
|
| 829 |
+
840
|
| 830 |
+
],
|
| 831 |
+
"page_idx": 6
|
| 832 |
+
},
|
| 833 |
+
{
|
| 834 |
+
"type": "header",
|
| 835 |
+
"text": "Monocular Occupancy Prediction for Scalable Indoor Scenes",
|
| 836 |
+
"bbox": [
|
| 837 |
+
328,
|
| 838 |
+
114,
|
| 839 |
+
730,
|
| 840 |
+
128
|
| 841 |
+
],
|
| 842 |
+
"page_idx": 6
|
| 843 |
+
},
|
| 844 |
+
{
|
| 845 |
+
"type": "page_number",
|
| 846 |
+
"text": "7",
|
| 847 |
+
"bbox": [
|
| 848 |
+
774,
|
| 849 |
+
116,
|
| 850 |
+
784,
|
| 851 |
+
126
|
| 852 |
+
],
|
| 853 |
+
"page_idx": 6
|
| 854 |
+
},
|
| 855 |
+
{
|
| 856 |
+
"type": "image",
|
| 857 |
+
"img_path": "images/c16bba8f8ca001189e36dac9dacbdb0c0ff618762516ff7dee305a17b86e6b84.jpg",
|
| 858 |
+
"image_caption": [
|
| 859 |
+
"Fig. 3: Samples Visualization in Occ-ScanNet Benchmark. The original RGB image is shown in column 1,3 and 5, the corresponding scene voxel labels is shown in column 2, 4 and 6. The first two rows are different views from different scenes and the last two rows each is three different views from the same scene."
|
| 860 |
+
],
|
| 861 |
+
"image_footnote": [],
|
| 862 |
+
"bbox": [
|
| 863 |
+
225,
|
| 864 |
+
152,
|
| 865 |
+
776,
|
| 866 |
+
428
|
| 867 |
+
],
|
| 868 |
+
"page_idx": 7
|
| 869 |
+
},
|
| 870 |
+
{
|
| 871 |
+
"type": "text",
|
| 872 |
+
"text": "Manual inspection. The manual inspection involves verifying camera poses, camera intrinsic, and voxel positions. We filter out erroneous samples to prevent convergence issues in the model training process.",
|
| 873 |
+
"bbox": [
|
| 874 |
+
212,
|
| 875 |
+
535,
|
| 876 |
+
787,
|
| 877 |
+
583
|
| 878 |
+
],
|
| 879 |
+
"page_idx": 7
|
| 880 |
+
},
|
| 881 |
+
{
|
| 882 |
+
"type": "text",
|
| 883 |
+
"text": "Voxel label assignment. For each frame, only a specific area in front of the camera is defined for analysis. Therefore, the selection of the voxel origin is paramount, as it determines the subsequent coordinates of other voxel. Subsequently, each voxel is assigned a label based on its nearest voxel in the CompleteScanNet dataset. Additionally, we excluded frames with a ratio of unknown or empty labels exceeding $95\\%$ or frames where the number of semantic label classes was less than two. Consequently, we obtained the generated 3D voxel labels for each frame.",
|
| 884 |
+
"bbox": [
|
| 885 |
+
212,
|
| 886 |
+
604,
|
| 887 |
+
787,
|
| 888 |
+
726
|
| 889 |
+
],
|
| 890 |
+
"page_idx": 7
|
| 891 |
+
},
|
| 892 |
+
{
|
| 893 |
+
"type": "text",
|
| 894 |
+
"text": "5 Experiments",
|
| 895 |
+
"text_level": 1,
|
| 896 |
+
"bbox": [
|
| 897 |
+
215,
|
| 898 |
+
750,
|
| 899 |
+
377,
|
| 900 |
+
767
|
| 901 |
+
],
|
| 902 |
+
"page_idx": 7
|
| 903 |
+
},
|
| 904 |
+
{
|
| 905 |
+
"type": "text",
|
| 906 |
+
"text": "5.1 Setup",
|
| 907 |
+
"text_level": 1,
|
| 908 |
+
"bbox": [
|
| 909 |
+
215,
|
| 910 |
+
782,
|
| 911 |
+
310,
|
| 912 |
+
799
|
| 913 |
+
],
|
| 914 |
+
"page_idx": 7
|
| 915 |
+
},
|
| 916 |
+
{
|
| 917 |
+
"type": "text",
|
| 918 |
+
"text": "Datasets. The NYUv2 [29] dataset provides scenes represented in $240 \\times 144 \\times 240$ voxels grids, labeled with 13 classes, including 1 for free space, 1 for unknown,",
|
| 919 |
+
"bbox": [
|
| 920 |
+
212,
|
| 921 |
+
809,
|
| 922 |
+
785,
|
| 923 |
+
840
|
| 924 |
+
],
|
| 925 |
+
"page_idx": 7
|
| 926 |
+
},
|
| 927 |
+
{
|
| 928 |
+
"type": "page_number",
|
| 929 |
+
"text": "8",
|
| 930 |
+
"bbox": [
|
| 931 |
+
217,
|
| 932 |
+
114,
|
| 933 |
+
228,
|
| 934 |
+
126
|
| 935 |
+
],
|
| 936 |
+
"page_idx": 7
|
| 937 |
+
},
|
| 938 |
+
{
|
| 939 |
+
"type": "header",
|
| 940 |
+
"text": "H. Yu et al.",
|
| 941 |
+
"bbox": [
|
| 942 |
+
271,
|
| 943 |
+
114,
|
| 944 |
+
351,
|
| 945 |
+
126
|
| 946 |
+
],
|
| 947 |
+
"page_idx": 7
|
| 948 |
+
},
|
| 949 |
+
{
|
| 950 |
+
"type": "image",
|
| 951 |
+
"img_path": "images/600bf412f151f2798f467dfc92b731ba91e8e72b94db45aad736c5d837165504.jpg",
|
| 952 |
+
"image_caption": [
|
| 953 |
+
"Fig. 4: Pipeline of Occ-ScanNet dataset label generation. Color images, depth images, camera intrinsic and poses are extracted from ScanNet scenes. For each scene, 100 frames were sampled and randomly split into training and validation sets with a $7/3$ ratio. Frames with invalid camera poses or exceeding scene boundaries were filtered out. Only the area in front of the camera was analyzed, necessitating careful selection of the voxel origin. Voxel were labeled based on their nearest voxel in the CompleteScanNet dataset. Frames with $>95\\%$ unknown/empty labels or $<2$ semantic classes were excluded, resulting in generated 3D voxel labels for each frame."
|
| 954 |
+
],
|
| 955 |
+
"image_footnote": [],
|
| 956 |
+
"bbox": [
|
| 957 |
+
222,
|
| 958 |
+
150,
|
| 959 |
+
777,
|
| 960 |
+
200
|
| 961 |
+
],
|
| 962 |
+
"page_idx": 8
|
| 963 |
+
},
|
| 964 |
+
{
|
| 965 |
+
"type": "text",
|
| 966 |
+
"text": "and 11 for specific semantics (ceiling, floor, wall, window, chair, bed, sofa, table, tvs, furniture, objects). The dataset consists of 795 / 654 scenes in the train / test splits. The model is trained and evaluated on down-sampled $60 \\times 36 \\times 60$ voxels.",
|
| 967 |
+
"bbox": [
|
| 968 |
+
212,
|
| 969 |
+
369,
|
| 970 |
+
787,
|
| 971 |
+
429
|
| 972 |
+
],
|
| 973 |
+
"page_idx": 8
|
| 974 |
+
},
|
| 975 |
+
{
|
| 976 |
+
"type": "text",
|
| 977 |
+
"text": "Occ-ScanNet dataset provides scenes represented in $60 \\times 60 \\times 36$ voxel grids, labeled with 12 classes including 1 for free space, and 11 for specific semantics (ceiling, floor, wall, window, chair, bed, sofa, table, tvs, furniture, objects). The dataset comprises 45,755 / 19,764 frames in the train / val splits. The model is trained and evaluated on the original scale.",
|
| 978 |
+
"bbox": [
|
| 979 |
+
212,
|
| 980 |
+
433,
|
| 981 |
+
787,
|
| 982 |
+
508
|
| 983 |
+
],
|
| 984 |
+
"page_idx": 8
|
| 985 |
+
},
|
| 986 |
+
{
|
| 987 |
+
"type": "text",
|
| 988 |
+
"text": "Occ-ScanNet-mini dataset has the same class number and voxel scene size as Occ-ScanNet, except that this mini dataset consists of 4,639 / 2,007 frames in the train / val splits. The model is also trained and evaluated on the original scale.",
|
| 989 |
+
"bbox": [
|
| 990 |
+
212,
|
| 991 |
+
512,
|
| 992 |
+
787,
|
| 993 |
+
571
|
| 994 |
+
],
|
| 995 |
+
"page_idx": 8
|
| 996 |
+
},
|
| 997 |
+
{
|
| 998 |
+
"type": "text",
|
| 999 |
+
"text": "Implementation Details. We employ a pre-trained EfficientNet-B7 [31] as the encoder in our 2D UNet architecture. In the depth branch, we utilize a pre-trained DepthAnything model [44], which remains frozen during training. Additionally, we integrate a depth loss specific to the depth branch, complementing the other losses outlined in Monoscene [5]. Our model is trained on two datasets: NYUv2 [29] and Occ-ScanNet. For NYUv2, ISO undergoes 30 epochs of training using AdamW optimizers. Initially, the learning rate is set to 5e-6 for the DepthNet and 1e-4 for other components, with a learning rate reduction by 0.1 at epoch 20. Training on NYUv2 takes approximately 7 hours using 2 NVIDIA L20 GPUs (2 items per GPU). On the Occ-ScanNet-mini dataset, ISO is trained for 60 epochs under similar learning rate settings, decreasing the rate by 0.1 at epoch 40. This training process takes around 12 hours using 8 L20 GPUs. For Occ-ScanNet, ISO is trained for 10 epochs with the same learning rate schedule as used for Occ-ScanNet-mini, requiring approximately a day using 8 L20 GPUs.",
|
| 1000 |
+
"bbox": [
|
| 1001 |
+
212,
|
| 1002 |
+
613,
|
| 1003 |
+
787,
|
| 1004 |
+
839
|
| 1005 |
+
],
|
| 1006 |
+
"page_idx": 8
|
| 1007 |
+
},
|
| 1008 |
+
{
|
| 1009 |
+
"type": "header",
|
| 1010 |
+
"text": "Monocular Occupancy Prediction for Scalable Indoor Scenes",
|
| 1011 |
+
"bbox": [
|
| 1012 |
+
328,
|
| 1013 |
+
114,
|
| 1014 |
+
730,
|
| 1015 |
+
128
|
| 1016 |
+
],
|
| 1017 |
+
"page_idx": 8
|
| 1018 |
+
},
|
| 1019 |
+
{
|
| 1020 |
+
"type": "page_number",
|
| 1021 |
+
"text": "9",
|
| 1022 |
+
"bbox": [
|
| 1023 |
+
774,
|
| 1024 |
+
116,
|
| 1025 |
+
784,
|
| 1026 |
+
126
|
| 1027 |
+
],
|
| 1028 |
+
"page_idx": 8
|
| 1029 |
+
},
|
| 1030 |
+
{
|
| 1031 |
+
"type": "text",
|
| 1032 |
+
"text": "5.2 Main Results",
|
| 1033 |
+
"text_level": 1,
|
| 1034 |
+
"bbox": [
|
| 1035 |
+
215,
|
| 1036 |
+
146,
|
| 1037 |
+
372,
|
| 1038 |
+
159
|
| 1039 |
+
],
|
| 1040 |
+
"page_idx": 9
|
| 1041 |
+
},
|
| 1042 |
+
{
|
| 1043 |
+
"type": "table",
|
| 1044 |
+
"img_path": "images/2a104e60b4f82f6c7f71fdac10f084e75fa5921047543bae1749de859d180f6f.jpg",
|
| 1045 |
+
"table_caption": [
|
| 1046 |
+
"Table 1: Performance on the Occ-ScanNet dataset"
|
| 1047 |
+
],
|
| 1048 |
+
"table_footnote": [],
|
| 1049 |
+
"table_body": "<table><tr><td>Method</td><td>Input</td><td>IoU</td><td>ceiling</td><td>floor</td><td>wall</td><td>window</td><td>chair</td><td>bed</td><td>sofa</td><td>table</td><td>tvs</td><td>furniture</td><td>objects</td></tr><tr><td>MonoScene* [5]</td><td>xrgb</td><td>41.60</td><td>15.17</td><td>44.71</td><td>22.41</td><td>12.55</td><td>26.11</td><td>27.03</td><td>35.91</td><td>28.32</td><td>6.57</td><td>32.16</td><td>19.84</td></tr><tr><td>ISO(Ours)</td><td>xrgb</td><td>42.16</td><td>19.88</td><td>41.88</td><td>22.37</td><td>16.98</td><td>29.09</td><td>42.43</td><td>42.00</td><td>29.60</td><td>10.62</td><td>36.36</td><td>24.61</td></tr></table>",
|
| 1050 |
+
"bbox": [
|
| 1051 |
+
228,
|
| 1052 |
+
227,
|
| 1053 |
+
772,
|
| 1054 |
+
338
|
| 1055 |
+
],
|
| 1056 |
+
"page_idx": 9
|
| 1057 |
+
},
|
| 1058 |
+
{
|
| 1059 |
+
"type": "text",
|
| 1060 |
+
"text": "Occ-ScanNet performance. We first evaluate our model's performance on the large-scale Occ-ScanNet dataset. As shown in Tab. 1, the results indicate that our method significantly outperforms MonoScene [5]. The * denotes results obtained using their code trained on our dataset. In Fig. 5, we also conducted qualitative visualization comparisons.",
|
| 1061 |
+
"bbox": [
|
| 1062 |
+
212,
|
| 1063 |
+
396,
|
| 1064 |
+
787,
|
| 1065 |
+
472
|
| 1066 |
+
],
|
| 1067 |
+
"page_idx": 9
|
| 1068 |
+
},
|
| 1069 |
+
{
|
| 1070 |
+
"type": "text",
|
| 1071 |
+
"text": "NYUv2 performance. The NYUv2 [29] dataset serves as a widely used benchmark for indoor scene evaluation. The results in Tab. 2 demonstrate that our method achieves state-of-the-art performance. Our method also demonstrates substantial performance improvement on the NYUv2 dataset, in addition to the gains observed on Occ-ScanNet. As illustrated in Fig. 6, our ISO demonstrates superior room layout prediction compared to [5], particularly evident in rows 1 and 2. This improvement can be attributed to incorporating depth knowledge into the model. Additionally, our method exhibits the capability of learning small objects, such as pictures on the wall in row 3.",
|
| 1072 |
+
"bbox": [
|
| 1073 |
+
212,
|
| 1074 |
+
500,
|
| 1075 |
+
789,
|
| 1076 |
+
635
|
| 1077 |
+
],
|
| 1078 |
+
"page_idx": 9
|
| 1079 |
+
},
|
| 1080 |
+
{
|
| 1081 |
+
"type": "table",
|
| 1082 |
+
"img_path": "images/f7253140a0003ccaf15ae943cff655b57cc59fb75729c5c34e91a9ab0459850d.jpg",
|
| 1083 |
+
"table_caption": [
|
| 1084 |
+
"Table 2: Performance on the NYUv2 dataset"
|
| 1085 |
+
],
|
| 1086 |
+
"table_footnote": [],
|
| 1087 |
+
"table_body": "<table><tr><td rowspan=\"2\">Method</td><td rowspan=\"2\">Input</td><td rowspan=\"2\">IoU</td><td colspan=\"2\">ceiling</td><td colspan=\"2\">floor</td><td colspan=\"2\">wall</td><td colspan=\"2\">window</td><td colspan=\"2\">chair</td><td colspan=\"2\">bed</td><td colspan=\"2\">sofa</td><td colspan=\"2\">table</td><td colspan=\"2\">tvs</td><td colspan=\"2\">furniture</td><td colspan=\"2\">objects</td><td></td></tr><tr><td colspan=\"2\">■</td><td colspan=\"2\">■</td><td colspan=\"2\">■</td><td colspan=\"2\">■</td><td colspan=\"2\">■</td><td colspan=\"2\">■</td><td colspan=\"2\">■</td><td colspan=\"2\">■</td><td colspan=\"2\">■</td><td colspan=\"2\">■</td><td colspan=\"2\">■</td><td></td></tr><tr><td>LMSCNetrgb [28]</td><td>xocc</td><td>33.93</td><td colspan=\"2\">4.49</td><td colspan=\"2\">88.41</td><td colspan=\"2\">4.63</td><td colspan=\"2\">0.25</td><td colspan=\"2\">3.94</td><td colspan=\"2\">32.03</td><td colspan=\"2\">15.44</td><td colspan=\"2\">6.57</td><td colspan=\"2\">0.02</td><td colspan=\"2\">14.51</td><td colspan=\"2\">4.39</td><td>15.88</td></tr><tr><td>AICNetrgb [18]</td><td>xrgb, xdepth</td><td>30.03</td><td colspan=\"2\">7.58</td><td colspan=\"2\">82.97</td><td colspan=\"2\">9.15</td><td colspan=\"2\">0.05</td><td colspan=\"2\">6.93</td><td colspan=\"2\">35.87</td><td colspan=\"2\">22.92</td><td colspan=\"2\">11.11</td><td colspan=\"2\">0.71</td><td colspan=\"2\">15.90</td><td colspan=\"2\">6.45</td><td>18.15</td></tr><tr><td>3DSketchrgb [6]</td><td>xrgb, xTSDF</td><td>38.64</td><td colspan=\"2\">8.53</td><td colspan=\"2\">90.45</td><td colspan=\"2\">9.94</td><td colspan=\"2\">5.67</td><td colspan=\"2\">10.64</td><td colspan=\"2\">42.29</td><td colspan=\"2\">29.21</td><td colspan=\"2\">13.88</td><td colspan=\"2\">9.38</td><td colspan=\"2\">23.83</td><td colspan=\"2\">8.19</td><td>22.91</td></tr><tr><td>MonoScene [5]</td><td>xrgb</td><td>42.51</td><td colspan=\"2\">8.89</td><td colspan=\"2\">93.50</td><td colspan=\"2\">12.06</td><td colspan=\"2\">12.57</td><td colspan=\"2\">13.72</td><td colspan=\"2\">48.19</td><td colspan=\"2\">36.11</td><td colspan=\"2\">15.13</td><td colspan=\"2\">15.22</td><td colspan=\"2\">27.96</td><td colspan=\"2\">12.94</td><td>26.94</td></tr><tr><td>NDC-Scene [45]</td><td>xrgb</td><td>44.17</td><td colspan=\"2\">12.02</td><td colspan=\"2\">93.51</td><td colspan=\"2\">13.11</td><td colspan=\"2\">13.77</td><td colspan=\"2\">15.83</td><td colspan=\"2\">49.57</td><td colspan=\"2\">39.87</td><td colspan=\"2\">17.17</td><td colspan=\"2\">24.57</td><td colspan=\"2\">31.00</td><td colspan=\"2\">14.96</td><td>29.03</td></tr><tr><td>ISO(Ours)</td><td>xrgb</td><td>47.11</td><td colspan=\"2\">14.21</td><td colspan=\"2\">93.47</td><td colspan=\"2\">15.89</td><td colspan=\"2\">15.14</td><td colspan=\"2\">18.35</td><td colspan=\"2\">50.01</td><td colspan=\"2\">40.82</td><td colspan=\"2\">18.25</td><td colspan=\"2\">25.90</td><td colspan=\"2\">34.08</td><td colspan=\"2\">17.67</td><td>31.25</td></tr></table>",
|
| 1088 |
+
"bbox": [
|
| 1089 |
+
215,
|
| 1090 |
+
689,
|
| 1091 |
+
792,
|
| 1092 |
+
809
|
| 1093 |
+
],
|
| 1094 |
+
"page_idx": 9
|
| 1095 |
+
},
|
| 1096 |
+
{
|
| 1097 |
+
"type": "page_number",
|
| 1098 |
+
"text": "10",
|
| 1099 |
+
"bbox": [
|
| 1100 |
+
217,
|
| 1101 |
+
114,
|
| 1102 |
+
235,
|
| 1103 |
+
126
|
| 1104 |
+
],
|
| 1105 |
+
"page_idx": 9
|
| 1106 |
+
},
|
| 1107 |
+
{
|
| 1108 |
+
"type": "header",
|
| 1109 |
+
"text": "H. Yu et al.",
|
| 1110 |
+
"bbox": [
|
| 1111 |
+
271,
|
| 1112 |
+
114,
|
| 1113 |
+
351,
|
| 1114 |
+
126
|
| 1115 |
+
],
|
| 1116 |
+
"page_idx": 9
|
| 1117 |
+
},
|
| 1118 |
+
{
|
| 1119 |
+
"type": "image",
|
| 1120 |
+
"img_path": "images/04e2c5714d704f272ab8c860434ace6dba790a5c5dbad8506865878756310863.jpg",
|
| 1121 |
+
"image_caption": [
|
| 1122 |
+
"Image"
|
| 1123 |
+
],
|
| 1124 |
+
"image_footnote": [],
|
| 1125 |
+
"bbox": [
|
| 1126 |
+
243,
|
| 1127 |
+
174,
|
| 1128 |
+
370,
|
| 1129 |
+
250
|
| 1130 |
+
],
|
| 1131 |
+
"page_idx": 10
|
| 1132 |
+
},
|
| 1133 |
+
{
|
| 1134 |
+
"type": "image",
|
| 1135 |
+
"img_path": "images/b56c565fcb1a40a72e382e16616d20db61fc9f40eeac725137d84d8009862ef3.jpg",
|
| 1136 |
+
"image_caption": [
|
| 1137 |
+
"MonoScene"
|
| 1138 |
+
],
|
| 1139 |
+
"image_footnote": [],
|
| 1140 |
+
"bbox": [
|
| 1141 |
+
377,
|
| 1142 |
+
176,
|
| 1143 |
+
491,
|
| 1144 |
+
247
|
| 1145 |
+
],
|
| 1146 |
+
"page_idx": 10
|
| 1147 |
+
},
|
| 1148 |
+
{
|
| 1149 |
+
"type": "image",
|
| 1150 |
+
"img_path": "images/d7fc5ba70886aa3bf20a8c1e7c897294e9dc3e9e11d5f7a942a351a5a6d7a1e0.jpg",
|
| 1151 |
+
"image_caption": [
|
| 1152 |
+
"Ours"
|
| 1153 |
+
],
|
| 1154 |
+
"image_footnote": [],
|
| 1155 |
+
"bbox": [
|
| 1156 |
+
506,
|
| 1157 |
+
175,
|
| 1158 |
+
622,
|
| 1159 |
+
247
|
| 1160 |
+
],
|
| 1161 |
+
"page_idx": 10
|
| 1162 |
+
},
|
| 1163 |
+
{
|
| 1164 |
+
"type": "image",
|
| 1165 |
+
"img_path": "images/cd182c385abf9d22677067034b20841ffaa1c7229ac8fee7df544e42ff29b5ee.jpg",
|
| 1166 |
+
"image_caption": [
|
| 1167 |
+
"Ground Truth"
|
| 1168 |
+
],
|
| 1169 |
+
"image_footnote": [],
|
| 1170 |
+
"bbox": [
|
| 1171 |
+
633,
|
| 1172 |
+
176,
|
| 1173 |
+
750,
|
| 1174 |
+
247
|
| 1175 |
+
],
|
| 1176 |
+
"page_idx": 10
|
| 1177 |
+
},
|
| 1178 |
+
{
|
| 1179 |
+
"type": "image",
|
| 1180 |
+
"img_path": "images/fdc44390d075e3fc6576247a1bd8f6e1d0312141487072c44920d171df5c8cc1.jpg",
|
| 1181 |
+
"image_caption": [],
|
| 1182 |
+
"image_footnote": [],
|
| 1183 |
+
"bbox": [
|
| 1184 |
+
243,
|
| 1185 |
+
270,
|
| 1186 |
+
370,
|
| 1187 |
+
345
|
| 1188 |
+
],
|
| 1189 |
+
"page_idx": 10
|
| 1190 |
+
},
|
| 1191 |
+
{
|
| 1192 |
+
"type": "image",
|
| 1193 |
+
"img_path": "images/1ebc98688a3d3fd3a3173e2685869d24e0e043b6de44b4a6d2f0e53ac916e198.jpg",
|
| 1194 |
+
"image_caption": [],
|
| 1195 |
+
"image_footnote": [],
|
| 1196 |
+
"bbox": [
|
| 1197 |
+
379,
|
| 1198 |
+
272,
|
| 1199 |
+
490,
|
| 1200 |
+
345
|
| 1201 |
+
],
|
| 1202 |
+
"page_idx": 10
|
| 1203 |
+
},
|
| 1204 |
+
{
|
| 1205 |
+
"type": "image",
|
| 1206 |
+
"img_path": "images/3ac2892e3db2e908440b580e701c4dc413377e7e64b972d05bf3372b2341814f.jpg",
|
| 1207 |
+
"image_caption": [],
|
| 1208 |
+
"image_footnote": [],
|
| 1209 |
+
"bbox": [
|
| 1210 |
+
504,
|
| 1211 |
+
273,
|
| 1212 |
+
622,
|
| 1213 |
+
343
|
| 1214 |
+
],
|
| 1215 |
+
"page_idx": 10
|
| 1216 |
+
},
|
| 1217 |
+
{
|
| 1218 |
+
"type": "image",
|
| 1219 |
+
"img_path": "images/d1400f44a4c56cceb48fbcb19e51e9f955459cff2e343a5887fad4d17eb142a0.jpg",
|
| 1220 |
+
"image_caption": [],
|
| 1221 |
+
"image_footnote": [],
|
| 1222 |
+
"bbox": [
|
| 1223 |
+
630,
|
| 1224 |
+
273,
|
| 1225 |
+
751,
|
| 1226 |
+
342
|
| 1227 |
+
],
|
| 1228 |
+
"page_idx": 10
|
| 1229 |
+
},
|
| 1230 |
+
{
|
| 1231 |
+
"type": "image",
|
| 1232 |
+
"img_path": "images/5cc79ad1a3418afb90628108bbd37d04b82c455cad379098b54a1c683844ec18.jpg",
|
| 1233 |
+
"image_caption": [],
|
| 1234 |
+
"image_footnote": [],
|
| 1235 |
+
"bbox": [
|
| 1236 |
+
243,
|
| 1237 |
+
366,
|
| 1238 |
+
370,
|
| 1239 |
+
441
|
| 1240 |
+
],
|
| 1241 |
+
"page_idx": 10
|
| 1242 |
+
},
|
| 1243 |
+
{
|
| 1244 |
+
"type": "image",
|
| 1245 |
+
"img_path": "images/15e5ef3661021822b1013b521dc01a98e033212ebacb07213fc7d9182cb23f7c.jpg",
|
| 1246 |
+
"image_caption": [],
|
| 1247 |
+
"image_footnote": [],
|
| 1248 |
+
"bbox": [
|
| 1249 |
+
375,
|
| 1250 |
+
371,
|
| 1251 |
+
493,
|
| 1252 |
+
435
|
| 1253 |
+
],
|
| 1254 |
+
"page_idx": 10
|
| 1255 |
+
},
|
| 1256 |
+
{
|
| 1257 |
+
"type": "image",
|
| 1258 |
+
"img_path": "images/6db8146507af64f3ad0e265ce1b04282f92a486f53ede1b967eb929a0a72a9d0.jpg",
|
| 1259 |
+
"image_caption": [],
|
| 1260 |
+
"image_footnote": [],
|
| 1261 |
+
"bbox": [
|
| 1262 |
+
503,
|
| 1263 |
+
371,
|
| 1264 |
+
625,
|
| 1265 |
+
436
|
| 1266 |
+
],
|
| 1267 |
+
"page_idx": 10
|
| 1268 |
+
},
|
| 1269 |
+
{
|
| 1270 |
+
"type": "image",
|
| 1271 |
+
"img_path": "images/94828630daa0a3789f33db9c02a70978ce6f2b4eab3a1c1da1a90f57e8435cee.jpg",
|
| 1272 |
+
"image_caption": [],
|
| 1273 |
+
"image_footnote": [],
|
| 1274 |
+
"bbox": [
|
| 1275 |
+
632,
|
| 1276 |
+
375,
|
| 1277 |
+
751,
|
| 1278 |
+
435
|
| 1279 |
+
],
|
| 1280 |
+
"page_idx": 10
|
| 1281 |
+
},
|
| 1282 |
+
{
|
| 1283 |
+
"type": "image",
|
| 1284 |
+
"img_path": "images/10396fe5d4ea378ddc1107ded1ed79d07052a88a88841e2e2835d29180f65811.jpg",
|
| 1285 |
+
"image_caption": [
|
| 1286 |
+
"Fig. 5: Qualitative Analysis on the Occ-ScanNet Dataset. The input image is displayed on the left, while the predicted scene is shown in the middle two columns, and ground truth on the right column."
|
| 1287 |
+
],
|
| 1288 |
+
"image_footnote": [],
|
| 1289 |
+
"bbox": [
|
| 1290 |
+
243,
|
| 1291 |
+
460,
|
| 1292 |
+
370,
|
| 1293 |
+
536
|
| 1294 |
+
],
|
| 1295 |
+
"page_idx": 10
|
| 1296 |
+
},
|
| 1297 |
+
{
|
| 1298 |
+
"type": "image",
|
| 1299 |
+
"img_path": "images/ac93ecc616302a419ebe3b5a45700ca5a7f49fc81925b8ef6119eae7527bce78.jpg",
|
| 1300 |
+
"image_caption": [],
|
| 1301 |
+
"image_footnote": [],
|
| 1302 |
+
"bbox": [
|
| 1303 |
+
377,
|
| 1304 |
+
467,
|
| 1305 |
+
496,
|
| 1306 |
+
532
|
| 1307 |
+
],
|
| 1308 |
+
"page_idx": 10
|
| 1309 |
+
},
|
| 1310 |
+
{
|
| 1311 |
+
"type": "image",
|
| 1312 |
+
"img_path": "images/f0898202e01e8619bc49b2aaa306d734b49a00b17d953826e5404469616c0a16.jpg",
|
| 1313 |
+
"image_caption": [],
|
| 1314 |
+
"image_footnote": [],
|
| 1315 |
+
"bbox": [
|
| 1316 |
+
504,
|
| 1317 |
+
460,
|
| 1318 |
+
625,
|
| 1319 |
+
539
|
| 1320 |
+
],
|
| 1321 |
+
"page_idx": 10
|
| 1322 |
+
},
|
| 1323 |
+
{
|
| 1324 |
+
"type": "image",
|
| 1325 |
+
"img_path": "images/30279eba2499b123db6f91a3fa0b0c4272933fffbf5a95082748f4553709b126.jpg",
|
| 1326 |
+
"image_caption": [],
|
| 1327 |
+
"image_footnote": [],
|
| 1328 |
+
"bbox": [
|
| 1329 |
+
632,
|
| 1330 |
+
460,
|
| 1331 |
+
751,
|
| 1332 |
+
537
|
| 1333 |
+
],
|
| 1334 |
+
"page_idx": 10
|
| 1335 |
+
},
|
| 1336 |
+
{
|
| 1337 |
+
"type": "text",
|
| 1338 |
+
"text": "5.3 Ablation Study",
|
| 1339 |
+
"text_level": 1,
|
| 1340 |
+
"bbox": [
|
| 1341 |
+
214,
|
| 1342 |
+
643,
|
| 1343 |
+
390,
|
| 1344 |
+
659
|
| 1345 |
+
],
|
| 1346 |
+
"page_idx": 10
|
| 1347 |
+
},
|
| 1348 |
+
{
|
| 1349 |
+
"type": "text",
|
| 1350 |
+
"text": "We conduct ablation studies on the NYUv2 [29] and Occ-ScanNet-mini datasets to validate the effectiveness of our model design.",
|
| 1351 |
+
"bbox": [
|
| 1352 |
+
212,
|
| 1353 |
+
667,
|
| 1354 |
+
787,
|
| 1355 |
+
698
|
| 1356 |
+
],
|
| 1357 |
+
"page_idx": 10
|
| 1358 |
+
},
|
| 1359 |
+
{
|
| 1360 |
+
"type": "text",
|
| 1361 |
+
"text": "Depth fusion. Inspired by BEVDepth [20], we initially adopt the integration of BEV depth information, denoted as $^{+}$ bev-depth'. However, we observed that while BEV depth works well for 3D detection in outdoor scenes, directly applying it to indoor scenes is less effective. We speculate that this discrepancy arises because height information is more crucial in indoor scenes, where indoor heights often correspond to different 3D structures and semantics. Therefore, we adopt voxel depth fusion, denoted as $^{+}$ voxel-depth'. Specifically, we project each 3D voxel centroid to a frustum grid to sample the depth distribution probability",
|
| 1362 |
+
"bbox": [
|
| 1363 |
+
212,
|
| 1364 |
+
718,
|
| 1365 |
+
787,
|
| 1366 |
+
840
|
| 1367 |
+
],
|
| 1368 |
+
"page_idx": 10
|
| 1369 |
+
},
|
| 1370 |
+
{
|
| 1371 |
+
"type": "header",
|
| 1372 |
+
"text": "Monocular Occupancy Prediction for Scalable Indoor Scenes",
|
| 1373 |
+
"bbox": [
|
| 1374 |
+
328,
|
| 1375 |
+
114,
|
| 1376 |
+
730,
|
| 1377 |
+
128
|
| 1378 |
+
],
|
| 1379 |
+
"page_idx": 10
|
| 1380 |
+
},
|
| 1381 |
+
{
|
| 1382 |
+
"type": "page_number",
|
| 1383 |
+
"text": "11",
|
| 1384 |
+
"bbox": [
|
| 1385 |
+
767,
|
| 1386 |
+
114,
|
| 1387 |
+
782,
|
| 1388 |
+
126
|
| 1389 |
+
],
|
| 1390 |
+
"page_idx": 10
|
| 1391 |
+
},
|
| 1392 |
+
{
|
| 1393 |
+
"type": "image",
|
| 1394 |
+
"img_path": "images/e1e790fceece663066ce4a35a679e75bee7a43e5396d4b9cfad477e8820e4e36.jpg",
|
| 1395 |
+
"image_caption": [
|
| 1396 |
+
"Fig. 6: Qualitative Analysis on the NYUv2 Dataset: The input is displayed on the left, with the camera viewing frustum shown in every image. ISO adeptly captures the scene layout and delineates various semantic instances more accurately. Notably, it excels in reconstructing the corners of walls (row 2) and accurately representing small objects like pictures on the wall (row 3)."
|
| 1397 |
+
],
|
| 1398 |
+
"image_footnote": [],
|
| 1399 |
+
"bbox": [
|
| 1400 |
+
238,
|
| 1401 |
+
148,
|
| 1402 |
+
764,
|
| 1403 |
+
378
|
| 1404 |
+
],
|
| 1405 |
+
"page_idx": 11
|
| 1406 |
+
},
|
| 1407 |
+
{
|
| 1408 |
+
"type": "text",
|
| 1409 |
+
"text": "of the voxel. We then weight the 2D feature at that pixel coordinate as the voxel feature. The results in Tab. 3 further validate that the voxel depth fusion approach is more effective.",
|
| 1410 |
+
"bbox": [
|
| 1411 |
+
212,
|
| 1412 |
+
496,
|
| 1413 |
+
784,
|
| 1414 |
+
541
|
| 1415 |
+
],
|
| 1416 |
+
"page_idx": 11
|
| 1417 |
+
},
|
| 1418 |
+
{
|
| 1419 |
+
"type": "text",
|
| 1420 |
+
"text": "The BEV method usually performs better in detecting ceilings and floors but performs poorly in detecting furniture such as windows, walls, televisions, beds, sofas, and tables. Ceilings and floors usually have relatively flat and regular geometric shapes, which makes their projections clearer and more consistent from a BEV perspective. In contrast, objects such as windows, walls, and televisions typically have more complex geometric shapes. In the BEV perspective, the projections of these objects may become blurry or overlapping, making it difficult to accurately distinguish and detect.",
|
| 1421 |
+
"bbox": [
|
| 1422 |
+
212,
|
| 1423 |
+
542,
|
| 1424 |
+
787,
|
| 1425 |
+
662
|
| 1426 |
+
],
|
| 1427 |
+
"page_idx": 11
|
| 1428 |
+
},
|
| 1429 |
+
{
|
| 1430 |
+
"type": "table",
|
| 1431 |
+
"img_path": "images/a0a0d6e1b1e3bccd7e8972fb9e8506736b3a0886d6898eb0af319b27a83b2b27.jpg",
|
| 1432 |
+
"table_caption": [
|
| 1433 |
+
"Table 3: Ablation study of depth fusion on the NYUv2 and Occ-ScanNet-mini dataset"
|
| 1434 |
+
],
|
| 1435 |
+
"table_footnote": [],
|
| 1436 |
+
"table_body": "<table><tr><td rowspan=\"2\">Method</td><td rowspan=\"2\">Input</td><td colspan=\"2\">NYUv2</td><td colspan=\"2\">Occ-ScanNet-mini</td></tr><tr><td>IoU</td><td>mIoU</td><td>IoU</td><td>mIoU</td></tr><tr><td>baseline</td><td>xrgb</td><td>42.27</td><td>27.13</td><td>50.94</td><td>38.95</td></tr><tr><td>+ bev-depth</td><td>xrgb</td><td>42.67</td><td>27.14</td><td>51.58</td><td>38.48</td></tr><tr><td>+ voxel-depth</td><td>xrgb</td><td>47.11</td><td>31.25</td><td>51.03</td><td>39.08</td></tr></table>",
|
| 1437 |
+
"bbox": [
|
| 1438 |
+
307,
|
| 1439 |
+
719,
|
| 1440 |
+
697,
|
| 1441 |
+
808
|
| 1442 |
+
],
|
| 1443 |
+
"page_idx": 11
|
| 1444 |
+
},
|
| 1445 |
+
{
|
| 1446 |
+
"type": "page_number",
|
| 1447 |
+
"text": "12",
|
| 1448 |
+
"bbox": [
|
| 1449 |
+
217,
|
| 1450 |
+
114,
|
| 1451 |
+
235,
|
| 1452 |
+
126
|
| 1453 |
+
],
|
| 1454 |
+
"page_idx": 11
|
| 1455 |
+
},
|
| 1456 |
+
{
|
| 1457 |
+
"type": "header",
|
| 1458 |
+
"text": "H. Yu et al.",
|
| 1459 |
+
"bbox": [
|
| 1460 |
+
271,
|
| 1461 |
+
114,
|
| 1462 |
+
351,
|
| 1463 |
+
126
|
| 1464 |
+
],
|
| 1465 |
+
"page_idx": 11
|
| 1466 |
+
},
|
| 1467 |
+
{
|
| 1468 |
+
"type": "table",
|
| 1469 |
+
"img_path": "images/9bbba87470286f0793b8ea3a417066382d7328d93c2cb7ab62af7c0492c2a150.jpg",
|
| 1470 |
+
"table_caption": [
|
| 1471 |
+
"Table 4: Ablation of depth information on the NYUv2 dataset"
|
| 1472 |
+
],
|
| 1473 |
+
"table_footnote": [],
|
| 1474 |
+
"table_body": "<table><tr><td>depth-method</td><td>learned</td><td>multi-scale</td><td>IoU</td><td>mIoU</td></tr><tr><td>GT</td><td></td><td>✓</td><td>53.98</td><td>34.47</td></tr><tr><td>ZoeDepth</td><td>✓</td><td></td><td>45.18</td><td>29.15</td></tr><tr><td>ZoeDepth</td><td>✓</td><td>✓</td><td>45.24</td><td>29.40</td></tr><tr><td>DepthAnything</td><td></td><td></td><td>44.77</td><td>29.28</td></tr><tr><td>DepthAnything</td><td></td><td>✓</td><td>45.48</td><td>29.57</td></tr><tr><td>DepthAnything</td><td>✓</td><td>✓</td><td>46.94</td><td>31.02</td></tr><tr><td>DepthAnything</td><td>✓</td><td></td><td>47.11</td><td>31.25</td></tr></table>",
|
| 1475 |
+
"bbox": [
|
| 1476 |
+
339,
|
| 1477 |
+
170,
|
| 1478 |
+
660,
|
| 1479 |
+
282
|
| 1480 |
+
],
|
| 1481 |
+
"page_idx": 12
|
| 1482 |
+
},
|
| 1483 |
+
{
|
| 1484 |
+
"type": "text",
|
| 1485 |
+
"text": "Different depth pre-trained models. We conducted ablation experiments on the depth method using the Occ-ScanNet-mini dataset. Initially, we observed the upper bound of model performance when ground truth (GT) depth information was provided. Subsequently, we initialized our depth module using ZoeDepth [3] and Depth-Anything [44]. Experimental results indicate that utilizing Depth-Anything [44] yields better performance. Moreover, our approach allows for further enhancement by fine-tuning a pre-trained depth module, resulting in improved performance.",
|
| 1486 |
+
"bbox": [
|
| 1487 |
+
215,
|
| 1488 |
+
315,
|
| 1489 |
+
785,
|
| 1490 |
+
436
|
| 1491 |
+
],
|
| 1492 |
+
"page_idx": 12
|
| 1493 |
+
},
|
| 1494 |
+
{
|
| 1495 |
+
"type": "text",
|
| 1496 |
+
"text": "Multi-scale depth. In Tab. 4, we investigate the impact of using multi-scale depth distribution on the NYUv2 [29] dataset. Specifically, the multi-scale depth distribution is used to weight the multi-scale 3D voxel feature. Meanwhile, we tried a method that only uses the single-scale depth distribution map to generate the 3D voxel depth matrix. The used depth distribution map has a scale of $\\frac{1}{8}$ original image scale. Only this single-scale 3D voxel matrix is used to weigh the summed-up 3D voxel feature out of the FLoSP module. Without multi-scale depth, all 3D features are simply added up and a single depth distribution is used for weighting. This method ignores the different importance that features at different scales may have, as all features at all scales are treated equally and all depth information is weighted by a unified depth distribution. With multiscale depth, the 3D features of each scale are first weighted with the depth distribution of their corresponding scale, and then these weighted features are added up. This method takes into account the differences in the importance of features at different scales, as each scale's feature is weighted based on its corresponding depth distribution.",
|
| 1497 |
+
"bbox": [
|
| 1498 |
+
215,
|
| 1499 |
+
474,
|
| 1500 |
+
785,
|
| 1501 |
+
715
|
| 1502 |
+
],
|
| 1503 |
+
"page_idx": 12
|
| 1504 |
+
},
|
| 1505 |
+
{
|
| 1506 |
+
"type": "text",
|
| 1507 |
+
"text": "In real-world scenarios, various objects and structures (such as ceilings, walls, furniture, etc.) often exhibit different scales. Hence, accounting for scale variations is crucial for accurately modeling these objects. Multi-scale depth can effectively capture these scale changes by assigning different weights to features at each scale. Depth information is indispensable for 3D scene completion. Leveraging the depth distribution information at each scale allows us to weight the corresponding 3D features, enabling the model to better understand and represent the objects in the scene and their relationships.",
|
| 1508 |
+
"bbox": [
|
| 1509 |
+
215,
|
| 1510 |
+
719,
|
| 1511 |
+
785,
|
| 1512 |
+
839
|
| 1513 |
+
],
|
| 1514 |
+
"page_idx": 12
|
| 1515 |
+
},
|
| 1516 |
+
{
|
| 1517 |
+
"type": "header",
|
| 1518 |
+
"text": "Monocular Occupancy Prediction for Scalable Indoor Scenes",
|
| 1519 |
+
"bbox": [
|
| 1520 |
+
328,
|
| 1521 |
+
114,
|
| 1522 |
+
730,
|
| 1523 |
+
128
|
| 1524 |
+
],
|
| 1525 |
+
"page_idx": 12
|
| 1526 |
+
},
|
| 1527 |
+
{
|
| 1528 |
+
"type": "page_number",
|
| 1529 |
+
"text": "13",
|
| 1530 |
+
"bbox": [
|
| 1531 |
+
767,
|
| 1532 |
+
116,
|
| 1533 |
+
784,
|
| 1534 |
+
126
|
| 1535 |
+
],
|
| 1536 |
+
"page_idx": 12
|
| 1537 |
+
},
|
| 1538 |
+
{
|
| 1539 |
+
"type": "text",
|
| 1540 |
+
"text": "5.4 Discussion",
|
| 1541 |
+
"text_level": 1,
|
| 1542 |
+
"bbox": [
|
| 1543 |
+
215,
|
| 1544 |
+
145,
|
| 1545 |
+
349,
|
| 1546 |
+
160
|
| 1547 |
+
],
|
| 1548 |
+
"page_idx": 13
|
| 1549 |
+
},
|
| 1550 |
+
{
|
| 1551 |
+
"type": "text",
|
| 1552 |
+
"text": "Data scaling up. In Tab. 5, we delve into the impact of data volume on our method's performance by scaling up the number of scene samples. We trained the model on Occ-ScanNet using $10\\%$ and $100\\%$ scene samples respectively, and tested the model performance. The results indicate that larger dataset exhibits more significant gains.",
|
| 1553 |
+
"bbox": [
|
| 1554 |
+
212,
|
| 1555 |
+
171,
|
| 1556 |
+
787,
|
| 1557 |
+
247
|
| 1558 |
+
],
|
| 1559 |
+
"page_idx": 13
|
| 1560 |
+
},
|
| 1561 |
+
{
|
| 1562 |
+
"type": "table",
|
| 1563 |
+
"img_path": "images/9081432db3b01e93e94c944b68deffaf5eb915ce6e250fce8593cf994bb1c78f.jpg",
|
| 1564 |
+
"table_caption": [
|
| 1565 |
+
"Table 5: Scaling up comparison on the Occ-ScanNet"
|
| 1566 |
+
],
|
| 1567 |
+
"table_footnote": [],
|
| 1568 |
+
"table_body": "<table><tr><td rowspan=\"2\">Data</td><td rowspan=\"2\">Depth Branch</td><td colspan=\"2\">Ours</td></tr><tr><td>IoU</td><td>mIoU</td></tr><tr><td>10%</td><td>✓</td><td>21.79</td><td>9.57</td></tr><tr><td>100%</td><td>✓</td><td>42.16</td><td>28.71</td></tr></table>",
|
| 1569 |
+
"bbox": [
|
| 1570 |
+
377,
|
| 1571 |
+
301,
|
| 1572 |
+
619,
|
| 1573 |
+
375
|
| 1574 |
+
],
|
| 1575 |
+
"page_idx": 13
|
| 1576 |
+
},
|
| 1577 |
+
{
|
| 1578 |
+
"type": "text",
|
| 1579 |
+
"text": "Limitations and future works. Although our model can estimate 3D occupancy effectively with the assistance of depth information, semantic learning still faces challenges due to class imbalance issues. Furthermore, our proposed Occ-ScanNet only considers 11 common semantic classes, which may not fully capture the diversity of categories present in real-world scenarios. Despite the significant increase in data volume compared to previous datasets, Occ-ScanNet remains limited in the number of scenes it covers. In future work, we will focus more on semantic exploration and conduct tests in more generalized scenarios.",
|
| 1580 |
+
"bbox": [
|
| 1581 |
+
212,
|
| 1582 |
+
424,
|
| 1583 |
+
787,
|
| 1584 |
+
546
|
| 1585 |
+
],
|
| 1586 |
+
"page_idx": 13
|
| 1587 |
+
},
|
| 1588 |
+
{
|
| 1589 |
+
"type": "text",
|
| 1590 |
+
"text": "6 Conclusion",
|
| 1591 |
+
"text_level": 1,
|
| 1592 |
+
"bbox": [
|
| 1593 |
+
215,
|
| 1594 |
+
569,
|
| 1595 |
+
359,
|
| 1596 |
+
585
|
| 1597 |
+
],
|
| 1598 |
+
"page_idx": 13
|
| 1599 |
+
},
|
| 1600 |
+
{
|
| 1601 |
+
"type": "text",
|
| 1602 |
+
"text": "In this paper, we introduce ISO, a novel method for monocular 3D occupancy prediction. ISO utilizes a D-FLoSP (Dual Feature Line of Sight Projection) and a multi-scale feature fusion strategy to address variations in scene and object sizes, specifically tailored for indoor scenes. Additionally, we present a new benchmark, Occ-ScanNet, aimed at fostering scalable studies of indoor scenes. We hope that this dataset will attract more attention to research on indoor occupancy prediction. While achieving promising results in 3D structure prediction, accurately identifying semantics remains a significant challenge in the future.",
|
| 1603 |
+
"bbox": [
|
| 1604 |
+
212,
|
| 1605 |
+
601,
|
| 1606 |
+
787,
|
| 1607 |
+
724
|
| 1608 |
+
],
|
| 1609 |
+
"page_idx": 13
|
| 1610 |
+
},
|
| 1611 |
+
{
|
| 1612 |
+
"type": "text",
|
| 1613 |
+
"text": "Acknowledgement",
|
| 1614 |
+
"text_level": 1,
|
| 1615 |
+
"bbox": [
|
| 1616 |
+
215,
|
| 1617 |
+
746,
|
| 1618 |
+
393,
|
| 1619 |
+
763
|
| 1620 |
+
],
|
| 1621 |
+
"page_idx": 13
|
| 1622 |
+
},
|
| 1623 |
+
{
|
| 1624 |
+
"type": "text",
|
| 1625 |
+
"text": "This work was supported in part by the National Key R&D Program of China (No. 2022ZD0160102), the National Natural Science Foundation of China (No. U21B2042, No. 62320106010), and in part by the 2035 Innovation Program of CAS, and the InnoHK program.",
|
| 1626 |
+
"bbox": [
|
| 1627 |
+
212,
|
| 1628 |
+
779,
|
| 1629 |
+
787,
|
| 1630 |
+
840
|
| 1631 |
+
],
|
| 1632 |
+
"page_idx": 13
|
| 1633 |
+
},
|
| 1634 |
+
{
|
| 1635 |
+
"type": "page_number",
|
| 1636 |
+
"text": "14",
|
| 1637 |
+
"bbox": [
|
| 1638 |
+
217,
|
| 1639 |
+
114,
|
| 1640 |
+
235,
|
| 1641 |
+
126
|
| 1642 |
+
],
|
| 1643 |
+
"page_idx": 13
|
| 1644 |
+
},
|
| 1645 |
+
{
|
| 1646 |
+
"type": "header",
|
| 1647 |
+
"text": "H. Yu et al.",
|
| 1648 |
+
"bbox": [
|
| 1649 |
+
271,
|
| 1650 |
+
114,
|
| 1651 |
+
351,
|
| 1652 |
+
126
|
| 1653 |
+
],
|
| 1654 |
+
"page_idx": 13
|
| 1655 |
+
},
|
| 1656 |
+
{
|
| 1657 |
+
"type": "text",
|
| 1658 |
+
"text": "References",
|
| 1659 |
+
"text_level": 1,
|
| 1660 |
+
"bbox": [
|
| 1661 |
+
215,
|
| 1662 |
+
143,
|
| 1663 |
+
321,
|
| 1664 |
+
159
|
| 1665 |
+
],
|
| 1666 |
+
"page_idx": 14
|
| 1667 |
+
},
|
| 1668 |
+
{
|
| 1669 |
+
"type": "list",
|
| 1670 |
+
"sub_type": "ref_text",
|
| 1671 |
+
"list_items": [
|
| 1672 |
+
"1. Arshad, M.S., Beksi, W.J.: List: Learning implicitly from spatial transformers for single-view 3d reconstruction. In: ICCV (2023)",
|
| 1673 |
+
"2. Azuma, R.T.: A survey of augmented reality. Presence: teleoperators & virtual environments (1997)",
|
| 1674 |
+
"3. Bhat, S.F., Birkl, R., Wofk, D., Wonka, P., Müller, M.: Zoedepth: Zero-shot transfer by combining relative and metric depth. arXiv preprint arXiv:2302.12288 (2023)",
|
| 1675 |
+
"4. Birkl, R., Wofk, D., Müller, M.: Midas v3. 1-a model zoo for robust monocular relative depth estimation. arXiv preprint arXiv:2307.14460 (2023)",
|
| 1676 |
+
"5. Cao, A.Q., de Charette, R.: Monoscene: Monocular 3d semantic scene completion. In: CVPR (2022)",
|
| 1677 |
+
"6. Chen, X., Lin, K.Y., Qian, C., Zeng, G., Li, H.: 3d sketch-aware semantic scene completion via semi-supervised structure prior. In: CVPR (2020)",
|
| 1678 |
+
"7. Dahnert, M., Hou, J., Nießner, M., Dai, A.: Panoptic 3d scene reconstruction from a single rgb image. NeurIPS (2021)",
|
| 1679 |
+
"8. Dai, A., Chang, A.X., Savva, M., Halber, M., Funkhouser, T., Nießner, M.: Scannet: Richly-annotated 3d reconstructions of indoor scenes. In: CVPR (2017)",
|
| 1680 |
+
"9. Denninger, M., Triebel, R.: 3d scene reconstruction from a single viewpoint. In: ECCV (2020)",
|
| 1681 |
+
"10. DeSouza, G.N., Kak, A.C.: Vision for mobile robot navigation: A survey. TPAMI (2002)",
|
| 1682 |
+
"1. Fan, H., Su, H., Guibas, L.J.: A point set generation network for 3d object reconstruction from a single image. In: CVPR (2017)",
|
| 1683 |
+
"2. Fieraru, M., Zanfir, M., Oneata, E., Popa, A.I., Olaru, V., Sminchisescu, C.: Reconstructing three-dimensional models of interacting humans. arXiv preprint arXiv:2308.01854 (2023)",
|
| 1684 |
+
"3. Goel, S., Pavlakos, G., Rajasegaran, J., Kanazawa, A., Malik, J.: Humans in 4d: Reconstructing and tracking humans with transformers. arXiv preprint arXiv:2305.20091 (2023)",
|
| 1685 |
+
"4. Guan, H., Song, C., Zhang, Z.: Gramo: geometric resampling augmentation for monocular 3d object detection. Frontiers of Computer Science (2024)",
|
| 1686 |
+
"5. Huang, J., Huang, G., Zhu, Z., Ye, Y., Du, D.: Bevdet: High-performance multicamera 3d object detection in bird-eye-view. arXiv preprint arXiv:2112.11790 (2021)",
|
| 1687 |
+
"6. Huang, Y., Zheng, W., Zhang, B., Zhou, J., Lu, J.: Selfocc: Self-supervised vision-based 3d occupancy prediction. In: CVPR (2024)",
|
| 1688 |
+
"7. Huang, Y., Zheng, W., Zhang, Y., Zhou, J., Lu, J.: Tri-perspective view for vision-based 3d semantic occupancy prediction. In: CVPR (2023)",
|
| 1689 |
+
"8. Li, J., Han, K., Wang, P., Liu, Y., Yuan, X.: Anisotropic convolutional networks for 3d semantic scene completion. In: CVPR (2020)",
|
| 1690 |
+
"9. Li, Y., Yu, Z., Choy, C., Xiao, C., Alvarez, J.M., Fidler, S., Feng, C., Anandkumar, A.: Voxformer: Sparse voxel transformer for camera-based 3d semantic scene completion. In: CVPR (2023)",
|
| 1691 |
+
"20. Li, Y., Ge, Z., Yu, G., Yang, J., Wang, Z., Shi, Y., Sun, J., Li, Z.: Bevdepth: Acquisition of reliable depth for multi-view 3d object detection. In: AAAI (2023)",
|
| 1692 |
+
"21. Li, Z., Wang, W., Li, H., Xie, E., Sima, C., Lu, T., Qiao, Y., Dai, J.: Bevformer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. In: ECCV (2022)"
|
| 1693 |
+
],
|
| 1694 |
+
"bbox": [
|
| 1695 |
+
225,
|
| 1696 |
+
178,
|
| 1697 |
+
784,
|
| 1698 |
+
839
|
| 1699 |
+
],
|
| 1700 |
+
"page_idx": 14
|
| 1701 |
+
},
|
| 1702 |
+
{
|
| 1703 |
+
"type": "header",
|
| 1704 |
+
"text": "Monocular Occupancy Prediction for Scalable Indoor Scenes",
|
| 1705 |
+
"bbox": [
|
| 1706 |
+
328,
|
| 1707 |
+
114,
|
| 1708 |
+
730,
|
| 1709 |
+
128
|
| 1710 |
+
],
|
| 1711 |
+
"page_idx": 14
|
| 1712 |
+
},
|
| 1713 |
+
{
|
| 1714 |
+
"type": "page_number",
|
| 1715 |
+
"text": "15",
|
| 1716 |
+
"bbox": [
|
| 1717 |
+
767,
|
| 1718 |
+
116,
|
| 1719 |
+
784,
|
| 1720 |
+
126
|
| 1721 |
+
],
|
| 1722 |
+
"page_idx": 14
|
| 1723 |
+
},
|
| 1724 |
+
{
|
| 1725 |
+
"type": "list",
|
| 1726 |
+
"sub_type": "ref_text",
|
| 1727 |
+
"list_items": [
|
| 1728 |
+
"22. Li, Z., Yu, Z., Austin, D., Fang, M., Lan, S., Kautz, J., Alvarez, J.M.: Fb-occ: 3d occupancy prediction based on forward-backward view transformation. arXiv preprint arXiv:2307.01492 (2023)",
|
| 1729 |
+
"23. Liu, S., Hu, Y., Zeng, Y., Tang, Q., Jin, B., Han, Y., Li, X.: See and think: Disentangling semantic scene completion. NeurIPS (2018)",
|
| 1730 |
+
"24. Murez, Z., Van As, T., Bartolozzi, J., Sinha, A., Badrinarayanan, V., Rabinovich, A.: Atlas: End-to-end 3d scene reconstruction from posed images. In: ECCV (2020)",
|
| 1731 |
+
"25. Pan, M., Liu, J., Zhang, R., Huang, P., Li, X., Liu, L., Zhang, S.: Renderocc: Visioncentric 3d occupancy prediction with 2d rendering supervision. arXiv preprint arXiv:2309.09502 (2023)",
|
| 1732 |
+
"26. Park, J.J., Florence, P., Straub, J., Newcombe, R., Lovegrove, S.: Deepsdf: Learning continuous signed distance functions for shape representation. In: CVPR (2019)",
|
| 1733 |
+
"27. Ranftl, R., Lasinger, K., Hafner, D., Schindler, K., Koltun, V.: Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer. TPAMI (2020)",
|
| 1734 |
+
"28. Roldao, L., de Charette, R., Verroust-Blondet, A.: Lmscnet: Lightweight multiscale 3d semantic completion. In: 3DV (2020)",
|
| 1735 |
+
"29. Silberman, N., Hoiem, D., Kohli, P., Fergus, R.: Indoor segmentation and support inference from rgbd images. In: ECCV (2012)",
|
| 1736 |
+
"30. Song, S., Yu, F., Zeng, A., Chang, A.X., Savva, M., Funkhouser, T.: Semantic scene completion from a single depth image. In: CVPR (2017)",
|
| 1737 |
+
"31. Tan, M., Le, Q.: Efficientnet: Rethinking model scaling for convolutional neural networks. In: ICML (2019)",
|
| 1738 |
+
"32. Tang, Y., Dorn, S., Savani, C.: Center3d: Center-based monocular 3d object detection with joint depth understanding. In: DAGM GCPR (2020)",
|
| 1739 |
+
"33. Tian, X., Jiang, T., Yun, L., Mao, Y., Yang, H., Wang, Y., Wang, Y., Zhao, H.: Occ3d: A large-scale 3d occupancy prediction benchmark for autonomous driving. NeurIPS (2023)",
|
| 1740 |
+
"34. Tong, W., Sima, C., Wang, T., Chen, L., Wu, S., Deng, H., Gu, Y., Lu, L., Luo, P., Lin, D., et al.: Scene as occupancy. In: ICCV (2023)",
|
| 1741 |
+
"35. Wang, S., Liu, Y., Wang, T., Li, Y., Zhang, X.: Exploring object-centric temporal modeling for efficient multi-view 3d object detection. In: ICCV (2023)",
|
| 1742 |
+
"36. Wang, X., Zhu, Z., Xu, W., Zhang, Y., Wei, Y., Chi, X., Ye, Y., Du, D., Lu, J., Wang, X.: Openoccupancy: A large scale benchmark for surrounding semantic occupancy perception. In: ICCV (2023)",
|
| 1743 |
+
"37. Wang, Y., Guizilini, V.C., Zhang, T., Wang, Y., Zhao, H., Solomon, J.: Detr3d: 3d object detection from multi-view images via 3d-to-2d queries. In: CoRL (2022)",
|
| 1744 |
+
"38. Wang, Y., Chen, Y., Liao, X., Fan, L., Zhang, Z.: Panoocc: Unified occupancy representation for camera-based 3d panoptic segmentation. In: CVPR (2024)",
|
| 1745 |
+
"39. Wang, Y., Chen, Y., Zhang, Z.: Frustumformer: Adaptive instance-aware resampling for multi-view 3d detection. In: CVPR (2023)",
|
| 1746 |
+
"40. Wei, Y., Zhao, L., Zheng, W., Zhu, Z., Zhou, J., Lu, J.: Surroundocc: Multi-camera 3d occupancy prediction for autonomous driving. In: ICCV (2023)",
|
| 1747 |
+
"41. Wu, Q., Wang, K., Li, K., Zheng, J., Cai, J.: Objectsdf++: Improved object-compositional neural implicit surfaces. In: ICCV (2023)",
|
| 1748 |
+
"42. Wu, S.C., Tateno, K., Navab, N., Tombari, F.: Scfusion: Real-time incremental scene reconstruction with semantic completion. In: 3DV (2020)",
|
| 1749 |
+
"43. Yang, C., Chen, Y., Tian, H., Tao, C., Zhu, X., Zhang, Z., Huang, G., Li, H., Qiao, Y., Lu, L., et al.: Bevformer v2: Adapting modern image backbones to bird's-eye-view recognition via perspective supervision. In: CVPR (2023)"
|
| 1750 |
+
],
|
| 1751 |
+
"bbox": [
|
| 1752 |
+
215,
|
| 1753 |
+
146,
|
| 1754 |
+
784,
|
| 1755 |
+
839
|
| 1756 |
+
],
|
| 1757 |
+
"page_idx": 15
|
| 1758 |
+
},
|
| 1759 |
+
{
|
| 1760 |
+
"type": "page_number",
|
| 1761 |
+
"text": "16",
|
| 1762 |
+
"bbox": [
|
| 1763 |
+
217,
|
| 1764 |
+
114,
|
| 1765 |
+
235,
|
| 1766 |
+
126
|
| 1767 |
+
],
|
| 1768 |
+
"page_idx": 15
|
| 1769 |
+
},
|
| 1770 |
+
{
|
| 1771 |
+
"type": "header",
|
| 1772 |
+
"text": "H. Yu et al.",
|
| 1773 |
+
"bbox": [
|
| 1774 |
+
271,
|
| 1775 |
+
114,
|
| 1776 |
+
351,
|
| 1777 |
+
126
|
| 1778 |
+
],
|
| 1779 |
+
"page_idx": 15
|
| 1780 |
+
},
|
| 1781 |
+
{
|
| 1782 |
+
"type": "list",
|
| 1783 |
+
"sub_type": "ref_text",
|
| 1784 |
+
"list_items": [
|
| 1785 |
+
"44. Yang, L., Kang, B., Huang, Z., Xu, X., Feng, J., Zhao, H.: Depth anything: Unleashing the power of large-scale unlabeled data. In: CVPR (2024)",
|
| 1786 |
+
"45. Yao, J., Li, C., Sun, K., Cai, Y., Li, H., Ouyang, W., Li, H.: Ndc-scene: Boost monocular 3d semantic scene completion in normalized device coordinates space. In: ICCV (2023)",
|
| 1787 |
+
"46. Yu, Z., Shu, C., Deng, J., Lu, K., Liu, Z., Yu, J., Yang, D., Li, H., Chen, Y.: Flashocc: Fast and memory-efficient occupancy prediction via channel-to-height plugin. arXiv preprint arXiv:2311.12058 (2023)",
|
| 1788 |
+
"47. Yurtsever, E., Lambert, J., Carballo, A., Takeda, K.: A survey of autonomous driving: Common practices and emerging technologies. IEEE access (2020)",
|
| 1789 |
+
"48. Zhang, J., Zhao, H., Yao, A., Chen, Y., Zhang, L., Liao, H.: Efficient semantic scene completion network with spatial group convolution. In: ECCV (2018)",
|
| 1790 |
+
"49. Zhang, X., Bi, S., Sunkavalli, K., Su, H., Xu, Z.: Nerfusion: Fusing radiance fields for large-scale scene reconstruction. In: CVPR (2022)",
|
| 1791 |
+
"50. Zheng, Z., Yu, T., Wei, Y., Dai, Q., Liu, Y.: Deephuman: 3d human reconstruction from a single image. In: ICCV (2019)",
|
| 1792 |
+
"51. Zhong, M., Zeng, G.: Semantic point completion network for 3d semantic scene completion. In: ECAI (2020)"
|
| 1793 |
+
],
|
| 1794 |
+
"bbox": [
|
| 1795 |
+
215,
|
| 1796 |
+
146,
|
| 1797 |
+
784,
|
| 1798 |
+
396
|
| 1799 |
+
],
|
| 1800 |
+
"page_idx": 16
|
| 1801 |
+
},
|
| 1802 |
+
{
|
| 1803 |
+
"type": "header",
|
| 1804 |
+
"text": "Monocular Occupancy Prediction for Scalable Indoor Scenes",
|
| 1805 |
+
"bbox": [
|
| 1806 |
+
328,
|
| 1807 |
+
114,
|
| 1808 |
+
730,
|
| 1809 |
+
128
|
| 1810 |
+
],
|
| 1811 |
+
"page_idx": 16
|
| 1812 |
+
},
|
| 1813 |
+
{
|
| 1814 |
+
"type": "page_number",
|
| 1815 |
+
"text": "17",
|
| 1816 |
+
"bbox": [
|
| 1817 |
+
767,
|
| 1818 |
+
116,
|
| 1819 |
+
784,
|
| 1820 |
+
126
|
| 1821 |
+
],
|
| 1822 |
+
"page_idx": 16
|
| 1823 |
+
}
|
| 1824 |
+
]
|
2024/Monocular Occupancy Prediction for Scalable Indoor Scenes/325d434b-3564-4d20-a82e-4ec4e345eb58_model.json
ADDED
|
@@ -0,0 +1,2456 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
[
|
| 3 |
+
{
|
| 4 |
+
"type": "title",
|
| 5 |
+
"bbox": [
|
| 6 |
+
0.24,
|
| 7 |
+
0.141,
|
| 8 |
+
0.764,
|
| 9 |
+
0.184
|
| 10 |
+
],
|
| 11 |
+
"angle": 0,
|
| 12 |
+
"content": "Monocular Occupancy Prediction for Scalable Indoor Scenes"
|
| 13 |
+
},
|
| 14 |
+
{
|
| 15 |
+
"type": "text",
|
| 16 |
+
"bbox": [
|
| 17 |
+
0.216,
|
| 18 |
+
0.212,
|
| 19 |
+
0.786,
|
| 20 |
+
0.229
|
| 21 |
+
],
|
| 22 |
+
"angle": 0,
|
| 23 |
+
"content": "Hongxiao \\(\\mathrm{Yu}^{1,2}\\) Yuqi Wang\\(^{1,2}\\), Yuntao Chen\\(^{3}\\), and Zhaoxiang Zhang\\(^{1,2,3}\\)"
|
| 24 |
+
},
|
| 25 |
+
{
|
| 26 |
+
"type": "text",
|
| 27 |
+
"bbox": [
|
| 28 |
+
0.217,
|
| 29 |
+
0.239,
|
| 30 |
+
0.784,
|
| 31 |
+
0.253
|
| 32 |
+
],
|
| 33 |
+
"angle": 0,
|
| 34 |
+
"content": "<sup>1</sup> School of Artificial Intelligence, University of Chinese Academy of Sciences (UCAS)"
|
| 35 |
+
},
|
| 36 |
+
{
|
| 37 |
+
"type": "text",
|
| 38 |
+
"bbox": [
|
| 39 |
+
0.23,
|
| 40 |
+
0.254,
|
| 41 |
+
0.771,
|
| 42 |
+
0.269
|
| 43 |
+
],
|
| 44 |
+
"angle": 0,
|
| 45 |
+
"content": "\\(^{2}\\) NLPR, MAIS, Institute of Automation, Chinese Academy of Sciences (CASIA)"
|
| 46 |
+
},
|
| 47 |
+
{
|
| 48 |
+
"type": "text",
|
| 49 |
+
"bbox": [
|
| 50 |
+
0.282,
|
| 51 |
+
0.269,
|
| 52 |
+
0.721,
|
| 53 |
+
0.31
|
| 54 |
+
],
|
| 55 |
+
"angle": 0,
|
| 56 |
+
"content": "<sup>3</sup> Centre for Artificial Intelligence and Robotics (HKISI_CAS) {yuhongxiao2023, wangyuqi2020, zhaoxiang.zhang}@ia.ac.cn, chenyuntao08@gmail.com"
|
| 57 |
+
},
|
| 58 |
+
{
|
| 59 |
+
"type": "text",
|
| 60 |
+
"bbox": [
|
| 61 |
+
0.261,
|
| 62 |
+
0.343,
|
| 63 |
+
0.74,
|
| 64 |
+
0.565
|
| 65 |
+
],
|
| 66 |
+
"angle": 0,
|
| 67 |
+
"content": "Abstract. Camera-based 3D occupancy prediction has recently garnered increasing attention in outdoor driving scenes. However, research in indoor scenes remains relatively unexplored. The core differences in indoor scenes lie in the complexity of scene scale and the variance in object size. In this paper, we propose a novel method, named ISO, for predicting indoor scene occupancy using monocular images. ISO harnesses the advantages of a pretrained depth model to achieve accurate depth predictions. Furthermore, we introduce the Dual Feature Line of Sight Projection (D-FLoSP) module within ISO, which enhances the learning of 3D voxel features. To foster further research in this domain, we introduce Occ-ScanNet, a large-scale occupancy benchmark for indoor scenes. With a dataset size 40 times larger than the NYUv2 dataset, it facilitates future scalable research in indoor scene analysis. Experimental results on both NYUv2 and Occ-ScanNet demonstrate that our method achieves state-of-the-art performance. The dataset and code are made publicly at https://github.com/hongxiaoy/ISO.git."
|
| 68 |
+
},
|
| 69 |
+
{
|
| 70 |
+
"type": "text",
|
| 71 |
+
"bbox": [
|
| 72 |
+
0.261,
|
| 73 |
+
0.577,
|
| 74 |
+
0.741,
|
| 75 |
+
0.606
|
| 76 |
+
],
|
| 77 |
+
"angle": 0,
|
| 78 |
+
"content": "Keywords: 3D occupancy prediction \\(\\cdot\\) Camera-based scene understanding \\(\\cdot\\) Semantic scene completion"
|
| 79 |
+
},
|
| 80 |
+
{
|
| 81 |
+
"type": "title",
|
| 82 |
+
"bbox": [
|
| 83 |
+
0.216,
|
| 84 |
+
0.629,
|
| 85 |
+
0.376,
|
| 86 |
+
0.645
|
| 87 |
+
],
|
| 88 |
+
"angle": 0,
|
| 89 |
+
"content": "1 Introduction"
|
| 90 |
+
},
|
| 91 |
+
{
|
| 92 |
+
"type": "text",
|
| 93 |
+
"bbox": [
|
| 94 |
+
0.214,
|
| 95 |
+
0.659,
|
| 96 |
+
0.788,
|
| 97 |
+
0.81
|
| 98 |
+
],
|
| 99 |
+
"angle": 0,
|
| 100 |
+
"content": "3D scene understanding is a crucial task in computer vision, becoming increasingly important for applications such as robotic navigation [10], augmented reality [2], and autonomous driving [47]. While humans possess a natural ability to comprehend 3D environments through vision, enabling this ability in computers poses a significant challenge due to limitations like restricted fields of view, sparse sensing capabilities, and measurement noise. However, with the rapid development of deep learning and the availability of large-scale 3D driving datasets, camera-based 3D object detection [14, 15, 20, 21, 35, 37, 39, 43] and occupancy prediction [5, 22, 33, 38, 40] have developed rapidly, achieving significant improvements in performance."
|
| 101 |
+
},
|
| 102 |
+
{
|
| 103 |
+
"type": "text",
|
| 104 |
+
"bbox": [
|
| 105 |
+
0.214,
|
| 106 |
+
0.81,
|
| 107 |
+
0.787,
|
| 108 |
+
0.84
|
| 109 |
+
],
|
| 110 |
+
"angle": 0,
|
| 111 |
+
"content": "Occupancy prediction has gained popularity recently due to its effectiveness in representing both background and foreground objects within a scene using a"
|
| 112 |
+
}
|
| 113 |
+
],
|
| 114 |
+
[
|
| 115 |
+
{
|
| 116 |
+
"type": "page_number",
|
| 117 |
+
"bbox": [
|
| 118 |
+
0.218,
|
| 119 |
+
0.116,
|
| 120 |
+
0.23,
|
| 121 |
+
0.127
|
| 122 |
+
],
|
| 123 |
+
"angle": 0,
|
| 124 |
+
"content": "2"
|
| 125 |
+
},
|
| 126 |
+
{
|
| 127 |
+
"type": "header",
|
| 128 |
+
"bbox": [
|
| 129 |
+
0.272,
|
| 130 |
+
0.115,
|
| 131 |
+
0.352,
|
| 132 |
+
0.128
|
| 133 |
+
],
|
| 134 |
+
"angle": 0,
|
| 135 |
+
"content": "H. Yu et al."
|
| 136 |
+
},
|
| 137 |
+
{
|
| 138 |
+
"type": "text",
|
| 139 |
+
"bbox": [
|
| 140 |
+
0.214,
|
| 141 |
+
0.147,
|
| 142 |
+
0.788,
|
| 143 |
+
0.403
|
| 144 |
+
],
|
| 145 |
+
"angle": 0,
|
| 146 |
+
"content": "unified representation. Although significant progress has been made in outdoor driving scenarios, research in indoor scenarios remains limited. Indoor scenes differ from outdoor driving scenes in two key aspects: (1) scene-scale complexity: Indoor rooms often exhibit a more diverse range of sizes compared to outdoor environments, where driving scenarios typically focus on a fixed 3D space for perception. This diversity, ranging from spacious living rooms to narrow kitchens, poses higher precision requirements for depth prediction. (2) object complexity: Indoor scenes feature a higher density and greater variety of objects. Unlike outdoor objects such as vehicles and pedestrians, which typically have consistent sizes within their respective categories and are relatively well-separated, indoor furniture and other objects often exhibit significant variations in scale and are closely positioned to each other. This increased complexity necessitates more sophisticated 3D perception techniques to accurately capture and understand the intricate geometry and relationships among objects within indoor spaces. Furthermore, existing works [5,45] focusing on indoor scenes primarily utilize the NYUv2 dataset [29], which lacks a more scalable and generalizable benchmark for comprehensive evaluation."
|
| 147 |
+
},
|
| 148 |
+
{
|
| 149 |
+
"type": "text",
|
| 150 |
+
"bbox": [
|
| 151 |
+
0.214,
|
| 152 |
+
0.406,
|
| 153 |
+
0.788,
|
| 154 |
+
0.603
|
| 155 |
+
],
|
| 156 |
+
"angle": 0,
|
| 157 |
+
"content": "To address the above issues, we firstly introduce a more scalable benchmark, Occ-ScanNet, for 3D occupancy prediction in indoor scenes. This benchmark builds upon the large-scale ScanNet [8] dataset, offering a 40 times more samples compared to the NYUv2 [29] dataset, thus greatly expanding the scope and diversity of indoor scene study. To effectively address the unique complexities of indoor scenes, we propose a novel method named ISO (Indoor Scene Occupancy). ISO leverages a powerful pretrained depth model and integrates a D-FLoSP (Dual Feature Line of Sight Projection) module. This module enables precise depth estimation and facilitates learning voxel features for accurate predictions. To accommodate the varying sizes of indoor objects, we introduce a multi-scale feature fusion module. This module enhances the learning of object features across different scales. Leveraging a powerful depth model, our method is adept at handling diverse indoor scenes, offering a robust and versatile solution."
|
| 158 |
+
},
|
| 159 |
+
{
|
| 160 |
+
"type": "text",
|
| 161 |
+
"bbox": [
|
| 162 |
+
0.24,
|
| 163 |
+
0.605,
|
| 164 |
+
0.633,
|
| 165 |
+
0.619
|
| 166 |
+
],
|
| 167 |
+
"angle": 0,
|
| 168 |
+
"content": "Our main contributions can be summarized as follows."
|
| 169 |
+
},
|
| 170 |
+
{
|
| 171 |
+
"type": "text",
|
| 172 |
+
"bbox": [
|
| 173 |
+
0.226,
|
| 174 |
+
0.639,
|
| 175 |
+
0.785,
|
| 176 |
+
0.699
|
| 177 |
+
],
|
| 178 |
+
"angle": 0,
|
| 179 |
+
"content": "- We introduce a new benchmark called Occ-ScanNet for monocular 3D occupancy prediction in indoor scenes. With a dataset size 40 times larger than the NYUv2 dataset, it significantly enhances the potential for scalable research in indoor scene analysis."
|
| 180 |
+
},
|
| 181 |
+
{
|
| 182 |
+
"type": "text",
|
| 183 |
+
"bbox": [
|
| 184 |
+
0.226,
|
| 185 |
+
0.702,
|
| 186 |
+
0.785,
|
| 187 |
+
0.776
|
| 188 |
+
],
|
| 189 |
+
"angle": 0,
|
| 190 |
+
"content": "- We propose a novel approach called \\( ISO \\), which primarily comprises the DFLoSP (Dual Feature Line of Sight Projection) module and a multi-scale feature fusion module. Together, these components effectively address the challenges posed by variations in scene and object sizes, enabling more accurate and robust 3D occupancy prediction."
|
| 191 |
+
},
|
| 192 |
+
{
|
| 193 |
+
"type": "text",
|
| 194 |
+
"bbox": [
|
| 195 |
+
0.225,
|
| 196 |
+
0.78,
|
| 197 |
+
0.785,
|
| 198 |
+
0.839
|
| 199 |
+
],
|
| 200 |
+
"angle": 0,
|
| 201 |
+
"content": "- Experiments on the Occ-ScanNet and NYUv2 datasets demonstrate that our approach achieves state-of-the-art performance in monocular 3D occupancy prediction. Furthermore, our method exhibits significant scalability potential."
|
| 202 |
+
},
|
| 203 |
+
{
|
| 204 |
+
"type": "list",
|
| 205 |
+
"bbox": [
|
| 206 |
+
0.225,
|
| 207 |
+
0.639,
|
| 208 |
+
0.785,
|
| 209 |
+
0.839
|
| 210 |
+
],
|
| 211 |
+
"angle": 0,
|
| 212 |
+
"content": null
|
| 213 |
+
}
|
| 214 |
+
],
|
| 215 |
+
[
|
| 216 |
+
{
|
| 217 |
+
"type": "header",
|
| 218 |
+
"bbox": [
|
| 219 |
+
0.33,
|
| 220 |
+
0.115,
|
| 221 |
+
0.732,
|
| 222 |
+
0.129
|
| 223 |
+
],
|
| 224 |
+
"angle": 0,
|
| 225 |
+
"content": "Monocular Occupancy Prediction for Scalable Indoor Scenes"
|
| 226 |
+
},
|
| 227 |
+
{
|
| 228 |
+
"type": "page_number",
|
| 229 |
+
"bbox": [
|
| 230 |
+
0.775,
|
| 231 |
+
0.117,
|
| 232 |
+
0.785,
|
| 233 |
+
0.127
|
| 234 |
+
],
|
| 235 |
+
"angle": 0,
|
| 236 |
+
"content": "3"
|
| 237 |
+
},
|
| 238 |
+
{
|
| 239 |
+
"type": "title",
|
| 240 |
+
"bbox": [
|
| 241 |
+
0.216,
|
| 242 |
+
0.145,
|
| 243 |
+
0.393,
|
| 244 |
+
0.161
|
| 245 |
+
],
|
| 246 |
+
"angle": 0,
|
| 247 |
+
"content": "2 Related works"
|
| 248 |
+
},
|
| 249 |
+
{
|
| 250 |
+
"type": "title",
|
| 251 |
+
"bbox": [
|
| 252 |
+
0.216,
|
| 253 |
+
0.177,
|
| 254 |
+
0.617,
|
| 255 |
+
0.192
|
| 256 |
+
],
|
| 257 |
+
"angle": 0,
|
| 258 |
+
"content": "2.1 Monocular 3D Semantic Scene Completion"
|
| 259 |
+
},
|
| 260 |
+
{
|
| 261 |
+
"type": "text",
|
| 262 |
+
"bbox": [
|
| 263 |
+
0.214,
|
| 264 |
+
0.202,
|
| 265 |
+
0.788,
|
| 266 |
+
0.399
|
| 267 |
+
],
|
| 268 |
+
"angle": 0,
|
| 269 |
+
"content": "Monocular 3D Semantic Scene Completion (SSC) aims to infer the complete 3D structure and corresponding semantics from a single image. This monocular setting was first introduced by Monoscene [5], advancing upon prior SSC methods [18,23,30,48,51] by relying solely on vision, without additional 3D inputs. Monoscene [5] introduces the Features Line of Sight Projection (FLoSP), inspired by optics, to obtain 3D features through ray projection. However, the shared 2D features lifted to 3D rays via FLoSP results in depth ambiguity. Consequently, subsequent efforts have placed greater emphasis on leveraging depth information. VoxFormer [19] proposed a novel query proposal network based on 2D convolutions, generating sparse queries from image depth, which showed impressive performance in driving scenes. Meanwhile, NDC-Scene [45] further devised a Depth-Adaptive Dual Decoder to concurrently upsample and merge the 2D and 3D feature maps, thereby enhancing overall performance."
|
| 270 |
+
},
|
| 271 |
+
{
|
| 272 |
+
"type": "title",
|
| 273 |
+
"bbox": [
|
| 274 |
+
0.216,
|
| 275 |
+
0.421,
|
| 276 |
+
0.566,
|
| 277 |
+
0.436
|
| 278 |
+
],
|
| 279 |
+
"angle": 0,
|
| 280 |
+
"content": "2.2 Multiview 3D Occupancy Prediction"
|
| 281 |
+
},
|
| 282 |
+
{
|
| 283 |
+
"type": "text",
|
| 284 |
+
"bbox": [
|
| 285 |
+
0.214,
|
| 286 |
+
0.446,
|
| 287 |
+
0.788,
|
| 288 |
+
0.643
|
| 289 |
+
],
|
| 290 |
+
"angle": 0,
|
| 291 |
+
"content": "Multiview 3D Occupancy Prediction predicts the semantic occupancy of the surrounding 3D scene given multiview images, has recently garnered significant attention in the field of autonomous driving. TPVFormer [17] pioneers exploration in driving scenes, employing sparse LiDAR labels for supervision. It introduces a tri-perspective view (TPV) representation that accompanies BEV with two additional perpendicular planes. However, due to the sparse supervision provided by LiDAR, subsequent works [33, 34, 36, 40] have focused on providing more dense occupancy benchmarks. FB-OCC [22] integrates the lifting of 2D to 3D and querying of 3D to 2D, achieving a more efficient 3D feature transformation. Meanwhile, PanoOcc [38] unifies detection and semantic occupancy tasks, enabling comprehensive panoramic scene understanding. Recent methods also explore improving model efficiency [46] and utilizing weaker forms of supervision [16, 25]."
|
| 292 |
+
},
|
| 293 |
+
{
|
| 294 |
+
"type": "title",
|
| 295 |
+
"bbox": [
|
| 296 |
+
0.216,
|
| 297 |
+
0.665,
|
| 298 |
+
0.519,
|
| 299 |
+
0.68
|
| 300 |
+
],
|
| 301 |
+
"angle": 0,
|
| 302 |
+
"content": "2.3 3D Reconstruction from Image"
|
| 303 |
+
},
|
| 304 |
+
{
|
| 305 |
+
"type": "text",
|
| 306 |
+
"bbox": [
|
| 307 |
+
0.214,
|
| 308 |
+
0.69,
|
| 309 |
+
0.788,
|
| 310 |
+
0.841
|
| 311 |
+
],
|
| 312 |
+
"angle": 0,
|
| 313 |
+
"content": "3D reconstruction is a technique that involves recovering a 3D representation of objects [1, 11, 26], scenes [9, 41, 49], or even human bodies [12, 13, 50] from camera images. The task of 3D reconstruction can be classified into two categories: monocular reconstruction and multi-view reconstruction, depending on the number of images utilized. In the context of indoor scenes, 3D scene reconstruction aims to determine the surface geometry of the entire scene, often without incorporating semantic information. Atlas [24] proposes an end-to-end 3D reconstruction framework using TSDF regression from RGB images, bypassing traditional depth map estimation for efficient semantic segmentation. Panoptic-Reconstruction [7] unifies geometric reconstruction, 3D semantic, and"
|
| 314 |
+
}
|
| 315 |
+
],
|
| 316 |
+
[
|
| 317 |
+
{
|
| 318 |
+
"type": "page_number",
|
| 319 |
+
"bbox": [
|
| 320 |
+
0.218,
|
| 321 |
+
0.116,
|
| 322 |
+
0.23,
|
| 323 |
+
0.127
|
| 324 |
+
],
|
| 325 |
+
"angle": 0,
|
| 326 |
+
"content": "4"
|
| 327 |
+
},
|
| 328 |
+
{
|
| 329 |
+
"type": "header",
|
| 330 |
+
"bbox": [
|
| 331 |
+
0.272,
|
| 332 |
+
0.115,
|
| 333 |
+
0.353,
|
| 334 |
+
0.127
|
| 335 |
+
],
|
| 336 |
+
"angle": 0,
|
| 337 |
+
"content": "H. Yu et al."
|
| 338 |
+
},
|
| 339 |
+
{
|
| 340 |
+
"type": "image",
|
| 341 |
+
"bbox": [
|
| 342 |
+
0.216,
|
| 343 |
+
0.151,
|
| 344 |
+
0.787,
|
| 345 |
+
0.289
|
| 346 |
+
],
|
| 347 |
+
"angle": 0,
|
| 348 |
+
"content": null
|
| 349 |
+
},
|
| 350 |
+
{
|
| 351 |
+
"type": "image_caption",
|
| 352 |
+
"bbox": [
|
| 353 |
+
0.214,
|
| 354 |
+
0.3,
|
| 355 |
+
0.788,
|
| 356 |
+
0.399
|
| 357 |
+
],
|
| 358 |
+
"angle": 0,
|
| 359 |
+
"content": "Fig. 1: The core design of ISO centers around the transformation of features from 2D to 3D spaces, encompassing the Depth Branch and the D-FLoSP module. A depth branch is initially integrated, it leverages a pre-trained depth model to estimate a pixel-wise depth map which is processed by the DepthNet to generate the final depth distribution. An element-wise multiplication between the voxel depth and features followed by summation are subsequently performed to derive the initial 3D voxel feature. The 3D feature is further processed to predict the 3D scene occupancy."
|
| 360 |
+
},
|
| 361 |
+
{
|
| 362 |
+
"type": "text",
|
| 363 |
+
"bbox": [
|
| 364 |
+
0.214,
|
| 365 |
+
0.43,
|
| 366 |
+
0.788,
|
| 367 |
+
0.522
|
| 368 |
+
],
|
| 369 |
+
"angle": 0,
|
| 370 |
+
"content": "instance segmentation tasks. It predicts complete geometric reconstruction, semantic, and instance segmentations from a single RGB image's camera frustum. SCFusion [42], a real-time scene reconstruction framework, integrates continuous depth data using neural architecture for occupancy maps. It efficiently combines semantic completion with voxel states for simultaneous scene reconstruction and semantic understanding in real-time 3D environments."
|
| 371 |
+
},
|
| 372 |
+
{
|
| 373 |
+
"type": "title",
|
| 374 |
+
"bbox": [
|
| 375 |
+
0.216,
|
| 376 |
+
0.549,
|
| 377 |
+
0.331,
|
| 378 |
+
0.566
|
| 379 |
+
],
|
| 380 |
+
"angle": 0,
|
| 381 |
+
"content": "3 Method"
|
| 382 |
+
},
|
| 383 |
+
{
|
| 384 |
+
"type": "text",
|
| 385 |
+
"bbox": [
|
| 386 |
+
0.214,
|
| 387 |
+
0.586,
|
| 388 |
+
0.788,
|
| 389 |
+
0.663
|
| 390 |
+
],
|
| 391 |
+
"angle": 0,
|
| 392 |
+
"content": "Problem definition. We focus on the problem of monocular 3D Occupancy Prediction. Specifically, this task takes a single RGB image \\(\\mathbf{I}^{RGB}\\) as input and output a voxel-wise occupancy along with semantic categories \\(\\mathbf{Y}^{X\\times Y\\times Z\\times C}\\). \\(X\\), \\(Y\\), and \\(Z\\) represent the dimensions of the predicted 3D scene, while \\(C\\) denotes the total number of semantic categories."
|
| 393 |
+
},
|
| 394 |
+
{
|
| 395 |
+
"type": "text",
|
| 396 |
+
"bbox": [
|
| 397 |
+
0.214,
|
| 398 |
+
0.689,
|
| 399 |
+
0.789,
|
| 400 |
+
0.842
|
| 401 |
+
],
|
| 402 |
+
"angle": 0,
|
| 403 |
+
"content": "Model overview. In this section, we introduce the overall architecture of ISO, as shown in Fig. 1. Following Monoscene [5], we utilize 2D Unet and 3D Unet architectures to handle 2D and 3D features. Our core design focuses on transforming features from 2D to 3D, incorporating the Depth Branch and D-FLoSP module. Specifically, we incorporate a depth branch to estimate the pixel-wise depth map using a pre-trained depth model. Processed by the DepthNet, the model outputs a refined depth distribution. Then, voxel depth and features are multiplied element-wise and summed to obtain the initial 3D voxel feature volume \\(\\mathbf{X}^{3\\mathbf{d}} \\in \\mathbb{R}^{X \\times Y \\times Z \\times C}\\). Finally, after processing the 3D voxel features, the model outputs the 3D scene occupancy."
|
| 404 |
+
}
|
| 405 |
+
],
|
| 406 |
+
[
|
| 407 |
+
{
|
| 408 |
+
"type": "header",
|
| 409 |
+
"bbox": [
|
| 410 |
+
0.329,
|
| 411 |
+
0.115,
|
| 412 |
+
0.733,
|
| 413 |
+
0.129
|
| 414 |
+
],
|
| 415 |
+
"angle": 0,
|
| 416 |
+
"content": "Monocular Occupancy Prediction for Scalable Indoor Scenes"
|
| 417 |
+
},
|
| 418 |
+
{
|
| 419 |
+
"type": "page_number",
|
| 420 |
+
"bbox": [
|
| 421 |
+
0.775,
|
| 422 |
+
0.117,
|
| 423 |
+
0.785,
|
| 424 |
+
0.127
|
| 425 |
+
],
|
| 426 |
+
"angle": 0,
|
| 427 |
+
"content": "5"
|
| 428 |
+
},
|
| 429 |
+
{
|
| 430 |
+
"type": "title",
|
| 431 |
+
"bbox": [
|
| 432 |
+
0.216,
|
| 433 |
+
0.147,
|
| 434 |
+
0.381,
|
| 435 |
+
0.162
|
| 436 |
+
],
|
| 437 |
+
"angle": 0,
|
| 438 |
+
"content": "3.1 Depth Branch"
|
| 439 |
+
},
|
| 440 |
+
{
|
| 441 |
+
"type": "text",
|
| 442 |
+
"bbox": [
|
| 443 |
+
0.214,
|
| 444 |
+
0.169,
|
| 445 |
+
0.784,
|
| 446 |
+
0.199
|
| 447 |
+
],
|
| 448 |
+
"angle": 0,
|
| 449 |
+
"content": "In this section, we introduce how to effectively estimate depth information from a single image."
|
| 450 |
+
},
|
| 451 |
+
{
|
| 452 |
+
"type": "text",
|
| 453 |
+
"bbox": [
|
| 454 |
+
0.214,
|
| 455 |
+
0.218,
|
| 456 |
+
0.784,
|
| 457 |
+
0.277
|
| 458 |
+
],
|
| 459 |
+
"angle": 0,
|
| 460 |
+
"content": "Coarse depth estimation. Learning depth from scratch can be quite challenging. However, thanks to the recent rapid advancements in depth estimation [3,4,27,44], we can leverage pre-trained depth models to initially estimate a coarse depth map."
|
| 461 |
+
},
|
| 462 |
+
{
|
| 463 |
+
"type": "text",
|
| 464 |
+
"bbox": [
|
| 465 |
+
0.214,
|
| 466 |
+
0.279,
|
| 467 |
+
0.785,
|
| 468 |
+
0.325
|
| 469 |
+
],
|
| 470 |
+
"angle": 0,
|
| 471 |
+
"content": "Compared to past models that could only estimate relative depth, we opted for Depth-Anything [44] as the pre-trained depth model, because it excels in predicting metric depth, as the Eq. (1) shows:"
|
| 472 |
+
},
|
| 473 |
+
{
|
| 474 |
+
"type": "equation",
|
| 475 |
+
"bbox": [
|
| 476 |
+
0.374,
|
| 477 |
+
0.329,
|
| 478 |
+
0.786,
|
| 479 |
+
0.349
|
| 480 |
+
],
|
| 481 |
+
"angle": 0,
|
| 482 |
+
"content": "\\[\n\\mathbf {D} ^ {\\mathrm {m e t r i c}} = \\mathbf {N} _ {\\mathrm {d e p t h}} (\\mathbf {I} ^ {\\mathrm {r g b}}) \\in \\mathbb {R} ^ {1 \\times H \\times W}, \\tag {1}\n\\]"
|
| 483 |
+
},
|
| 484 |
+
{
|
| 485 |
+
"type": "text",
|
| 486 |
+
"bbox": [
|
| 487 |
+
0.215,
|
| 488 |
+
0.356,
|
| 489 |
+
0.784,
|
| 490 |
+
0.387
|
| 491 |
+
],
|
| 492 |
+
"angle": 0,
|
| 493 |
+
"content": "where \\(\\mathbf{N}_{\\mathrm{depth}}\\) denotes the pre-trained depth model [3, 44], and \\(\\mathbf{I}^{\\mathrm{rgb}}\\) represents the input image. \\(H\\) and \\(W\\) represents the height and width of the input image."
|
| 494 |
+
},
|
| 495 |
+
{
|
| 496 |
+
"type": "text",
|
| 497 |
+
"bbox": [
|
| 498 |
+
0.215,
|
| 499 |
+
0.405,
|
| 500 |
+
0.79,
|
| 501 |
+
0.497
|
| 502 |
+
],
|
| 503 |
+
"angle": 0,
|
| 504 |
+
"content": "Depth refinement. The coarse depth estimation is subsequently refined through model learning. The depth from pre-trained depth model is not precise enough for a higher mIoU, so we design a fine-tuning strategy. Specifically, the predicted metric depth \\(\\mathbf{D}^{\\mathrm{metric}}\\) is then concatenated with the image feature \\(\\mathbf{X}\\) that has the same spatial scale. The augmented feature is continue processed by a following DepthNet to get a refined depth distribution \\(\\mathbf{D}_{s=1}^{\\mathrm{dist}}\\), as Eq. (2) shows:"
|
| 505 |
+
},
|
| 506 |
+
{
|
| 507 |
+
"type": "equation",
|
| 508 |
+
"bbox": [
|
| 509 |
+
0.269,
|
| 510 |
+
0.515,
|
| 511 |
+
0.785,
|
| 512 |
+
0.548
|
| 513 |
+
],
|
| 514 |
+
"angle": 0,
|
| 515 |
+
"content": "\\[\n\\mathbf {D} _ {\\mathrm {s} = 1} ^ {\\text {d i s t}} = \\mathbf {F} _ {\\text {d e p t h}} \\left(\\operatorname {C o n c a t} \\left(\\mathbf {D} ^ {\\text {m e t r i c}}, \\mathbf {X} _ {\\mathrm {s} = 1} ^ {2 \\mathrm {d}}\\right)\\right), \\quad \\mathbf {D} _ {\\mathrm {s} = 1} ^ {\\text {d i s t}} \\in \\mathbb {R} ^ {N _ {\\text {b i n s}} \\times H \\times W}, \\tag {2}\n\\]"
|
| 516 |
+
},
|
| 517 |
+
{
|
| 518 |
+
"type": "text",
|
| 519 |
+
"bbox": [
|
| 520 |
+
0.214,
|
| 521 |
+
0.555,
|
| 522 |
+
0.784,
|
| 523 |
+
0.601
|
| 524 |
+
],
|
| 525 |
+
"angle": 0,
|
| 526 |
+
"content": "where \\(\\mathbf{F}_{\\mathrm{depth}}\\) is the DepthNet, \\(\\mathbf{X}_{s=1}^{2d}\\) is the 2D image feature, and \\(N_{\\mathrm{bins}}\\) is the number of discrete depth bins. \\(s\\) denotes the level of feature scale, where \\(s = 1\\) implies that the feature is at a 1:1 original image ratio."
|
| 527 |
+
},
|
| 528 |
+
{
|
| 529 |
+
"type": "text",
|
| 530 |
+
"bbox": [
|
| 531 |
+
0.214,
|
| 532 |
+
0.601,
|
| 533 |
+
0.784,
|
| 534 |
+
0.632
|
| 535 |
+
],
|
| 536 |
+
"angle": 0,
|
| 537 |
+
"content": "We down-samples the scale \\( \\frac{1}{1} \\) distribution to get other smaller-scale 2D image feature maps' depth distribution. We represent the procedure as Eq. (3):"
|
| 538 |
+
},
|
| 539 |
+
{
|
| 540 |
+
"type": "equation",
|
| 541 |
+
"bbox": [
|
| 542 |
+
0.348,
|
| 543 |
+
0.637,
|
| 544 |
+
0.785,
|
| 545 |
+
0.656
|
| 546 |
+
],
|
| 547 |
+
"angle": 0,
|
| 548 |
+
"content": "\\[\n\\mathbf {D} _ {\\mathrm {s} = k} ^ {\\text {d i s t}} = \\operatorname {D o w n S a m p l e} \\left(\\mathbf {D} _ {\\mathrm {s} = 1} ^ {\\text {d i s t}}\\right), k = 2, 4, 8. \\tag {3}\n\\]"
|
| 549 |
+
},
|
| 550 |
+
{
|
| 551 |
+
"type": "text",
|
| 552 |
+
"bbox": [
|
| 553 |
+
0.214,
|
| 554 |
+
0.663,
|
| 555 |
+
0.784,
|
| 556 |
+
0.725
|
| 557 |
+
],
|
| 558 |
+
"angle": 0,
|
| 559 |
+
"content": "The refined depth is supervised by the ground truth \\(\\mathbf{D}^{\\mathrm{GT}}\\). The ground truth has single depth value at each pixel, so we convert the single value to a one-hot vector of length \\(N_{\\mathrm{bins}}\\). The refined depth prediction can be optimized using the BCE loss function as Eq. (4) shows."
|
| 560 |
+
},
|
| 561 |
+
{
|
| 562 |
+
"type": "equation",
|
| 563 |
+
"bbox": [
|
| 564 |
+
0.23,
|
| 565 |
+
0.746,
|
| 566 |
+
0.786,
|
| 567 |
+
0.844
|
| 568 |
+
],
|
| 569 |
+
"angle": 0,
|
| 570 |
+
"content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {\\mathrm {d e p t h}} = - \\frac {1}{N _ {\\mathrm {b i n s}} \\times H \\times W} \\\\ \\sum_ {d = 1} ^ {N _ {\\mathrm {b i n s}}} \\sum_ {h = 1} ^ {H} \\sum_ {w = 1} ^ {W} \\left[ \\mathbf {D} _ {N _ {\\mathrm {b i n s}}, d, h, w} ^ {\\mathrm {G T}} \\cdot \\log \\left(\\mathbf {D} _ {N _ {\\mathrm {b i n s}}, d, h, w} ^ {\\mathrm {d i s t}}\\right) \\right. \\\\ \\left. + \\left(1 - \\mathbf {D} _ {N _ {\\mathrm {b i n s}}, d, h, w} ^ {\\mathrm {G T}}\\right) \\cdot \\log \\left(1 - \\mathbf {D} _ {N _ {\\mathrm {b i n s}}, d, h, w} ^ {\\mathrm {d i s t}}\\right) \\right] \\quad (4) \\\\ \\end{array}\n\\]"
|
| 571 |
+
}
|
| 572 |
+
],
|
| 573 |
+
[
|
| 574 |
+
{
|
| 575 |
+
"type": "page_number",
|
| 576 |
+
"bbox": [
|
| 577 |
+
0.218,
|
| 578 |
+
0.116,
|
| 579 |
+
0.23,
|
| 580 |
+
0.127
|
| 581 |
+
],
|
| 582 |
+
"angle": 0,
|
| 583 |
+
"content": "6"
|
| 584 |
+
},
|
| 585 |
+
{
|
| 586 |
+
"type": "header",
|
| 587 |
+
"bbox": [
|
| 588 |
+
0.272,
|
| 589 |
+
0.115,
|
| 590 |
+
0.352,
|
| 591 |
+
0.128
|
| 592 |
+
],
|
| 593 |
+
"angle": 0,
|
| 594 |
+
"content": "H. Yu et al."
|
| 595 |
+
},
|
| 596 |
+
{
|
| 597 |
+
"type": "title",
|
| 598 |
+
"bbox": [
|
| 599 |
+
0.216,
|
| 600 |
+
0.147,
|
| 601 |
+
0.672,
|
| 602 |
+
0.164
|
| 603 |
+
],
|
| 604 |
+
"angle": 0,
|
| 605 |
+
"content": "3.2 Dual Feature Line of Sight Projection (D-FLoSP)"
|
| 606 |
+
},
|
| 607 |
+
{
|
| 608 |
+
"type": "text",
|
| 609 |
+
"bbox": [
|
| 610 |
+
0.214,
|
| 611 |
+
0.181,
|
| 612 |
+
0.784,
|
| 613 |
+
0.242
|
| 614 |
+
],
|
| 615 |
+
"angle": 0,
|
| 616 |
+
"content": "Compared to the FLoSP module in MonoScene [5], which projects 3D voxels to 2D, and 3D feature vectors are the corresponding pixel feature vectors. It only considers the projection of 3D rays and overlooks depth information, our proposed D-FLoSP module can more effectively integrate depth information."
|
| 617 |
+
},
|
| 618 |
+
{
|
| 619 |
+
"type": "text",
|
| 620 |
+
"bbox": [
|
| 621 |
+
0.214,
|
| 622 |
+
0.244,
|
| 623 |
+
0.785,
|
| 624 |
+
0.319
|
| 625 |
+
],
|
| 626 |
+
"angle": 0,
|
| 627 |
+
"content": "Specifically, after getting depth distribution in the camera frame, we implement the FLoSP module to depth distribution. Each 3D voxel centroid position \\( x^{c} \\) in the world frame can be projected to the camera frame and the pixel frame using camera pose and intrinsic matrix. We assign the projected depth values to discrete depth bins index by Eqs. (5) and (6), according to [32],"
|
| 628 |
+
},
|
| 629 |
+
{
|
| 630 |
+
"type": "equation",
|
| 631 |
+
"bbox": [
|
| 632 |
+
0.383,
|
| 633 |
+
0.334,
|
| 634 |
+
0.785,
|
| 635 |
+
0.366
|
| 636 |
+
],
|
| 637 |
+
"angle": 0,
|
| 638 |
+
"content": "\\[\n\\delta = \\frac {2 \\left(d _ {\\max } - d _ {\\min }\\right)}{N _ {b i n s} \\left(1 + N _ {b i n s}\\right)}, \\tag {5}\n\\]"
|
| 639 |
+
},
|
| 640 |
+
{
|
| 641 |
+
"type": "equation",
|
| 642 |
+
"bbox": [
|
| 643 |
+
0.386,
|
| 644 |
+
0.37,
|
| 645 |
+
0.785,
|
| 646 |
+
0.402
|
| 647 |
+
],
|
| 648 |
+
"angle": 0,
|
| 649 |
+
"content": "\\[\nl = - 0. 5 + 0. 5 \\sqrt {1 + \\frac {8 \\left(d - d _ {\\operatorname* {m i n}}\\right)}{\\delta}}, \\tag {6}\n\\]"
|
| 650 |
+
},
|
| 651 |
+
{
|
| 652 |
+
"type": "text",
|
| 653 |
+
"bbox": [
|
| 654 |
+
0.214,
|
| 655 |
+
0.416,
|
| 656 |
+
0.784,
|
| 657 |
+
0.446
|
| 658 |
+
],
|
| 659 |
+
"angle": 0,
|
| 660 |
+
"content": "and the depth distribution probability it lies in serves as the depth probability of that voxel. The process is illustrated by Eq. (7):"
|
| 661 |
+
},
|
| 662 |
+
{
|
| 663 |
+
"type": "equation",
|
| 664 |
+
"bbox": [
|
| 665 |
+
0.324,
|
| 666 |
+
0.46,
|
| 667 |
+
0.785,
|
| 668 |
+
0.48
|
| 669 |
+
],
|
| 670 |
+
"angle": 0,
|
| 671 |
+
"content": "\\[\n\\mathbf {D} _ {\\mathrm {s} = k} ^ {\\mathrm {3 d}} = \\Phi_ {\\rho \\left(x ^ {c}\\right)} ^ {\\mathrm {3 d}} \\left(\\mathbf {D} _ {\\mathrm {s} = k} ^ {\\mathrm {d i s t}}\\right) \\in \\mathbb {R} ^ {X \\times Y \\times Z \\times C}, k = 1, 2, 4, 8, \\tag {7}\n\\]"
|
| 672 |
+
},
|
| 673 |
+
{
|
| 674 |
+
"type": "text",
|
| 675 |
+
"bbox": [
|
| 676 |
+
0.214,
|
| 677 |
+
0.495,
|
| 678 |
+
0.785,
|
| 679 |
+
0.585
|
| 680 |
+
],
|
| 681 |
+
"angle": 0,
|
| 682 |
+
"content": "where \\(\\Phi_{a}^{3\\mathrm{d}}(b)\\) is the 3D sampling of b at coordinates a, and \\(\\rho (\\cdot)\\) is the perspective projection. Thus, the depth distribution is lifted from 2D to 3D. The bin-based representation with quantization is used for mapping continuous depth value to a discrete form, each bin will be assigned a different depth distribution score to make voxel aware of depth. The original depth precision is a continuous value, and hard to integrate it with the discrete form of voxel."
|
| 683 |
+
},
|
| 684 |
+
{
|
| 685 |
+
"type": "text",
|
| 686 |
+
"bbox": [
|
| 687 |
+
0.214,
|
| 688 |
+
0.587,
|
| 689 |
+
0.785,
|
| 690 |
+
0.649
|
| 691 |
+
],
|
| 692 |
+
"angle": 0,
|
| 693 |
+
"content": "Following [5], we also employ the FLoSP module to generate a 3D voxel feature from a 2D feature map. Similar to what we did with the voxel centroids, the projected voxel can sample corresponding \\(1:k\\) scaled feature maps \\(F_{1:k}\\) from the 2D UNet decoder. This process is illustrated by Equation Eq. (8):"
|
| 694 |
+
},
|
| 695 |
+
{
|
| 696 |
+
"type": "equation",
|
| 697 |
+
"bbox": [
|
| 698 |
+
0.324,
|
| 699 |
+
0.663,
|
| 700 |
+
0.785,
|
| 701 |
+
0.682
|
| 702 |
+
],
|
| 703 |
+
"angle": 0,
|
| 704 |
+
"content": "\\[\n\\mathbf {X} _ {\\mathrm {s} = k} ^ {3 \\mathrm {d}} = \\Phi_ {\\rho \\left(x ^ {c}\\right)} ^ {2 \\mathrm {d}} \\left(\\mathbf {X} _ {\\mathrm {s} = k} ^ {2 \\mathrm {d}}\\right) \\in \\mathbb {R} ^ {X \\times Y \\times Z \\times C}, k = 1, 2, 4, 8, \\tag {8}\n\\]"
|
| 705 |
+
},
|
| 706 |
+
{
|
| 707 |
+
"type": "text",
|
| 708 |
+
"bbox": [
|
| 709 |
+
0.214,
|
| 710 |
+
0.697,
|
| 711 |
+
0.784,
|
| 712 |
+
0.728
|
| 713 |
+
],
|
| 714 |
+
"angle": 0,
|
| 715 |
+
"content": "where \\(\\Phi_{a}^{2\\mathrm{d}}(b)\\) is the 2D sampling of b at coordinates a, and \\(\\rho (\\cdot)\\) is the perspective projection."
|
| 716 |
+
},
|
| 717 |
+
{
|
| 718 |
+
"type": "text",
|
| 719 |
+
"bbox": [
|
| 720 |
+
0.239,
|
| 721 |
+
0.73,
|
| 722 |
+
0.632,
|
| 723 |
+
0.745
|
| 724 |
+
],
|
| 725 |
+
"angle": 0,
|
| 726 |
+
"content": "The final 3D feature map \\(\\mathbf{X}^{3\\mathrm{d}}\\) are summed in Eq. (9):"
|
| 727 |
+
},
|
| 728 |
+
{
|
| 729 |
+
"type": "equation",
|
| 730 |
+
"bbox": [
|
| 731 |
+
0.393,
|
| 732 |
+
0.761,
|
| 733 |
+
0.785,
|
| 734 |
+
0.793
|
| 735 |
+
],
|
| 736 |
+
"angle": 0,
|
| 737 |
+
"content": "\\[\n\\mathbf {X} ^ {\\mathrm {3 d}} = \\sum_ {\\mathrm {s} \\in \\{1, 2, 4, 8 \\}} \\mathbf {X} _ {\\mathrm {s} = k} ^ {\\mathrm {3 d}} \\odot \\mathbf {D} _ {\\mathrm {s} = k} ^ {\\mathrm {3 d}}, \\tag {9}\n\\]"
|
| 738 |
+
},
|
| 739 |
+
{
|
| 740 |
+
"type": "text",
|
| 741 |
+
"bbox": [
|
| 742 |
+
0.214,
|
| 743 |
+
0.81,
|
| 744 |
+
0.784,
|
| 745 |
+
0.84
|
| 746 |
+
],
|
| 747 |
+
"angle": 0,
|
| 748 |
+
"content": "where \\(\\odot\\) is the element-wise multiplication. The output map \\(\\mathbf{X}^{3\\mathrm{d}}\\) is then served as the input of 3D UNet."
|
| 749 |
+
}
|
| 750 |
+
],
|
| 751 |
+
[
|
| 752 |
+
{
|
| 753 |
+
"type": "header",
|
| 754 |
+
"bbox": [
|
| 755 |
+
0.33,
|
| 756 |
+
0.115,
|
| 757 |
+
0.732,
|
| 758 |
+
0.129
|
| 759 |
+
],
|
| 760 |
+
"angle": 0,
|
| 761 |
+
"content": "Monocular Occupancy Prediction for Scalable Indoor Scenes"
|
| 762 |
+
},
|
| 763 |
+
{
|
| 764 |
+
"type": "page_number",
|
| 765 |
+
"bbox": [
|
| 766 |
+
0.775,
|
| 767 |
+
0.117,
|
| 768 |
+
0.785,
|
| 769 |
+
0.127
|
| 770 |
+
],
|
| 771 |
+
"angle": 0,
|
| 772 |
+
"content": "7"
|
| 773 |
+
},
|
| 774 |
+
{
|
| 775 |
+
"type": "image",
|
| 776 |
+
"bbox": [
|
| 777 |
+
0.272,
|
| 778 |
+
0.163,
|
| 779 |
+
0.431,
|
| 780 |
+
0.288
|
| 781 |
+
],
|
| 782 |
+
"angle": 0,
|
| 783 |
+
"content": null
|
| 784 |
+
},
|
| 785 |
+
{
|
| 786 |
+
"type": "image_caption",
|
| 787 |
+
"bbox": [
|
| 788 |
+
0.216,
|
| 789 |
+
0.299,
|
| 790 |
+
0.498,
|
| 791 |
+
0.321
|
| 792 |
+
],
|
| 793 |
+
"angle": 0,
|
| 794 |
+
"content": "(a) Depth ranges in NYUv2 and our Occ-ScanNet Benchmark"
|
| 795 |
+
},
|
| 796 |
+
{
|
| 797 |
+
"type": "image",
|
| 798 |
+
"bbox": [
|
| 799 |
+
0.56,
|
| 800 |
+
0.163,
|
| 801 |
+
0.718,
|
| 802 |
+
0.287
|
| 803 |
+
],
|
| 804 |
+
"angle": 0,
|
| 805 |
+
"content": null
|
| 806 |
+
},
|
| 807 |
+
{
|
| 808 |
+
"type": "image_caption",
|
| 809 |
+
"bbox": [
|
| 810 |
+
0.506,
|
| 811 |
+
0.299,
|
| 812 |
+
0.786,
|
| 813 |
+
0.321
|
| 814 |
+
],
|
| 815 |
+
"angle": 0,
|
| 816 |
+
"content": "(b) Dataset attributes of NYUv2 and our OccScanNet Benchmark"
|
| 817 |
+
},
|
| 818 |
+
{
|
| 819 |
+
"type": "image_caption",
|
| 820 |
+
"bbox": [
|
| 821 |
+
0.214,
|
| 822 |
+
0.332,
|
| 823 |
+
0.788,
|
| 824 |
+
0.404
|
| 825 |
+
],
|
| 826 |
+
"angle": 0,
|
| 827 |
+
"content": "Fig. 2: Comparison of NYUv2 and Occ-ScanNet Benchmark. In (a), the depth ranges of NYUv2 and Occ-ScanNet are distinguished by dark and light green, respectively, with the horizontal axis indicating the minimum depth and the vertical axis showing the maximum depth of scenes. (b) quantitatively demonstrates that Occ-ScanNet possesses a significantly larger data scale compared to the NYUv2 dataset."
|
| 828 |
+
},
|
| 829 |
+
{
|
| 830 |
+
"type": "title",
|
| 831 |
+
"bbox": [
|
| 832 |
+
0.216,
|
| 833 |
+
0.428,
|
| 834 |
+
0.495,
|
| 835 |
+
0.443
|
| 836 |
+
],
|
| 837 |
+
"angle": 0,
|
| 838 |
+
"content": "4 Occ-ScanNet Benchmark"
|
| 839 |
+
},
|
| 840 |
+
{
|
| 841 |
+
"type": "text",
|
| 842 |
+
"bbox": [
|
| 843 |
+
0.214,
|
| 844 |
+
0.456,
|
| 845 |
+
0.788,
|
| 846 |
+
0.547
|
| 847 |
+
],
|
| 848 |
+
"angle": 0,
|
| 849 |
+
"content": "Compared to the previous widely used indoor scene benchmark NYUv2 [29], which contains only 795 / 654 for train / test samples, our benchmark boasts 45,755 / 19,764 samples. Our benchmark significantly surpasses NYUv2 [29] in both data quantity and richness of scene depth, as illustrated in Fig. 2. This dataset is available at https://huggingface.co/datasets/hongxiaoy/OccScanNet."
|
| 850 |
+
},
|
| 851 |
+
{
|
| 852 |
+
"type": "title",
|
| 853 |
+
"bbox": [
|
| 854 |
+
0.216,
|
| 855 |
+
0.566,
|
| 856 |
+
0.341,
|
| 857 |
+
0.58
|
| 858 |
+
],
|
| 859 |
+
"angle": 0,
|
| 860 |
+
"content": "4.1 Overview"
|
| 861 |
+
},
|
| 862 |
+
{
|
| 863 |
+
"type": "text",
|
| 864 |
+
"bbox": [
|
| 865 |
+
0.214,
|
| 866 |
+
0.588,
|
| 867 |
+
0.787,
|
| 868 |
+
0.664
|
| 869 |
+
],
|
| 870 |
+
"angle": 0,
|
| 871 |
+
"content": "Occ-ScanNet benchmark features a train/Validation split of 45,755 / 19,764 samples. As shown in Fig. 3, Occ-ScanNet exhibits rich diversity in scenes and viewpoints. This diverse dataset not only challenges the task of predicting scene occupancy but also fosters future research endeavors towards developing larger-scale and more versatile occupancy models."
|
| 872 |
+
},
|
| 873 |
+
{
|
| 874 |
+
"type": "title",
|
| 875 |
+
"bbox": [
|
| 876 |
+
0.216,
|
| 877 |
+
0.683,
|
| 878 |
+
0.504,
|
| 879 |
+
0.699
|
| 880 |
+
],
|
| 881 |
+
"angle": 0,
|
| 882 |
+
"content": "4.2 Occupancy Label Generation"
|
| 883 |
+
},
|
| 884 |
+
{
|
| 885 |
+
"type": "text",
|
| 886 |
+
"bbox": [
|
| 887 |
+
0.214,
|
| 888 |
+
0.705,
|
| 889 |
+
0.787,
|
| 890 |
+
0.764
|
| 891 |
+
],
|
| 892 |
+
"angle": 0,
|
| 893 |
+
"content": "We adhere to the data formulation used in NYUv2 [29] dataset. The pipeline of occupancy generation process is showed in Fig. 4. For generating 3D voxel labels, we initially followed the generation process in CompleteScanNet and then employed manual inspection."
|
| 894 |
+
},
|
| 895 |
+
{
|
| 896 |
+
"type": "text",
|
| 897 |
+
"bbox": [
|
| 898 |
+
0.214,
|
| 899 |
+
0.765,
|
| 900 |
+
0.787,
|
| 901 |
+
0.841
|
| 902 |
+
],
|
| 903 |
+
"angle": 0,
|
| 904 |
+
"content": "More specifically, from the official ScanNet [8] scenes, we extracted various data components, including color images, depth images, camera intrinsic matrices, and camera poses. A total of 100 frames were sampled from each scene and were randomly divided into training and validation sets with a \\(7/3\\) split ratio per scene."
|
| 905 |
+
}
|
| 906 |
+
],
|
| 907 |
+
[
|
| 908 |
+
{
|
| 909 |
+
"type": "page_number",
|
| 910 |
+
"bbox": [
|
| 911 |
+
0.218,
|
| 912 |
+
0.116,
|
| 913 |
+
0.23,
|
| 914 |
+
0.127
|
| 915 |
+
],
|
| 916 |
+
"angle": 0,
|
| 917 |
+
"content": "8"
|
| 918 |
+
},
|
| 919 |
+
{
|
| 920 |
+
"type": "header",
|
| 921 |
+
"bbox": [
|
| 922 |
+
0.272,
|
| 923 |
+
0.115,
|
| 924 |
+
0.352,
|
| 925 |
+
0.127
|
| 926 |
+
],
|
| 927 |
+
"angle": 0,
|
| 928 |
+
"content": "H. Yu et al."
|
| 929 |
+
},
|
| 930 |
+
{
|
| 931 |
+
"type": "image",
|
| 932 |
+
"bbox": [
|
| 933 |
+
0.226,
|
| 934 |
+
0.154,
|
| 935 |
+
0.777,
|
| 936 |
+
0.429
|
| 937 |
+
],
|
| 938 |
+
"angle": 0,
|
| 939 |
+
"content": null
|
| 940 |
+
},
|
| 941 |
+
{
|
| 942 |
+
"type": "image_caption",
|
| 943 |
+
"bbox": [
|
| 944 |
+
0.214,
|
| 945 |
+
0.449,
|
| 946 |
+
0.788,
|
| 947 |
+
0.507
|
| 948 |
+
],
|
| 949 |
+
"angle": 0,
|
| 950 |
+
"content": "Fig. 3: Samples Visualization in Occ-ScanNet Benchmark. The original RGB image is shown in column 1,3 and 5, the corresponding scene voxel labels is shown in column 2, 4 and 6. The first two rows are different views from different scenes and the last two rows each is three different views from the same scene."
|
| 951 |
+
},
|
| 952 |
+
{
|
| 953 |
+
"type": "text",
|
| 954 |
+
"bbox": [
|
| 955 |
+
0.214,
|
| 956 |
+
0.536,
|
| 957 |
+
0.788,
|
| 958 |
+
0.584
|
| 959 |
+
],
|
| 960 |
+
"angle": 0,
|
| 961 |
+
"content": "Manual inspection. The manual inspection involves verifying camera poses, camera intrinsic, and voxel positions. We filter out erroneous samples to prevent convergence issues in the model training process."
|
| 962 |
+
},
|
| 963 |
+
{
|
| 964 |
+
"type": "text",
|
| 965 |
+
"bbox": [
|
| 966 |
+
0.214,
|
| 967 |
+
0.605,
|
| 968 |
+
0.788,
|
| 969 |
+
0.727
|
| 970 |
+
],
|
| 971 |
+
"angle": 0,
|
| 972 |
+
"content": "Voxel label assignment. For each frame, only a specific area in front of the camera is defined for analysis. Therefore, the selection of the voxel origin is paramount, as it determines the subsequent coordinates of other voxel. Subsequently, each voxel is assigned a label based on its nearest voxel in the CompleteScanNet dataset. Additionally, we excluded frames with a ratio of unknown or empty labels exceeding \\(95\\%\\) or frames where the number of semantic label classes was less than two. Consequently, we obtained the generated 3D voxel labels for each frame."
|
| 973 |
+
},
|
| 974 |
+
{
|
| 975 |
+
"type": "title",
|
| 976 |
+
"bbox": [
|
| 977 |
+
0.216,
|
| 978 |
+
0.751,
|
| 979 |
+
0.378,
|
| 980 |
+
0.768
|
| 981 |
+
],
|
| 982 |
+
"angle": 0,
|
| 983 |
+
"content": "5 Experiments"
|
| 984 |
+
},
|
| 985 |
+
{
|
| 986 |
+
"type": "title",
|
| 987 |
+
"bbox": [
|
| 988 |
+
0.216,
|
| 989 |
+
0.784,
|
| 990 |
+
0.312,
|
| 991 |
+
0.8
|
| 992 |
+
],
|
| 993 |
+
"angle": 0,
|
| 994 |
+
"content": "5.1 Setup"
|
| 995 |
+
},
|
| 996 |
+
{
|
| 997 |
+
"type": "text",
|
| 998 |
+
"bbox": [
|
| 999 |
+
0.214,
|
| 1000 |
+
0.81,
|
| 1001 |
+
0.787,
|
| 1002 |
+
0.842
|
| 1003 |
+
],
|
| 1004 |
+
"angle": 0,
|
| 1005 |
+
"content": "Datasets. The NYUv2 [29] dataset provides scenes represented in \\( 240 \\times 144 \\times 240 \\) voxels grids, labeled with 13 classes, including 1 for free space, 1 for unknown,"
|
| 1006 |
+
}
|
| 1007 |
+
],
|
| 1008 |
+
[
|
| 1009 |
+
{
|
| 1010 |
+
"type": "header",
|
| 1011 |
+
"bbox": [
|
| 1012 |
+
0.33,
|
| 1013 |
+
0.115,
|
| 1014 |
+
0.732,
|
| 1015 |
+
0.129
|
| 1016 |
+
],
|
| 1017 |
+
"angle": 0,
|
| 1018 |
+
"content": "Monocular Occupancy Prediction for Scalable Indoor Scenes"
|
| 1019 |
+
},
|
| 1020 |
+
{
|
| 1021 |
+
"type": "page_number",
|
| 1022 |
+
"bbox": [
|
| 1023 |
+
0.775,
|
| 1024 |
+
0.117,
|
| 1025 |
+
0.785,
|
| 1026 |
+
0.127
|
| 1027 |
+
],
|
| 1028 |
+
"angle": 0,
|
| 1029 |
+
"content": "9"
|
| 1030 |
+
},
|
| 1031 |
+
{
|
| 1032 |
+
"type": "image",
|
| 1033 |
+
"bbox": [
|
| 1034 |
+
0.223,
|
| 1035 |
+
0.151,
|
| 1036 |
+
0.778,
|
| 1037 |
+
0.202
|
| 1038 |
+
],
|
| 1039 |
+
"angle": 0,
|
| 1040 |
+
"content": null
|
| 1041 |
+
},
|
| 1042 |
+
{
|
| 1043 |
+
"type": "image_caption",
|
| 1044 |
+
"bbox": [
|
| 1045 |
+
0.214,
|
| 1046 |
+
0.222,
|
| 1047 |
+
0.788,
|
| 1048 |
+
0.333
|
| 1049 |
+
],
|
| 1050 |
+
"angle": 0,
|
| 1051 |
+
"content": "Fig. 4: Pipeline of Occ-ScanNet dataset label generation. Color images, depth images, camera intrinsic and poses are extracted from ScanNet scenes. For each scene, 100 frames were sampled and randomly split into training and validation sets with a \\(7/3\\) ratio. Frames with invalid camera poses or exceeding scene boundaries were filtered out. Only the area in front of the camera was analyzed, necessitating careful selection of the voxel origin. Voxel were labeled based on their nearest voxel in the CompleteScanNet dataset. Frames with \\(>95\\%\\) unknown/empty labels or \\(<2\\) semantic classes were excluded, resulting in generated 3D voxel labels for each frame."
|
| 1052 |
+
},
|
| 1053 |
+
{
|
| 1054 |
+
"type": "text",
|
| 1055 |
+
"bbox": [
|
| 1056 |
+
0.214,
|
| 1057 |
+
0.37,
|
| 1058 |
+
0.788,
|
| 1059 |
+
0.43
|
| 1060 |
+
],
|
| 1061 |
+
"angle": 0,
|
| 1062 |
+
"content": "and 11 for specific semantics (ceiling, floor, wall, window, chair, bed, sofa, table, tvs, furniture, objects). The dataset consists of 795 / 654 scenes in the train / test splits. The model is trained and evaluated on down-sampled \\(60 \\times 36 \\times 60\\) voxels."
|
| 1063 |
+
},
|
| 1064 |
+
{
|
| 1065 |
+
"type": "text",
|
| 1066 |
+
"bbox": [
|
| 1067 |
+
0.214,
|
| 1068 |
+
0.434,
|
| 1069 |
+
0.788,
|
| 1070 |
+
0.51
|
| 1071 |
+
],
|
| 1072 |
+
"angle": 0,
|
| 1073 |
+
"content": "Occ-ScanNet dataset provides scenes represented in \\(60 \\times 60 \\times 36\\) voxel grids, labeled with 12 classes including 1 for free space, and 11 for specific semantics (ceiling, floor, wall, window, chair, bed, sofa, table, tvs, furniture, objects). The dataset comprises 45,755 / 19,764 frames in the train / val splits. The model is trained and evaluated on the original scale."
|
| 1074 |
+
},
|
| 1075 |
+
{
|
| 1076 |
+
"type": "text",
|
| 1077 |
+
"bbox": [
|
| 1078 |
+
0.214,
|
| 1079 |
+
0.513,
|
| 1080 |
+
0.788,
|
| 1081 |
+
0.573
|
| 1082 |
+
],
|
| 1083 |
+
"angle": 0,
|
| 1084 |
+
"content": "Occ-ScanNet-mini dataset has the same class number and voxel scene size as Occ-ScanNet, except that this mini dataset consists of 4,639 / 2,007 frames in the train / val splits. The model is also trained and evaluated on the original scale."
|
| 1085 |
+
},
|
| 1086 |
+
{
|
| 1087 |
+
"type": "text",
|
| 1088 |
+
"bbox": [
|
| 1089 |
+
0.214,
|
| 1090 |
+
0.614,
|
| 1091 |
+
0.789,
|
| 1092 |
+
0.84
|
| 1093 |
+
],
|
| 1094 |
+
"angle": 0,
|
| 1095 |
+
"content": "Implementation Details. We employ a pre-trained EfficientNet-B7 [31] as the encoder in our 2D UNet architecture. In the depth branch, we utilize a pre-trained DepthAnything model [44], which remains frozen during training. Additionally, we integrate a depth loss specific to the depth branch, complementing the other losses outlined in Monoscene [5]. Our model is trained on two datasets: NYUv2 [29] and Occ-ScanNet. For NYUv2, ISO undergoes 30 epochs of training using AdamW optimizers. Initially, the learning rate is set to 5e-6 for the DepthNet and 1e-4 for other components, with a learning rate reduction by 0.1 at epoch 20. Training on NYUv2 takes approximately 7 hours using 2 NVIDIA L20 GPUs (2 items per GPU). On the Occ-ScanNet-mini dataset, ISO is trained for 60 epochs under similar learning rate settings, decreasing the rate by 0.1 at epoch 40. This training process takes around 12 hours using 8 L20 GPUs. For Occ-ScanNet, ISO is trained for 10 epochs with the same learning rate schedule as used for Occ-ScanNet-mini, requiring approximately a day using 8 L20 GPUs."
|
| 1096 |
+
}
|
| 1097 |
+
],
|
| 1098 |
+
[
|
| 1099 |
+
{
|
| 1100 |
+
"type": "page_number",
|
| 1101 |
+
"bbox": [
|
| 1102 |
+
0.218,
|
| 1103 |
+
0.116,
|
| 1104 |
+
0.236,
|
| 1105 |
+
0.127
|
| 1106 |
+
],
|
| 1107 |
+
"angle": 0,
|
| 1108 |
+
"content": "10"
|
| 1109 |
+
},
|
| 1110 |
+
{
|
| 1111 |
+
"type": "header",
|
| 1112 |
+
"bbox": [
|
| 1113 |
+
0.272,
|
| 1114 |
+
0.115,
|
| 1115 |
+
0.352,
|
| 1116 |
+
0.127
|
| 1117 |
+
],
|
| 1118 |
+
"angle": 0,
|
| 1119 |
+
"content": "H. Yu et al."
|
| 1120 |
+
},
|
| 1121 |
+
{
|
| 1122 |
+
"type": "title",
|
| 1123 |
+
"bbox": [
|
| 1124 |
+
0.216,
|
| 1125 |
+
0.147,
|
| 1126 |
+
0.373,
|
| 1127 |
+
0.16
|
| 1128 |
+
],
|
| 1129 |
+
"angle": 0,
|
| 1130 |
+
"content": "5.2 Main Results"
|
| 1131 |
+
},
|
| 1132 |
+
{
|
| 1133 |
+
"type": "table_caption",
|
| 1134 |
+
"bbox": [
|
| 1135 |
+
0.33,
|
| 1136 |
+
0.202,
|
| 1137 |
+
0.674,
|
| 1138 |
+
0.216
|
| 1139 |
+
],
|
| 1140 |
+
"angle": 0,
|
| 1141 |
+
"content": "Table 1: Performance on the Occ-ScanNet dataset"
|
| 1142 |
+
},
|
| 1143 |
+
{
|
| 1144 |
+
"type": "table",
|
| 1145 |
+
"bbox": [
|
| 1146 |
+
0.229,
|
| 1147 |
+
0.228,
|
| 1148 |
+
0.773,
|
| 1149 |
+
0.339
|
| 1150 |
+
],
|
| 1151 |
+
"angle": 0,
|
| 1152 |
+
"content": "<table><tr><td>Method</td><td>Input</td><td>IoU</td><td>ceiling</td><td>floor</td><td>wall</td><td>window</td><td>chair</td><td>bed</td><td>sofa</td><td>table</td><td>tvs</td><td>furniture</td><td>objects</td></tr><tr><td>MonoScene* [5]</td><td>xrgb</td><td>41.60</td><td>15.17</td><td>44.71</td><td>22.41</td><td>12.55</td><td>26.11</td><td>27.03</td><td>35.91</td><td>28.32</td><td>6.57</td><td>32.16</td><td>19.84</td></tr><tr><td>ISO(Ours)</td><td>xrgb</td><td>42.16</td><td>19.88</td><td>41.88</td><td>22.37</td><td>16.98</td><td>29.09</td><td>42.43</td><td>42.00</td><td>29.60</td><td>10.62</td><td>36.36</td><td>24.61</td></tr></table>"
|
| 1153 |
+
},
|
| 1154 |
+
{
|
| 1155 |
+
"type": "text",
|
| 1156 |
+
"bbox": [
|
| 1157 |
+
0.214,
|
| 1158 |
+
0.397,
|
| 1159 |
+
0.789,
|
| 1160 |
+
0.473
|
| 1161 |
+
],
|
| 1162 |
+
"angle": 0,
|
| 1163 |
+
"content": "Occ-ScanNet performance. We first evaluate our model's performance on the large-scale Occ-ScanNet dataset. As shown in Tab. 1, the results indicate that our method significantly outperforms MonoScene [5]. The * denotes results obtained using their code trained on our dataset. In Fig. 5, we also conducted qualitative visualization comparisons."
|
| 1164 |
+
},
|
| 1165 |
+
{
|
| 1166 |
+
"type": "text",
|
| 1167 |
+
"bbox": [
|
| 1168 |
+
0.214,
|
| 1169 |
+
0.5,
|
| 1170 |
+
0.79,
|
| 1171 |
+
0.636
|
| 1172 |
+
],
|
| 1173 |
+
"angle": 0,
|
| 1174 |
+
"content": "NYUv2 performance. The NYUv2 [29] dataset serves as a widely used benchmark for indoor scene evaluation. The results in Tab. 2 demonstrate that our method achieves state-of-the-art performance. Our method also demonstrates substantial performance improvement on the NYUv2 dataset, in addition to the gains observed on Occ-ScanNet. As illustrated in Fig. 6, our ISO demonstrates superior room layout prediction compared to [5], particularly evident in rows 1 and 2. This improvement can be attributed to incorporating depth knowledge into the model. Additionally, our method exhibits the capability of learning small objects, such as pictures on the wall in row 3."
|
| 1175 |
+
},
|
| 1176 |
+
{
|
| 1177 |
+
"type": "table_caption",
|
| 1178 |
+
"bbox": [
|
| 1179 |
+
0.348,
|
| 1180 |
+
0.664,
|
| 1181 |
+
0.656,
|
| 1182 |
+
0.677
|
| 1183 |
+
],
|
| 1184 |
+
"angle": 0,
|
| 1185 |
+
"content": "Table 2: Performance on the NYUv2 dataset"
|
| 1186 |
+
},
|
| 1187 |
+
{
|
| 1188 |
+
"type": "table",
|
| 1189 |
+
"bbox": [
|
| 1190 |
+
0.216,
|
| 1191 |
+
0.69,
|
| 1192 |
+
0.794,
|
| 1193 |
+
0.81
|
| 1194 |
+
],
|
| 1195 |
+
"angle": 0,
|
| 1196 |
+
"content": "<table><tr><td rowspan=\"2\">Method</td><td rowspan=\"2\">Input</td><td rowspan=\"2\">IoU</td><td colspan=\"2\">ceiling</td><td colspan=\"2\">floor</td><td colspan=\"2\">wall</td><td colspan=\"2\">window</td><td colspan=\"2\">chair</td><td colspan=\"2\">bed</td><td colspan=\"2\">sofa</td><td colspan=\"2\">table</td><td colspan=\"2\">tvs</td><td colspan=\"2\">furniture</td><td colspan=\"2\">objects</td><td></td></tr><tr><td colspan=\"2\">■</td><td colspan=\"2\">■</td><td colspan=\"2\">■</td><td colspan=\"2\">■</td><td colspan=\"2\">■</td><td colspan=\"2\">■</td><td colspan=\"2\">■</td><td colspan=\"2\">■</td><td colspan=\"2\">■</td><td colspan=\"2\">■</td><td colspan=\"2\">■</td><td></td></tr><tr><td>LMSCNetrgb [28]</td><td>xocc</td><td>33.93</td><td colspan=\"2\">4.49</td><td colspan=\"2\">88.41</td><td colspan=\"2\">4.63</td><td colspan=\"2\">0.25</td><td colspan=\"2\">3.94</td><td colspan=\"2\">32.03</td><td colspan=\"2\">15.44</td><td colspan=\"2\">6.57</td><td colspan=\"2\">0.02</td><td colspan=\"2\">14.51</td><td colspan=\"2\">4.39</td><td>15.88</td></tr><tr><td>AICNetrgb [18]</td><td>xrgb, xdepth</td><td>30.03</td><td colspan=\"2\">7.58</td><td colspan=\"2\">82.97</td><td colspan=\"2\">9.15</td><td colspan=\"2\">0.05</td><td colspan=\"2\">6.93</td><td colspan=\"2\">35.87</td><td colspan=\"2\">22.92</td><td colspan=\"2\">11.11</td><td colspan=\"2\">0.71</td><td colspan=\"2\">15.90</td><td colspan=\"2\">6.45</td><td>18.15</td></tr><tr><td>3DSketchrgb [6]</td><td>xrgb, xTSDF</td><td>38.64</td><td colspan=\"2\">8.53</td><td colspan=\"2\">90.45</td><td colspan=\"2\">9.94</td><td colspan=\"2\">5.67</td><td colspan=\"2\">10.64</td><td colspan=\"2\">42.29</td><td colspan=\"2\">29.21</td><td colspan=\"2\">13.88</td><td colspan=\"2\">9.38</td><td colspan=\"2\">23.83</td><td colspan=\"2\">8.19</td><td>22.91</td></tr><tr><td>MonoScene [5]</td><td>xrgb</td><td>42.51</td><td colspan=\"2\">8.89</td><td colspan=\"2\">93.50</td><td colspan=\"2\">12.06</td><td colspan=\"2\">12.57</td><td colspan=\"2\">13.72</td><td colspan=\"2\">48.19</td><td colspan=\"2\">36.11</td><td colspan=\"2\">15.13</td><td colspan=\"2\">15.22</td><td colspan=\"2\">27.96</td><td colspan=\"2\">12.94</td><td>26.94</td></tr><tr><td>NDC-Scene [45]</td><td>xrgb</td><td>44.17</td><td colspan=\"2\">12.02</td><td colspan=\"2\">93.51</td><td colspan=\"2\">13.11</td><td colspan=\"2\">13.77</td><td colspan=\"2\">15.83</td><td colspan=\"2\">49.57</td><td colspan=\"2\">39.87</td><td colspan=\"2\">17.17</td><td colspan=\"2\">24.57</td><td colspan=\"2\">31.00</td><td colspan=\"2\">14.96</td><td>29.03</td></tr><tr><td>ISO(Ours)</td><td>xrgb</td><td>47.11</td><td colspan=\"2\">14.21</td><td colspan=\"2\">93.47</td><td colspan=\"2\">15.89</td><td colspan=\"2\">15.14</td><td colspan=\"2\">18.35</td><td colspan=\"2\">50.01</td><td colspan=\"2\">40.82</td><td colspan=\"2\">18.25</td><td colspan=\"2\">25.90</td><td colspan=\"2\">34.08</td><td colspan=\"2\">17.67</td><td>31.25</td></tr></table>"
|
| 1197 |
+
}
|
| 1198 |
+
],
|
| 1199 |
+
[
|
| 1200 |
+
{
|
| 1201 |
+
"type": "header",
|
| 1202 |
+
"bbox": [
|
| 1203 |
+
0.33,
|
| 1204 |
+
0.115,
|
| 1205 |
+
0.732,
|
| 1206 |
+
0.129
|
| 1207 |
+
],
|
| 1208 |
+
"angle": 0,
|
| 1209 |
+
"content": "Monocular Occupancy Prediction for Scalable Indoor Scenes"
|
| 1210 |
+
},
|
| 1211 |
+
{
|
| 1212 |
+
"type": "page_number",
|
| 1213 |
+
"bbox": [
|
| 1214 |
+
0.768,
|
| 1215 |
+
0.116,
|
| 1216 |
+
0.784,
|
| 1217 |
+
0.127
|
| 1218 |
+
],
|
| 1219 |
+
"angle": 0,
|
| 1220 |
+
"content": "11"
|
| 1221 |
+
},
|
| 1222 |
+
{
|
| 1223 |
+
"type": "image_caption",
|
| 1224 |
+
"bbox": [
|
| 1225 |
+
0.279,
|
| 1226 |
+
0.154,
|
| 1227 |
+
0.327,
|
| 1228 |
+
0.168
|
| 1229 |
+
],
|
| 1230 |
+
"angle": 0,
|
| 1231 |
+
"content": "Image"
|
| 1232 |
+
},
|
| 1233 |
+
{
|
| 1234 |
+
"type": "image",
|
| 1235 |
+
"bbox": [
|
| 1236 |
+
0.245,
|
| 1237 |
+
0.175,
|
| 1238 |
+
0.372,
|
| 1239 |
+
0.25
|
| 1240 |
+
],
|
| 1241 |
+
"angle": 0,
|
| 1242 |
+
"content": null
|
| 1243 |
+
},
|
| 1244 |
+
{
|
| 1245 |
+
"type": "image_caption",
|
| 1246 |
+
"bbox": [
|
| 1247 |
+
0.39,
|
| 1248 |
+
0.154,
|
| 1249 |
+
0.475,
|
| 1250 |
+
0.167
|
| 1251 |
+
],
|
| 1252 |
+
"angle": 0,
|
| 1253 |
+
"content": "MonoScene"
|
| 1254 |
+
},
|
| 1255 |
+
{
|
| 1256 |
+
"type": "image",
|
| 1257 |
+
"bbox": [
|
| 1258 |
+
0.379,
|
| 1259 |
+
0.177,
|
| 1260 |
+
0.492,
|
| 1261 |
+
0.248
|
| 1262 |
+
],
|
| 1263 |
+
"angle": 0,
|
| 1264 |
+
"content": null
|
| 1265 |
+
},
|
| 1266 |
+
{
|
| 1267 |
+
"type": "image_caption",
|
| 1268 |
+
"bbox": [
|
| 1269 |
+
0.538,
|
| 1270 |
+
0.154,
|
| 1271 |
+
0.574,
|
| 1272 |
+
0.166
|
| 1273 |
+
],
|
| 1274 |
+
"angle": 0,
|
| 1275 |
+
"content": "Ours"
|
| 1276 |
+
},
|
| 1277 |
+
{
|
| 1278 |
+
"type": "image",
|
| 1279 |
+
"bbox": [
|
| 1280 |
+
0.508,
|
| 1281 |
+
0.176,
|
| 1282 |
+
0.623,
|
| 1283 |
+
0.248
|
| 1284 |
+
],
|
| 1285 |
+
"angle": 0,
|
| 1286 |
+
"content": null
|
| 1287 |
+
},
|
| 1288 |
+
{
|
| 1289 |
+
"type": "image_caption",
|
| 1290 |
+
"bbox": [
|
| 1291 |
+
0.638,
|
| 1292 |
+
0.153,
|
| 1293 |
+
0.736,
|
| 1294 |
+
0.167
|
| 1295 |
+
],
|
| 1296 |
+
"angle": 0,
|
| 1297 |
+
"content": "Ground Truth"
|
| 1298 |
+
},
|
| 1299 |
+
{
|
| 1300 |
+
"type": "image",
|
| 1301 |
+
"bbox": [
|
| 1302 |
+
0.635,
|
| 1303 |
+
0.177,
|
| 1304 |
+
0.75,
|
| 1305 |
+
0.248
|
| 1306 |
+
],
|
| 1307 |
+
"angle": 0,
|
| 1308 |
+
"content": null
|
| 1309 |
+
},
|
| 1310 |
+
{
|
| 1311 |
+
"type": "image",
|
| 1312 |
+
"bbox": [
|
| 1313 |
+
0.245,
|
| 1314 |
+
0.271,
|
| 1315 |
+
0.372,
|
| 1316 |
+
0.346
|
| 1317 |
+
],
|
| 1318 |
+
"angle": 0,
|
| 1319 |
+
"content": null
|
| 1320 |
+
},
|
| 1321 |
+
{
|
| 1322 |
+
"type": "image",
|
| 1323 |
+
"bbox": [
|
| 1324 |
+
0.38,
|
| 1325 |
+
0.273,
|
| 1326 |
+
0.491,
|
| 1327 |
+
0.346
|
| 1328 |
+
],
|
| 1329 |
+
"angle": 0,
|
| 1330 |
+
"content": null
|
| 1331 |
+
},
|
| 1332 |
+
{
|
| 1333 |
+
"type": "image",
|
| 1334 |
+
"bbox": [
|
| 1335 |
+
0.505,
|
| 1336 |
+
0.274,
|
| 1337 |
+
0.624,
|
| 1338 |
+
0.344
|
| 1339 |
+
],
|
| 1340 |
+
"angle": 0,
|
| 1341 |
+
"content": null
|
| 1342 |
+
},
|
| 1343 |
+
{
|
| 1344 |
+
"type": "image",
|
| 1345 |
+
"bbox": [
|
| 1346 |
+
0.632,
|
| 1347 |
+
0.275,
|
| 1348 |
+
0.752,
|
| 1349 |
+
0.343
|
| 1350 |
+
],
|
| 1351 |
+
"angle": 0,
|
| 1352 |
+
"content": null
|
| 1353 |
+
},
|
| 1354 |
+
{
|
| 1355 |
+
"type": "image",
|
| 1356 |
+
"bbox": [
|
| 1357 |
+
0.245,
|
| 1358 |
+
0.367,
|
| 1359 |
+
0.372,
|
| 1360 |
+
0.442
|
| 1361 |
+
],
|
| 1362 |
+
"angle": 0,
|
| 1363 |
+
"content": null
|
| 1364 |
+
},
|
| 1365 |
+
{
|
| 1366 |
+
"type": "image",
|
| 1367 |
+
"bbox": [
|
| 1368 |
+
0.376,
|
| 1369 |
+
0.372,
|
| 1370 |
+
0.495,
|
| 1371 |
+
0.436
|
| 1372 |
+
],
|
| 1373 |
+
"angle": 0,
|
| 1374 |
+
"content": null
|
| 1375 |
+
},
|
| 1376 |
+
{
|
| 1377 |
+
"type": "image",
|
| 1378 |
+
"bbox": [
|
| 1379 |
+
0.504,
|
| 1380 |
+
0.372,
|
| 1381 |
+
0.626,
|
| 1382 |
+
0.437
|
| 1383 |
+
],
|
| 1384 |
+
"angle": 0,
|
| 1385 |
+
"content": null
|
| 1386 |
+
},
|
| 1387 |
+
{
|
| 1388 |
+
"type": "image",
|
| 1389 |
+
"bbox": [
|
| 1390 |
+
0.633,
|
| 1391 |
+
0.376,
|
| 1392 |
+
0.752,
|
| 1393 |
+
0.436
|
| 1394 |
+
],
|
| 1395 |
+
"angle": 0,
|
| 1396 |
+
"content": null
|
| 1397 |
+
},
|
| 1398 |
+
{
|
| 1399 |
+
"type": "image",
|
| 1400 |
+
"bbox": [
|
| 1401 |
+
0.245,
|
| 1402 |
+
0.462,
|
| 1403 |
+
0.372,
|
| 1404 |
+
0.537
|
| 1405 |
+
],
|
| 1406 |
+
"angle": 0,
|
| 1407 |
+
"content": null
|
| 1408 |
+
},
|
| 1409 |
+
{
|
| 1410 |
+
"type": "image",
|
| 1411 |
+
"bbox": [
|
| 1412 |
+
0.378,
|
| 1413 |
+
0.468,
|
| 1414 |
+
0.497,
|
| 1415 |
+
0.533
|
| 1416 |
+
],
|
| 1417 |
+
"angle": 0,
|
| 1418 |
+
"content": null
|
| 1419 |
+
},
|
| 1420 |
+
{
|
| 1421 |
+
"type": "image",
|
| 1422 |
+
"bbox": [
|
| 1423 |
+
0.506,
|
| 1424 |
+
0.462,
|
| 1425 |
+
0.626,
|
| 1426 |
+
0.54
|
| 1427 |
+
],
|
| 1428 |
+
"angle": 0,
|
| 1429 |
+
"content": null
|
| 1430 |
+
},
|
| 1431 |
+
{
|
| 1432 |
+
"type": "image",
|
| 1433 |
+
"bbox": [
|
| 1434 |
+
0.633,
|
| 1435 |
+
0.462,
|
| 1436 |
+
0.752,
|
| 1437 |
+
0.538
|
| 1438 |
+
],
|
| 1439 |
+
"angle": 0,
|
| 1440 |
+
"content": null
|
| 1441 |
+
},
|
| 1442 |
+
{
|
| 1443 |
+
"type": "image_caption",
|
| 1444 |
+
"bbox": [
|
| 1445 |
+
0.214,
|
| 1446 |
+
0.574,
|
| 1447 |
+
0.788,
|
| 1448 |
+
0.618
|
| 1449 |
+
],
|
| 1450 |
+
"angle": 0,
|
| 1451 |
+
"content": "Fig. 5: Qualitative Analysis on the Occ-ScanNet Dataset. The input image is displayed on the left, while the predicted scene is shown in the middle two columns, and ground truth on the right column."
|
| 1452 |
+
},
|
| 1453 |
+
{
|
| 1454 |
+
"type": "title",
|
| 1455 |
+
"bbox": [
|
| 1456 |
+
0.215,
|
| 1457 |
+
0.645,
|
| 1458 |
+
0.391,
|
| 1459 |
+
0.66
|
| 1460 |
+
],
|
| 1461 |
+
"angle": 0,
|
| 1462 |
+
"content": "5.3 Ablation Study"
|
| 1463 |
+
},
|
| 1464 |
+
{
|
| 1465 |
+
"type": "text",
|
| 1466 |
+
"bbox": [
|
| 1467 |
+
0.214,
|
| 1468 |
+
0.668,
|
| 1469 |
+
0.788,
|
| 1470 |
+
0.699
|
| 1471 |
+
],
|
| 1472 |
+
"angle": 0,
|
| 1473 |
+
"content": "We conduct ablation studies on the NYUv2 [29] and Occ-ScanNet-mini datasets to validate the effectiveness of our model design."
|
| 1474 |
+
},
|
| 1475 |
+
{
|
| 1476 |
+
"type": "text",
|
| 1477 |
+
"bbox": [
|
| 1478 |
+
0.214,
|
| 1479 |
+
0.719,
|
| 1480 |
+
0.789,
|
| 1481 |
+
0.841
|
| 1482 |
+
],
|
| 1483 |
+
"angle": 0,
|
| 1484 |
+
"content": "Depth fusion. Inspired by BEVDepth [20], we initially adopt the integration of BEV depth information, denoted as \\( ^{+} \\)bev-depth'. However, we observed that while BEV depth works well for 3D detection in outdoor scenes, directly applying it to indoor scenes is less effective. We speculate that this discrepancy arises because height information is more crucial in indoor scenes, where indoor heights often correspond to different 3D structures and semantics. Therefore, we adopt voxel depth fusion, denoted as \\( ^{+} \\)voxel-depth'. Specifically, we project each 3D voxel centroid to a frustum grid to sample the depth distribution probability"
|
| 1485 |
+
}
|
| 1486 |
+
],
|
| 1487 |
+
[
|
| 1488 |
+
{
|
| 1489 |
+
"type": "page_number",
|
| 1490 |
+
"bbox": [
|
| 1491 |
+
0.218,
|
| 1492 |
+
0.116,
|
| 1493 |
+
0.236,
|
| 1494 |
+
0.127
|
| 1495 |
+
],
|
| 1496 |
+
"angle": 0,
|
| 1497 |
+
"content": "12"
|
| 1498 |
+
},
|
| 1499 |
+
{
|
| 1500 |
+
"type": "header",
|
| 1501 |
+
"bbox": [
|
| 1502 |
+
0.272,
|
| 1503 |
+
0.115,
|
| 1504 |
+
0.352,
|
| 1505 |
+
0.127
|
| 1506 |
+
],
|
| 1507 |
+
"angle": 0,
|
| 1508 |
+
"content": "H. Yu et al."
|
| 1509 |
+
},
|
| 1510 |
+
{
|
| 1511 |
+
"type": "image",
|
| 1512 |
+
"bbox": [
|
| 1513 |
+
0.239,
|
| 1514 |
+
0.149,
|
| 1515 |
+
0.765,
|
| 1516 |
+
0.379
|
| 1517 |
+
],
|
| 1518 |
+
"angle": 0,
|
| 1519 |
+
"content": null
|
| 1520 |
+
},
|
| 1521 |
+
{
|
| 1522 |
+
"type": "image_caption",
|
| 1523 |
+
"bbox": [
|
| 1524 |
+
0.214,
|
| 1525 |
+
0.395,
|
| 1526 |
+
0.788,
|
| 1527 |
+
0.466
|
| 1528 |
+
],
|
| 1529 |
+
"angle": 0,
|
| 1530 |
+
"content": "Fig. 6: Qualitative Analysis on the NYUv2 Dataset: The input is displayed on the left, with the camera viewing frustum shown in every image. ISO adeptly captures the scene layout and delineates various semantic instances more accurately. Notably, it excels in reconstructing the corners of walls (row 2) and accurately representing small objects like pictures on the wall (row 3)."
|
| 1531 |
+
},
|
| 1532 |
+
{
|
| 1533 |
+
"type": "text",
|
| 1534 |
+
"bbox": [
|
| 1535 |
+
0.214,
|
| 1536 |
+
0.497,
|
| 1537 |
+
0.785,
|
| 1538 |
+
0.542
|
| 1539 |
+
],
|
| 1540 |
+
"angle": 0,
|
| 1541 |
+
"content": "of the voxel. We then weight the 2D feature at that pixel coordinate as the voxel feature. The results in Tab. 3 further validate that the voxel depth fusion approach is more effective."
|
| 1542 |
+
},
|
| 1543 |
+
{
|
| 1544 |
+
"type": "text",
|
| 1545 |
+
"bbox": [
|
| 1546 |
+
0.214,
|
| 1547 |
+
0.544,
|
| 1548 |
+
0.788,
|
| 1549 |
+
0.664
|
| 1550 |
+
],
|
| 1551 |
+
"angle": 0,
|
| 1552 |
+
"content": "The BEV method usually performs better in detecting ceilings and floors but performs poorly in detecting furniture such as windows, walls, televisions, beds, sofas, and tables. Ceilings and floors usually have relatively flat and regular geometric shapes, which makes their projections clearer and more consistent from a BEV perspective. In contrast, objects such as windows, walls, and televisions typically have more complex geometric shapes. In the BEV perspective, the projections of these objects may become blurry or overlapping, making it difficult to accurately distinguish and detect."
|
| 1553 |
+
},
|
| 1554 |
+
{
|
| 1555 |
+
"type": "table_caption",
|
| 1556 |
+
"bbox": [
|
| 1557 |
+
0.216,
|
| 1558 |
+
0.694,
|
| 1559 |
+
0.787,
|
| 1560 |
+
0.709
|
| 1561 |
+
],
|
| 1562 |
+
"angle": 0,
|
| 1563 |
+
"content": "Table 3: Ablation study of depth fusion on the NYUv2 and Occ-ScanNet-mini dataset"
|
| 1564 |
+
},
|
| 1565 |
+
{
|
| 1566 |
+
"type": "table",
|
| 1567 |
+
"bbox": [
|
| 1568 |
+
0.308,
|
| 1569 |
+
0.72,
|
| 1570 |
+
0.698,
|
| 1571 |
+
0.809
|
| 1572 |
+
],
|
| 1573 |
+
"angle": 0,
|
| 1574 |
+
"content": "<table><tr><td rowspan=\"2\">Method</td><td rowspan=\"2\">Input</td><td colspan=\"2\">NYUv2</td><td colspan=\"2\">Occ-ScanNet-mini</td></tr><tr><td>IoU</td><td>mIoU</td><td>IoU</td><td>mIoU</td></tr><tr><td>baseline</td><td>xrgb</td><td>42.27</td><td>27.13</td><td>50.94</td><td>38.95</td></tr><tr><td>+ bev-depth</td><td>xrgb</td><td>42.67</td><td>27.14</td><td>51.58</td><td>38.48</td></tr><tr><td>+ voxel-depth</td><td>xrgb</td><td>47.11</td><td>31.25</td><td>51.03</td><td>39.08</td></tr></table>"
|
| 1575 |
+
}
|
| 1576 |
+
],
|
| 1577 |
+
[
|
| 1578 |
+
{
|
| 1579 |
+
"type": "header",
|
| 1580 |
+
"bbox": [
|
| 1581 |
+
0.33,
|
| 1582 |
+
0.115,
|
| 1583 |
+
0.732,
|
| 1584 |
+
0.129
|
| 1585 |
+
],
|
| 1586 |
+
"angle": 0,
|
| 1587 |
+
"content": "Monocular Occupancy Prediction for Scalable Indoor Scenes"
|
| 1588 |
+
},
|
| 1589 |
+
{
|
| 1590 |
+
"type": "page_number",
|
| 1591 |
+
"bbox": [
|
| 1592 |
+
0.769,
|
| 1593 |
+
0.117,
|
| 1594 |
+
0.785,
|
| 1595 |
+
0.127
|
| 1596 |
+
],
|
| 1597 |
+
"angle": 0,
|
| 1598 |
+
"content": "13"
|
| 1599 |
+
},
|
| 1600 |
+
{
|
| 1601 |
+
"type": "table_caption",
|
| 1602 |
+
"bbox": [
|
| 1603 |
+
0.29,
|
| 1604 |
+
0.145,
|
| 1605 |
+
0.712,
|
| 1606 |
+
0.159
|
| 1607 |
+
],
|
| 1608 |
+
"angle": 0,
|
| 1609 |
+
"content": "Table 4: Ablation of depth information on the NYUv2 dataset"
|
| 1610 |
+
},
|
| 1611 |
+
{
|
| 1612 |
+
"type": "table",
|
| 1613 |
+
"bbox": [
|
| 1614 |
+
0.341,
|
| 1615 |
+
0.171,
|
| 1616 |
+
0.661,
|
| 1617 |
+
0.283
|
| 1618 |
+
],
|
| 1619 |
+
"angle": 0,
|
| 1620 |
+
"content": "<table><tr><td>depth-method</td><td>learned</td><td>multi-scale</td><td>IoU</td><td>mIoU</td></tr><tr><td>GT</td><td></td><td>✓</td><td>53.98</td><td>34.47</td></tr><tr><td>ZoeDepth</td><td>✓</td><td></td><td>45.18</td><td>29.15</td></tr><tr><td>ZoeDepth</td><td>✓</td><td>✓</td><td>45.24</td><td>29.40</td></tr><tr><td>DepthAnything</td><td></td><td></td><td>44.77</td><td>29.28</td></tr><tr><td>DepthAnything</td><td></td><td>✓</td><td>45.48</td><td>29.57</td></tr><tr><td>DepthAnything</td><td>✓</td><td>✓</td><td>46.94</td><td>31.02</td></tr><tr><td>DepthAnything</td><td>✓</td><td></td><td>47.11</td><td>31.25</td></tr></table>"
|
| 1621 |
+
},
|
| 1622 |
+
{
|
| 1623 |
+
"type": "text",
|
| 1624 |
+
"bbox": [
|
| 1625 |
+
0.216,
|
| 1626 |
+
0.316,
|
| 1627 |
+
0.787,
|
| 1628 |
+
0.438
|
| 1629 |
+
],
|
| 1630 |
+
"angle": 0,
|
| 1631 |
+
"content": "Different depth pre-trained models. We conducted ablation experiments on the depth method using the Occ-ScanNet-mini dataset. Initially, we observed the upper bound of model performance when ground truth (GT) depth information was provided. Subsequently, we initialized our depth module using ZoeDepth [3] and Depth-Anything [44]. Experimental results indicate that utilizing Depth-Anything [44] yields better performance. Moreover, our approach allows for further enhancement by fine-tuning a pre-trained depth module, resulting in improved performance."
|
| 1632 |
+
},
|
| 1633 |
+
{
|
| 1634 |
+
"type": "text",
|
| 1635 |
+
"bbox": [
|
| 1636 |
+
0.216,
|
| 1637 |
+
0.475,
|
| 1638 |
+
0.787,
|
| 1639 |
+
0.717
|
| 1640 |
+
],
|
| 1641 |
+
"angle": 0,
|
| 1642 |
+
"content": "Multi-scale depth. In Tab. 4, we investigate the impact of using multi-scale depth distribution on the NYUv2 [29] dataset. Specifically, the multi-scale depth distribution is used to weight the multi-scale 3D voxel feature. Meanwhile, we tried a method that only uses the single-scale depth distribution map to generate the 3D voxel depth matrix. The used depth distribution map has a scale of \\(\\frac{1}{8}\\) original image scale. Only this single-scale 3D voxel matrix is used to weigh the summed-up 3D voxel feature out of the FLoSP module. Without multi-scale depth, all 3D features are simply added up and a single depth distribution is used for weighting. This method ignores the different importance that features at different scales may have, as all features at all scales are treated equally and all depth information is weighted by a unified depth distribution. With multiscale depth, the 3D features of each scale are first weighted with the depth distribution of their corresponding scale, and then these weighted features are added up. This method takes into account the differences in the importance of features at different scales, as each scale's feature is weighted based on its corresponding depth distribution."
|
| 1643 |
+
},
|
| 1644 |
+
{
|
| 1645 |
+
"type": "text",
|
| 1646 |
+
"bbox": [
|
| 1647 |
+
0.216,
|
| 1648 |
+
0.72,
|
| 1649 |
+
0.787,
|
| 1650 |
+
0.84
|
| 1651 |
+
],
|
| 1652 |
+
"angle": 0,
|
| 1653 |
+
"content": "In real-world scenarios, various objects and structures (such as ceilings, walls, furniture, etc.) often exhibit different scales. Hence, accounting for scale variations is crucial for accurately modeling these objects. Multi-scale depth can effectively capture these scale changes by assigning different weights to features at each scale. Depth information is indispensable for 3D scene completion. Leveraging the depth distribution information at each scale allows us to weight the corresponding 3D features, enabling the model to better understand and represent the objects in the scene and their relationships."
|
| 1654 |
+
}
|
| 1655 |
+
],
|
| 1656 |
+
[
|
| 1657 |
+
{
|
| 1658 |
+
"type": "page_number",
|
| 1659 |
+
"bbox": [
|
| 1660 |
+
0.218,
|
| 1661 |
+
0.116,
|
| 1662 |
+
0.236,
|
| 1663 |
+
0.127
|
| 1664 |
+
],
|
| 1665 |
+
"angle": 0,
|
| 1666 |
+
"content": "14"
|
| 1667 |
+
},
|
| 1668 |
+
{
|
| 1669 |
+
"type": "header",
|
| 1670 |
+
"bbox": [
|
| 1671 |
+
0.272,
|
| 1672 |
+
0.115,
|
| 1673 |
+
0.352,
|
| 1674 |
+
0.127
|
| 1675 |
+
],
|
| 1676 |
+
"angle": 0,
|
| 1677 |
+
"content": "H. Yu et al."
|
| 1678 |
+
},
|
| 1679 |
+
{
|
| 1680 |
+
"type": "title",
|
| 1681 |
+
"bbox": [
|
| 1682 |
+
0.217,
|
| 1683 |
+
0.146,
|
| 1684 |
+
0.35,
|
| 1685 |
+
0.161
|
| 1686 |
+
],
|
| 1687 |
+
"angle": 0,
|
| 1688 |
+
"content": "5.4 Discussion"
|
| 1689 |
+
},
|
| 1690 |
+
{
|
| 1691 |
+
"type": "text",
|
| 1692 |
+
"bbox": [
|
| 1693 |
+
0.214,
|
| 1694 |
+
0.172,
|
| 1695 |
+
0.788,
|
| 1696 |
+
0.248
|
| 1697 |
+
],
|
| 1698 |
+
"angle": 0,
|
| 1699 |
+
"content": "Data scaling up. In Tab. 5, we delve into the impact of data volume on our method's performance by scaling up the number of scene samples. We trained the model on Occ-ScanNet using \\(10\\%\\) and \\(100\\%\\) scene samples respectively, and tested the model performance. The results indicate that larger dataset exhibits more significant gains."
|
| 1700 |
+
},
|
| 1701 |
+
{
|
| 1702 |
+
"type": "table_caption",
|
| 1703 |
+
"bbox": [
|
| 1704 |
+
0.323,
|
| 1705 |
+
0.274,
|
| 1706 |
+
0.681,
|
| 1707 |
+
0.29
|
| 1708 |
+
],
|
| 1709 |
+
"angle": 0,
|
| 1710 |
+
"content": "Table 5: Scaling up comparison on the Occ-ScanNet"
|
| 1711 |
+
},
|
| 1712 |
+
{
|
| 1713 |
+
"type": "table",
|
| 1714 |
+
"bbox": [
|
| 1715 |
+
0.379,
|
| 1716 |
+
0.302,
|
| 1717 |
+
0.62,
|
| 1718 |
+
0.375
|
| 1719 |
+
],
|
| 1720 |
+
"angle": 0,
|
| 1721 |
+
"content": "<table><tr><td rowspan=\"2\">Data</td><td rowspan=\"2\">Depth Branch</td><td colspan=\"2\">Ours</td></tr><tr><td>IoU</td><td>mIoU</td></tr><tr><td>10%</td><td>✓</td><td>21.79</td><td>9.57</td></tr><tr><td>100%</td><td>✓</td><td>42.16</td><td>28.71</td></tr></table>"
|
| 1722 |
+
},
|
| 1723 |
+
{
|
| 1724 |
+
"type": "text",
|
| 1725 |
+
"bbox": [
|
| 1726 |
+
0.214,
|
| 1727 |
+
0.425,
|
| 1728 |
+
0.788,
|
| 1729 |
+
0.547
|
| 1730 |
+
],
|
| 1731 |
+
"angle": 0,
|
| 1732 |
+
"content": "Limitations and future works. Although our model can estimate 3D occupancy effectively with the assistance of depth information, semantic learning still faces challenges due to class imbalance issues. Furthermore, our proposed Occ-ScanNet only considers 11 common semantic classes, which may not fully capture the diversity of categories present in real-world scenarios. Despite the significant increase in data volume compared to previous datasets, Occ-ScanNet remains limited in the number of scenes it covers. In future work, we will focus more on semantic exploration and conduct tests in more generalized scenarios."
|
| 1733 |
+
},
|
| 1734 |
+
{
|
| 1735 |
+
"type": "title",
|
| 1736 |
+
"bbox": [
|
| 1737 |
+
0.216,
|
| 1738 |
+
0.57,
|
| 1739 |
+
0.36,
|
| 1740 |
+
0.587
|
| 1741 |
+
],
|
| 1742 |
+
"angle": 0,
|
| 1743 |
+
"content": "6 Conclusion"
|
| 1744 |
+
},
|
| 1745 |
+
{
|
| 1746 |
+
"type": "text",
|
| 1747 |
+
"bbox": [
|
| 1748 |
+
0.214,
|
| 1749 |
+
0.602,
|
| 1750 |
+
0.788,
|
| 1751 |
+
0.725
|
| 1752 |
+
],
|
| 1753 |
+
"angle": 0,
|
| 1754 |
+
"content": "In this paper, we introduce ISO, a novel method for monocular 3D occupancy prediction. ISO utilizes a D-FLoSP (Dual Feature Line of Sight Projection) and a multi-scale feature fusion strategy to address variations in scene and object sizes, specifically tailored for indoor scenes. Additionally, we present a new benchmark, Occ-ScanNet, aimed at fostering scalable studies of indoor scenes. We hope that this dataset will attract more attention to research on indoor occupancy prediction. While achieving promising results in 3D structure prediction, accurately identifying semantics remains a significant challenge in the future."
|
| 1755 |
+
},
|
| 1756 |
+
{
|
| 1757 |
+
"type": "title",
|
| 1758 |
+
"bbox": [
|
| 1759 |
+
0.216,
|
| 1760 |
+
0.747,
|
| 1761 |
+
0.395,
|
| 1762 |
+
0.765
|
| 1763 |
+
],
|
| 1764 |
+
"angle": 0,
|
| 1765 |
+
"content": "Acknowledgement"
|
| 1766 |
+
},
|
| 1767 |
+
{
|
| 1768 |
+
"type": "text",
|
| 1769 |
+
"bbox": [
|
| 1770 |
+
0.214,
|
| 1771 |
+
0.78,
|
| 1772 |
+
0.788,
|
| 1773 |
+
0.841
|
| 1774 |
+
],
|
| 1775 |
+
"angle": 0,
|
| 1776 |
+
"content": "This work was supported in part by the National Key R&D Program of China (No. 2022ZD0160102), the National Natural Science Foundation of China (No. U21B2042, No. 62320106010), and in part by the 2035 Innovation Program of CAS, and the InnoHK program."
|
| 1777 |
+
}
|
| 1778 |
+
],
|
| 1779 |
+
[
|
| 1780 |
+
{
|
| 1781 |
+
"type": "header",
|
| 1782 |
+
"bbox": [
|
| 1783 |
+
0.329,
|
| 1784 |
+
0.115,
|
| 1785 |
+
0.732,
|
| 1786 |
+
0.129
|
| 1787 |
+
],
|
| 1788 |
+
"angle": 0,
|
| 1789 |
+
"content": "Monocular Occupancy Prediction for Scalable Indoor Scenes"
|
| 1790 |
+
},
|
| 1791 |
+
{
|
| 1792 |
+
"type": "page_number",
|
| 1793 |
+
"bbox": [
|
| 1794 |
+
0.768,
|
| 1795 |
+
0.117,
|
| 1796 |
+
0.785,
|
| 1797 |
+
0.127
|
| 1798 |
+
],
|
| 1799 |
+
"angle": 0,
|
| 1800 |
+
"content": "15"
|
| 1801 |
+
},
|
| 1802 |
+
{
|
| 1803 |
+
"type": "title",
|
| 1804 |
+
"bbox": [
|
| 1805 |
+
0.217,
|
| 1806 |
+
0.145,
|
| 1807 |
+
0.323,
|
| 1808 |
+
0.16
|
| 1809 |
+
],
|
| 1810 |
+
"angle": 0,
|
| 1811 |
+
"content": "References"
|
| 1812 |
+
},
|
| 1813 |
+
{
|
| 1814 |
+
"type": "ref_text",
|
| 1815 |
+
"bbox": [
|
| 1816 |
+
0.226,
|
| 1817 |
+
0.179,
|
| 1818 |
+
0.785,
|
| 1819 |
+
0.207
|
| 1820 |
+
],
|
| 1821 |
+
"angle": 0,
|
| 1822 |
+
"content": "1. Arshad, M.S., Beksi, W.J.: List: Learning implicitly from spatial transformers for single-view 3d reconstruction. In: ICCV (2023)"
|
| 1823 |
+
},
|
| 1824 |
+
{
|
| 1825 |
+
"type": "ref_text",
|
| 1826 |
+
"bbox": [
|
| 1827 |
+
0.226,
|
| 1828 |
+
0.209,
|
| 1829 |
+
0.785,
|
| 1830 |
+
0.235
|
| 1831 |
+
],
|
| 1832 |
+
"angle": 0,
|
| 1833 |
+
"content": "2. Azuma, R.T.: A survey of augmented reality. Presence: teleoperators & virtual environments (1997)"
|
| 1834 |
+
},
|
| 1835 |
+
{
|
| 1836 |
+
"type": "ref_text",
|
| 1837 |
+
"bbox": [
|
| 1838 |
+
0.226,
|
| 1839 |
+
0.236,
|
| 1840 |
+
0.785,
|
| 1841 |
+
0.263
|
| 1842 |
+
],
|
| 1843 |
+
"angle": 0,
|
| 1844 |
+
"content": "3. Bhat, S.F., Birkl, R., Wofk, D., Wonka, P., Müller, M.: Zoedepth: Zero-shot transfer by combining relative and metric depth. arXiv preprint arXiv:2302.12288 (2023)"
|
| 1845 |
+
},
|
| 1846 |
+
{
|
| 1847 |
+
"type": "ref_text",
|
| 1848 |
+
"bbox": [
|
| 1849 |
+
0.226,
|
| 1850 |
+
0.265,
|
| 1851 |
+
0.785,
|
| 1852 |
+
0.292
|
| 1853 |
+
],
|
| 1854 |
+
"angle": 0,
|
| 1855 |
+
"content": "4. Birkl, R., Wofk, D., Müller, M.: Midas v3. 1-a model zoo for robust monocular relative depth estimation. arXiv preprint arXiv:2307.14460 (2023)"
|
| 1856 |
+
},
|
| 1857 |
+
{
|
| 1858 |
+
"type": "ref_text",
|
| 1859 |
+
"bbox": [
|
| 1860 |
+
0.226,
|
| 1861 |
+
0.294,
|
| 1862 |
+
0.785,
|
| 1863 |
+
0.32
|
| 1864 |
+
],
|
| 1865 |
+
"angle": 0,
|
| 1866 |
+
"content": "5. Cao, A.Q., de Charette, R.: Monoscene: Monocular 3d semantic scene completion. In: CVPR (2022)"
|
| 1867 |
+
},
|
| 1868 |
+
{
|
| 1869 |
+
"type": "ref_text",
|
| 1870 |
+
"bbox": [
|
| 1871 |
+
0.226,
|
| 1872 |
+
0.321,
|
| 1873 |
+
0.785,
|
| 1874 |
+
0.348
|
| 1875 |
+
],
|
| 1876 |
+
"angle": 0,
|
| 1877 |
+
"content": "6. Chen, X., Lin, K.Y., Qian, C., Zeng, G., Li, H.: 3d sketch-aware semantic scene completion via semi-supervised structure prior. In: CVPR (2020)"
|
| 1878 |
+
},
|
| 1879 |
+
{
|
| 1880 |
+
"type": "ref_text",
|
| 1881 |
+
"bbox": [
|
| 1882 |
+
0.226,
|
| 1883 |
+
0.349,
|
| 1884 |
+
0.785,
|
| 1885 |
+
0.376
|
| 1886 |
+
],
|
| 1887 |
+
"angle": 0,
|
| 1888 |
+
"content": "7. Dahnert, M., Hou, J., Nießner, M., Dai, A.: Panoptic 3d scene reconstruction from a single rgb image. NeurIPS (2021)"
|
| 1889 |
+
},
|
| 1890 |
+
{
|
| 1891 |
+
"type": "ref_text",
|
| 1892 |
+
"bbox": [
|
| 1893 |
+
0.226,
|
| 1894 |
+
0.378,
|
| 1895 |
+
0.785,
|
| 1896 |
+
0.404
|
| 1897 |
+
],
|
| 1898 |
+
"angle": 0,
|
| 1899 |
+
"content": "8. Dai, A., Chang, A.X., Savva, M., Halber, M., Funkhouser, T., Nießner, M.: Scannet: Richly-annotated 3d reconstructions of indoor scenes. In: CVPR (2017)"
|
| 1900 |
+
},
|
| 1901 |
+
{
|
| 1902 |
+
"type": "ref_text",
|
| 1903 |
+
"bbox": [
|
| 1904 |
+
0.226,
|
| 1905 |
+
0.406,
|
| 1906 |
+
0.785,
|
| 1907 |
+
0.432
|
| 1908 |
+
],
|
| 1909 |
+
"angle": 0,
|
| 1910 |
+
"content": "9. Denninger, M., Triebel, R.: 3d scene reconstruction from a single viewpoint. In: ECCV (2020)"
|
| 1911 |
+
},
|
| 1912 |
+
{
|
| 1913 |
+
"type": "ref_text",
|
| 1914 |
+
"bbox": [
|
| 1915 |
+
0.226,
|
| 1916 |
+
0.434,
|
| 1917 |
+
0.785,
|
| 1918 |
+
0.46
|
| 1919 |
+
],
|
| 1920 |
+
"angle": 0,
|
| 1921 |
+
"content": "10. DeSouza, G.N., Kak, A.C.: Vision for mobile robot navigation: A survey. TPAMI (2002)"
|
| 1922 |
+
},
|
| 1923 |
+
{
|
| 1924 |
+
"type": "ref_text",
|
| 1925 |
+
"bbox": [
|
| 1926 |
+
0.226,
|
| 1927 |
+
0.462,
|
| 1928 |
+
0.785,
|
| 1929 |
+
0.489
|
| 1930 |
+
],
|
| 1931 |
+
"angle": 0,
|
| 1932 |
+
"content": "1. Fan, H., Su, H., Guibas, L.J.: A point set generation network for 3d object reconstruction from a single image. In: CVPR (2017)"
|
| 1933 |
+
},
|
| 1934 |
+
{
|
| 1935 |
+
"type": "ref_text",
|
| 1936 |
+
"bbox": [
|
| 1937 |
+
0.226,
|
| 1938 |
+
0.49,
|
| 1939 |
+
0.785,
|
| 1940 |
+
0.531
|
| 1941 |
+
],
|
| 1942 |
+
"angle": 0,
|
| 1943 |
+
"content": "2. Fieraru, M., Zanfir, M., Oneata, E., Popa, A.I., Olaru, V., Sminchisescu, C.: Reconstructing three-dimensional models of interacting humans. arXiv preprint arXiv:2308.01854 (2023)"
|
| 1944 |
+
},
|
| 1945 |
+
{
|
| 1946 |
+
"type": "ref_text",
|
| 1947 |
+
"bbox": [
|
| 1948 |
+
0.226,
|
| 1949 |
+
0.532,
|
| 1950 |
+
0.785,
|
| 1951 |
+
0.573
|
| 1952 |
+
],
|
| 1953 |
+
"angle": 0,
|
| 1954 |
+
"content": "3. Goel, S., Pavlakos, G., Rajasegaran, J., Kanazawa, A., Malik, J.: Humans in 4d: Reconstructing and tracking humans with transformers. arXiv preprint arXiv:2305.20091 (2023)"
|
| 1955 |
+
},
|
| 1956 |
+
{
|
| 1957 |
+
"type": "ref_text",
|
| 1958 |
+
"bbox": [
|
| 1959 |
+
0.226,
|
| 1960 |
+
0.575,
|
| 1961 |
+
0.785,
|
| 1962 |
+
0.601
|
| 1963 |
+
],
|
| 1964 |
+
"angle": 0,
|
| 1965 |
+
"content": "4. Guan, H., Song, C., Zhang, Z.: Gramo: geometric resampling augmentation for monocular 3d object detection. Frontiers of Computer Science (2024)"
|
| 1966 |
+
},
|
| 1967 |
+
{
|
| 1968 |
+
"type": "ref_text",
|
| 1969 |
+
"bbox": [
|
| 1970 |
+
0.226,
|
| 1971 |
+
0.602,
|
| 1972 |
+
0.785,
|
| 1973 |
+
0.643
|
| 1974 |
+
],
|
| 1975 |
+
"angle": 0,
|
| 1976 |
+
"content": "5. Huang, J., Huang, G., Zhu, Z., Ye, Y., Du, D.: Bevdet: High-performance multicamera 3d object detection in bird-eye-view. arXiv preprint arXiv:2112.11790 (2021)"
|
| 1977 |
+
},
|
| 1978 |
+
{
|
| 1979 |
+
"type": "ref_text",
|
| 1980 |
+
"bbox": [
|
| 1981 |
+
0.226,
|
| 1982 |
+
0.645,
|
| 1983 |
+
0.785,
|
| 1984 |
+
0.672
|
| 1985 |
+
],
|
| 1986 |
+
"angle": 0,
|
| 1987 |
+
"content": "6. Huang, Y., Zheng, W., Zhang, B., Zhou, J., Lu, J.: Selfocc: Self-supervised vision-based 3d occupancy prediction. In: CVPR (2024)"
|
| 1988 |
+
},
|
| 1989 |
+
{
|
| 1990 |
+
"type": "ref_text",
|
| 1991 |
+
"bbox": [
|
| 1992 |
+
0.226,
|
| 1993 |
+
0.673,
|
| 1994 |
+
0.785,
|
| 1995 |
+
0.699
|
| 1996 |
+
],
|
| 1997 |
+
"angle": 0,
|
| 1998 |
+
"content": "7. Huang, Y., Zheng, W., Zhang, Y., Zhou, J., Lu, J.: Tri-perspective view for vision-based 3d semantic occupancy prediction. In: CVPR (2023)"
|
| 1999 |
+
},
|
| 2000 |
+
{
|
| 2001 |
+
"type": "ref_text",
|
| 2002 |
+
"bbox": [
|
| 2003 |
+
0.226,
|
| 2004 |
+
0.701,
|
| 2005 |
+
0.785,
|
| 2006 |
+
0.728
|
| 2007 |
+
],
|
| 2008 |
+
"angle": 0,
|
| 2009 |
+
"content": "8. Li, J., Han, K., Wang, P., Liu, Y., Yuan, X.: Anisotropic convolutional networks for 3d semantic scene completion. In: CVPR (2020)"
|
| 2010 |
+
},
|
| 2011 |
+
{
|
| 2012 |
+
"type": "ref_text",
|
| 2013 |
+
"bbox": [
|
| 2014 |
+
0.226,
|
| 2015 |
+
0.729,
|
| 2016 |
+
0.785,
|
| 2017 |
+
0.77
|
| 2018 |
+
],
|
| 2019 |
+
"angle": 0,
|
| 2020 |
+
"content": "9. Li, Y., Yu, Z., Choy, C., Xiao, C., Alvarez, J.M., Fidler, S., Feng, C., Anandkumar, A.: Voxformer: Sparse voxel transformer for camera-based 3d semantic scene completion. In: CVPR (2023)"
|
| 2021 |
+
},
|
| 2022 |
+
{
|
| 2023 |
+
"type": "ref_text",
|
| 2024 |
+
"bbox": [
|
| 2025 |
+
0.226,
|
| 2026 |
+
0.771,
|
| 2027 |
+
0.785,
|
| 2028 |
+
0.798
|
| 2029 |
+
],
|
| 2030 |
+
"angle": 0,
|
| 2031 |
+
"content": "20. Li, Y., Ge, Z., Yu, G., Yang, J., Wang, Z., Shi, Y., Sun, J., Li, Z.: Bevdepth: Acquisition of reliable depth for multi-view 3d object detection. In: AAAI (2023)"
|
| 2032 |
+
},
|
| 2033 |
+
{
|
| 2034 |
+
"type": "ref_text",
|
| 2035 |
+
"bbox": [
|
| 2036 |
+
0.226,
|
| 2037 |
+
0.799,
|
| 2038 |
+
0.785,
|
| 2039 |
+
0.84
|
| 2040 |
+
],
|
| 2041 |
+
"angle": 0,
|
| 2042 |
+
"content": "21. Li, Z., Wang, W., Li, H., Xie, E., Sima, C., Lu, T., Qiao, Y., Dai, J.: Bevformer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. In: ECCV (2022)"
|
| 2043 |
+
},
|
| 2044 |
+
{
|
| 2045 |
+
"type": "list",
|
| 2046 |
+
"bbox": [
|
| 2047 |
+
0.226,
|
| 2048 |
+
0.179,
|
| 2049 |
+
0.785,
|
| 2050 |
+
0.84
|
| 2051 |
+
],
|
| 2052 |
+
"angle": 0,
|
| 2053 |
+
"content": null
|
| 2054 |
+
}
|
| 2055 |
+
],
|
| 2056 |
+
[
|
| 2057 |
+
{
|
| 2058 |
+
"type": "page_number",
|
| 2059 |
+
"bbox": [
|
| 2060 |
+
0.218,
|
| 2061 |
+
0.116,
|
| 2062 |
+
0.236,
|
| 2063 |
+
0.127
|
| 2064 |
+
],
|
| 2065 |
+
"angle": 0,
|
| 2066 |
+
"content": "16"
|
| 2067 |
+
},
|
| 2068 |
+
{
|
| 2069 |
+
"type": "header",
|
| 2070 |
+
"bbox": [
|
| 2071 |
+
0.272,
|
| 2072 |
+
0.115,
|
| 2073 |
+
0.352,
|
| 2074 |
+
0.127
|
| 2075 |
+
],
|
| 2076 |
+
"angle": 0,
|
| 2077 |
+
"content": "H. Yu et al."
|
| 2078 |
+
},
|
| 2079 |
+
{
|
| 2080 |
+
"type": "ref_text",
|
| 2081 |
+
"bbox": [
|
| 2082 |
+
0.217,
|
| 2083 |
+
0.147,
|
| 2084 |
+
0.785,
|
| 2085 |
+
0.189
|
| 2086 |
+
],
|
| 2087 |
+
"angle": 0,
|
| 2088 |
+
"content": "22. Li, Z., Yu, Z., Austin, D., Fang, M., Lan, S., Kautz, J., Alvarez, J.M.: Fb-occ: 3d occupancy prediction based on forward-backward view transformation. arXiv preprint arXiv:2307.01492 (2023)"
|
| 2089 |
+
},
|
| 2090 |
+
{
|
| 2091 |
+
"type": "ref_text",
|
| 2092 |
+
"bbox": [
|
| 2093 |
+
0.217,
|
| 2094 |
+
0.19,
|
| 2095 |
+
0.785,
|
| 2096 |
+
0.217
|
| 2097 |
+
],
|
| 2098 |
+
"angle": 0,
|
| 2099 |
+
"content": "23. Liu, S., Hu, Y., Zeng, Y., Tang, Q., Jin, B., Han, Y., Li, X.: See and think: Disentangling semantic scene completion. NeurIPS (2018)"
|
| 2100 |
+
},
|
| 2101 |
+
{
|
| 2102 |
+
"type": "ref_text",
|
| 2103 |
+
"bbox": [
|
| 2104 |
+
0.217,
|
| 2105 |
+
0.218,
|
| 2106 |
+
0.785,
|
| 2107 |
+
0.244
|
| 2108 |
+
],
|
| 2109 |
+
"angle": 0,
|
| 2110 |
+
"content": "24. Murez, Z., Van As, T., Bartolozzi, J., Sinha, A., Badrinarayanan, V., Rabinovich, A.: Atlas: End-to-end 3d scene reconstruction from posed images. In: ECCV (2020)"
|
| 2111 |
+
},
|
| 2112 |
+
{
|
| 2113 |
+
"type": "ref_text",
|
| 2114 |
+
"bbox": [
|
| 2115 |
+
0.217,
|
| 2116 |
+
0.245,
|
| 2117 |
+
0.785,
|
| 2118 |
+
0.285
|
| 2119 |
+
],
|
| 2120 |
+
"angle": 0,
|
| 2121 |
+
"content": "25. Pan, M., Liu, J., Zhang, R., Huang, P., Li, X., Liu, L., Zhang, S.: Renderocc: Visioncentric 3d occupancy prediction with 2d rendering supervision. arXiv preprint arXiv:2309.09502 (2023)"
|
| 2122 |
+
},
|
| 2123 |
+
{
|
| 2124 |
+
"type": "ref_text",
|
| 2125 |
+
"bbox": [
|
| 2126 |
+
0.217,
|
| 2127 |
+
0.287,
|
| 2128 |
+
0.785,
|
| 2129 |
+
0.314
|
| 2130 |
+
],
|
| 2131 |
+
"angle": 0,
|
| 2132 |
+
"content": "26. Park, J.J., Florence, P., Straub, J., Newcombe, R., Lovegrove, S.: Deepsdf: Learning continuous signed distance functions for shape representation. In: CVPR (2019)"
|
| 2133 |
+
},
|
| 2134 |
+
{
|
| 2135 |
+
"type": "ref_text",
|
| 2136 |
+
"bbox": [
|
| 2137 |
+
0.217,
|
| 2138 |
+
0.315,
|
| 2139 |
+
0.785,
|
| 2140 |
+
0.355
|
| 2141 |
+
],
|
| 2142 |
+
"angle": 0,
|
| 2143 |
+
"content": "27. Ranftl, R., Lasinger, K., Hafner, D., Schindler, K., Koltun, V.: Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer. TPAMI (2020)"
|
| 2144 |
+
},
|
| 2145 |
+
{
|
| 2146 |
+
"type": "ref_text",
|
| 2147 |
+
"bbox": [
|
| 2148 |
+
0.217,
|
| 2149 |
+
0.357,
|
| 2150 |
+
0.785,
|
| 2151 |
+
0.383
|
| 2152 |
+
],
|
| 2153 |
+
"angle": 0,
|
| 2154 |
+
"content": "28. Roldao, L., de Charette, R., Verroust-Blondet, A.: Lmscnet: Lightweight multiscale 3d semantic completion. In: 3DV (2020)"
|
| 2155 |
+
},
|
| 2156 |
+
{
|
| 2157 |
+
"type": "ref_text",
|
| 2158 |
+
"bbox": [
|
| 2159 |
+
0.217,
|
| 2160 |
+
0.384,
|
| 2161 |
+
0.785,
|
| 2162 |
+
0.411
|
| 2163 |
+
],
|
| 2164 |
+
"angle": 0,
|
| 2165 |
+
"content": "29. Silberman, N., Hoiem, D., Kohli, P., Fergus, R.: Indoor segmentation and support inference from rgbd images. In: ECCV (2012)"
|
| 2166 |
+
},
|
| 2167 |
+
{
|
| 2168 |
+
"type": "ref_text",
|
| 2169 |
+
"bbox": [
|
| 2170 |
+
0.217,
|
| 2171 |
+
0.412,
|
| 2172 |
+
0.785,
|
| 2173 |
+
0.438
|
| 2174 |
+
],
|
| 2175 |
+
"angle": 0,
|
| 2176 |
+
"content": "30. Song, S., Yu, F., Zeng, A., Chang, A.X., Savva, M., Funkhouser, T.: Semantic scene completion from a single depth image. In: CVPR (2017)"
|
| 2177 |
+
},
|
| 2178 |
+
{
|
| 2179 |
+
"type": "ref_text",
|
| 2180 |
+
"bbox": [
|
| 2181 |
+
0.217,
|
| 2182 |
+
0.44,
|
| 2183 |
+
0.785,
|
| 2184 |
+
0.466
|
| 2185 |
+
],
|
| 2186 |
+
"angle": 0,
|
| 2187 |
+
"content": "31. Tan, M., Le, Q.: Efficientnet: Rethinking model scaling for convolutional neural networks. In: ICML (2019)"
|
| 2188 |
+
},
|
| 2189 |
+
{
|
| 2190 |
+
"type": "ref_text",
|
| 2191 |
+
"bbox": [
|
| 2192 |
+
0.217,
|
| 2193 |
+
0.467,
|
| 2194 |
+
0.785,
|
| 2195 |
+
0.493
|
| 2196 |
+
],
|
| 2197 |
+
"angle": 0,
|
| 2198 |
+
"content": "32. Tang, Y., Dorn, S., Savani, C.: Center3d: Center-based monocular 3d object detection with joint depth understanding. In: DAGM GCPR (2020)"
|
| 2199 |
+
},
|
| 2200 |
+
{
|
| 2201 |
+
"type": "ref_text",
|
| 2202 |
+
"bbox": [
|
| 2203 |
+
0.217,
|
| 2204 |
+
0.495,
|
| 2205 |
+
0.785,
|
| 2206 |
+
0.535
|
| 2207 |
+
],
|
| 2208 |
+
"angle": 0,
|
| 2209 |
+
"content": "33. Tian, X., Jiang, T., Yun, L., Mao, Y., Yang, H., Wang, Y., Wang, Y., Zhao, H.: Occ3d: A large-scale 3d occupancy prediction benchmark for autonomous driving. NeurIPS (2023)"
|
| 2210 |
+
},
|
| 2211 |
+
{
|
| 2212 |
+
"type": "ref_text",
|
| 2213 |
+
"bbox": [
|
| 2214 |
+
0.217,
|
| 2215 |
+
0.536,
|
| 2216 |
+
0.785,
|
| 2217 |
+
0.563
|
| 2218 |
+
],
|
| 2219 |
+
"angle": 0,
|
| 2220 |
+
"content": "34. Tong, W., Sima, C., Wang, T., Chen, L., Wu, S., Deng, H., Gu, Y., Lu, L., Luo, P., Lin, D., et al.: Scene as occupancy. In: ICCV (2023)"
|
| 2221 |
+
},
|
| 2222 |
+
{
|
| 2223 |
+
"type": "ref_text",
|
| 2224 |
+
"bbox": [
|
| 2225 |
+
0.217,
|
| 2226 |
+
0.565,
|
| 2227 |
+
0.785,
|
| 2228 |
+
0.591
|
| 2229 |
+
],
|
| 2230 |
+
"angle": 0,
|
| 2231 |
+
"content": "35. Wang, S., Liu, Y., Wang, T., Li, Y., Zhang, X.: Exploring object-centric temporal modeling for efficient multi-view 3d object detection. In: ICCV (2023)"
|
| 2232 |
+
},
|
| 2233 |
+
{
|
| 2234 |
+
"type": "ref_text",
|
| 2235 |
+
"bbox": [
|
| 2236 |
+
0.217,
|
| 2237 |
+
0.592,
|
| 2238 |
+
0.785,
|
| 2239 |
+
0.632
|
| 2240 |
+
],
|
| 2241 |
+
"angle": 0,
|
| 2242 |
+
"content": "36. Wang, X., Zhu, Z., Xu, W., Zhang, Y., Wei, Y., Chi, X., Ye, Y., Du, D., Lu, J., Wang, X.: Openoccupancy: A large scale benchmark for surrounding semantic occupancy perception. In: ICCV (2023)"
|
| 2243 |
+
},
|
| 2244 |
+
{
|
| 2245 |
+
"type": "ref_text",
|
| 2246 |
+
"bbox": [
|
| 2247 |
+
0.217,
|
| 2248 |
+
0.633,
|
| 2249 |
+
0.785,
|
| 2250 |
+
0.66
|
| 2251 |
+
],
|
| 2252 |
+
"angle": 0,
|
| 2253 |
+
"content": "37. Wang, Y., Guizilini, V.C., Zhang, T., Wang, Y., Zhao, H., Solomon, J.: Detr3d: 3d object detection from multi-view images via 3d-to-2d queries. In: CoRL (2022)"
|
| 2254 |
+
},
|
| 2255 |
+
{
|
| 2256 |
+
"type": "ref_text",
|
| 2257 |
+
"bbox": [
|
| 2258 |
+
0.217,
|
| 2259 |
+
0.661,
|
| 2260 |
+
0.785,
|
| 2261 |
+
0.688
|
| 2262 |
+
],
|
| 2263 |
+
"angle": 0,
|
| 2264 |
+
"content": "38. Wang, Y., Chen, Y., Liao, X., Fan, L., Zhang, Z.: Panoocc: Unified occupancy representation for camera-based 3d panoptic segmentation. In: CVPR (2024)"
|
| 2265 |
+
},
|
| 2266 |
+
{
|
| 2267 |
+
"type": "ref_text",
|
| 2268 |
+
"bbox": [
|
| 2269 |
+
0.217,
|
| 2270 |
+
0.689,
|
| 2271 |
+
0.785,
|
| 2272 |
+
0.715
|
| 2273 |
+
],
|
| 2274 |
+
"angle": 0,
|
| 2275 |
+
"content": "39. Wang, Y., Chen, Y., Zhang, Z.: Frustumformer: Adaptive instance-aware resampling for multi-view 3d detection. In: CVPR (2023)"
|
| 2276 |
+
},
|
| 2277 |
+
{
|
| 2278 |
+
"type": "ref_text",
|
| 2279 |
+
"bbox": [
|
| 2280 |
+
0.217,
|
| 2281 |
+
0.716,
|
| 2282 |
+
0.785,
|
| 2283 |
+
0.743
|
| 2284 |
+
],
|
| 2285 |
+
"angle": 0,
|
| 2286 |
+
"content": "40. Wei, Y., Zhao, L., Zheng, W., Zhu, Z., Zhou, J., Lu, J.: Surroundocc: Multi-camera 3d occupancy prediction for autonomous driving. In: ICCV (2023)"
|
| 2287 |
+
},
|
| 2288 |
+
{
|
| 2289 |
+
"type": "ref_text",
|
| 2290 |
+
"bbox": [
|
| 2291 |
+
0.217,
|
| 2292 |
+
0.744,
|
| 2293 |
+
0.785,
|
| 2294 |
+
0.771
|
| 2295 |
+
],
|
| 2296 |
+
"angle": 0,
|
| 2297 |
+
"content": "41. Wu, Q., Wang, K., Li, K., Zheng, J., Cai, J.: Objectsdf++: Improved object-compositional neural implicit surfaces. In: ICCV (2023)"
|
| 2298 |
+
},
|
| 2299 |
+
{
|
| 2300 |
+
"type": "ref_text",
|
| 2301 |
+
"bbox": [
|
| 2302 |
+
0.217,
|
| 2303 |
+
0.772,
|
| 2304 |
+
0.785,
|
| 2305 |
+
0.798
|
| 2306 |
+
],
|
| 2307 |
+
"angle": 0,
|
| 2308 |
+
"content": "42. Wu, S.C., Tateno, K., Navab, N., Tombari, F.: Scfusion: Real-time incremental scene reconstruction with semantic completion. In: 3DV (2020)"
|
| 2309 |
+
},
|
| 2310 |
+
{
|
| 2311 |
+
"type": "ref_text",
|
| 2312 |
+
"bbox": [
|
| 2313 |
+
0.217,
|
| 2314 |
+
0.799,
|
| 2315 |
+
0.785,
|
| 2316 |
+
0.84
|
| 2317 |
+
],
|
| 2318 |
+
"angle": 0,
|
| 2319 |
+
"content": "43. Yang, C., Chen, Y., Tian, H., Tao, C., Zhu, X., Zhang, Z., Huang, G., Li, H., Qiao, Y., Lu, L., et al.: Bevformer v2: Adapting modern image backbones to bird's-eye-view recognition via perspective supervision. In: CVPR (2023)"
|
| 2320 |
+
},
|
| 2321 |
+
{
|
| 2322 |
+
"type": "list",
|
| 2323 |
+
"bbox": [
|
| 2324 |
+
0.217,
|
| 2325 |
+
0.147,
|
| 2326 |
+
0.785,
|
| 2327 |
+
0.84
|
| 2328 |
+
],
|
| 2329 |
+
"angle": 0,
|
| 2330 |
+
"content": null
|
| 2331 |
+
}
|
| 2332 |
+
],
|
| 2333 |
+
[
|
| 2334 |
+
{
|
| 2335 |
+
"type": "header",
|
| 2336 |
+
"bbox": [
|
| 2337 |
+
0.33,
|
| 2338 |
+
0.115,
|
| 2339 |
+
0.732,
|
| 2340 |
+
0.129
|
| 2341 |
+
],
|
| 2342 |
+
"angle": 0,
|
| 2343 |
+
"content": "Monocular Occupancy Prediction for Scalable Indoor Scenes"
|
| 2344 |
+
},
|
| 2345 |
+
{
|
| 2346 |
+
"type": "page_number",
|
| 2347 |
+
"bbox": [
|
| 2348 |
+
0.768,
|
| 2349 |
+
0.117,
|
| 2350 |
+
0.785,
|
| 2351 |
+
0.127
|
| 2352 |
+
],
|
| 2353 |
+
"angle": 0,
|
| 2354 |
+
"content": "17"
|
| 2355 |
+
},
|
| 2356 |
+
{
|
| 2357 |
+
"type": "ref_text",
|
| 2358 |
+
"bbox": [
|
| 2359 |
+
0.217,
|
| 2360 |
+
0.147,
|
| 2361 |
+
0.785,
|
| 2362 |
+
0.175
|
| 2363 |
+
],
|
| 2364 |
+
"angle": 0,
|
| 2365 |
+
"content": "44. Yang, L., Kang, B., Huang, Z., Xu, X., Feng, J., Zhao, H.: Depth anything: Unleashing the power of large-scale unlabeled data. In: CVPR (2024)"
|
| 2366 |
+
},
|
| 2367 |
+
{
|
| 2368 |
+
"type": "ref_text",
|
| 2369 |
+
"bbox": [
|
| 2370 |
+
0.217,
|
| 2371 |
+
0.175,
|
| 2372 |
+
0.785,
|
| 2373 |
+
0.217
|
| 2374 |
+
],
|
| 2375 |
+
"angle": 0,
|
| 2376 |
+
"content": "45. Yao, J., Li, C., Sun, K., Cai, Y., Li, H., Ouyang, W., Li, H.: Ndc-scene: Boost monocular 3d semantic scene completion in normalized device coordinates space. In: ICCV (2023)"
|
| 2377 |
+
},
|
| 2378 |
+
{
|
| 2379 |
+
"type": "ref_text",
|
| 2380 |
+
"bbox": [
|
| 2381 |
+
0.217,
|
| 2382 |
+
0.218,
|
| 2383 |
+
0.785,
|
| 2384 |
+
0.259
|
| 2385 |
+
],
|
| 2386 |
+
"angle": 0,
|
| 2387 |
+
"content": "46. Yu, Z., Shu, C., Deng, J., Lu, K., Liu, Z., Yu, J., Yang, D., Li, H., Chen, Y.: Flashocc: Fast and memory-efficient occupancy prediction via channel-to-height plugin. arXiv preprint arXiv:2311.12058 (2023)"
|
| 2388 |
+
},
|
| 2389 |
+
{
|
| 2390 |
+
"type": "ref_text",
|
| 2391 |
+
"bbox": [
|
| 2392 |
+
0.217,
|
| 2393 |
+
0.259,
|
| 2394 |
+
0.785,
|
| 2395 |
+
0.286
|
| 2396 |
+
],
|
| 2397 |
+
"angle": 0,
|
| 2398 |
+
"content": "47. Yurtsever, E., Lambert, J., Carballo, A., Takeda, K.: A survey of autonomous driving: Common practices and emerging technologies. IEEE access (2020)"
|
| 2399 |
+
},
|
| 2400 |
+
{
|
| 2401 |
+
"type": "ref_text",
|
| 2402 |
+
"bbox": [
|
| 2403 |
+
0.217,
|
| 2404 |
+
0.286,
|
| 2405 |
+
0.785,
|
| 2406 |
+
0.314
|
| 2407 |
+
],
|
| 2408 |
+
"angle": 0,
|
| 2409 |
+
"content": "48. Zhang, J., Zhao, H., Yao, A., Chen, Y., Zhang, L., Liao, H.: Efficient semantic scene completion network with spatial group convolution. In: ECCV (2018)"
|
| 2410 |
+
},
|
| 2411 |
+
{
|
| 2412 |
+
"type": "ref_text",
|
| 2413 |
+
"bbox": [
|
| 2414 |
+
0.217,
|
| 2415 |
+
0.314,
|
| 2416 |
+
0.785,
|
| 2417 |
+
0.342
|
| 2418 |
+
],
|
| 2419 |
+
"angle": 0,
|
| 2420 |
+
"content": "49. Zhang, X., Bi, S., Sunkavalli, K., Su, H., Xu, Z.: Nerfusion: Fusing radiance fields for large-scale scene reconstruction. In: CVPR (2022)"
|
| 2421 |
+
},
|
| 2422 |
+
{
|
| 2423 |
+
"type": "ref_text",
|
| 2424 |
+
"bbox": [
|
| 2425 |
+
0.217,
|
| 2426 |
+
0.342,
|
| 2427 |
+
0.785,
|
| 2428 |
+
0.37
|
| 2429 |
+
],
|
| 2430 |
+
"angle": 0,
|
| 2431 |
+
"content": "50. Zheng, Z., Yu, T., Wei, Y., Dai, Q., Liu, Y.: Deephuman: 3d human reconstruction from a single image. In: ICCV (2019)"
|
| 2432 |
+
},
|
| 2433 |
+
{
|
| 2434 |
+
"type": "ref_text",
|
| 2435 |
+
"bbox": [
|
| 2436 |
+
0.217,
|
| 2437 |
+
0.37,
|
| 2438 |
+
0.785,
|
| 2439 |
+
0.397
|
| 2440 |
+
],
|
| 2441 |
+
"angle": 0,
|
| 2442 |
+
"content": "51. Zhong, M., Zeng, G.: Semantic point completion network for 3d semantic scene completion. In: ECAI (2020)"
|
| 2443 |
+
},
|
| 2444 |
+
{
|
| 2445 |
+
"type": "list",
|
| 2446 |
+
"bbox": [
|
| 2447 |
+
0.217,
|
| 2448 |
+
0.147,
|
| 2449 |
+
0.785,
|
| 2450 |
+
0.397
|
| 2451 |
+
],
|
| 2452 |
+
"angle": 0,
|
| 2453 |
+
"content": null
|
| 2454 |
+
}
|
| 2455 |
+
]
|
| 2456 |
+
]
|
2024/Monocular Occupancy Prediction for Scalable Indoor Scenes/325d434b-3564-4d20-a82e-4ec4e345eb58_origin.pdf
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:d58f7e7baa7239ee9b1c2a6feedb53420ef603a1d417acbf68f7451da60da1aa
|
| 3 |
+
size 3336218
|
2024/Monocular Occupancy Prediction for Scalable Indoor Scenes/full.md
ADDED
|
@@ -0,0 +1,323 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Monocular Occupancy Prediction for Scalable Indoor Scenes
|
| 2 |
+
|
| 3 |
+
Hongxiao $\mathrm{Yu}^{1,2}$ Yuqi Wang $^{1,2}$ , Yuntao Chen $^{3}$ , and Zhaoxiang Zhang $^{1,2,3}$
|
| 4 |
+
|
| 5 |
+
<sup>1</sup> School of Artificial Intelligence, University of Chinese Academy of Sciences (UCAS)
|
| 6 |
+
|
| 7 |
+
$^{2}$ NLPR, MAIS, Institute of Automation, Chinese Academy of Sciences (CASIA)
|
| 8 |
+
|
| 9 |
+
<sup>3</sup> Centre for Artificial Intelligence and Robotics (HKISI_CAS) {yuhongxiao2023, wangyuqi2020, zhaoxiang.zhang}@ia.ac.cn, chenyuntao08@gmail.com
|
| 10 |
+
|
| 11 |
+
Abstract. Camera-based 3D occupancy prediction has recently garnered increasing attention in outdoor driving scenes. However, research in indoor scenes remains relatively unexplored. The core differences in indoor scenes lie in the complexity of scene scale and the variance in object size. In this paper, we propose a novel method, named ISO, for predicting indoor scene occupancy using monocular images. ISO harnesses the advantages of a pretrained depth model to achieve accurate depth predictions. Furthermore, we introduce the Dual Feature Line of Sight Projection (D-FLoSP) module within ISO, which enhances the learning of 3D voxel features. To foster further research in this domain, we introduce Occ-ScanNet, a large-scale occupancy benchmark for indoor scenes. With a dataset size 40 times larger than the NYUv2 dataset, it facilitates future scalable research in indoor scene analysis. Experimental results on both NYUv2 and Occ-ScanNet demonstrate that our method achieves state-of-the-art performance. The dataset and code are made publicly at https://github.com/hongxiaoy/ISO.git.
|
| 12 |
+
|
| 13 |
+
Keywords: 3D occupancy prediction $\cdot$ Camera-based scene understanding $\cdot$ Semantic scene completion
|
| 14 |
+
|
| 15 |
+
# 1 Introduction
|
| 16 |
+
|
| 17 |
+
3D scene understanding is a crucial task in computer vision, becoming increasingly important for applications such as robotic navigation [10], augmented reality [2], and autonomous driving [47]. While humans possess a natural ability to comprehend 3D environments through vision, enabling this ability in computers poses a significant challenge due to limitations like restricted fields of view, sparse sensing capabilities, and measurement noise. However, with the rapid development of deep learning and the availability of large-scale 3D driving datasets, camera-based 3D object detection [14, 15, 20, 21, 35, 37, 39, 43] and occupancy prediction [5, 22, 33, 38, 40] have developed rapidly, achieving significant improvements in performance.
|
| 18 |
+
|
| 19 |
+
Occupancy prediction has gained popularity recently due to its effectiveness in representing both background and foreground objects within a scene using a
|
| 20 |
+
|
| 21 |
+
unified representation. Although significant progress has been made in outdoor driving scenarios, research in indoor scenarios remains limited. Indoor scenes differ from outdoor driving scenes in two key aspects: (1) scene-scale complexity: Indoor rooms often exhibit a more diverse range of sizes compared to outdoor environments, where driving scenarios typically focus on a fixed 3D space for perception. This diversity, ranging from spacious living rooms to narrow kitchens, poses higher precision requirements for depth prediction. (2) object complexity: Indoor scenes feature a higher density and greater variety of objects. Unlike outdoor objects such as vehicles and pedestrians, which typically have consistent sizes within their respective categories and are relatively well-separated, indoor furniture and other objects often exhibit significant variations in scale and are closely positioned to each other. This increased complexity necessitates more sophisticated 3D perception techniques to accurately capture and understand the intricate geometry and relationships among objects within indoor spaces. Furthermore, existing works [5,45] focusing on indoor scenes primarily utilize the NYUv2 dataset [29], which lacks a more scalable and generalizable benchmark for comprehensive evaluation.
|
| 22 |
+
|
| 23 |
+
To address the above issues, we firstly introduce a more scalable benchmark, Occ-ScanNet, for 3D occupancy prediction in indoor scenes. This benchmark builds upon the large-scale ScanNet [8] dataset, offering a 40 times more samples compared to the NYUv2 [29] dataset, thus greatly expanding the scope and diversity of indoor scene study. To effectively address the unique complexities of indoor scenes, we propose a novel method named ISO (Indoor Scene Occupancy). ISO leverages a powerful pretrained depth model and integrates a D-FLoSP (Dual Feature Line of Sight Projection) module. This module enables precise depth estimation and facilitates learning voxel features for accurate predictions. To accommodate the varying sizes of indoor objects, we introduce a multi-scale feature fusion module. This module enhances the learning of object features across different scales. Leveraging a powerful depth model, our method is adept at handling diverse indoor scenes, offering a robust and versatile solution.
|
| 24 |
+
|
| 25 |
+
Our main contributions can be summarized as follows.
|
| 26 |
+
|
| 27 |
+
- We introduce a new benchmark called Occ-ScanNet for monocular 3D occupancy prediction in indoor scenes. With a dataset size 40 times larger than the NYUv2 dataset, it significantly enhances the potential for scalable research in indoor scene analysis.
|
| 28 |
+
- We propose a novel approach called $ISO$ , which primarily comprises the DFLoSP (Dual Feature Line of Sight Projection) module and a multi-scale feature fusion module. Together, these components effectively address the challenges posed by variations in scene and object sizes, enabling more accurate and robust 3D occupancy prediction.
|
| 29 |
+
- Experiments on the Occ-ScanNet and NYUv2 datasets demonstrate that our approach achieves state-of-the-art performance in monocular 3D occupancy prediction. Furthermore, our method exhibits significant scalability potential.
|
| 30 |
+
|
| 31 |
+
# 2 Related works
|
| 32 |
+
|
| 33 |
+
# 2.1 Monocular 3D Semantic Scene Completion
|
| 34 |
+
|
| 35 |
+
Monocular 3D Semantic Scene Completion (SSC) aims to infer the complete 3D structure and corresponding semantics from a single image. This monocular setting was first introduced by Monoscene [5], advancing upon prior SSC methods [18,23,30,48,51] by relying solely on vision, without additional 3D inputs. Monoscene [5] introduces the Features Line of Sight Projection (FLoSP), inspired by optics, to obtain 3D features through ray projection. However, the shared 2D features lifted to 3D rays via FLoSP results in depth ambiguity. Consequently, subsequent efforts have placed greater emphasis on leveraging depth information. VoxFormer [19] proposed a novel query proposal network based on 2D convolutions, generating sparse queries from image depth, which showed impressive performance in driving scenes. Meanwhile, NDC-Scene [45] further devised a Depth-Adaptive Dual Decoder to concurrently upsample and merge the 2D and 3D feature maps, thereby enhancing overall performance.
|
| 36 |
+
|
| 37 |
+
# 2.2 Multiview 3D Occupancy Prediction
|
| 38 |
+
|
| 39 |
+
Multiview 3D Occupancy Prediction predicts the semantic occupancy of the surrounding 3D scene given multiview images, has recently garnered significant attention in the field of autonomous driving. TPVFormer [17] pioneers exploration in driving scenes, employing sparse LiDAR labels for supervision. It introduces a tri-perspective view (TPV) representation that accompanies BEV with two additional perpendicular planes. However, due to the sparse supervision provided by LiDAR, subsequent works [33, 34, 36, 40] have focused on providing more dense occupancy benchmarks. FB-OCC [22] integrates the lifting of 2D to 3D and querying of 3D to 2D, achieving a more efficient 3D feature transformation. Meanwhile, PanoOcc [38] unifies detection and semantic occupancy tasks, enabling comprehensive panoramic scene understanding. Recent methods also explore improving model efficiency [46] and utilizing weaker forms of supervision [16, 25].
|
| 40 |
+
|
| 41 |
+
# 2.3 3D Reconstruction from Image
|
| 42 |
+
|
| 43 |
+
3D reconstruction is a technique that involves recovering a 3D representation of objects [1, 11, 26], scenes [9, 41, 49], or even human bodies [12, 13, 50] from camera images. The task of 3D reconstruction can be classified into two categories: monocular reconstruction and multi-view reconstruction, depending on the number of images utilized. In the context of indoor scenes, 3D scene reconstruction aims to determine the surface geometry of the entire scene, often without incorporating semantic information. Atlas [24] proposes an end-to-end 3D reconstruction framework using TSDF regression from RGB images, bypassing traditional depth map estimation for efficient semantic segmentation. Panoptic-Reconstruction [7] unifies geometric reconstruction, 3D semantic, and
|
| 44 |
+
|
| 45 |
+

|
| 46 |
+
Fig. 1: The core design of ISO centers around the transformation of features from 2D to 3D spaces, encompassing the Depth Branch and the D-FLoSP module. A depth branch is initially integrated, it leverages a pre-trained depth model to estimate a pixel-wise depth map which is processed by the DepthNet to generate the final depth distribution. An element-wise multiplication between the voxel depth and features followed by summation are subsequently performed to derive the initial 3D voxel feature. The 3D feature is further processed to predict the 3D scene occupancy.
|
| 47 |
+
|
| 48 |
+
instance segmentation tasks. It predicts complete geometric reconstruction, semantic, and instance segmentations from a single RGB image's camera frustum. SCFusion [42], a real-time scene reconstruction framework, integrates continuous depth data using neural architecture for occupancy maps. It efficiently combines semantic completion with voxel states for simultaneous scene reconstruction and semantic understanding in real-time 3D environments.
|
| 49 |
+
|
| 50 |
+
# 3 Method
|
| 51 |
+
|
| 52 |
+
Problem definition. We focus on the problem of monocular 3D Occupancy Prediction. Specifically, this task takes a single RGB image $\mathbf{I}^{RGB}$ as input and output a voxel-wise occupancy along with semantic categories $\mathbf{Y}^{X\times Y\times Z\times C}$ . $X$ , $Y$ , and $Z$ represent the dimensions of the predicted 3D scene, while $C$ denotes the total number of semantic categories.
|
| 53 |
+
|
| 54 |
+
Model overview. In this section, we introduce the overall architecture of ISO, as shown in Fig. 1. Following Monoscene [5], we utilize 2D Unet and 3D Unet architectures to handle 2D and 3D features. Our core design focuses on transforming features from 2D to 3D, incorporating the Depth Branch and D-FLoSP module. Specifically, we incorporate a depth branch to estimate the pixel-wise depth map using a pre-trained depth model. Processed by the DepthNet, the model outputs a refined depth distribution. Then, voxel depth and features are multiplied element-wise and summed to obtain the initial 3D voxel feature volume $\mathbf{X}^{3\mathbf{d}} \in \mathbb{R}^{X \times Y \times Z \times C}$ . Finally, after processing the 3D voxel features, the model outputs the 3D scene occupancy.
|
| 55 |
+
|
| 56 |
+
# 3.1 Depth Branch
|
| 57 |
+
|
| 58 |
+
In this section, we introduce how to effectively estimate depth information from a single image.
|
| 59 |
+
|
| 60 |
+
Coarse depth estimation. Learning depth from scratch can be quite challenging. However, thanks to the recent rapid advancements in depth estimation [3,4,27,44], we can leverage pre-trained depth models to initially estimate a coarse depth map.
|
| 61 |
+
|
| 62 |
+
Compared to past models that could only estimate relative depth, we opted for Depth-Anything [44] as the pre-trained depth model, because it excels in predicting metric depth, as the Eq. (1) shows:
|
| 63 |
+
|
| 64 |
+
$$
|
| 65 |
+
\mathbf {D} ^ {\mathrm {m e t r i c}} = \mathbf {N} _ {\mathrm {d e p t h}} (\mathbf {I} ^ {\mathrm {r g b}}) \in \mathbb {R} ^ {1 \times H \times W}, \tag {1}
|
| 66 |
+
$$
|
| 67 |
+
|
| 68 |
+
where $\mathbf{N}_{\mathrm{depth}}$ denotes the pre-trained depth model [3, 44], and $\mathbf{I}^{\mathrm{rgb}}$ represents the input image. $H$ and $W$ represents the height and width of the input image.
|
| 69 |
+
|
| 70 |
+
Depth refinement. The coarse depth estimation is subsequently refined through model learning. The depth from pre-trained depth model is not precise enough for a higher mIoU, so we design a fine-tuning strategy. Specifically, the predicted metric depth $\mathbf{D}^{\mathrm{metric}}$ is then concatenated with the image feature $\mathbf{X}$ that has the same spatial scale. The augmented feature is continue processed by a following DepthNet to get a refined depth distribution $\mathbf{D}_{s=1}^{\mathrm{dist}}$ , as Eq. (2) shows:
|
| 71 |
+
|
| 72 |
+
$$
|
| 73 |
+
\mathbf {D} _ {\mathrm {s} = 1} ^ {\text {d i s t}} = \mathbf {F} _ {\text {d e p t h}} \left(\operatorname {C o n c a t} \left(\mathbf {D} ^ {\text {m e t r i c}}, \mathbf {X} _ {\mathrm {s} = 1} ^ {2 \mathrm {d}}\right)\right), \quad \mathbf {D} _ {\mathrm {s} = 1} ^ {\text {d i s t}} \in \mathbb {R} ^ {N _ {\text {b i n s}} \times H \times W}, \tag {2}
|
| 74 |
+
$$
|
| 75 |
+
|
| 76 |
+
where $\mathbf{F}_{\mathrm{depth}}$ is the DepthNet, $\mathbf{X}_{s=1}^{2d}$ is the 2D image feature, and $N_{\mathrm{bins}}$ is the number of discrete depth bins. $s$ denotes the level of feature scale, where $s = 1$ implies that the feature is at a 1:1 original image ratio.
|
| 77 |
+
|
| 78 |
+
We down-samples the scale $\frac{1}{1}$ distribution to get other smaller-scale 2D image feature maps' depth distribution. We represent the procedure as Eq. (3):
|
| 79 |
+
|
| 80 |
+
$$
|
| 81 |
+
\mathbf {D} _ {\mathrm {s} = k} ^ {\text {d i s t}} = \operatorname {D o w n S a m p l e} \left(\mathbf {D} _ {\mathrm {s} = 1} ^ {\text {d i s t}}\right), k = 2, 4, 8. \tag {3}
|
| 82 |
+
$$
|
| 83 |
+
|
| 84 |
+
The refined depth is supervised by the ground truth $\mathbf{D}^{\mathrm{GT}}$ . The ground truth has single depth value at each pixel, so we convert the single value to a one-hot vector of length $N_{\mathrm{bins}}$ . The refined depth prediction can be optimized using the BCE loss function as Eq. (4) shows.
|
| 85 |
+
|
| 86 |
+
$$
|
| 87 |
+
\begin{array}{l} \mathcal {L} _ {\mathrm {d e p t h}} = - \frac {1}{N _ {\mathrm {b i n s}} \times H \times W} \\ \sum_ {d = 1} ^ {N _ {\mathrm {b i n s}}} \sum_ {h = 1} ^ {H} \sum_ {w = 1} ^ {W} \left[ \mathbf {D} _ {N _ {\mathrm {b i n s}}, d, h, w} ^ {\mathrm {G T}} \cdot \log \left(\mathbf {D} _ {N _ {\mathrm {b i n s}}, d, h, w} ^ {\mathrm {d i s t}}\right) \right. \\ \left. + \left(1 - \mathbf {D} _ {N _ {\mathrm {b i n s}}, d, h, w} ^ {\mathrm {G T}}\right) \cdot \log \left(1 - \mathbf {D} _ {N _ {\mathrm {b i n s}}, d, h, w} ^ {\mathrm {d i s t}}\right) \right] \quad (4) \\ \end{array}
|
| 88 |
+
$$
|
| 89 |
+
|
| 90 |
+
# 3.2 Dual Feature Line of Sight Projection (D-FLoSP)
|
| 91 |
+
|
| 92 |
+
Compared to the FLoSP module in MonoScene [5], which projects 3D voxels to 2D, and 3D feature vectors are the corresponding pixel feature vectors. It only considers the projection of 3D rays and overlooks depth information, our proposed D-FLoSP module can more effectively integrate depth information.
|
| 93 |
+
|
| 94 |
+
Specifically, after getting depth distribution in the camera frame, we implement the FLoSP module to depth distribution. Each 3D voxel centroid position $x^{c}$ in the world frame can be projected to the camera frame and the pixel frame using camera pose and intrinsic matrix. We assign the projected depth values to discrete depth bins index by Eqs. (5) and (6), according to [32],
|
| 95 |
+
|
| 96 |
+
$$
|
| 97 |
+
\delta = \frac {2 \left(d _ {\max } - d _ {\min }\right)}{N _ {b i n s} \left(1 + N _ {b i n s}\right)}, \tag {5}
|
| 98 |
+
$$
|
| 99 |
+
|
| 100 |
+
$$
|
| 101 |
+
l = - 0. 5 + 0. 5 \sqrt {1 + \frac {8 \left(d - d _ {\operatorname* {m i n}}\right)}{\delta}}, \tag {6}
|
| 102 |
+
$$
|
| 103 |
+
|
| 104 |
+
and the depth distribution probability it lies in serves as the depth probability of that voxel. The process is illustrated by Eq. (7):
|
| 105 |
+
|
| 106 |
+
$$
|
| 107 |
+
\mathbf {D} _ {\mathrm {s} = k} ^ {\mathrm {3 d}} = \Phi_ {\rho \left(x ^ {c}\right)} ^ {\mathrm {3 d}} \left(\mathbf {D} _ {\mathrm {s} = k} ^ {\mathrm {d i s t}}\right) \in \mathbb {R} ^ {X \times Y \times Z \times C}, k = 1, 2, 4, 8, \tag {7}
|
| 108 |
+
$$
|
| 109 |
+
|
| 110 |
+
where $\Phi_{a}^{3\mathrm{d}}(b)$ is the 3D sampling of b at coordinates a, and $\rho (\cdot)$ is the perspective projection. Thus, the depth distribution is lifted from 2D to 3D. The bin-based representation with quantization is used for mapping continuous depth value to a discrete form, each bin will be assigned a different depth distribution score to make voxel aware of depth. The original depth precision is a continuous value, and hard to integrate it with the discrete form of voxel.
|
| 111 |
+
|
| 112 |
+
Following [5], we also employ the FLoSP module to generate a 3D voxel feature from a 2D feature map. Similar to what we did with the voxel centroids, the projected voxel can sample corresponding $1:k$ scaled feature maps $F_{1:k}$ from the 2D UNet decoder. This process is illustrated by Equation Eq. (8):
|
| 113 |
+
|
| 114 |
+
$$
|
| 115 |
+
\mathbf {X} _ {\mathrm {s} = k} ^ {3 \mathrm {d}} = \Phi_ {\rho \left(x ^ {c}\right)} ^ {2 \mathrm {d}} \left(\mathbf {X} _ {\mathrm {s} = k} ^ {2 \mathrm {d}}\right) \in \mathbb {R} ^ {X \times Y \times Z \times C}, k = 1, 2, 4, 8, \tag {8}
|
| 116 |
+
$$
|
| 117 |
+
|
| 118 |
+
where $\Phi_{a}^{2\mathrm{d}}(b)$ is the 2D sampling of b at coordinates a, and $\rho (\cdot)$ is the perspective projection.
|
| 119 |
+
|
| 120 |
+
The final 3D feature map $\mathbf{X}^{3\mathrm{d}}$ are summed in Eq. (9):
|
| 121 |
+
|
| 122 |
+
$$
|
| 123 |
+
\mathbf {X} ^ {\mathrm {3 d}} = \sum_ {\mathrm {s} \in \{1, 2, 4, 8 \}} \mathbf {X} _ {\mathrm {s} = k} ^ {\mathrm {3 d}} \odot \mathbf {D} _ {\mathrm {s} = k} ^ {\mathrm {3 d}}, \tag {9}
|
| 124 |
+
$$
|
| 125 |
+
|
| 126 |
+
where $\odot$ is the element-wise multiplication. The output map $\mathbf{X}^{3\mathrm{d}}$ is then served as the input of 3D UNet.
|
| 127 |
+
|
| 128 |
+

|
| 129 |
+
(a) Depth ranges in NYUv2 and our Occ-ScanNet Benchmark
|
| 130 |
+
|
| 131 |
+

|
| 132 |
+
(b) Dataset attributes of NYUv2 and our OccScanNet Benchmark
|
| 133 |
+
Fig. 2: Comparison of NYUv2 and Occ-ScanNet Benchmark. In (a), the depth ranges of NYUv2 and Occ-ScanNet are distinguished by dark and light green, respectively, with the horizontal axis indicating the minimum depth and the vertical axis showing the maximum depth of scenes. (b) quantitatively demonstrates that Occ-ScanNet possesses a significantly larger data scale compared to the NYUv2 dataset.
|
| 134 |
+
|
| 135 |
+
# 4 Occ-ScanNet Benchmark
|
| 136 |
+
|
| 137 |
+
Compared to the previous widely used indoor scene benchmark NYUv2 [29], which contains only 795 / 654 for train / test samples, our benchmark boasts 45,755 / 19,764 samples. Our benchmark significantly surpasses NYUv2 [29] in both data quantity and richness of scene depth, as illustrated in Fig. 2. This dataset is available at https://huggingface.co/datasets/hongxiaoy/OccScanNet.
|
| 138 |
+
|
| 139 |
+
# 4.1 Overview
|
| 140 |
+
|
| 141 |
+
Occ-ScanNet benchmark features a train/Validation split of 45,755 / 19,764 samples. As shown in Fig. 3, Occ-ScanNet exhibits rich diversity in scenes and viewpoints. This diverse dataset not only challenges the task of predicting scene occupancy but also fosters future research endeavors towards developing larger-scale and more versatile occupancy models.
|
| 142 |
+
|
| 143 |
+
# 4.2 Occupancy Label Generation
|
| 144 |
+
|
| 145 |
+
We adhere to the data formulation used in NYUv2 [29] dataset. The pipeline of occupancy generation process is showed in Fig. 4. For generating 3D voxel labels, we initially followed the generation process in CompleteScanNet and then employed manual inspection.
|
| 146 |
+
|
| 147 |
+
More specifically, from the official ScanNet [8] scenes, we extracted various data components, including color images, depth images, camera intrinsic matrices, and camera poses. A total of 100 frames were sampled from each scene and were randomly divided into training and validation sets with a $7/3$ split ratio per scene.
|
| 148 |
+
|
| 149 |
+

|
| 150 |
+
Fig. 3: Samples Visualization in Occ-ScanNet Benchmark. The original RGB image is shown in column 1,3 and 5, the corresponding scene voxel labels is shown in column 2, 4 and 6. The first two rows are different views from different scenes and the last two rows each is three different views from the same scene.
|
| 151 |
+
|
| 152 |
+
Manual inspection. The manual inspection involves verifying camera poses, camera intrinsic, and voxel positions. We filter out erroneous samples to prevent convergence issues in the model training process.
|
| 153 |
+
|
| 154 |
+
Voxel label assignment. For each frame, only a specific area in front of the camera is defined for analysis. Therefore, the selection of the voxel origin is paramount, as it determines the subsequent coordinates of other voxel. Subsequently, each voxel is assigned a label based on its nearest voxel in the CompleteScanNet dataset. Additionally, we excluded frames with a ratio of unknown or empty labels exceeding $95\%$ or frames where the number of semantic label classes was less than two. Consequently, we obtained the generated 3D voxel labels for each frame.
|
| 155 |
+
|
| 156 |
+
# 5 Experiments
|
| 157 |
+
|
| 158 |
+
# 5.1 Setup
|
| 159 |
+
|
| 160 |
+
Datasets. The NYUv2 [29] dataset provides scenes represented in $240 \times 144 \times 240$ voxels grids, labeled with 13 classes, including 1 for free space, 1 for unknown,
|
| 161 |
+
|
| 162 |
+

|
| 163 |
+
Fig. 4: Pipeline of Occ-ScanNet dataset label generation. Color images, depth images, camera intrinsic and poses are extracted from ScanNet scenes. For each scene, 100 frames were sampled and randomly split into training and validation sets with a $7/3$ ratio. Frames with invalid camera poses or exceeding scene boundaries were filtered out. Only the area in front of the camera was analyzed, necessitating careful selection of the voxel origin. Voxel were labeled based on their nearest voxel in the CompleteScanNet dataset. Frames with $>95\%$ unknown/empty labels or $<2$ semantic classes were excluded, resulting in generated 3D voxel labels for each frame.
|
| 164 |
+
|
| 165 |
+
and 11 for specific semantics (ceiling, floor, wall, window, chair, bed, sofa, table, tvs, furniture, objects). The dataset consists of 795 / 654 scenes in the train / test splits. The model is trained and evaluated on down-sampled $60 \times 36 \times 60$ voxels.
|
| 166 |
+
|
| 167 |
+
Occ-ScanNet dataset provides scenes represented in $60 \times 60 \times 36$ voxel grids, labeled with 12 classes including 1 for free space, and 11 for specific semantics (ceiling, floor, wall, window, chair, bed, sofa, table, tvs, furniture, objects). The dataset comprises 45,755 / 19,764 frames in the train / val splits. The model is trained and evaluated on the original scale.
|
| 168 |
+
|
| 169 |
+
Occ-ScanNet-mini dataset has the same class number and voxel scene size as Occ-ScanNet, except that this mini dataset consists of 4,639 / 2,007 frames in the train / val splits. The model is also trained and evaluated on the original scale.
|
| 170 |
+
|
| 171 |
+
Implementation Details. We employ a pre-trained EfficientNet-B7 [31] as the encoder in our 2D UNet architecture. In the depth branch, we utilize a pre-trained DepthAnything model [44], which remains frozen during training. Additionally, we integrate a depth loss specific to the depth branch, complementing the other losses outlined in Monoscene [5]. Our model is trained on two datasets: NYUv2 [29] and Occ-ScanNet. For NYUv2, ISO undergoes 30 epochs of training using AdamW optimizers. Initially, the learning rate is set to 5e-6 for the DepthNet and 1e-4 for other components, with a learning rate reduction by 0.1 at epoch 20. Training on NYUv2 takes approximately 7 hours using 2 NVIDIA L20 GPUs (2 items per GPU). On the Occ-ScanNet-mini dataset, ISO is trained for 60 epochs under similar learning rate settings, decreasing the rate by 0.1 at epoch 40. This training process takes around 12 hours using 8 L20 GPUs. For Occ-ScanNet, ISO is trained for 10 epochs with the same learning rate schedule as used for Occ-ScanNet-mini, requiring approximately a day using 8 L20 GPUs.
|
| 172 |
+
|
| 173 |
+
# 5.2 Main Results
|
| 174 |
+
|
| 175 |
+
Table 1: Performance on the Occ-ScanNet dataset
|
| 176 |
+
|
| 177 |
+
<table><tr><td>Method</td><td>Input</td><td>IoU</td><td>ceiling</td><td>floor</td><td>wall</td><td>window</td><td>chair</td><td>bed</td><td>sofa</td><td>table</td><td>tvs</td><td>furniture</td><td>objects</td></tr><tr><td>MonoScene* [5]</td><td>xrgb</td><td>41.60</td><td>15.17</td><td>44.71</td><td>22.41</td><td>12.55</td><td>26.11</td><td>27.03</td><td>35.91</td><td>28.32</td><td>6.57</td><td>32.16</td><td>19.84</td></tr><tr><td>ISO(Ours)</td><td>xrgb</td><td>42.16</td><td>19.88</td><td>41.88</td><td>22.37</td><td>16.98</td><td>29.09</td><td>42.43</td><td>42.00</td><td>29.60</td><td>10.62</td><td>36.36</td><td>24.61</td></tr></table>
|
| 178 |
+
|
| 179 |
+
Occ-ScanNet performance. We first evaluate our model's performance on the large-scale Occ-ScanNet dataset. As shown in Tab. 1, the results indicate that our method significantly outperforms MonoScene [5]. The * denotes results obtained using their code trained on our dataset. In Fig. 5, we also conducted qualitative visualization comparisons.
|
| 180 |
+
|
| 181 |
+
NYUv2 performance. The NYUv2 [29] dataset serves as a widely used benchmark for indoor scene evaluation. The results in Tab. 2 demonstrate that our method achieves state-of-the-art performance. Our method also demonstrates substantial performance improvement on the NYUv2 dataset, in addition to the gains observed on Occ-ScanNet. As illustrated in Fig. 6, our ISO demonstrates superior room layout prediction compared to [5], particularly evident in rows 1 and 2. This improvement can be attributed to incorporating depth knowledge into the model. Additionally, our method exhibits the capability of learning small objects, such as pictures on the wall in row 3.
|
| 182 |
+
|
| 183 |
+
Table 2: Performance on the NYUv2 dataset
|
| 184 |
+
|
| 185 |
+
<table><tr><td rowspan="2">Method</td><td rowspan="2">Input</td><td rowspan="2">IoU</td><td colspan="2">ceiling</td><td colspan="2">floor</td><td colspan="2">wall</td><td colspan="2">window</td><td colspan="2">chair</td><td colspan="2">bed</td><td colspan="2">sofa</td><td colspan="2">table</td><td colspan="2">tvs</td><td colspan="2">furniture</td><td colspan="2">objects</td><td></td></tr><tr><td colspan="2">■</td><td colspan="2">■</td><td colspan="2">■</td><td colspan="2">■</td><td colspan="2">■</td><td colspan="2">■</td><td colspan="2">■</td><td colspan="2">■</td><td colspan="2">■</td><td colspan="2">■</td><td colspan="2">■</td><td></td></tr><tr><td>LMSCNetrgb [28]</td><td>xocc</td><td>33.93</td><td colspan="2">4.49</td><td colspan="2">88.41</td><td colspan="2">4.63</td><td colspan="2">0.25</td><td colspan="2">3.94</td><td colspan="2">32.03</td><td colspan="2">15.44</td><td colspan="2">6.57</td><td colspan="2">0.02</td><td colspan="2">14.51</td><td colspan="2">4.39</td><td>15.88</td></tr><tr><td>AICNetrgb [18]</td><td>xrgb, xdepth</td><td>30.03</td><td colspan="2">7.58</td><td colspan="2">82.97</td><td colspan="2">9.15</td><td colspan="2">0.05</td><td colspan="2">6.93</td><td colspan="2">35.87</td><td colspan="2">22.92</td><td colspan="2">11.11</td><td colspan="2">0.71</td><td colspan="2">15.90</td><td colspan="2">6.45</td><td>18.15</td></tr><tr><td>3DSketchrgb [6]</td><td>xrgb, xTSDF</td><td>38.64</td><td colspan="2">8.53</td><td colspan="2">90.45</td><td colspan="2">9.94</td><td colspan="2">5.67</td><td colspan="2">10.64</td><td colspan="2">42.29</td><td colspan="2">29.21</td><td colspan="2">13.88</td><td colspan="2">9.38</td><td colspan="2">23.83</td><td colspan="2">8.19</td><td>22.91</td></tr><tr><td>MonoScene [5]</td><td>xrgb</td><td>42.51</td><td colspan="2">8.89</td><td colspan="2">93.50</td><td colspan="2">12.06</td><td colspan="2">12.57</td><td colspan="2">13.72</td><td colspan="2">48.19</td><td colspan="2">36.11</td><td colspan="2">15.13</td><td colspan="2">15.22</td><td colspan="2">27.96</td><td colspan="2">12.94</td><td>26.94</td></tr><tr><td>NDC-Scene [45]</td><td>xrgb</td><td>44.17</td><td colspan="2">12.02</td><td colspan="2">93.51</td><td colspan="2">13.11</td><td colspan="2">13.77</td><td colspan="2">15.83</td><td colspan="2">49.57</td><td colspan="2">39.87</td><td colspan="2">17.17</td><td colspan="2">24.57</td><td colspan="2">31.00</td><td colspan="2">14.96</td><td>29.03</td></tr><tr><td>ISO(Ours)</td><td>xrgb</td><td>47.11</td><td colspan="2">14.21</td><td colspan="2">93.47</td><td colspan="2">15.89</td><td colspan="2">15.14</td><td colspan="2">18.35</td><td colspan="2">50.01</td><td colspan="2">40.82</td><td colspan="2">18.25</td><td colspan="2">25.90</td><td colspan="2">34.08</td><td colspan="2">17.67</td><td>31.25</td></tr></table>
|
| 186 |
+
|
| 187 |
+

|
| 188 |
+
Image
|
| 189 |
+
|
| 190 |
+

|
| 191 |
+
MonoScene
|
| 192 |
+
|
| 193 |
+

|
| 194 |
+
Ours
|
| 195 |
+
|
| 196 |
+

|
| 197 |
+
Ground Truth
|
| 198 |
+
|
| 199 |
+

|
| 200 |
+
|
| 201 |
+

|
| 202 |
+
|
| 203 |
+

|
| 204 |
+
|
| 205 |
+

|
| 206 |
+
|
| 207 |
+

|
| 208 |
+
|
| 209 |
+

|
| 210 |
+
|
| 211 |
+

|
| 212 |
+
|
| 213 |
+

|
| 214 |
+
|
| 215 |
+

|
| 216 |
+
Fig. 5: Qualitative Analysis on the Occ-ScanNet Dataset. The input image is displayed on the left, while the predicted scene is shown in the middle two columns, and ground truth on the right column.
|
| 217 |
+
|
| 218 |
+

|
| 219 |
+
|
| 220 |
+

|
| 221 |
+
|
| 222 |
+

|
| 223 |
+
|
| 224 |
+
# 5.3 Ablation Study
|
| 225 |
+
|
| 226 |
+
We conduct ablation studies on the NYUv2 [29] and Occ-ScanNet-mini datasets to validate the effectiveness of our model design.
|
| 227 |
+
|
| 228 |
+
Depth fusion. Inspired by BEVDepth [20], we initially adopt the integration of BEV depth information, denoted as $^{+}$ bev-depth'. However, we observed that while BEV depth works well for 3D detection in outdoor scenes, directly applying it to indoor scenes is less effective. We speculate that this discrepancy arises because height information is more crucial in indoor scenes, where indoor heights often correspond to different 3D structures and semantics. Therefore, we adopt voxel depth fusion, denoted as $^{+}$ voxel-depth'. Specifically, we project each 3D voxel centroid to a frustum grid to sample the depth distribution probability
|
| 229 |
+
|
| 230 |
+

|
| 231 |
+
Fig. 6: Qualitative Analysis on the NYUv2 Dataset: The input is displayed on the left, with the camera viewing frustum shown in every image. ISO adeptly captures the scene layout and delineates various semantic instances more accurately. Notably, it excels in reconstructing the corners of walls (row 2) and accurately representing small objects like pictures on the wall (row 3).
|
| 232 |
+
|
| 233 |
+
of the voxel. We then weight the 2D feature at that pixel coordinate as the voxel feature. The results in Tab. 3 further validate that the voxel depth fusion approach is more effective.
|
| 234 |
+
|
| 235 |
+
The BEV method usually performs better in detecting ceilings and floors but performs poorly in detecting furniture such as windows, walls, televisions, beds, sofas, and tables. Ceilings and floors usually have relatively flat and regular geometric shapes, which makes their projections clearer and more consistent from a BEV perspective. In contrast, objects such as windows, walls, and televisions typically have more complex geometric shapes. In the BEV perspective, the projections of these objects may become blurry or overlapping, making it difficult to accurately distinguish and detect.
|
| 236 |
+
|
| 237 |
+
Table 3: Ablation study of depth fusion on the NYUv2 and Occ-ScanNet-mini dataset
|
| 238 |
+
|
| 239 |
+
<table><tr><td rowspan="2">Method</td><td rowspan="2">Input</td><td colspan="2">NYUv2</td><td colspan="2">Occ-ScanNet-mini</td></tr><tr><td>IoU</td><td>mIoU</td><td>IoU</td><td>mIoU</td></tr><tr><td>baseline</td><td>xrgb</td><td>42.27</td><td>27.13</td><td>50.94</td><td>38.95</td></tr><tr><td>+ bev-depth</td><td>xrgb</td><td>42.67</td><td>27.14</td><td>51.58</td><td>38.48</td></tr><tr><td>+ voxel-depth</td><td>xrgb</td><td>47.11</td><td>31.25</td><td>51.03</td><td>39.08</td></tr></table>
|
| 240 |
+
|
| 241 |
+
Table 4: Ablation of depth information on the NYUv2 dataset
|
| 242 |
+
|
| 243 |
+
<table><tr><td>depth-method</td><td>learned</td><td>multi-scale</td><td>IoU</td><td>mIoU</td></tr><tr><td>GT</td><td></td><td>✓</td><td>53.98</td><td>34.47</td></tr><tr><td>ZoeDepth</td><td>✓</td><td></td><td>45.18</td><td>29.15</td></tr><tr><td>ZoeDepth</td><td>✓</td><td>✓</td><td>45.24</td><td>29.40</td></tr><tr><td>DepthAnything</td><td></td><td></td><td>44.77</td><td>29.28</td></tr><tr><td>DepthAnything</td><td></td><td>✓</td><td>45.48</td><td>29.57</td></tr><tr><td>DepthAnything</td><td>✓</td><td>✓</td><td>46.94</td><td>31.02</td></tr><tr><td>DepthAnything</td><td>✓</td><td></td><td>47.11</td><td>31.25</td></tr></table>
|
| 244 |
+
|
| 245 |
+
Different depth pre-trained models. We conducted ablation experiments on the depth method using the Occ-ScanNet-mini dataset. Initially, we observed the upper bound of model performance when ground truth (GT) depth information was provided. Subsequently, we initialized our depth module using ZoeDepth [3] and Depth-Anything [44]. Experimental results indicate that utilizing Depth-Anything [44] yields better performance. Moreover, our approach allows for further enhancement by fine-tuning a pre-trained depth module, resulting in improved performance.
|
| 246 |
+
|
| 247 |
+
Multi-scale depth. In Tab. 4, we investigate the impact of using multi-scale depth distribution on the NYUv2 [29] dataset. Specifically, the multi-scale depth distribution is used to weight the multi-scale 3D voxel feature. Meanwhile, we tried a method that only uses the single-scale depth distribution map to generate the 3D voxel depth matrix. The used depth distribution map has a scale of $\frac{1}{8}$ original image scale. Only this single-scale 3D voxel matrix is used to weigh the summed-up 3D voxel feature out of the FLoSP module. Without multi-scale depth, all 3D features are simply added up and a single depth distribution is used for weighting. This method ignores the different importance that features at different scales may have, as all features at all scales are treated equally and all depth information is weighted by a unified depth distribution. With multiscale depth, the 3D features of each scale are first weighted with the depth distribution of their corresponding scale, and then these weighted features are added up. This method takes into account the differences in the importance of features at different scales, as each scale's feature is weighted based on its corresponding depth distribution.
|
| 248 |
+
|
| 249 |
+
In real-world scenarios, various objects and structures (such as ceilings, walls, furniture, etc.) often exhibit different scales. Hence, accounting for scale variations is crucial for accurately modeling these objects. Multi-scale depth can effectively capture these scale changes by assigning different weights to features at each scale. Depth information is indispensable for 3D scene completion. Leveraging the depth distribution information at each scale allows us to weight the corresponding 3D features, enabling the model to better understand and represent the objects in the scene and their relationships.
|
| 250 |
+
|
| 251 |
+
# 5.4 Discussion
|
| 252 |
+
|
| 253 |
+
Data scaling up. In Tab. 5, we delve into the impact of data volume on our method's performance by scaling up the number of scene samples. We trained the model on Occ-ScanNet using $10\%$ and $100\%$ scene samples respectively, and tested the model performance. The results indicate that larger dataset exhibits more significant gains.
|
| 254 |
+
|
| 255 |
+
Table 5: Scaling up comparison on the Occ-ScanNet
|
| 256 |
+
|
| 257 |
+
<table><tr><td rowspan="2">Data</td><td rowspan="2">Depth Branch</td><td colspan="2">Ours</td></tr><tr><td>IoU</td><td>mIoU</td></tr><tr><td>10%</td><td>✓</td><td>21.79</td><td>9.57</td></tr><tr><td>100%</td><td>✓</td><td>42.16</td><td>28.71</td></tr></table>
|
| 258 |
+
|
| 259 |
+
Limitations and future works. Although our model can estimate 3D occupancy effectively with the assistance of depth information, semantic learning still faces challenges due to class imbalance issues. Furthermore, our proposed Occ-ScanNet only considers 11 common semantic classes, which may not fully capture the diversity of categories present in real-world scenarios. Despite the significant increase in data volume compared to previous datasets, Occ-ScanNet remains limited in the number of scenes it covers. In future work, we will focus more on semantic exploration and conduct tests in more generalized scenarios.
|
| 260 |
+
|
| 261 |
+
# 6 Conclusion
|
| 262 |
+
|
| 263 |
+
In this paper, we introduce ISO, a novel method for monocular 3D occupancy prediction. ISO utilizes a D-FLoSP (Dual Feature Line of Sight Projection) and a multi-scale feature fusion strategy to address variations in scene and object sizes, specifically tailored for indoor scenes. Additionally, we present a new benchmark, Occ-ScanNet, aimed at fostering scalable studies of indoor scenes. We hope that this dataset will attract more attention to research on indoor occupancy prediction. While achieving promising results in 3D structure prediction, accurately identifying semantics remains a significant challenge in the future.
|
| 264 |
+
|
| 265 |
+
# Acknowledgement
|
| 266 |
+
|
| 267 |
+
This work was supported in part by the National Key R&D Program of China (No. 2022ZD0160102), the National Natural Science Foundation of China (No. U21B2042, No. 62320106010), and in part by the 2035 Innovation Program of CAS, and the InnoHK program.
|
| 268 |
+
|
| 269 |
+
# References
|
| 270 |
+
|
| 271 |
+
1. Arshad, M.S., Beksi, W.J.: List: Learning implicitly from spatial transformers for single-view 3d reconstruction. In: ICCV (2023)
|
| 272 |
+
2. Azuma, R.T.: A survey of augmented reality. Presence: teleoperators & virtual environments (1997)
|
| 273 |
+
3. Bhat, S.F., Birkl, R., Wofk, D., Wonka, P., Müller, M.: Zoedepth: Zero-shot transfer by combining relative and metric depth. arXiv preprint arXiv:2302.12288 (2023)
|
| 274 |
+
4. Birkl, R., Wofk, D., Müller, M.: Midas v3. 1-a model zoo for robust monocular relative depth estimation. arXiv preprint arXiv:2307.14460 (2023)
|
| 275 |
+
5. Cao, A.Q., de Charette, R.: Monoscene: Monocular 3d semantic scene completion. In: CVPR (2022)
|
| 276 |
+
6. Chen, X., Lin, K.Y., Qian, C., Zeng, G., Li, H.: 3d sketch-aware semantic scene completion via semi-supervised structure prior. In: CVPR (2020)
|
| 277 |
+
7. Dahnert, M., Hou, J., Nießner, M., Dai, A.: Panoptic 3d scene reconstruction from a single rgb image. NeurIPS (2021)
|
| 278 |
+
8. Dai, A., Chang, A.X., Savva, M., Halber, M., Funkhouser, T., Nießner, M.: Scannet: Richly-annotated 3d reconstructions of indoor scenes. In: CVPR (2017)
|
| 279 |
+
9. Denninger, M., Triebel, R.: 3d scene reconstruction from a single viewpoint. In: ECCV (2020)
|
| 280 |
+
10. DeSouza, G.N., Kak, A.C.: Vision for mobile robot navigation: A survey. TPAMI (2002)
|
| 281 |
+
1. Fan, H., Su, H., Guibas, L.J.: A point set generation network for 3d object reconstruction from a single image. In: CVPR (2017)
|
| 282 |
+
2. Fieraru, M., Zanfir, M., Oneata, E., Popa, A.I., Olaru, V., Sminchisescu, C.: Reconstructing three-dimensional models of interacting humans. arXiv preprint arXiv:2308.01854 (2023)
|
| 283 |
+
3. Goel, S., Pavlakos, G., Rajasegaran, J., Kanazawa, A., Malik, J.: Humans in 4d: Reconstructing and tracking humans with transformers. arXiv preprint arXiv:2305.20091 (2023)
|
| 284 |
+
4. Guan, H., Song, C., Zhang, Z.: Gramo: geometric resampling augmentation for monocular 3d object detection. Frontiers of Computer Science (2024)
|
| 285 |
+
5. Huang, J., Huang, G., Zhu, Z., Ye, Y., Du, D.: Bevdet: High-performance multicamera 3d object detection in bird-eye-view. arXiv preprint arXiv:2112.11790 (2021)
|
| 286 |
+
6. Huang, Y., Zheng, W., Zhang, B., Zhou, J., Lu, J.: Selfocc: Self-supervised vision-based 3d occupancy prediction. In: CVPR (2024)
|
| 287 |
+
7. Huang, Y., Zheng, W., Zhang, Y., Zhou, J., Lu, J.: Tri-perspective view for vision-based 3d semantic occupancy prediction. In: CVPR (2023)
|
| 288 |
+
8. Li, J., Han, K., Wang, P., Liu, Y., Yuan, X.: Anisotropic convolutional networks for 3d semantic scene completion. In: CVPR (2020)
|
| 289 |
+
9. Li, Y., Yu, Z., Choy, C., Xiao, C., Alvarez, J.M., Fidler, S., Feng, C., Anandkumar, A.: Voxformer: Sparse voxel transformer for camera-based 3d semantic scene completion. In: CVPR (2023)
|
| 290 |
+
20. Li, Y., Ge, Z., Yu, G., Yang, J., Wang, Z., Shi, Y., Sun, J., Li, Z.: Bevdepth: Acquisition of reliable depth for multi-view 3d object detection. In: AAAI (2023)
|
| 291 |
+
21. Li, Z., Wang, W., Li, H., Xie, E., Sima, C., Lu, T., Qiao, Y., Dai, J.: Bevformer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. In: ECCV (2022)
|
| 292 |
+
|
| 293 |
+
22. Li, Z., Yu, Z., Austin, D., Fang, M., Lan, S., Kautz, J., Alvarez, J.M.: Fb-occ: 3d occupancy prediction based on forward-backward view transformation. arXiv preprint arXiv:2307.01492 (2023)
|
| 294 |
+
23. Liu, S., Hu, Y., Zeng, Y., Tang, Q., Jin, B., Han, Y., Li, X.: See and think: Disentangling semantic scene completion. NeurIPS (2018)
|
| 295 |
+
24. Murez, Z., Van As, T., Bartolozzi, J., Sinha, A., Badrinarayanan, V., Rabinovich, A.: Atlas: End-to-end 3d scene reconstruction from posed images. In: ECCV (2020)
|
| 296 |
+
25. Pan, M., Liu, J., Zhang, R., Huang, P., Li, X., Liu, L., Zhang, S.: Renderocc: Visioncentric 3d occupancy prediction with 2d rendering supervision. arXiv preprint arXiv:2309.09502 (2023)
|
| 297 |
+
26. Park, J.J., Florence, P., Straub, J., Newcombe, R., Lovegrove, S.: Deepsdf: Learning continuous signed distance functions for shape representation. In: CVPR (2019)
|
| 298 |
+
27. Ranftl, R., Lasinger, K., Hafner, D., Schindler, K., Koltun, V.: Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer. TPAMI (2020)
|
| 299 |
+
28. Roldao, L., de Charette, R., Verroust-Blondet, A.: Lmscnet: Lightweight multiscale 3d semantic completion. In: 3DV (2020)
|
| 300 |
+
29. Silberman, N., Hoiem, D., Kohli, P., Fergus, R.: Indoor segmentation and support inference from rgbd images. In: ECCV (2012)
|
| 301 |
+
30. Song, S., Yu, F., Zeng, A., Chang, A.X., Savva, M., Funkhouser, T.: Semantic scene completion from a single depth image. In: CVPR (2017)
|
| 302 |
+
31. Tan, M., Le, Q.: Efficientnet: Rethinking model scaling for convolutional neural networks. In: ICML (2019)
|
| 303 |
+
32. Tang, Y., Dorn, S., Savani, C.: Center3d: Center-based monocular 3d object detection with joint depth understanding. In: DAGM GCPR (2020)
|
| 304 |
+
33. Tian, X., Jiang, T., Yun, L., Mao, Y., Yang, H., Wang, Y., Wang, Y., Zhao, H.: Occ3d: A large-scale 3d occupancy prediction benchmark for autonomous driving. NeurIPS (2023)
|
| 305 |
+
34. Tong, W., Sima, C., Wang, T., Chen, L., Wu, S., Deng, H., Gu, Y., Lu, L., Luo, P., Lin, D., et al.: Scene as occupancy. In: ICCV (2023)
|
| 306 |
+
35. Wang, S., Liu, Y., Wang, T., Li, Y., Zhang, X.: Exploring object-centric temporal modeling for efficient multi-view 3d object detection. In: ICCV (2023)
|
| 307 |
+
36. Wang, X., Zhu, Z., Xu, W., Zhang, Y., Wei, Y., Chi, X., Ye, Y., Du, D., Lu, J., Wang, X.: Openoccupancy: A large scale benchmark for surrounding semantic occupancy perception. In: ICCV (2023)
|
| 308 |
+
37. Wang, Y., Guizilini, V.C., Zhang, T., Wang, Y., Zhao, H., Solomon, J.: Detr3d: 3d object detection from multi-view images via 3d-to-2d queries. In: CoRL (2022)
|
| 309 |
+
38. Wang, Y., Chen, Y., Liao, X., Fan, L., Zhang, Z.: Panoocc: Unified occupancy representation for camera-based 3d panoptic segmentation. In: CVPR (2024)
|
| 310 |
+
39. Wang, Y., Chen, Y., Zhang, Z.: Frustumformer: Adaptive instance-aware resampling for multi-view 3d detection. In: CVPR (2023)
|
| 311 |
+
40. Wei, Y., Zhao, L., Zheng, W., Zhu, Z., Zhou, J., Lu, J.: Surroundocc: Multi-camera 3d occupancy prediction for autonomous driving. In: ICCV (2023)
|
| 312 |
+
41. Wu, Q., Wang, K., Li, K., Zheng, J., Cai, J.: Objectsdf++: Improved object-compositional neural implicit surfaces. In: ICCV (2023)
|
| 313 |
+
42. Wu, S.C., Tateno, K., Navab, N., Tombari, F.: Scfusion: Real-time incremental scene reconstruction with semantic completion. In: 3DV (2020)
|
| 314 |
+
43. Yang, C., Chen, Y., Tian, H., Tao, C., Zhu, X., Zhang, Z., Huang, G., Li, H., Qiao, Y., Lu, L., et al.: Bevformer v2: Adapting modern image backbones to bird's-eye-view recognition via perspective supervision. In: CVPR (2023)
|
| 315 |
+
|
| 316 |
+
44. Yang, L., Kang, B., Huang, Z., Xu, X., Feng, J., Zhao, H.: Depth anything: Unleashing the power of large-scale unlabeled data. In: CVPR (2024)
|
| 317 |
+
45. Yao, J., Li, C., Sun, K., Cai, Y., Li, H., Ouyang, W., Li, H.: Ndc-scene: Boost monocular 3d semantic scene completion in normalized device coordinates space. In: ICCV (2023)
|
| 318 |
+
46. Yu, Z., Shu, C., Deng, J., Lu, K., Liu, Z., Yu, J., Yang, D., Li, H., Chen, Y.: Flashocc: Fast and memory-efficient occupancy prediction via channel-to-height plugin. arXiv preprint arXiv:2311.12058 (2023)
|
| 319 |
+
47. Yurtsever, E., Lambert, J., Carballo, A., Takeda, K.: A survey of autonomous driving: Common practices and emerging technologies. IEEE access (2020)
|
| 320 |
+
48. Zhang, J., Zhao, H., Yao, A., Chen, Y., Zhang, L., Liao, H.: Efficient semantic scene completion network with spatial group convolution. In: ECCV (2018)
|
| 321 |
+
49. Zhang, X., Bi, S., Sunkavalli, K., Su, H., Xu, Z.: Nerfusion: Fusing radiance fields for large-scale scene reconstruction. In: CVPR (2022)
|
| 322 |
+
50. Zheng, Z., Yu, T., Wei, Y., Dai, Q., Liu, Y.: Deephuman: 3d human reconstruction from a single image. In: ICCV (2019)
|
| 323 |
+
51. Zhong, M., Zeng, G.: Semantic point completion network for 3d semantic scene completion. In: ECAI (2020)
|
2024/Monocular Occupancy Prediction for Scalable Indoor Scenes/images.zip
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:ce7b980313a267490a07d464220bbb8b2d92f000251265a535f97ffcf1e3cd32
|
| 3 |
+
size 541269
|
2024/Monocular Occupancy Prediction for Scalable Indoor Scenes/layout.json
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
2024/Motion Aware Event Representation-driven Image Deblurring/a9105a02-b537-456a-9778-80b745c423ab_content_list.json
ADDED
|
@@ -0,0 +1,2161 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
{
|
| 3 |
+
"type": "text",
|
| 4 |
+
"text": "Motion Aware Event Representation-driven Image Deblurring",
|
| 5 |
+
"text_level": 1,
|
| 6 |
+
"bbox": [
|
| 7 |
+
254,
|
| 8 |
+
140,
|
| 9 |
+
750,
|
| 10 |
+
186
|
| 11 |
+
],
|
| 12 |
+
"page_idx": 0
|
| 13 |
+
},
|
| 14 |
+
{
|
| 15 |
+
"type": "text",
|
| 16 |
+
"text": "Zhijing Sun, Xueyang Fu, Longzhuo Huang, Aiping Liu, and Zheng-Jun Zha*",
|
| 17 |
+
"bbox": [
|
| 18 |
+
223,
|
| 19 |
+
212,
|
| 20 |
+
777,
|
| 21 |
+
228
|
| 22 |
+
],
|
| 23 |
+
"page_idx": 0
|
| 24 |
+
},
|
| 25 |
+
{
|
| 26 |
+
"type": "text",
|
| 27 |
+
"text": "School of Information Science and Technology and MoE Key Laboratory of Brain-inspired Intelligent Perception and Cognition, University of Science and Technology of China, Hefei, 230026, China {sunzhijing, hlz_0}@mail.ustc.edu.cn {xyfu, aipingl, zhazj}@ustc.edu.cn",
|
| 28 |
+
"bbox": [
|
| 29 |
+
250,
|
| 30 |
+
239,
|
| 31 |
+
753,
|
| 32 |
+
310
|
| 33 |
+
],
|
| 34 |
+
"page_idx": 0
|
| 35 |
+
},
|
| 36 |
+
{
|
| 37 |
+
"type": "text",
|
| 38 |
+
"text": "Abstract. Traditional image deblurring struggles with high-quality reconstruction due to limited motion data from single blurred images. Excitingly, the high-temporal resolution of event cameras records motion more precisely in a different modality, transforming image deblurring. However, many event camera-based methods, which only care about the final value of the polarity accumulation, ignore the influence of the absolute intensity change where events generate so fall short in perceiving motion patterns and effectively aiding image reconstruction. To overcome this, in this work, we propose a new event preprocessing technique that accumulates the deviation from the initial moment each time the event is updated. This process can distinguish the order of events to improve the perception of object motion patterns. To complement our proposed event representation, we create a recurrent module designed to meticulously extract motion features across local and global time scales. To further facilitate the event feature and image feature integration, which assists in image reconstruction, we develop a bi-directional feature alignment and fusion module. This module works to lessen inter-modal inconsistencies. Our approach has been thoroughly tested through rigorous experiments carried out on several datasets with different distributions. These trials have delivered promising results, with our method achieving top-tier performance in both quantitative and qualitative assessments. Code is available at https://github.com/ZhijingS/DA_event_deblur.",
|
| 39 |
+
"bbox": [
|
| 40 |
+
261,
|
| 41 |
+
348,
|
| 42 |
+
738,
|
| 43 |
+
654
|
| 44 |
+
],
|
| 45 |
+
"page_idx": 0
|
| 46 |
+
},
|
| 47 |
+
{
|
| 48 |
+
"type": "text",
|
| 49 |
+
"text": "Keywords: Image deblurring $\\cdot$ Event camera $\\cdot$ Event representation",
|
| 50 |
+
"bbox": [
|
| 51 |
+
261,
|
| 52 |
+
666,
|
| 53 |
+
723,
|
| 54 |
+
681
|
| 55 |
+
],
|
| 56 |
+
"page_idx": 0
|
| 57 |
+
},
|
| 58 |
+
{
|
| 59 |
+
"type": "text",
|
| 60 |
+
"text": "1 Introduction",
|
| 61 |
+
"text_level": 1,
|
| 62 |
+
"bbox": [
|
| 63 |
+
217,
|
| 64 |
+
708,
|
| 65 |
+
374,
|
| 66 |
+
723
|
| 67 |
+
],
|
| 68 |
+
"page_idx": 0
|
| 69 |
+
},
|
| 70 |
+
{
|
| 71 |
+
"type": "text",
|
| 72 |
+
"text": "Frame-based cameras often produce blurred images due to factors like camera shake or object motion, given their configured exposure time. Such blur, a common form of image degradation, occurs across various scenes. While photographers typically aim for clear and focused photos, clear inputs are equally critical for numerous computer vision tasks. These tasks include super-resolution, target",
|
| 73 |
+
"bbox": [
|
| 74 |
+
212,
|
| 75 |
+
739,
|
| 76 |
+
787,
|
| 77 |
+
816
|
| 78 |
+
],
|
| 79 |
+
"page_idx": 0
|
| 80 |
+
},
|
| 81 |
+
{
|
| 82 |
+
"type": "page_footnote",
|
| 83 |
+
"text": "*Corresponding author",
|
| 84 |
+
"bbox": [
|
| 85 |
+
232,
|
| 86 |
+
825,
|
| 87 |
+
388,
|
| 88 |
+
839
|
| 89 |
+
],
|
| 90 |
+
"page_idx": 0
|
| 91 |
+
},
|
| 92 |
+
{
|
| 93 |
+
"type": "image",
|
| 94 |
+
"img_path": "images/5bfd42e747bf5dbc6e4d8cc6284f047ed99a601db36a9a27c56293f41bf9dedb.jpg",
|
| 95 |
+
"image_caption": [
|
| 96 |
+
"(a) GT (PSNR |SSIM)"
|
| 97 |
+
],
|
| 98 |
+
"image_footnote": [],
|
| 99 |
+
"bbox": [
|
| 100 |
+
217,
|
| 101 |
+
143,
|
| 102 |
+
312,
|
| 103 |
+
215
|
| 104 |
+
],
|
| 105 |
+
"page_idx": 1
|
| 106 |
+
},
|
| 107 |
+
{
|
| 108 |
+
"type": "image",
|
| 109 |
+
"img_path": "images/17adc58bc8c7075408d48df65f7dadc8389c5148ab985309ad083c9396351c0f.jpg",
|
| 110 |
+
"image_caption": [
|
| 111 |
+
"(b) Blurry (28.62dB |0.819)"
|
| 112 |
+
],
|
| 113 |
+
"image_footnote": [],
|
| 114 |
+
"bbox": [
|
| 115 |
+
313,
|
| 116 |
+
143,
|
| 117 |
+
406,
|
| 118 |
+
215
|
| 119 |
+
],
|
| 120 |
+
"page_idx": 1
|
| 121 |
+
},
|
| 122 |
+
{
|
| 123 |
+
"type": "image",
|
| 124 |
+
"img_path": "images/9296ea44d1b5bdf6a1adb5c182b71bd016571a194d8379b57fef5dee10f83353.jpg",
|
| 125 |
+
"image_caption": [
|
| 126 |
+
"(c) HINet [4] (29.14dB |0.837)",
|
| 127 |
+
"Fig. 1: A deblurring example on HS-ERGB [37] dataset via recent state-of-the-art methods. Compared to other methods, our approach yields a more distinct result owing to its enhanced motion perception capability."
|
| 128 |
+
],
|
| 129 |
+
"image_footnote": [],
|
| 130 |
+
"bbox": [
|
| 131 |
+
408,
|
| 132 |
+
143,
|
| 133 |
+
500,
|
| 134 |
+
215
|
| 135 |
+
],
|
| 136 |
+
"page_idx": 1
|
| 137 |
+
},
|
| 138 |
+
{
|
| 139 |
+
"type": "image",
|
| 140 |
+
"img_path": "images/13044559a5ade1896c43bb1813ecb568a3239ac67268cd28c59ff07c03370df3.jpg",
|
| 141 |
+
"image_caption": [
|
| 142 |
+
"(d) NAFNet [3] (30.41dB |0.855)"
|
| 143 |
+
],
|
| 144 |
+
"image_footnote": [],
|
| 145 |
+
"bbox": [
|
| 146 |
+
501,
|
| 147 |
+
143,
|
| 148 |
+
596,
|
| 149 |
+
215
|
| 150 |
+
],
|
| 151 |
+
"page_idx": 1
|
| 152 |
+
},
|
| 153 |
+
{
|
| 154 |
+
"type": "image",
|
| 155 |
+
"img_path": "images/473f75e245093e693cdddd4d784fb33482225b41569de5edb3f7f85fa30dcde8.jpg",
|
| 156 |
+
"image_caption": [
|
| 157 |
+
"(e) EFNet [33] (30.76dB |0.856)"
|
| 158 |
+
],
|
| 159 |
+
"image_footnote": [],
|
| 160 |
+
"bbox": [
|
| 161 |
+
598,
|
| 162 |
+
143,
|
| 163 |
+
691,
|
| 164 |
+
215
|
| 165 |
+
],
|
| 166 |
+
"page_idx": 1
|
| 167 |
+
},
|
| 168 |
+
{
|
| 169 |
+
"type": "image",
|
| 170 |
+
"img_path": "images/0ad319576978f72a3e13af2d0e95f04fc6e4cb45953156a234d058cded602fea.jpg",
|
| 171 |
+
"image_caption": [
|
| 172 |
+
"(f) Ours (31.01dB [0.858]"
|
| 173 |
+
],
|
| 174 |
+
"image_footnote": [],
|
| 175 |
+
"bbox": [
|
| 176 |
+
694,
|
| 177 |
+
143,
|
| 178 |
+
785,
|
| 179 |
+
215
|
| 180 |
+
],
|
| 181 |
+
"page_idx": 1
|
| 182 |
+
},
|
| 183 |
+
{
|
| 184 |
+
"type": "text",
|
| 185 |
+
"text": "detection, and semantic segmentation, to name a few. Hence, image deblurring serves as a fundamental task in the realm of low-level computer vision, reinforcing the image quality and supporting the proper function of various algorithms.",
|
| 186 |
+
"bbox": [
|
| 187 |
+
212,
|
| 188 |
+
337,
|
| 189 |
+
784,
|
| 190 |
+
383
|
| 191 |
+
],
|
| 192 |
+
"page_idx": 1
|
| 193 |
+
},
|
| 194 |
+
{
|
| 195 |
+
"type": "text",
|
| 196 |
+
"text": "Traditional image-based techniques often attempt to derive a clear image from a single blurred one. They do this by leveraging natural image priors or making assumptions about blur operations [6, 8, 13, 15, 34, 44]. However, in capturing a single blurred image, much motion information is lost. This lack of motion information during exposure creates undesirable artifacts in the deblurred results, especially in complex motion scenarios, as illustrated in Figures 1c and 1d. To counter this challenge, recent methodologies use more temporal information, such as differently exposed frames [2, 46, 48] and neighboring frames [16, 28, 32]. However, these approaches have their limitations. For instance, the use of differently exposed frames is constrained by the low temporal resolution of frame-based cameras, improving the reconstruction effect primarily in slow-motion scenes. The method that uses adjacent frames to estimate the current frame's motion information can lead to accumulative errors. In summary, traditional deblurring techniques encounter hurdles due to the inability of frame-based cameras to capture motion information during exposure.",
|
| 197 |
+
"bbox": [
|
| 198 |
+
212,
|
| 199 |
+
385,
|
| 200 |
+
787,
|
| 201 |
+
611
|
| 202 |
+
],
|
| 203 |
+
"page_idx": 1
|
| 204 |
+
},
|
| 205 |
+
{
|
| 206 |
+
"type": "text",
|
| 207 |
+
"text": "As opposed to frame-based cameras, event cameras, inspired by biological systems, offer a promising solution to this issue. These cameras' microsecond-level temporal resolution $(\\mu \\mathrm{s})$ and asynchronous output architecture allow the recording of fast motions. This capability enables event streams to store pixel intensity changes and provide accurate motion information to deblur the corresponding frame. By using event frames as additional input, some methods attempt to guide the deblurring process through these event cameras [9,25,39,43,49]. One method [45] subdivides the event stream into ternary 2D data and generates a mask to localize high-blur regions, thereby enhancing the network's ability to distinguish different blur levels. Another method [33] symmetrically accumulates event polarities and fuses event and image features using an attention mechanism, enabling the integration of motion information into image reconstruction. However, these methods primarily employ the event stream in a manner akin to polarity accumulation. In essence, they consider only the final accumulation result of relative intensity change between subsequent moments, neglecting the ab",
|
| 208 |
+
"bbox": [
|
| 209 |
+
212,
|
| 210 |
+
613,
|
| 211 |
+
787,
|
| 212 |
+
840
|
| 213 |
+
],
|
| 214 |
+
"page_idx": 1
|
| 215 |
+
},
|
| 216 |
+
{
|
| 217 |
+
"type": "page_number",
|
| 218 |
+
"text": "2",
|
| 219 |
+
"bbox": [
|
| 220 |
+
217,
|
| 221 |
+
114,
|
| 222 |
+
228,
|
| 223 |
+
126
|
| 224 |
+
],
|
| 225 |
+
"page_idx": 1
|
| 226 |
+
},
|
| 227 |
+
{
|
| 228 |
+
"type": "header",
|
| 229 |
+
"text": "Z. Sun et al.",
|
| 230 |
+
"bbox": [
|
| 231 |
+
271,
|
| 232 |
+
114,
|
| 233 |
+
357,
|
| 234 |
+
127
|
| 235 |
+
],
|
| 236 |
+
"page_idx": 1
|
| 237 |
+
},
|
| 238 |
+
{
|
| 239 |
+
"type": "text",
|
| 240 |
+
"text": "solute intensity change caused by historical events. Consequently, this approach fails to differentiate between various motion patterns, injecting unnecessary ambiguity into the deblurring process, and leading to edge artifacts as displayed in Figure 1e.",
|
| 241 |
+
"bbox": [
|
| 242 |
+
212,
|
| 243 |
+
146,
|
| 244 |
+
782,
|
| 245 |
+
205
|
| 246 |
+
],
|
| 247 |
+
"page_idx": 2
|
| 248 |
+
},
|
| 249 |
+
{
|
| 250 |
+
"type": "text",
|
| 251 |
+
"text": "Given the limited perception of motion during exposure offered by polarity accumulation, we posit that the deviation from the initial moment provides more precise motion information for image deblurring. Each time the event generates, the absolute light intensity changes. The deviation from the initial moment enables us to consider the impact of changes in absolute light intensity, even when the absolute light intensity at the initial moment is unknown. Thus, the deviation, to a degree, implies the order of arrival of the events, offering different information for varying motion patterns. As a result, we propose the Deviation Accumulation (DA) method to enhance motion perception. Specifically, every time a new event arrives, our method accumulates the deviation relative to the initial moment, thus encompassing more comprehensive motion information. Furthermore, we find that longer time scales for event accumulation can capture global motion patterns, while shorter scales precisely depict local fast-motion patterns. To better perceive motion at both global and local time scales, we have designed a Recurrent Motion Extraction (RME) module that works with corresponding event representations. Moreover, in line with multimodal learning [18-20, 29, 36, 42], we present an effective Feature Alignment and Fusion (FAF) module. This module serves to address the issue that simple fusion methods struggle with the disparate nature of event and frame modalities. As a result, these two modalities align, ensuring better information interaction. We integrate these modules into a single image deblurring network. This network extracts the motion features within the exposure time from the event stream and combines them with the texture information supplied by the image to complete the motion blur removal, as illustrated in Figure 1f.",
|
| 252 |
+
"bbox": [
|
| 253 |
+
212,
|
| 254 |
+
207,
|
| 255 |
+
787,
|
| 256 |
+
568
|
| 257 |
+
],
|
| 258 |
+
"page_idx": 2
|
| 259 |
+
},
|
| 260 |
+
{
|
| 261 |
+
"type": "text",
|
| 262 |
+
"text": "The key contributions of this paper are as follows:",
|
| 263 |
+
"bbox": [
|
| 264 |
+
238,
|
| 265 |
+
569,
|
| 266 |
+
601,
|
| 267 |
+
583
|
| 268 |
+
],
|
| 269 |
+
"page_idx": 2
|
| 270 |
+
},
|
| 271 |
+
{
|
| 272 |
+
"type": "list",
|
| 273 |
+
"sub_type": "text",
|
| 274 |
+
"list_items": [
|
| 275 |
+
"- We introduce an event preprocessing method designed specifically for the deblurring problem. The deviation accumulation is capable of responding distinctively to various motion patterns, thereby adding accurate motion information for image deblurring tasks.",
|
| 276 |
+
"- We develop a network equipped with recurrent motion extraction modules, designed to better extract motion features on local and global time scales. This design enables the network to handle multiple composite motion scenarios effectively.",
|
| 277 |
+
"- We propose a feature alignment and fusion module to mitigate the impacts of inter-modal inconsistencies and ensure that motion information more effectively guides the image deblurring process."
|
| 278 |
+
],
|
| 279 |
+
"bbox": [
|
| 280 |
+
227,
|
| 281 |
+
592,
|
| 282 |
+
785,
|
| 283 |
+
756
|
| 284 |
+
],
|
| 285 |
+
"page_idx": 2
|
| 286 |
+
},
|
| 287 |
+
{
|
| 288 |
+
"type": "text",
|
| 289 |
+
"text": "Our experimental results indicate that our method achieves state-of-the-art performance on synthetic, semi-synthetic, and real-world datasets, including a 0.61dB improvement in PSNR on the GoPro dataset compared to previous studies [33]. Moreover, our method produces more distinct subjective results, as demonstrated in Figure 1.",
|
| 290 |
+
"bbox": [
|
| 291 |
+
212,
|
| 292 |
+
763,
|
| 293 |
+
787,
|
| 294 |
+
839
|
| 295 |
+
],
|
| 296 |
+
"page_idx": 2
|
| 297 |
+
},
|
| 298 |
+
{
|
| 299 |
+
"type": "header",
|
| 300 |
+
"text": "Motion Aware Event Representation-driven Image Deblurring",
|
| 301 |
+
"bbox": [
|
| 302 |
+
321,
|
| 303 |
+
114,
|
| 304 |
+
730,
|
| 305 |
+
128
|
| 306 |
+
],
|
| 307 |
+
"page_idx": 2
|
| 308 |
+
},
|
| 309 |
+
{
|
| 310 |
+
"type": "page_number",
|
| 311 |
+
"text": "3",
|
| 312 |
+
"bbox": [
|
| 313 |
+
774,
|
| 314 |
+
116,
|
| 315 |
+
784,
|
| 316 |
+
126
|
| 317 |
+
],
|
| 318 |
+
"page_idx": 2
|
| 319 |
+
},
|
| 320 |
+
{
|
| 321 |
+
"type": "text",
|
| 322 |
+
"text": "2 Related Work",
|
| 323 |
+
"text_level": 1,
|
| 324 |
+
"bbox": [
|
| 325 |
+
215,
|
| 326 |
+
143,
|
| 327 |
+
388,
|
| 328 |
+
160
|
| 329 |
+
],
|
| 330 |
+
"page_idx": 3
|
| 331 |
+
},
|
| 332 |
+
{
|
| 333 |
+
"type": "text",
|
| 334 |
+
"text": "2.1 Event-based Motion Deblurring",
|
| 335 |
+
"text_level": 1,
|
| 336 |
+
"bbox": [
|
| 337 |
+
215,
|
| 338 |
+
176,
|
| 339 |
+
526,
|
| 340 |
+
193
|
| 341 |
+
],
|
| 342 |
+
"page_idx": 3
|
| 343 |
+
},
|
| 344 |
+
{
|
| 345 |
+
"type": "text",
|
| 346 |
+
"text": "The utilization of event information to assist in deblurring tasks has become increasingly common in recent research [9, 12, 25, 31, 39, 43, 49]. The high temporal resolution and low latency of event cameras not only capture edge positions but also embed moving temporal information within the event stream. This information is crucial to reconstruct the latent sharp image. Many methods have been attempted to explore the relationship between event information, blurred images, and sharp images. One such method [25] includes the Event-based Double Integral (EDI) model, which leverages the principles of event generation to formulate an optimization problem that establishes a correlation between a blurry image and a latent sharp image. In recent studies, the event stream is typically converted into event frames [30, 38], which simplifies the processing of event data within traditional CNN frameworks. [9] combines the motion feature extracted from event data with the image feature, using them as the input for the decoder. [50] employs a Multi-Layer Perceptron (MLP) to predict the fused feature value from event and image features. [33] integrates them through a cross-modal attention method, which generates a query matrix from the image features, but derives key and value matrices from the event features.",
|
| 347 |
+
"bbox": [
|
| 348 |
+
212,
|
| 349 |
+
202,
|
| 350 |
+
787,
|
| 351 |
+
458
|
| 352 |
+
],
|
| 353 |
+
"page_idx": 3
|
| 354 |
+
},
|
| 355 |
+
{
|
| 356 |
+
"type": "text",
|
| 357 |
+
"text": "Confronting the prevalent issue of integrating image and event features is the problem of inter-modal inconsistencies, which traditional fusion techniques cannot easily mitigate. To address this, our approach seeks to align these two modalities, optimizing the use of event data. Through this alignment, our goal is to minimize the inconsistencies spawned when merging differing modalities, thus amplifying the efficacy of the deblurring process.",
|
| 358 |
+
"bbox": [
|
| 359 |
+
212,
|
| 360 |
+
459,
|
| 361 |
+
789,
|
| 362 |
+
550
|
| 363 |
+
],
|
| 364 |
+
"page_idx": 3
|
| 365 |
+
},
|
| 366 |
+
{
|
| 367 |
+
"type": "text",
|
| 368 |
+
"text": "2.2 Event Representation",
|
| 369 |
+
"text_level": 1,
|
| 370 |
+
"bbox": [
|
| 371 |
+
215,
|
| 372 |
+
571,
|
| 373 |
+
441,
|
| 374 |
+
588
|
| 375 |
+
],
|
| 376 |
+
"page_idx": 3
|
| 377 |
+
},
|
| 378 |
+
{
|
| 379 |
+
"type": "text",
|
| 380 |
+
"text": "Event data, characterized by its spatial sparsity and strong temporal continuity due to its asynchronous generation, necessitates a custom-designed event representation to deliver the required information. Methods [11, 26, 41] that process events individually maximize the use of high temporal resolution, as suggested by [26], which proposed an adaptive spiking neuron model to manage the fluctuating event input. The effectiveness of these algorithms, however, can be significantly hampered by the noise within the event due to independent processing. A Time Surface [1, 14, 22, 51] is created using the timestamp of the last event at each pixel, which prominently reveals the rich temporal information contained within the event. Furthermore, the event frame [5, 7, 21, 23, 45] transforms the original 4D event data into a 2D frame by accumulating events on the number or the polarity pixel-wise, facilitating extraction with conventional image processing techniques. In [45], the authors utilize ternary data for their event frames to allocate high blur region of the image.",
|
| 381 |
+
"bbox": [
|
| 382 |
+
212,
|
| 383 |
+
598,
|
| 384 |
+
787,
|
| 385 |
+
809
|
| 386 |
+
],
|
| 387 |
+
"page_idx": 3
|
| 388 |
+
},
|
| 389 |
+
{
|
| 390 |
+
"type": "text",
|
| 391 |
+
"text": "However, existing event frame approaches result in abundant light intensity change information being lost in the squeezing process, making it impossible to",
|
| 392 |
+
"bbox": [
|
| 393 |
+
215,
|
| 394 |
+
809,
|
| 395 |
+
787,
|
| 396 |
+
839
|
| 397 |
+
],
|
| 398 |
+
"page_idx": 3
|
| 399 |
+
},
|
| 400 |
+
{
|
| 401 |
+
"type": "page_number",
|
| 402 |
+
"text": "4",
|
| 403 |
+
"bbox": [
|
| 404 |
+
217,
|
| 405 |
+
114,
|
| 406 |
+
228,
|
| 407 |
+
126
|
| 408 |
+
],
|
| 409 |
+
"page_idx": 3
|
| 410 |
+
},
|
| 411 |
+
{
|
| 412 |
+
"type": "header",
|
| 413 |
+
"text": "Z. Sun et al.",
|
| 414 |
+
"bbox": [
|
| 415 |
+
271,
|
| 416 |
+
114,
|
| 417 |
+
357,
|
| 418 |
+
127
|
| 419 |
+
],
|
| 420 |
+
"page_idx": 3
|
| 421 |
+
},
|
| 422 |
+
{
|
| 423 |
+
"type": "text",
|
| 424 |
+
"text": "distinguish the order of event arrivals and lacking the ability to detect multiple motion patterns. Some techniques attempt to recover some of this lost data by dividing the exposure time into smaller segments, though this essentially equates to sampling the original data. Consequently, we design a motion-aware event representation specifically tailored for motion deblurring.",
|
| 425 |
+
"bbox": [
|
| 426 |
+
212,
|
| 427 |
+
146,
|
| 428 |
+
787,
|
| 429 |
+
223
|
| 430 |
+
],
|
| 431 |
+
"page_idx": 4
|
| 432 |
+
},
|
| 433 |
+
{
|
| 434 |
+
"type": "text",
|
| 435 |
+
"text": "3 Method",
|
| 436 |
+
"text_level": 1,
|
| 437 |
+
"bbox": [
|
| 438 |
+
215,
|
| 439 |
+
250,
|
| 440 |
+
330,
|
| 441 |
+
265
|
| 442 |
+
],
|
| 443 |
+
"page_idx": 4
|
| 444 |
+
},
|
| 445 |
+
{
|
| 446 |
+
"type": "text",
|
| 447 |
+
"text": "3.1 Overall Framework",
|
| 448 |
+
"text_level": 1,
|
| 449 |
+
"bbox": [
|
| 450 |
+
215,
|
| 451 |
+
284,
|
| 452 |
+
419,
|
| 453 |
+
299
|
| 454 |
+
],
|
| 455 |
+
"page_idx": 4
|
| 456 |
+
},
|
| 457 |
+
{
|
| 458 |
+
"type": "text",
|
| 459 |
+
"text": "The overall framework of our proposed method is depicted in Figure 2. We adopt a classical encoder-decoder architecture for our approach. Initially, the blurred image is inputted into the image branch, where we use the baseline conv block mentioned in [3] as our fundamental block, to extract relevant features. Simultaneously, the corresponding event frames generated by our deviation accumulation are fed into the Recurrent Motion Extraction (RME) module, which facilitates the extraction of features covering motion in local and global time scales. Following each block of the two branches, the Feature Alignment and Fusion (FAF) module is employed to eliminate the inconsistencies between the two modalities. By performing down-sampling operations, we are able to obtain both shallow and semantic information. In the decoder, skip connections are utilized to incorporate the extracted features for image reconstruction. The architecture of the conv block and the decoder is shown in the supplementary materials. The entire training process is conducted in an end-to-end manner, with the PSNR Loss function [4] employed to optimize the network's parameters.",
|
| 460 |
+
"bbox": [
|
| 461 |
+
212,
|
| 462 |
+
313,
|
| 463 |
+
787,
|
| 464 |
+
541
|
| 465 |
+
],
|
| 466 |
+
"page_idx": 4
|
| 467 |
+
},
|
| 468 |
+
{
|
| 469 |
+
"type": "text",
|
| 470 |
+
"text": "3.2 Deviation Accumulation Event Representation",
|
| 471 |
+
"text_level": 1,
|
| 472 |
+
"bbox": [
|
| 473 |
+
215,
|
| 474 |
+
566,
|
| 475 |
+
645,
|
| 476 |
+
580
|
| 477 |
+
],
|
| 478 |
+
"page_idx": 4
|
| 479 |
+
},
|
| 480 |
+
{
|
| 481 |
+
"type": "text",
|
| 482 |
+
"text": "Event cameras are bio-inspired sensors which asynchronously output signals. They trigger an event while the logarithm of the intensity change exceeds the preset threshold $c$ at any pixel, which is formulated as",
|
| 483 |
+
"bbox": [
|
| 484 |
+
212,
|
| 485 |
+
595,
|
| 486 |
+
787,
|
| 487 |
+
641
|
| 488 |
+
],
|
| 489 |
+
"page_idx": 4
|
| 490 |
+
},
|
| 491 |
+
{
|
| 492 |
+
"type": "equation",
|
| 493 |
+
"text": "\n$$\np = \\left\\{ \\begin{array}{l l} + 1 & \\text {i f} \\log \\left(I _ {x y} (t)\\right) - \\log \\left(I _ {x y} (t - \\Delta t)\\right) \\geq c, \\\\ - 1 & \\text {i f} \\log \\left(I _ {x y} (t)\\right) - \\log \\left(I _ {x y} (t - \\Delta t)\\right) \\leq - c, \\end{array} \\right. \\tag {1}\n$$\n",
|
| 494 |
+
"text_format": "latex",
|
| 495 |
+
"bbox": [
|
| 496 |
+
321,
|
| 497 |
+
654,
|
| 498 |
+
785,
|
| 499 |
+
694
|
| 500 |
+
],
|
| 501 |
+
"page_idx": 4
|
| 502 |
+
},
|
| 503 |
+
{
|
| 504 |
+
"type": "text",
|
| 505 |
+
"text": "where $p$ is the event's polarity, $I_{xy}(t)$ and $I_{xy}(t - \\Delta t)$ represent the intensity at the pixel coordinate $(x,y)$ at time $t$ and $t - \\Delta t$ , respectively. An event is a quaternion value $(x,y,t,p)$ where $x,y$ denotes the pixel coordinate, $t$ represents the time of change, the polarity $p \\in \\{+1, -1\\}$ indicates the increase and decrease. For a given blurry image $B$ , it can be formulated as the average of latent intensity images $I(t)$ during the exposure time [25]:",
|
| 506 |
+
"bbox": [
|
| 507 |
+
212,
|
| 508 |
+
705,
|
| 509 |
+
787,
|
| 510 |
+
797
|
| 511 |
+
],
|
| 512 |
+
"page_idx": 4
|
| 513 |
+
},
|
| 514 |
+
{
|
| 515 |
+
"type": "equation",
|
| 516 |
+
"text": "\n$$\nB = \\frac {1}{T} \\int_ {0} ^ {T} I (t) d t, \\tag {2}\n$$\n",
|
| 517 |
+
"text_format": "latex",
|
| 518 |
+
"bbox": [
|
| 519 |
+
437,
|
| 520 |
+
809,
|
| 521 |
+
785,
|
| 522 |
+
844
|
| 523 |
+
],
|
| 524 |
+
"page_idx": 4
|
| 525 |
+
},
|
| 526 |
+
{
|
| 527 |
+
"type": "header",
|
| 528 |
+
"text": "Motion Aware Event Representation-driven Image Deblurring",
|
| 529 |
+
"bbox": [
|
| 530 |
+
321,
|
| 531 |
+
114,
|
| 532 |
+
732,
|
| 533 |
+
130
|
| 534 |
+
],
|
| 535 |
+
"page_idx": 4
|
| 536 |
+
},
|
| 537 |
+
{
|
| 538 |
+
"type": "page_number",
|
| 539 |
+
"text": "5",
|
| 540 |
+
"bbox": [
|
| 541 |
+
774,
|
| 542 |
+
116,
|
| 543 |
+
784,
|
| 544 |
+
126
|
| 545 |
+
],
|
| 546 |
+
"page_idx": 4
|
| 547 |
+
},
|
| 548 |
+
{
|
| 549 |
+
"type": "image",
|
| 550 |
+
"img_path": "images/727fbf1276279060979dc2c666d494d80bcb46f3f42f97cec3fb0bcc253b654c.jpg",
|
| 551 |
+
"image_caption": [
|
| 552 |
+
"Fig. 2: Overview of the architecture of our proposed method for image deblurring. The blurred image and related event frames within the exposure time are fed into the corresponding network branches respectively. The core module of the event branch: (a) Recurrent Motion Extraction (RME) module processes event frames from longer time scales to shorter, merging old features into new features through weighted sum, which results in perceiving various motion patterns under different time ranges. The image branch extracts the features with conv block and fuses them with event features using (b) Feature Alignment and Fusion (FAF) module. FAF module eliminates the inter-modal inconsistencies in a bi-directional way ensuring better information interaction."
|
| 553 |
+
],
|
| 554 |
+
"image_footnote": [],
|
| 555 |
+
"bbox": [
|
| 556 |
+
232,
|
| 557 |
+
143,
|
| 558 |
+
772,
|
| 559 |
+
324
|
| 560 |
+
],
|
| 561 |
+
"page_idx": 5
|
| 562 |
+
},
|
| 563 |
+
{
|
| 564 |
+
"type": "text",
|
| 565 |
+
"text": "where $T$ is the exposure time. For simplicity of understanding, we now only consider the pixel at $(x,y)$ then combining Equations (1) and (2), we have",
|
| 566 |
+
"bbox": [
|
| 567 |
+
214,
|
| 568 |
+
486,
|
| 569 |
+
787,
|
| 570 |
+
518
|
| 571 |
+
],
|
| 572 |
+
"page_idx": 5
|
| 573 |
+
},
|
| 574 |
+
{
|
| 575 |
+
"type": "equation",
|
| 576 |
+
"text": "\n$$\n\\begin{array}{l} B _ {x y} = \\frac {1}{T} \\int_ {0} ^ {T} \\exp (\\log (I _ {x y} (t))) d t (3) \\\\ = \\frac {1}{T} \\int_ {0} ^ {T} e x p (\\log (I _ {x y} (0)) + c \\int_ {0} ^ {t} p (s) d s) d t. (4) \\\\ \\end{array}\n$$\n",
|
| 577 |
+
"text_format": "latex",
|
| 578 |
+
"bbox": [
|
| 579 |
+
331,
|
| 580 |
+
526,
|
| 581 |
+
785,
|
| 582 |
+
599
|
| 583 |
+
],
|
| 584 |
+
"page_idx": 5
|
| 585 |
+
},
|
| 586 |
+
{
|
| 587 |
+
"type": "text",
|
| 588 |
+
"text": "The internal part of the exponential operation is the logarithmic value of the light intensity of the pixel at time $t$ , while the external part of the exponential operation is the integration and averaging of each moment within the exposure time. Based on the event data, we can discretize Equation (4):",
|
| 589 |
+
"bbox": [
|
| 590 |
+
214,
|
| 591 |
+
607,
|
| 592 |
+
787,
|
| 593 |
+
669
|
| 594 |
+
],
|
| 595 |
+
"page_idx": 5
|
| 596 |
+
},
|
| 597 |
+
{
|
| 598 |
+
"type": "equation",
|
| 599 |
+
"text": "\n$$\n\\begin{array}{l} B _ {x y} = \\frac {1}{N} \\sum_ {t ^ {\\prime} \\in \\mathcal {T}} \\exp \\left(\\log \\left(I _ {x y} (0)\\right) + c \\sum_ {t = 0} ^ {t ^ {\\prime}} p _ {t}\\right) (5) \\\\ = \\frac {I _ {x y} (0)}{N} \\sum_ {t ^ {\\prime} \\in \\mathcal {T}} e x p \\left(c \\sum_ {t = 0} ^ {t ^ {\\prime}} p _ {t}\\right), (6) \\\\ \\end{array}\n$$\n",
|
| 600 |
+
"text_format": "latex",
|
| 601 |
+
"bbox": [
|
| 602 |
+
351,
|
| 603 |
+
680,
|
| 604 |
+
785,
|
| 605 |
+
768
|
| 606 |
+
],
|
| 607 |
+
"page_idx": 5
|
| 608 |
+
},
|
| 609 |
+
{
|
| 610 |
+
"type": "text",
|
| 611 |
+
"text": "where $\\mathcal{T} = \\{t_1, t_2, \\dots, t_N\\}$ indicates the set of time events occurs, $N$ is the number of events, and $p_t$ means the polarity of the event at time $t$ .",
|
| 612 |
+
"bbox": [
|
| 613 |
+
212,
|
| 614 |
+
779,
|
| 615 |
+
787,
|
| 616 |
+
809
|
| 617 |
+
],
|
| 618 |
+
"page_idx": 5
|
| 619 |
+
},
|
| 620 |
+
{
|
| 621 |
+
"type": "text",
|
| 622 |
+
"text": "The aforementioned derivation prompts us to consider not only the polarity of the event but also the impact of the alteration in the absolute value after",
|
| 623 |
+
"bbox": [
|
| 624 |
+
212,
|
| 625 |
+
810,
|
| 626 |
+
787,
|
| 627 |
+
839
|
| 628 |
+
],
|
| 629 |
+
"page_idx": 5
|
| 630 |
+
},
|
| 631 |
+
{
|
| 632 |
+
"type": "page_number",
|
| 633 |
+
"text": "6",
|
| 634 |
+
"bbox": [
|
| 635 |
+
217,
|
| 636 |
+
114,
|
| 637 |
+
228,
|
| 638 |
+
126
|
| 639 |
+
],
|
| 640 |
+
"page_idx": 5
|
| 641 |
+
},
|
| 642 |
+
{
|
| 643 |
+
"type": "header",
|
| 644 |
+
"text": "Z. Sun et al.",
|
| 645 |
+
"bbox": [
|
| 646 |
+
271,
|
| 647 |
+
114,
|
| 648 |
+
357,
|
| 649 |
+
127
|
| 650 |
+
],
|
| 651 |
+
"page_idx": 5
|
| 652 |
+
},
|
| 653 |
+
{
|
| 654 |
+
"type": "text",
|
| 655 |
+
"text": "Algorithm 1 Deviation Accumulation (DA)",
|
| 656 |
+
"text_level": 1,
|
| 657 |
+
"bbox": [
|
| 658 |
+
217,
|
| 659 |
+
146,
|
| 660 |
+
540,
|
| 661 |
+
162
|
| 662 |
+
],
|
| 663 |
+
"page_idx": 6
|
| 664 |
+
},
|
| 665 |
+
{
|
| 666 |
+
"type": "text",
|
| 667 |
+
"text": "Input: taking pixel at $(x,y)$ for an example, the events stream $(x,y,t_i,p_i)$ in which $i\\in [1,N]$ , deviation from the initial moment $D$ , accumulation of deviation $R$ , Output: the value of DA event representation $V$ .",
|
| 668 |
+
"bbox": [
|
| 669 |
+
215,
|
| 670 |
+
166,
|
| 671 |
+
785,
|
| 672 |
+
208
|
| 673 |
+
],
|
| 674 |
+
"page_idx": 6
|
| 675 |
+
},
|
| 676 |
+
{
|
| 677 |
+
"type": "list",
|
| 678 |
+
"sub_type": "text",
|
| 679 |
+
"list_items": [
|
| 680 |
+
"1: $D = 0,R = 0$",
|
| 681 |
+
"2: Counting the number of events $N$ ;",
|
| 682 |
+
"3: for $(i = 1$ to $N)$ do",
|
| 683 |
+
"4: $D = D + p_{i}$",
|
| 684 |
+
"5: $R = R + D$",
|
| 685 |
+
"6: end for",
|
| 686 |
+
"7: $V = R / N$"
|
| 687 |
+
],
|
| 688 |
+
"bbox": [
|
| 689 |
+
222,
|
| 690 |
+
209,
|
| 691 |
+
472,
|
| 692 |
+
305
|
| 693 |
+
],
|
| 694 |
+
"page_idx": 6
|
| 695 |
+
},
|
| 696 |
+
{
|
| 697 |
+
"type": "text",
|
| 698 |
+
"text": "each accumulation when considering the problem of deblurring. However, since event data cannot provide $I_{xy}(0)$ , the absolute light intensity is not accessible. But we can indirectly capture this information through the deviation from the initial moment. Therefore, we design a new event representation method for the deblurring problem presented in Algorithm 1. First, we accumulate all the events in the exposure time, so we get a 2D matrix named frequency matrix representing the number of the event occurs at each pixel position. We deal with the events considering their polarities subsequently, recording the deviation at each pixel by an additional tensor. Upon the arrival of the subsequent event, we update the deviation according to polarity and then add the deviation to the corresponding position. In the end, by dividing the cumulative matrix by the frequency matrix, we obtain the average deviation change over the exposure period as the new event representation named Deviation Accumulation (DA).",
|
| 699 |
+
"bbox": [
|
| 700 |
+
212,
|
| 701 |
+
340,
|
| 702 |
+
785,
|
| 703 |
+
537
|
| 704 |
+
],
|
| 705 |
+
"page_idx": 6
|
| 706 |
+
},
|
| 707 |
+
{
|
| 708 |
+
"type": "text",
|
| 709 |
+
"text": "Traditional polarity accumulation methods directly accumulating event polarities lead to identical outputs whenever the number of positive and negative events within a given time frame $([0, T])$ remains unchanged. Nevertheless, by considering the deviation brought by historical events, our method can generate distinct values for events arriving in different orders, thereby improving the ability to perceive motion patterns and providing more precise motion information for the deblurring procedure. For example, as illustrated in Figure 3, (a) shows event signals generated by a motion pattern which represents that the light intensity at that pixel has risen twice and then fallen twice while (d) shows another pattern. Facing these two patterns, (b) and (e) depict the changing process of the outputs of traditional method and our method after each event is accumulated. After processing all four events, traditional polarity accumulation gets the same output of 0 for both patterns, but our DA representation yields 1 and 0.5 respectively which contains richer motion information. More examples of motion patterns our method can distinguish are shown in supplementary materials.",
|
| 710 |
+
"bbox": [
|
| 711 |
+
212,
|
| 712 |
+
537,
|
| 713 |
+
785,
|
| 714 |
+
763
|
| 715 |
+
],
|
| 716 |
+
"page_idx": 6
|
| 717 |
+
},
|
| 718 |
+
{
|
| 719 |
+
"type": "text",
|
| 720 |
+
"text": "Overall, the phenomenon that polarity accumulation methods represent some different motion patterns as the same value introduces ambiguity into the deblurring process. Our method has a better perception of the motion patterns and offers a more distinct motion edge than conventional approaches from the visual effect as shown in Figure 3 (c) and (f). Inspired by [33], directly segmenting",
|
| 721 |
+
"bbox": [
|
| 722 |
+
212,
|
| 723 |
+
765,
|
| 724 |
+
785,
|
| 725 |
+
840
|
| 726 |
+
],
|
| 727 |
+
"page_idx": 6
|
| 728 |
+
},
|
| 729 |
+
{
|
| 730 |
+
"type": "header",
|
| 731 |
+
"text": "Motion Aware Event Representation-driven Image Deblurring",
|
| 732 |
+
"bbox": [
|
| 733 |
+
321,
|
| 734 |
+
114,
|
| 735 |
+
732,
|
| 736 |
+
130
|
| 737 |
+
],
|
| 738 |
+
"page_idx": 6
|
| 739 |
+
},
|
| 740 |
+
{
|
| 741 |
+
"type": "page_number",
|
| 742 |
+
"text": "7",
|
| 743 |
+
"bbox": [
|
| 744 |
+
774,
|
| 745 |
+
116,
|
| 746 |
+
784,
|
| 747 |
+
126
|
| 748 |
+
],
|
| 749 |
+
"page_idx": 6
|
| 750 |
+
},
|
| 751 |
+
{
|
| 752 |
+
"type": "image",
|
| 753 |
+
"img_path": "images/f27df0ad0618d5db90f55bedf00ddfb7b27b5a54a5d9f563739f621712895c21.jpg",
|
| 754 |
+
"image_caption": [],
|
| 755 |
+
"image_footnote": [],
|
| 756 |
+
"bbox": [
|
| 757 |
+
263,
|
| 758 |
+
148,
|
| 759 |
+
411,
|
| 760 |
+
255
|
| 761 |
+
],
|
| 762 |
+
"page_idx": 7
|
| 763 |
+
},
|
| 764 |
+
{
|
| 765 |
+
"type": "image",
|
| 766 |
+
"img_path": "images/42187f67118f8b940e6d4101d46ceb7e90f1c0c1d5e25fc0efcf2e6b161825e5.jpg",
|
| 767 |
+
"image_caption": [],
|
| 768 |
+
"image_footnote": [],
|
| 769 |
+
"bbox": [
|
| 770 |
+
416,
|
| 771 |
+
145,
|
| 772 |
+
594,
|
| 773 |
+
253
|
| 774 |
+
],
|
| 775 |
+
"page_idx": 7
|
| 776 |
+
},
|
| 777 |
+
{
|
| 778 |
+
"type": "image",
|
| 779 |
+
"img_path": "images/2594bfdcceba4383dab92ecf16d109c444ee6f9000430d2f4182e0ca07f17411.jpg",
|
| 780 |
+
"image_caption": [
|
| 781 |
+
"Fig. 3: (a) and (d) are event signals of two different motion patterns. (b) shows the Polarity Accumulation (PA) can't distinguish (a) and (d) as it gets the same output of 0 after processing all events while (e) shows our Deviation Accumulation (DA) can. DA perceives more motion information than PA as shown in (c) and (f)."
|
| 782 |
+
],
|
| 783 |
+
"image_footnote": [],
|
| 784 |
+
"bbox": [
|
| 785 |
+
263,
|
| 786 |
+
257,
|
| 787 |
+
411,
|
| 788 |
+
359
|
| 789 |
+
],
|
| 790 |
+
"page_idx": 7
|
| 791 |
+
},
|
| 792 |
+
{
|
| 793 |
+
"type": "image",
|
| 794 |
+
"img_path": "images/1c9c90660ce6ee749ed03a3071e895c9dbd79e1abc3052138dd0644cf4112f44.jpg",
|
| 795 |
+
"image_caption": [],
|
| 796 |
+
"image_footnote": [],
|
| 797 |
+
"bbox": [
|
| 798 |
+
418,
|
| 799 |
+
255,
|
| 800 |
+
593,
|
| 801 |
+
359
|
| 802 |
+
],
|
| 803 |
+
"page_idx": 7
|
| 804 |
+
},
|
| 805 |
+
{
|
| 806 |
+
"type": "image",
|
| 807 |
+
"img_path": "images/16ebf709d3c72c1877985cdb2c25867915be52e3966f58db11e80140a702ae40.jpg",
|
| 808 |
+
"image_caption": [],
|
| 809 |
+
"image_footnote": [],
|
| 810 |
+
"bbox": [
|
| 811 |
+
617,
|
| 812 |
+
147,
|
| 813 |
+
736,
|
| 814 |
+
255
|
| 815 |
+
],
|
| 816 |
+
"page_idx": 7
|
| 817 |
+
},
|
| 818 |
+
{
|
| 819 |
+
"type": "image",
|
| 820 |
+
"img_path": "images/a6efd72a2e849957ba7e7139a202dbda424f44cc76c8a7f95b59c21f5a3f3d94.jpg",
|
| 821 |
+
"image_caption": [],
|
| 822 |
+
"image_footnote": [],
|
| 823 |
+
"bbox": [
|
| 824 |
+
617,
|
| 825 |
+
256,
|
| 826 |
+
736,
|
| 827 |
+
359
|
| 828 |
+
],
|
| 829 |
+
"page_idx": 7
|
| 830 |
+
},
|
| 831 |
+
{
|
| 832 |
+
"type": "text",
|
| 833 |
+
"text": "the entire exposure time averagely is not suitable for motion deblurring, which focuses on capturing rapid motion but falls short in adequately modeling the continuous movement throughout the exposure. In this paper, we symmetrically accumulate events along the $T / 2$ axis using the DA method. Taking $[0,T / 2]$ as an example, we accumulate events on different time scales. The longest accumulating range is $T / 2$ , followed by $T / 3$ , and the shortest is $T / 6$ . A detailed schematic is provided in the supplementary materials.",
|
| 834 |
+
"bbox": [
|
| 835 |
+
212,
|
| 836 |
+
454,
|
| 837 |
+
787,
|
| 838 |
+
561
|
| 839 |
+
],
|
| 840 |
+
"page_idx": 7
|
| 841 |
+
},
|
| 842 |
+
{
|
| 843 |
+
"type": "text",
|
| 844 |
+
"text": "3.3 Recurrent Motion Extraction Module",
|
| 845 |
+
"text_level": 1,
|
| 846 |
+
"bbox": [
|
| 847 |
+
214,
|
| 848 |
+
579,
|
| 849 |
+
573,
|
| 850 |
+
593
|
| 851 |
+
],
|
| 852 |
+
"page_idx": 7
|
| 853 |
+
},
|
| 854 |
+
{
|
| 855 |
+
"type": "text",
|
| 856 |
+
"text": "The goal of the Recurrent Motion Extraction (RME) module is to cater to both global slow motion and local fast motion in different time scales. Based on our observations of the motion process and event representations, the focus of the event representations varies in different time scales. In shorter time intervals, the global slow motion is not significant in the event representation due to the smaller number of events triggered. On the contrary slow motion can be clearly recorded in longer intervals, but at the same time, the violent moving part will produce edge spreading artifact.",
|
| 857 |
+
"bbox": [
|
| 858 |
+
212,
|
| 859 |
+
599,
|
| 860 |
+
787,
|
| 861 |
+
720
|
| 862 |
+
],
|
| 863 |
+
"page_idx": 7
|
| 864 |
+
},
|
| 865 |
+
{
|
| 866 |
+
"type": "text",
|
| 867 |
+
"text": "To enable the network to accurately model the various in-homogeneous motions, we devised an RME module that incorporates a recurrent mechanism. As shown in Figure 2(a), the event features are fed into the processing module from the longest time scales to the shortest and we use $e1$ , $e2$ , and $e3$ to denote them. Formally, the RME module can be presented as",
|
| 868 |
+
"bbox": [
|
| 869 |
+
212,
|
| 870 |
+
720,
|
| 871 |
+
787,
|
| 872 |
+
797
|
| 873 |
+
],
|
| 874 |
+
"page_idx": 7
|
| 875 |
+
},
|
| 876 |
+
{
|
| 877 |
+
"type": "equation",
|
| 878 |
+
"text": "\n$$\ne _ {i} ^ {\\prime} = \\left\\{ \\begin{array}{l l} f _ {e x} \\left(e _ {i}\\right) & \\text {i f} i = 1, \\\\ \\left(1 - \\alpha\\right) \\times f _ {e x} \\left(e _ {i}\\right) + \\alpha \\times f _ {r e} \\left(e _ {i - 1} ^ {\\prime}\\right) & \\text {i f} i = 2, 3, \\end{array} \\right. \\tag {7}\n$$\n",
|
| 879 |
+
"text_format": "latex",
|
| 880 |
+
"bbox": [
|
| 881 |
+
313,
|
| 882 |
+
803,
|
| 883 |
+
785,
|
| 884 |
+
844
|
| 885 |
+
],
|
| 886 |
+
"page_idx": 7
|
| 887 |
+
},
|
| 888 |
+
{
|
| 889 |
+
"type": "page_number",
|
| 890 |
+
"text": "8",
|
| 891 |
+
"bbox": [
|
| 892 |
+
217,
|
| 893 |
+
114,
|
| 894 |
+
228,
|
| 895 |
+
126
|
| 896 |
+
],
|
| 897 |
+
"page_idx": 7
|
| 898 |
+
},
|
| 899 |
+
{
|
| 900 |
+
"type": "header",
|
| 901 |
+
"text": "Z. Sun et al.",
|
| 902 |
+
"bbox": [
|
| 903 |
+
271,
|
| 904 |
+
114,
|
| 905 |
+
357,
|
| 906 |
+
127
|
| 907 |
+
],
|
| 908 |
+
"page_idx": 7
|
| 909 |
+
},
|
| 910 |
+
{
|
| 911 |
+
"type": "text",
|
| 912 |
+
"text": "where $f_{ex}(\\cdot)$ and $f_{re}(\\cdot)$ are two baseline conv blocks, $f_{ex}(\\cdot)$ is used to extract information from the current representation, $f_{re}(\\cdot)$ is used to retain useful information from previous features, $\\alpha$ is the learnable weight as a trade-off between two features. By sharing the parameters of $f_{ex}(\\cdot)$ and $f_{re}(\\cdot)$ for different time scales, we force the network to be able to perceive motions at varying time scales. Each input feature is symmetric on the time axis so that the priori information about the motion direction is simultaneously fed into the network.",
|
| 913 |
+
"bbox": [
|
| 914 |
+
212,
|
| 915 |
+
146,
|
| 916 |
+
787,
|
| 917 |
+
252
|
| 918 |
+
],
|
| 919 |
+
"page_idx": 8
|
| 920 |
+
},
|
| 921 |
+
{
|
| 922 |
+
"type": "text",
|
| 923 |
+
"text": "3.4 Feature Alignment and Fusion Module",
|
| 924 |
+
"text_level": 1,
|
| 925 |
+
"bbox": [
|
| 926 |
+
214,
|
| 927 |
+
273,
|
| 928 |
+
581,
|
| 929 |
+
289
|
| 930 |
+
],
|
| 931 |
+
"page_idx": 8
|
| 932 |
+
},
|
| 933 |
+
{
|
| 934 |
+
"type": "text",
|
| 935 |
+
"text": "Although the information captured by the event camera exhibits a high correlation with the frame camera during the equivalent exposure period, there is still an inevitable gap originating from the modal dissimilarities between them. This phenomenon is frequently mentioned in multi-modal learning [18,42], but in event-guided image restoration tasks people usually [9,33,45] fuse features from two modalities directly using convolutions or attention mechanisms. Therefore, we propose an effective Feature Alignment and Fusion (FAF) module to facilitate enhanced feature fusion in a more coherent environment. Concretely, as shown in Figure 2(b), this module consists of a feature alignment module and a multi-head attention module. We use the operations of multiplication and addition successively to merge the raw information of one modality into the changing process of the other modality, which fills the gap of inter-modal inconsistency. The aligned features are fed into the attention module. In each fusion process, attention plays a bi-directional role, we not only generate the Q matrix through the frame features and the K and V matrix through the event features but also use the event features to perform look-ups on the frame features. The weights of the attention module are shared so that we can fuse the motion information under a more consistent space. Therefore, FAF module can be expressed as",
|
| 936 |
+
"bbox": [
|
| 937 |
+
212,
|
| 938 |
+
299,
|
| 939 |
+
787,
|
| 940 |
+
571
|
| 941 |
+
],
|
| 942 |
+
"page_idx": 8
|
| 943 |
+
},
|
| 944 |
+
{
|
| 945 |
+
"type": "equation",
|
| 946 |
+
"text": "\n$$\nx ^ {\\prime} = \\operatorname {C o n v} _ {2} (x) + (\\operatorname {C o n v} _ {1} (x) \\times e + e), \\tag {8}\n$$\n",
|
| 947 |
+
"text_format": "latex",
|
| 948 |
+
"bbox": [
|
| 949 |
+
227,
|
| 950 |
+
582,
|
| 951 |
+
784,
|
| 952 |
+
599
|
| 953 |
+
],
|
| 954 |
+
"page_idx": 8
|
| 955 |
+
},
|
| 956 |
+
{
|
| 957 |
+
"type": "equation",
|
| 958 |
+
"text": "\n$$\ne ^ {\\prime} = \\operatorname {C o n v} _ {2} (e) + (\\operatorname {C o n v} _ {1} (e) \\times x + x), \\tag {9}\n$$\n",
|
| 959 |
+
"text_format": "latex",
|
| 960 |
+
"bbox": [
|
| 961 |
+
228,
|
| 962 |
+
604,
|
| 963 |
+
784,
|
| 964 |
+
619
|
| 965 |
+
],
|
| 966 |
+
"page_idx": 8
|
| 967 |
+
},
|
| 968 |
+
{
|
| 969 |
+
"type": "equation",
|
| 970 |
+
"text": "\n$$\nF = \\operatorname {C o n v} _ {3} (\\operatorname {C o n c a t} ([ \\text {A t t e n t i o n} (Q _ {x ^ {\\prime}}, K _ {e ^ {\\prime}}, V _ {e ^ {\\prime}}), \\text {A t t e n t i o n} (Q _ {e ^ {\\prime}}, K _ {x ^ {\\prime}}, V _ {x ^ {\\prime}}) ])), \\tag {10}\n$$\n",
|
| 971 |
+
"text_format": "latex",
|
| 972 |
+
"bbox": [
|
| 973 |
+
228,
|
| 974 |
+
625,
|
| 975 |
+
784,
|
| 976 |
+
640
|
| 977 |
+
],
|
| 978 |
+
"page_idx": 8
|
| 979 |
+
},
|
| 980 |
+
{
|
| 981 |
+
"type": "text",
|
| 982 |
+
"text": "where $x'$ and $e'$ are aligned image features $x$ and event feature $e$ respectively, Attention(\\cdot) denotes the multi-head attention operation, Concat(\\cdot) denotes the concatenation operation along the channel dimension.",
|
| 983 |
+
"bbox": [
|
| 984 |
+
214,
|
| 985 |
+
652,
|
| 986 |
+
787,
|
| 987 |
+
700
|
| 988 |
+
],
|
| 989 |
+
"page_idx": 8
|
| 990 |
+
},
|
| 991 |
+
{
|
| 992 |
+
"type": "text",
|
| 993 |
+
"text": "4 Experiments and Analysis",
|
| 994 |
+
"text_level": 1,
|
| 995 |
+
"bbox": [
|
| 996 |
+
214,
|
| 997 |
+
722,
|
| 998 |
+
504,
|
| 999 |
+
739
|
| 1000 |
+
],
|
| 1001 |
+
"page_idx": 8
|
| 1002 |
+
},
|
| 1003 |
+
{
|
| 1004 |
+
"type": "text",
|
| 1005 |
+
"text": "4.1 Experimental Settings",
|
| 1006 |
+
"text_level": 1,
|
| 1007 |
+
"bbox": [
|
| 1008 |
+
214,
|
| 1009 |
+
753,
|
| 1010 |
+
447,
|
| 1011 |
+
768
|
| 1012 |
+
],
|
| 1013 |
+
"page_idx": 8
|
| 1014 |
+
},
|
| 1015 |
+
{
|
| 1016 |
+
"type": "text",
|
| 1017 |
+
"text": "Datasets. We use three different datasets containing synthetic, semi-synthetic, and real-world frames and events to evaluate the proposed method.",
|
| 1018 |
+
"bbox": [
|
| 1019 |
+
212,
|
| 1020 |
+
779,
|
| 1021 |
+
782,
|
| 1022 |
+
808
|
| 1023 |
+
],
|
| 1024 |
+
"page_idx": 8
|
| 1025 |
+
},
|
| 1026 |
+
{
|
| 1027 |
+
"type": "text",
|
| 1028 |
+
"text": "GoPro: We train the network on the GoPro dataset which is the benchmark dataset for the image motion deblurring [24]. The blurry image is averaged from",
|
| 1029 |
+
"bbox": [
|
| 1030 |
+
212,
|
| 1031 |
+
809,
|
| 1032 |
+
785,
|
| 1033 |
+
839
|
| 1034 |
+
],
|
| 1035 |
+
"page_idx": 8
|
| 1036 |
+
},
|
| 1037 |
+
{
|
| 1038 |
+
"type": "header",
|
| 1039 |
+
"text": "Motion Aware Event Representation-driven Image Deblurring",
|
| 1040 |
+
"bbox": [
|
| 1041 |
+
320,
|
| 1042 |
+
114,
|
| 1043 |
+
732,
|
| 1044 |
+
130
|
| 1045 |
+
],
|
| 1046 |
+
"page_idx": 8
|
| 1047 |
+
},
|
| 1048 |
+
{
|
| 1049 |
+
"type": "page_number",
|
| 1050 |
+
"text": "9",
|
| 1051 |
+
"bbox": [
|
| 1052 |
+
774,
|
| 1053 |
+
116,
|
| 1054 |
+
784,
|
| 1055 |
+
126
|
| 1056 |
+
],
|
| 1057 |
+
"page_idx": 8
|
| 1058 |
+
},
|
| 1059 |
+
{
|
| 1060 |
+
"type": "table",
|
| 1061 |
+
"img_path": "images/c910256561819228f874ee586a3a0c12b3b6df3b608f13ac97ba3cd6c9d9c00b.jpg",
|
| 1062 |
+
"table_caption": [
|
| 1063 |
+
"Table 1: The quantitative results on GoPro, HS-ERGB, and REBlur test datasets. The best values are highlighted in bold. HINet+ and UFPNet+ are event-enhanced versions of HINet and UFPNet by concatenating our DA to their input."
|
| 1064 |
+
],
|
| 1065 |
+
"table_footnote": [],
|
| 1066 |
+
"table_body": "<table><tr><td rowspan=\"2\">Method</td><td rowspan=\"2\">Input</td><td colspan=\"2\">GoPro</td><td colspan=\"2\">HS-ERGB</td><td colspan=\"2\">REBlur</td><td rowspan=\"2\">Params(M)</td></tr><tr><td>PSNR</td><td>SSIM</td><td>PSNR</td><td>SSIM</td><td>PSNR</td><td>SSIM</td></tr><tr><td>LEVS [10]</td><td>F</td><td>20.84</td><td>0.547</td><td>22.13</td><td>0.555</td><td>-</td><td>-</td><td>18.2</td></tr><tr><td>EDI [25]</td><td>F+E</td><td>29.06</td><td>0.943</td><td>23.93</td><td>0.704</td><td>36.52</td><td>0.964</td><td>0.5</td></tr><tr><td>SRN [35]</td><td>F</td><td>30.26</td><td>0.934</td><td>-</td><td>-</td><td>35.10</td><td>0.961</td><td>10.2</td></tr><tr><td>EVDI [49]</td><td>F+E</td><td>30.40</td><td>0.906</td><td>25.13</td><td>0.707</td><td>-</td><td>-</td><td>0.4</td></tr><tr><td>HINet [4]</td><td>F</td><td>32.71</td><td>0.959</td><td>27.32</td><td>0.807</td><td>35.58</td><td>0.965</td><td>88.7</td></tr><tr><td>Restormer [47]</td><td>F</td><td>32.92</td><td>0.961</td><td>27.55</td><td>0.808</td><td>35.50</td><td>0.959</td><td>26.1</td></tr><tr><td>MSDI-Net [17]</td><td>F</td><td>33.28</td><td>0.964</td><td>27.46</td><td>0.809</td><td>36.14</td><td>0.968</td><td>241.3</td></tr><tr><td>NAFNet [3]</td><td>F</td><td>33.71</td><td>0.967</td><td>27.64</td><td>0.811</td><td>36.15</td><td>0.969</td><td>67.8</td></tr><tr><td>UFPNet [6]</td><td>F</td><td>34.06</td><td>0.968</td><td>27.64</td><td>0.809</td><td>36.11</td><td>0.968</td><td>80.3</td></tr><tr><td>HINet+ [4]</td><td>F+E</td><td>34.63</td><td>0.968</td><td>27.66</td><td>0.808</td><td>37.92</td><td>0.976</td><td>88.7</td></tr><tr><td>UFPNet+ [6]</td><td>F+E</td><td>35.22</td><td>0.972</td><td>27.68</td><td>0.809</td><td>37.97</td><td>0.976</td><td>80.3</td></tr><tr><td>EFNet [33]</td><td>F+E</td><td>35.46</td><td>0.972</td><td>26.68</td><td>0.800</td><td>38.12</td><td>0.975</td><td>8.5</td></tr><tr><td>Ours</td><td>F+E</td><td>36.07</td><td>0.976</td><td>27.93</td><td>0.812</td><td>38.47</td><td>0.978</td><td>13.9</td></tr></table>",
|
| 1067 |
+
"bbox": [
|
| 1068 |
+
228,
|
| 1069 |
+
199,
|
| 1070 |
+
771,
|
| 1071 |
+
422
|
| 1072 |
+
],
|
| 1073 |
+
"page_idx": 9
|
| 1074 |
+
},
|
| 1075 |
+
{
|
| 1076 |
+
"type": "text",
|
| 1077 |
+
"text": "adjacent sharp frames, by doing so the dataset contains 3214 pairs of blurry and sharp images. According to the standard division, we adopt 2103 pairs for training and 1111 pairs for testing. For the event data, we use the GoPro raw event dataset provided by [33] which synthesizes events with a random threshold $c$ following the Gaussian distribution by ESIM simulator [27]. Following the DA method, the raw event data is preprocessed into 6 event frames for each image.",
|
| 1078 |
+
"bbox": [
|
| 1079 |
+
212,
|
| 1080 |
+
462,
|
| 1081 |
+
784,
|
| 1082 |
+
551
|
| 1083 |
+
],
|
| 1084 |
+
"page_idx": 9
|
| 1085 |
+
},
|
| 1086 |
+
{
|
| 1087 |
+
"type": "text",
|
| 1088 |
+
"text": "HS-ERGB: The HS-ERGB dataset [37] consists of sharp videos and real-world events, we use [50] released normal blur version which synthesizes blurry frames by averaging 49 interpolating images. Each blurry frame in the dataset has a known exposure time, allowing us to generate the DA representation for each frame within its respective exposure duration.",
|
| 1089 |
+
"bbox": [
|
| 1090 |
+
212,
|
| 1091 |
+
551,
|
| 1092 |
+
785,
|
| 1093 |
+
628
|
| 1094 |
+
],
|
| 1095 |
+
"page_idx": 9
|
| 1096 |
+
},
|
| 1097 |
+
{
|
| 1098 |
+
"type": "text",
|
| 1099 |
+
"text": "REBlur: The REBlur dataset collects a number of sequences of real-world event data, corresponding with blurry images and reliable ground-truth sharp images [33]. It contains 12 kinds of linear and nonlinear motions. Following the original setting, we use 486 pairs of blurry-sharp images with associated events for training and 983 pairs for testing.",
|
| 1100 |
+
"bbox": [
|
| 1101 |
+
212,
|
| 1102 |
+
628,
|
| 1103 |
+
785,
|
| 1104 |
+
703
|
| 1105 |
+
],
|
| 1106 |
+
"page_idx": 9
|
| 1107 |
+
},
|
| 1108 |
+
{
|
| 1109 |
+
"type": "text",
|
| 1110 |
+
"text": "Implementation details. We train the end-to-end network without pre-training on a single Tesla V100 GPU. We extract patches of size $256 \\times 256$ for training images and corresponding event frames, and the batch size is set to 2 by default. The AdamW optimizer is employed ( $\\beta_{1} = 0.9$ and $\\beta_{2} = 0.99$ ) with the initial learning rate 0.0001 with the cosine annealing schedule where the $T_{max}$ is 400K iteration. For data augmentation, we perform horizontal and vertical flips, the random noise and the hot pixels are added into event frames in order to simulate the real-world situation. For HS-ERGB and REBlur datasets, We fine-tune the model trained on GoPro with the training set of them. The fine-tuning process",
|
| 1111 |
+
"bbox": [
|
| 1112 |
+
212,
|
| 1113 |
+
704,
|
| 1114 |
+
787,
|
| 1115 |
+
839
|
| 1116 |
+
],
|
| 1117 |
+
"page_idx": 9
|
| 1118 |
+
},
|
| 1119 |
+
{
|
| 1120 |
+
"type": "page_number",
|
| 1121 |
+
"text": "10",
|
| 1122 |
+
"bbox": [
|
| 1123 |
+
217,
|
| 1124 |
+
114,
|
| 1125 |
+
235,
|
| 1126 |
+
126
|
| 1127 |
+
],
|
| 1128 |
+
"page_idx": 9
|
| 1129 |
+
},
|
| 1130 |
+
{
|
| 1131 |
+
"type": "header",
|
| 1132 |
+
"text": "Z. Sun et al.",
|
| 1133 |
+
"bbox": [
|
| 1134 |
+
271,
|
| 1135 |
+
114,
|
| 1136 |
+
354,
|
| 1137 |
+
126
|
| 1138 |
+
],
|
| 1139 |
+
"page_idx": 9
|
| 1140 |
+
},
|
| 1141 |
+
{
|
| 1142 |
+
"type": "image",
|
| 1143 |
+
"img_path": "images/20a836d4e920f402f1279f0912bc458bdbdbfc258f2ce5277dd485ce5c876056.jpg",
|
| 1144 |
+
"image_caption": [
|
| 1145 |
+
"(a) Blur image in GoPro testset"
|
| 1146 |
+
],
|
| 1147 |
+
"image_footnote": [],
|
| 1148 |
+
"bbox": [
|
| 1149 |
+
259,
|
| 1150 |
+
146,
|
| 1151 |
+
431,
|
| 1152 |
+
277
|
| 1153 |
+
],
|
| 1154 |
+
"page_idx": 10
|
| 1155 |
+
},
|
| 1156 |
+
{
|
| 1157 |
+
"type": "text",
|
| 1158 |
+
"text": "(a) Blur image in GoPro testset",
|
| 1159 |
+
"bbox": [
|
| 1160 |
+
261,
|
| 1161 |
+
279,
|
| 1162 |
+
429,
|
| 1163 |
+
290
|
| 1164 |
+
],
|
| 1165 |
+
"page_idx": 10
|
| 1166 |
+
},
|
| 1167 |
+
{
|
| 1168 |
+
"type": "image",
|
| 1169 |
+
"img_path": "images/090bccd521fd0515526e1fdddda2a1d27d3b2e0539b66ab88db218c349e85371.jpg",
|
| 1170 |
+
"image_caption": [
|
| 1171 |
+
"(b) HINet"
|
| 1172 |
+
],
|
| 1173 |
+
"image_footnote": [],
|
| 1174 |
+
"bbox": [
|
| 1175 |
+
434,
|
| 1176 |
+
146,
|
| 1177 |
+
509,
|
| 1178 |
+
204
|
| 1179 |
+
],
|
| 1180 |
+
"page_idx": 10
|
| 1181 |
+
},
|
| 1182 |
+
{
|
| 1183 |
+
"type": "image",
|
| 1184 |
+
"img_path": "images/de240c96888191bad7fd23d6f5d1fd42c49aab6b15b4ad351ec42ba2462276a4.jpg",
|
| 1185 |
+
"image_caption": [
|
| 1186 |
+
"(c) UFPNet"
|
| 1187 |
+
],
|
| 1188 |
+
"image_footnote": [],
|
| 1189 |
+
"bbox": [
|
| 1190 |
+
511,
|
| 1191 |
+
146,
|
| 1192 |
+
586,
|
| 1193 |
+
204
|
| 1194 |
+
],
|
| 1195 |
+
"page_idx": 10
|
| 1196 |
+
},
|
| 1197 |
+
{
|
| 1198 |
+
"type": "image",
|
| 1199 |
+
"img_path": "images/855a3626dcaa19b3a6f81c7a38abba6894d8210cbb0351136d26db20c5ded06a.jpg",
|
| 1200 |
+
"image_caption": [
|
| 1201 |
+
"(d) EFNet"
|
| 1202 |
+
],
|
| 1203 |
+
"image_footnote": [],
|
| 1204 |
+
"bbox": [
|
| 1205 |
+
589,
|
| 1206 |
+
146,
|
| 1207 |
+
663,
|
| 1208 |
+
204
|
| 1209 |
+
],
|
| 1210 |
+
"page_idx": 10
|
| 1211 |
+
},
|
| 1212 |
+
{
|
| 1213 |
+
"type": "image",
|
| 1214 |
+
"img_path": "images/9d59b2a9ca86a9e8467a6a4624297134ae54fe4303e5a9a48ebb08b3760aa938.jpg",
|
| 1215 |
+
"image_caption": [
|
| 1216 |
+
"(e) NAFNet"
|
| 1217 |
+
],
|
| 1218 |
+
"image_footnote": [],
|
| 1219 |
+
"bbox": [
|
| 1220 |
+
666,
|
| 1221 |
+
146,
|
| 1222 |
+
741,
|
| 1223 |
+
204
|
| 1224 |
+
],
|
| 1225 |
+
"page_idx": 10
|
| 1226 |
+
},
|
| 1227 |
+
{
|
| 1228 |
+
"type": "image",
|
| 1229 |
+
"img_path": "images/33624c427a9def027a5e47ac4beec07fe892844f016eb3e4c139b6b2dff08fd6.jpg",
|
| 1230 |
+
"image_caption": [
|
| 1231 |
+
"(f) HINet+"
|
| 1232 |
+
],
|
| 1233 |
+
"image_footnote": [],
|
| 1234 |
+
"bbox": [
|
| 1235 |
+
434,
|
| 1236 |
+
220,
|
| 1237 |
+
509,
|
| 1238 |
+
277
|
| 1239 |
+
],
|
| 1240 |
+
"page_idx": 10
|
| 1241 |
+
},
|
| 1242 |
+
{
|
| 1243 |
+
"type": "image",
|
| 1244 |
+
"img_path": "images/b844597b0b2628d9b494fbb49079e9538da6c357ce669b0189408d786989910e.jpg",
|
| 1245 |
+
"image_caption": [
|
| 1246 |
+
"(g) UFPNet+"
|
| 1247 |
+
],
|
| 1248 |
+
"image_footnote": [],
|
| 1249 |
+
"bbox": [
|
| 1250 |
+
511,
|
| 1251 |
+
220,
|
| 1252 |
+
586,
|
| 1253 |
+
277
|
| 1254 |
+
],
|
| 1255 |
+
"page_idx": 10
|
| 1256 |
+
},
|
| 1257 |
+
{
|
| 1258 |
+
"type": "image",
|
| 1259 |
+
"img_path": "images/525fd323a7e502680d99c7d455846726cc7a4848aa71c023a1ac82cea2cdfb8e.jpg",
|
| 1260 |
+
"image_caption": [
|
| 1261 |
+
"(h) Ours"
|
| 1262 |
+
],
|
| 1263 |
+
"image_footnote": [],
|
| 1264 |
+
"bbox": [
|
| 1265 |
+
589,
|
| 1266 |
+
220,
|
| 1267 |
+
663,
|
| 1268 |
+
277
|
| 1269 |
+
],
|
| 1270 |
+
"page_idx": 10
|
| 1271 |
+
},
|
| 1272 |
+
{
|
| 1273 |
+
"type": "image",
|
| 1274 |
+
"img_path": "images/1352bdc01951e52d0cda0812c9706632ddf092d3b3b441a507c212605ee64399.jpg",
|
| 1275 |
+
"image_caption": [
|
| 1276 |
+
"(i) GT"
|
| 1277 |
+
],
|
| 1278 |
+
"image_footnote": [],
|
| 1279 |
+
"bbox": [
|
| 1280 |
+
668,
|
| 1281 |
+
220,
|
| 1282 |
+
743,
|
| 1283 |
+
277
|
| 1284 |
+
],
|
| 1285 |
+
"page_idx": 10
|
| 1286 |
+
},
|
| 1287 |
+
{
|
| 1288 |
+
"type": "image",
|
| 1289 |
+
"img_path": "images/0843f6c080552a51c92468fb751bc0e4305eec693b7bb36b6c1b2a4b492e83be.jpg",
|
| 1290 |
+
"image_caption": [
|
| 1291 |
+
"Fig. 4: Qualitative comparisons on the GoPro dataset. The notation is the same as in Table 1. Zoom in for better view.",
|
| 1292 |
+
"(a) Blur image in REBlur testset",
|
| 1293 |
+
": Qualitative comparisons on the REBlur dataset. The notation is the same as Table 1. Zoom in for better view."
|
| 1294 |
+
],
|
| 1295 |
+
"image_footnote": [],
|
| 1296 |
+
"bbox": [
|
| 1297 |
+
261,
|
| 1298 |
+
348,
|
| 1299 |
+
431,
|
| 1300 |
+
479
|
| 1301 |
+
],
|
| 1302 |
+
"page_idx": 10
|
| 1303 |
+
},
|
| 1304 |
+
{
|
| 1305 |
+
"type": "image",
|
| 1306 |
+
"img_path": "images/e3739c1a5a39ccec854698002b1ed5041ff11d891789b8f9a04cefc4b762dc8d.jpg",
|
| 1307 |
+
"image_caption": [
|
| 1308 |
+
"(b) HINet"
|
| 1309 |
+
],
|
| 1310 |
+
"image_footnote": [],
|
| 1311 |
+
"bbox": [
|
| 1312 |
+
434,
|
| 1313 |
+
348,
|
| 1314 |
+
509,
|
| 1315 |
+
406
|
| 1316 |
+
],
|
| 1317 |
+
"page_idx": 10
|
| 1318 |
+
},
|
| 1319 |
+
{
|
| 1320 |
+
"type": "image",
|
| 1321 |
+
"img_path": "images/1cee75ff1bd3522ee52b1985f367cea90277b0673d9543de97be08f469dad2dc.jpg",
|
| 1322 |
+
"image_caption": [
|
| 1323 |
+
"(c) UFPNet"
|
| 1324 |
+
],
|
| 1325 |
+
"image_footnote": [],
|
| 1326 |
+
"bbox": [
|
| 1327 |
+
514,
|
| 1328 |
+
348,
|
| 1329 |
+
586,
|
| 1330 |
+
406
|
| 1331 |
+
],
|
| 1332 |
+
"page_idx": 10
|
| 1333 |
+
},
|
| 1334 |
+
{
|
| 1335 |
+
"type": "image",
|
| 1336 |
+
"img_path": "images/49d03350932205f4c7456d7b8004e01f93a025b15a3769694659b63a66aebda9.jpg",
|
| 1337 |
+
"image_caption": [
|
| 1338 |
+
"(d) EFNet"
|
| 1339 |
+
],
|
| 1340 |
+
"image_footnote": [],
|
| 1341 |
+
"bbox": [
|
| 1342 |
+
589,
|
| 1343 |
+
348,
|
| 1344 |
+
665,
|
| 1345 |
+
406
|
| 1346 |
+
],
|
| 1347 |
+
"page_idx": 10
|
| 1348 |
+
},
|
| 1349 |
+
{
|
| 1350 |
+
"type": "image",
|
| 1351 |
+
"img_path": "images/b60a52f299b5b271275e91e328b201333ba9fc08023bb258ff4899f4394d7338.jpg",
|
| 1352 |
+
"image_caption": [
|
| 1353 |
+
"(e) NAFNet"
|
| 1354 |
+
],
|
| 1355 |
+
"image_footnote": [],
|
| 1356 |
+
"bbox": [
|
| 1357 |
+
668,
|
| 1358 |
+
348,
|
| 1359 |
+
741,
|
| 1360 |
+
406
|
| 1361 |
+
],
|
| 1362 |
+
"page_idx": 10
|
| 1363 |
+
},
|
| 1364 |
+
{
|
| 1365 |
+
"type": "text",
|
| 1366 |
+
"text": "consisted of 4000 iterations while keeping the other configurations consistent with the previous experiments.",
|
| 1367 |
+
"bbox": [
|
| 1368 |
+
214,
|
| 1369 |
+
563,
|
| 1370 |
+
784,
|
| 1371 |
+
594
|
| 1372 |
+
],
|
| 1373 |
+
"page_idx": 10
|
| 1374 |
+
},
|
| 1375 |
+
{
|
| 1376 |
+
"type": "text",
|
| 1377 |
+
"text": "4.2 Comparison with State-of-the-Art Methods",
|
| 1378 |
+
"text_level": 1,
|
| 1379 |
+
"bbox": [
|
| 1380 |
+
214,
|
| 1381 |
+
617,
|
| 1382 |
+
619,
|
| 1383 |
+
633
|
| 1384 |
+
],
|
| 1385 |
+
"page_idx": 10
|
| 1386 |
+
},
|
| 1387 |
+
{
|
| 1388 |
+
"type": "text",
|
| 1389 |
+
"text": "Table 1 shows the comparison results of our method and other SOTA methods on GoPro [24], HS-ERGB [50] and REBlur [33]. We compare our network with not only event-based methods, but also with some SOTA image-only deblurring networks enhancing with event input. The results illustrate that: (1) Our method outperforms all other methods in synthetic, semi-synthetic and real-world datasets, which indicates that with the proper utilization of events the designed approach is able to handle blur with different distributions. Specifically, the average gains of our method over the second-best one are 0.61dB, 0.25dB and 0.35dB in terms of PSNR on GoPro, HS-ERGB and REBlur datasets, respectively. (2) Through the enhancement of our proposed event representation, the HINet+ and the UFPNet+ increase 1.92dB and 1.16dB in PSNR on the benchmark GoPro dataset compared with their image-only version, respectively. Taking HINet as an example, with the same operation of [33], HINet with DA",
|
| 1390 |
+
"bbox": [
|
| 1391 |
+
212,
|
| 1392 |
+
643,
|
| 1393 |
+
787,
|
| 1394 |
+
840
|
| 1395 |
+
],
|
| 1396 |
+
"page_idx": 10
|
| 1397 |
+
},
|
| 1398 |
+
{
|
| 1399 |
+
"type": "header",
|
| 1400 |
+
"text": "Motion Aware Event Representation-driven Image Deblurring",
|
| 1401 |
+
"bbox": [
|
| 1402 |
+
321,
|
| 1403 |
+
114,
|
| 1404 |
+
732,
|
| 1405 |
+
128
|
| 1406 |
+
],
|
| 1407 |
+
"page_idx": 10
|
| 1408 |
+
},
|
| 1409 |
+
{
|
| 1410 |
+
"type": "page_number",
|
| 1411 |
+
"text": "11",
|
| 1412 |
+
"bbox": [
|
| 1413 |
+
767,
|
| 1414 |
+
114,
|
| 1415 |
+
782,
|
| 1416 |
+
126
|
| 1417 |
+
],
|
| 1418 |
+
"page_idx": 10
|
| 1419 |
+
},
|
| 1420 |
+
{
|
| 1421 |
+
"type": "image",
|
| 1422 |
+
"img_path": "images/734fd1b1a209722f46554a3beff3f560d60b9d4c52ce867556462ddd29643728.jpg",
|
| 1423 |
+
"image_caption": [
|
| 1424 |
+
"(a) GT"
|
| 1425 |
+
],
|
| 1426 |
+
"image_footnote": [],
|
| 1427 |
+
"bbox": [
|
| 1428 |
+
267,
|
| 1429 |
+
143,
|
| 1430 |
+
343,
|
| 1431 |
+
202
|
| 1432 |
+
],
|
| 1433 |
+
"page_idx": 11
|
| 1434 |
+
},
|
| 1435 |
+
{
|
| 1436 |
+
"type": "image",
|
| 1437 |
+
"img_path": "images/23665851adff685957098865bc95f18a700e064978f64467d82b3f335ad0e2c5.jpg",
|
| 1438 |
+
"image_caption": [
|
| 1439 |
+
"(b) Blurry"
|
| 1440 |
+
],
|
| 1441 |
+
"image_footnote": [],
|
| 1442 |
+
"bbox": [
|
| 1443 |
+
344,
|
| 1444 |
+
143,
|
| 1445 |
+
421,
|
| 1446 |
+
202
|
| 1447 |
+
],
|
| 1448 |
+
"page_idx": 11
|
| 1449 |
+
},
|
| 1450 |
+
{
|
| 1451 |
+
"type": "image",
|
| 1452 |
+
"img_path": "images/e436f01102b7a8802bc8f0f4fda16ca5e7ee47f048512df78a52a0737fd0df72.jpg",
|
| 1453 |
+
"image_caption": [
|
| 1454 |
+
"(c) UFPNet"
|
| 1455 |
+
],
|
| 1456 |
+
"image_footnote": [],
|
| 1457 |
+
"bbox": [
|
| 1458 |
+
424,
|
| 1459 |
+
143,
|
| 1460 |
+
500,
|
| 1461 |
+
202
|
| 1462 |
+
],
|
| 1463 |
+
"page_idx": 11
|
| 1464 |
+
},
|
| 1465 |
+
{
|
| 1466 |
+
"type": "image",
|
| 1467 |
+
"img_path": "images/116209ed0274009ad194fa824ab167736e9ae06fc2d09cdee0ae569b63bfb553.jpg",
|
| 1468 |
+
"image_caption": [
|
| 1469 |
+
"(d) NAFNet"
|
| 1470 |
+
],
|
| 1471 |
+
"image_footnote": [],
|
| 1472 |
+
"bbox": [
|
| 1473 |
+
501,
|
| 1474 |
+
143,
|
| 1475 |
+
578,
|
| 1476 |
+
202
|
| 1477 |
+
],
|
| 1478 |
+
"page_idx": 11
|
| 1479 |
+
},
|
| 1480 |
+
{
|
| 1481 |
+
"type": "image",
|
| 1482 |
+
"img_path": "images/75b923889f3b670525027bcd31c23cc3db0dd31392f76380b22b20f16a83888f.jpg",
|
| 1483 |
+
"image_caption": [
|
| 1484 |
+
"(e) EFNet"
|
| 1485 |
+
],
|
| 1486 |
+
"image_footnote": [],
|
| 1487 |
+
"bbox": [
|
| 1488 |
+
581,
|
| 1489 |
+
143,
|
| 1490 |
+
656,
|
| 1491 |
+
202
|
| 1492 |
+
],
|
| 1493 |
+
"page_idx": 11
|
| 1494 |
+
},
|
| 1495 |
+
{
|
| 1496 |
+
"type": "image",
|
| 1497 |
+
"img_path": "images/1e7c2eabaea83f858b8621cc979ffad7fca20e9e2d345effe15da8aa95aa0aa0.jpg",
|
| 1498 |
+
"image_caption": [
|
| 1499 |
+
"(f) Ours"
|
| 1500 |
+
],
|
| 1501 |
+
"image_footnote": [],
|
| 1502 |
+
"bbox": [
|
| 1503 |
+
658,
|
| 1504 |
+
143,
|
| 1505 |
+
735,
|
| 1506 |
+
202
|
| 1507 |
+
],
|
| 1508 |
+
"page_idx": 11
|
| 1509 |
+
},
|
| 1510 |
+
{
|
| 1511 |
+
"type": "image",
|
| 1512 |
+
"img_path": "images/41e0651326b76cccb5d5710c1442995461b95c4ddbe1fcacc18b94d134bc4605.jpg",
|
| 1513 |
+
"image_caption": [
|
| 1514 |
+
"Fig. 6: Qualitative comparisons on the HS-ERGB dataset."
|
| 1515 |
+
],
|
| 1516 |
+
"image_footnote": [],
|
| 1517 |
+
"bbox": [
|
| 1518 |
+
267,
|
| 1519 |
+
256,
|
| 1520 |
+
382,
|
| 1521 |
+
345
|
| 1522 |
+
],
|
| 1523 |
+
"page_idx": 11
|
| 1524 |
+
},
|
| 1525 |
+
{
|
| 1526 |
+
"type": "image",
|
| 1527 |
+
"img_path": "images/a720aa58896079b0e0b39e7458067a7fbec6479614a71146e05ad3f99b7d9700.jpg",
|
| 1528 |
+
"image_caption": [
|
| 1529 |
+
"(a) Blurry"
|
| 1530 |
+
],
|
| 1531 |
+
"image_footnote": [],
|
| 1532 |
+
"bbox": [
|
| 1533 |
+
267,
|
| 1534 |
+
348,
|
| 1535 |
+
380,
|
| 1536 |
+
434
|
| 1537 |
+
],
|
| 1538 |
+
"page_idx": 11
|
| 1539 |
+
},
|
| 1540 |
+
{
|
| 1541 |
+
"type": "image",
|
| 1542 |
+
"img_path": "images/68e51ec54743ea9c8e64884178afd1a439dff5366a8d45a80122461dcc428e57.jpg",
|
| 1543 |
+
"image_caption": [],
|
| 1544 |
+
"image_footnote": [],
|
| 1545 |
+
"bbox": [
|
| 1546 |
+
383,
|
| 1547 |
+
257,
|
| 1548 |
+
500,
|
| 1549 |
+
345
|
| 1550 |
+
],
|
| 1551 |
+
"page_idx": 11
|
| 1552 |
+
},
|
| 1553 |
+
{
|
| 1554 |
+
"type": "image",
|
| 1555 |
+
"img_path": "images/aca0016fe3ce7794d334f49c234da809ebe1c6ab4418653c1f8df20c1c5fd89a.jpg",
|
| 1556 |
+
"image_caption": [
|
| 1557 |
+
"(b) SCER [33]"
|
| 1558 |
+
],
|
| 1559 |
+
"image_footnote": [],
|
| 1560 |
+
"bbox": [
|
| 1561 |
+
385,
|
| 1562 |
+
348,
|
| 1563 |
+
500,
|
| 1564 |
+
433
|
| 1565 |
+
],
|
| 1566 |
+
"page_idx": 11
|
| 1567 |
+
},
|
| 1568 |
+
{
|
| 1569 |
+
"type": "image",
|
| 1570 |
+
"img_path": "images/c83f0d2d8eb537da6533e4b9f86be810fee0e31af3b33665379c2553ef6c7061.jpg",
|
| 1571 |
+
"image_caption": [],
|
| 1572 |
+
"image_footnote": [],
|
| 1573 |
+
"bbox": [
|
| 1574 |
+
501,
|
| 1575 |
+
257,
|
| 1576 |
+
617,
|
| 1577 |
+
345
|
| 1578 |
+
],
|
| 1579 |
+
"page_idx": 11
|
| 1580 |
+
},
|
| 1581 |
+
{
|
| 1582 |
+
"type": "image",
|
| 1583 |
+
"img_path": "images/e006b54c00ea90658069b0fd4689d9fe01929f4c59460730601fbe9ea5ab731a.jpg",
|
| 1584 |
+
"image_caption": [
|
| 1585 |
+
"(c) DA (Ours)"
|
| 1586 |
+
],
|
| 1587 |
+
"image_footnote": [],
|
| 1588 |
+
"bbox": [
|
| 1589 |
+
503,
|
| 1590 |
+
348,
|
| 1591 |
+
617,
|
| 1592 |
+
434
|
| 1593 |
+
],
|
| 1594 |
+
"page_idx": 11
|
| 1595 |
+
},
|
| 1596 |
+
{
|
| 1597 |
+
"type": "image",
|
| 1598 |
+
"img_path": "images/2355dc71134c668061af47e87b50aa1f1aa24daa40fa40843e9198ed91743256.jpg",
|
| 1599 |
+
"image_caption": [],
|
| 1600 |
+
"image_footnote": [],
|
| 1601 |
+
"bbox": [
|
| 1602 |
+
620,
|
| 1603 |
+
257,
|
| 1604 |
+
733,
|
| 1605 |
+
345
|
| 1606 |
+
],
|
| 1607 |
+
"page_idx": 11
|
| 1608 |
+
},
|
| 1609 |
+
{
|
| 1610 |
+
"type": "image",
|
| 1611 |
+
"img_path": "images/f1a387eda2a305341be0a7284808fbc25adaa1a60547ed3ebee3eb97710f9f44.jpg",
|
| 1612 |
+
"image_caption": [
|
| 1613 |
+
"(d) GT",
|
| 1614 |
+
"Fig. 7: Visual comparisons on GoPro dataset with different event representation. Compared to SCER representation, our DA representation reduces edge artifacts."
|
| 1615 |
+
],
|
| 1616 |
+
"image_footnote": [],
|
| 1617 |
+
"bbox": [
|
| 1618 |
+
620,
|
| 1619 |
+
348,
|
| 1620 |
+
733,
|
| 1621 |
+
434
|
| 1622 |
+
],
|
| 1623 |
+
"page_idx": 11
|
| 1624 |
+
},
|
| 1625 |
+
{
|
| 1626 |
+
"type": "text",
|
| 1627 |
+
"text": "event representation exceeds 0.94dB over the result in [33], which presents the fact that DA extracts more conducive information for motion deblurring.",
|
| 1628 |
+
"bbox": [
|
| 1629 |
+
212,
|
| 1630 |
+
517,
|
| 1631 |
+
784,
|
| 1632 |
+
546
|
| 1633 |
+
],
|
| 1634 |
+
"page_idx": 11
|
| 1635 |
+
},
|
| 1636 |
+
{
|
| 1637 |
+
"type": "text",
|
| 1638 |
+
"text": "We also show the visual comparison results to verify the effectiveness of our method in Figure 4, Figure 5 and Figure 6 with representative samples from each dataset. As illustrated in the figures, our model reconstructs better details and has minor spatial distortions compared with other methods. This indicates that our event representation perceives more motion information and the modules effectively embed it to reconstruct a high-quality image with enhanced details.",
|
| 1639 |
+
"bbox": [
|
| 1640 |
+
212,
|
| 1641 |
+
547,
|
| 1642 |
+
784,
|
| 1643 |
+
638
|
| 1644 |
+
],
|
| 1645 |
+
"page_idx": 11
|
| 1646 |
+
},
|
| 1647 |
+
{
|
| 1648 |
+
"type": "text",
|
| 1649 |
+
"text": "4.3 Ablation Studies",
|
| 1650 |
+
"text_level": 1,
|
| 1651 |
+
"bbox": [
|
| 1652 |
+
214,
|
| 1653 |
+
662,
|
| 1654 |
+
400,
|
| 1655 |
+
676
|
| 1656 |
+
],
|
| 1657 |
+
"page_idx": 11
|
| 1658 |
+
},
|
| 1659 |
+
{
|
| 1660 |
+
"type": "text",
|
| 1661 |
+
"text": "We conduct plenty ablation experiments on GoPro dataset to verify the effectiveness of our event representation and proposed method.",
|
| 1662 |
+
"bbox": [
|
| 1663 |
+
212,
|
| 1664 |
+
688,
|
| 1665 |
+
784,
|
| 1666 |
+
718
|
| 1667 |
+
],
|
| 1668 |
+
"page_idx": 11
|
| 1669 |
+
},
|
| 1670 |
+
{
|
| 1671 |
+
"type": "text",
|
| 1672 |
+
"text": "Effectiveness of Event Representation. To demonstrate the adaptability of our event representation for motion deblurring tasks, as shown in Table 2, we replace our DA representation with several other types of representations and trained the network accordingly. The results reveal certain shortcomings of the alternative representations. Firstly, the stack representation simply accumulates all events on their polarities which leaves out the temporal information. Secondly, the SBT approach [40], which subdivides the entire exposure time equally, proves to be inadequate for modeling the varying degrees of blur. Thirdly, the SCER",
|
| 1673 |
+
"bbox": [
|
| 1674 |
+
212,
|
| 1675 |
+
719,
|
| 1676 |
+
785,
|
| 1677 |
+
839
|
| 1678 |
+
],
|
| 1679 |
+
"page_idx": 11
|
| 1680 |
+
},
|
| 1681 |
+
{
|
| 1682 |
+
"type": "page_number",
|
| 1683 |
+
"text": "12",
|
| 1684 |
+
"bbox": [
|
| 1685 |
+
217,
|
| 1686 |
+
114,
|
| 1687 |
+
235,
|
| 1688 |
+
126
|
| 1689 |
+
],
|
| 1690 |
+
"page_idx": 11
|
| 1691 |
+
},
|
| 1692 |
+
{
|
| 1693 |
+
"type": "header",
|
| 1694 |
+
"text": "Z. Sun et al.",
|
| 1695 |
+
"bbox": [
|
| 1696 |
+
271,
|
| 1697 |
+
114,
|
| 1698 |
+
354,
|
| 1699 |
+
126
|
| 1700 |
+
],
|
| 1701 |
+
"page_idx": 11
|
| 1702 |
+
},
|
| 1703 |
+
{
|
| 1704 |
+
"type": "table",
|
| 1705 |
+
"img_path": "images/5d434a7baeecb04e0bb393f3d63371deeecd9117414eb116d1158d1296fd0d7d.jpg",
|
| 1706 |
+
"table_caption": [
|
| 1707 |
+
"Table 2: Ablation study of each module effects."
|
| 1708 |
+
],
|
| 1709 |
+
"table_footnote": [],
|
| 1710 |
+
"table_body": "<table><tr><td>Event Representation</td><td>RME</td><td>Fusion</td><td>PSNR</td><td>SSIM</td></tr><tr><td>Stack</td><td></td><td></td><td>33.24</td><td>0.953</td></tr><tr><td>SBT [40]</td><td>✓</td><td>FAF(Ours)</td><td>35.63</td><td>0.973</td></tr><tr><td>SCER [33]</td><td></td><td></td><td>35.81</td><td>0.975</td></tr><tr><td></td><td></td><td>Concat.</td><td>34.02</td><td>0.967</td></tr><tr><td></td><td></td><td>Add</td><td>34.74</td><td>0.970</td></tr><tr><td></td><td></td><td>Attention</td><td>35.75</td><td>0.974</td></tr><tr><td>ACA(Ours)</td><td>✓</td><td>Only Align Event+Attention</td><td>35.78</td><td>0.975</td></tr><tr><td></td><td></td><td>Only Align Image+Attention</td><td>35.74</td><td>0.974</td></tr><tr><td></td><td></td><td></td><td>35.42</td><td>0.973</td></tr><tr><td>ACA(Ours)</td><td>✗</td><td>FAF(Ours)</td><td>36.07</td><td>0.976</td></tr><tr><td></td><td>✓</td><td></td><td></td><td></td></tr></table>",
|
| 1711 |
+
"bbox": [
|
| 1712 |
+
241,
|
| 1713 |
+
171,
|
| 1714 |
+
756,
|
| 1715 |
+
339
|
| 1716 |
+
],
|
| 1717 |
+
"page_idx": 12
|
| 1718 |
+
},
|
| 1719 |
+
{
|
| 1720 |
+
"type": "image",
|
| 1721 |
+
"img_path": "images/81f7908c63a90b433e90594dc13fa912cb37f5504f5f993a8d50ce93a37c0d8f.jpg",
|
| 1722 |
+
"image_caption": [
|
| 1723 |
+
"(a) w/o RME"
|
| 1724 |
+
],
|
| 1725 |
+
"image_footnote": [],
|
| 1726 |
+
"bbox": [
|
| 1727 |
+
248,
|
| 1728 |
+
353,
|
| 1729 |
+
415,
|
| 1730 |
+
426
|
| 1731 |
+
],
|
| 1732 |
+
"page_idx": 12
|
| 1733 |
+
},
|
| 1734 |
+
{
|
| 1735 |
+
"type": "image",
|
| 1736 |
+
"img_path": "images/11be934ee6cb38b10dfb35c3fc2d91777a02b74c7898ac526a9795643b474ccb.jpg",
|
| 1737 |
+
"image_caption": [
|
| 1738 |
+
"(b) w/o RME"
|
| 1739 |
+
],
|
| 1740 |
+
"image_footnote": [],
|
| 1741 |
+
"bbox": [
|
| 1742 |
+
416,
|
| 1743 |
+
353,
|
| 1744 |
+
584,
|
| 1745 |
+
426
|
| 1746 |
+
],
|
| 1747 |
+
"page_idx": 12
|
| 1748 |
+
},
|
| 1749 |
+
{
|
| 1750 |
+
"type": "image",
|
| 1751 |
+
"img_path": "images/1f18276d5b5d8ae61f7cddeb2945d238a52e8902dde8319fdaa11eaef552ba1d.jpg",
|
| 1752 |
+
"image_caption": [
|
| 1753 |
+
"(c) w/o RME"
|
| 1754 |
+
],
|
| 1755 |
+
"image_footnote": [],
|
| 1756 |
+
"bbox": [
|
| 1757 |
+
586,
|
| 1758 |
+
353,
|
| 1759 |
+
754,
|
| 1760 |
+
426
|
| 1761 |
+
],
|
| 1762 |
+
"page_idx": 12
|
| 1763 |
+
},
|
| 1764 |
+
{
|
| 1765 |
+
"type": "image",
|
| 1766 |
+
"img_path": "images/7815bdf963277e3713ad059dbd509e05af04828114043170576ce6d39efadbb0.jpg",
|
| 1767 |
+
"image_caption": [
|
| 1768 |
+
"(d) w/RME",
|
| 1769 |
+
"Fig. 8: Visual comparison results of w/ and w/o RME module. All of these scenarios involve movement in different directions and speeds, RME module can successfully capture motion features in these complex motion scenarios. Zoom in for better view."
|
| 1770 |
+
],
|
| 1771 |
+
"image_footnote": [],
|
| 1772 |
+
"bbox": [
|
| 1773 |
+
250,
|
| 1774 |
+
440,
|
| 1775 |
+
415,
|
| 1776 |
+
513
|
| 1777 |
+
],
|
| 1778 |
+
"page_idx": 12
|
| 1779 |
+
},
|
| 1780 |
+
{
|
| 1781 |
+
"type": "image",
|
| 1782 |
+
"img_path": "images/7a552d1a9227850a52be4560f38c65a17ba9c1bf2257581d64e68c82f2f56e03.jpg",
|
| 1783 |
+
"image_caption": [
|
| 1784 |
+
"(e) w/RME"
|
| 1785 |
+
],
|
| 1786 |
+
"image_footnote": [],
|
| 1787 |
+
"bbox": [
|
| 1788 |
+
416,
|
| 1789 |
+
440,
|
| 1790 |
+
584,
|
| 1791 |
+
513
|
| 1792 |
+
],
|
| 1793 |
+
"page_idx": 12
|
| 1794 |
+
},
|
| 1795 |
+
{
|
| 1796 |
+
"type": "image",
|
| 1797 |
+
"img_path": "images/a03e2927b922c9edc7b8ca0699305e75b5d34138e48b931449aba8642488c107.jpg",
|
| 1798 |
+
"image_caption": [
|
| 1799 |
+
"(f) w/RME"
|
| 1800 |
+
],
|
| 1801 |
+
"image_footnote": [],
|
| 1802 |
+
"bbox": [
|
| 1803 |
+
586,
|
| 1804 |
+
440,
|
| 1805 |
+
753,
|
| 1806 |
+
513
|
| 1807 |
+
],
|
| 1808 |
+
"page_idx": 12
|
| 1809 |
+
},
|
| 1810 |
+
{
|
| 1811 |
+
"type": "text",
|
| 1812 |
+
"text": "method [33], which directly accumulates event polarities, is found to be incapable of capturing the influence of absolute intensity changes. This limitation is illustrated in Figure 3. In contrast, our DA representation not only possesses a solid physical foundation for event generation but also perceives more motion patterns. As a result, it achieves a 0.26dB improvement in performance compared to the best alternative representation. We also provide the visual comparison on the image reconstruction results in Figure 7. Our DA representation reduces motion-generated edge artifacts, resulting in a better deblurring effect.",
|
| 1813 |
+
"bbox": [
|
| 1814 |
+
212,
|
| 1815 |
+
611,
|
| 1816 |
+
787,
|
| 1817 |
+
733
|
| 1818 |
+
],
|
| 1819 |
+
"page_idx": 12
|
| 1820 |
+
},
|
| 1821 |
+
{
|
| 1822 |
+
"type": "text",
|
| 1823 |
+
"text": "Effectiveness of Recurrent Motion Extraction Module. To assess the RME module's contribution, we replace it with a baseline conv block and adjust the parameters for fair comparison. In this approach, we concatenate event frames of different time scales and feed them into the conv block. The findings, as presented in the last two rows of Table 2, clearly indicate that the contribution of $0.65\\mathrm{dB}$ increment in PSNR stems from the network structure design, rather than an increase in parameter quantity. This shows the importance of incorpo",
|
| 1824 |
+
"bbox": [
|
| 1825 |
+
212,
|
| 1826 |
+
734,
|
| 1827 |
+
787,
|
| 1828 |
+
840
|
| 1829 |
+
],
|
| 1830 |
+
"page_idx": 12
|
| 1831 |
+
},
|
| 1832 |
+
{
|
| 1833 |
+
"type": "header",
|
| 1834 |
+
"text": "Motion Aware Event Representation-driven Image Deblurring",
|
| 1835 |
+
"bbox": [
|
| 1836 |
+
321,
|
| 1837 |
+
114,
|
| 1838 |
+
730,
|
| 1839 |
+
128
|
| 1840 |
+
],
|
| 1841 |
+
"page_idx": 12
|
| 1842 |
+
},
|
| 1843 |
+
{
|
| 1844 |
+
"type": "page_number",
|
| 1845 |
+
"text": "13",
|
| 1846 |
+
"bbox": [
|
| 1847 |
+
767,
|
| 1848 |
+
116,
|
| 1849 |
+
784,
|
| 1850 |
+
126
|
| 1851 |
+
],
|
| 1852 |
+
"page_idx": 12
|
| 1853 |
+
},
|
| 1854 |
+
{
|
| 1855 |
+
"type": "text",
|
| 1856 |
+
"text": "rating the module for effectively leveraging motion information in multi-timescales and improving performance in motion deblurring tasks. We also present the visual comparisons when removing the RME module in Figure 8, the RME module adeptly captures both fast local and slow global motion characteristics.",
|
| 1857 |
+
"bbox": [
|
| 1858 |
+
212,
|
| 1859 |
+
146,
|
| 1860 |
+
787,
|
| 1861 |
+
205
|
| 1862 |
+
],
|
| 1863 |
+
"page_idx": 13
|
| 1864 |
+
},
|
| 1865 |
+
{
|
| 1866 |
+
"type": "text",
|
| 1867 |
+
"text": "Effectiveness of Feature Alignment and Fusion Module. To illustrate the effectiveness of our FAF module, we compare it with several common fusion methods, including concatenation, addition, and cross-modality attention. As shown in Table 2, it is arduous to utilize the motion information contained in event feature through the direct concatenation or addition. While attention mechanisms can capture relative parts in the two features, the inherent inconsistencies within the two modalities still lead to a 0.32dB performance gap when compared to our FAF module. We also use a t-SNE visualization in Figure 9 to demonstrate the effectiveness of our feature alignment, where two originally in-congruent modal features are mixed up together after the FAF module.",
|
| 1868 |
+
"bbox": [
|
| 1869 |
+
212,
|
| 1870 |
+
207,
|
| 1871 |
+
787,
|
| 1872 |
+
358
|
| 1873 |
+
],
|
| 1874 |
+
"page_idx": 13
|
| 1875 |
+
},
|
| 1876 |
+
{
|
| 1877 |
+
"type": "text",
|
| 1878 |
+
"text": "Furthermore, we also verify the importance of bi-directional alignment by conducting experiments solely on aligning event or image features. Since both events and frames provide crucial information for image restoration, such as edge motion information and textual information, respectively, it is unreasonable to let one modality dominate the alignment process. As shown in Table 2, the bi-directional alignment outperforms the highest score achieved by one-way correction by 0.29dB. This demonstrates the necessity of fully utilizing the mutual information between the two modalities in order to achieve better image reconstruction results.",
|
| 1879 |
+
"bbox": [
|
| 1880 |
+
212,
|
| 1881 |
+
359,
|
| 1882 |
+
488,
|
| 1883 |
+
630
|
| 1884 |
+
],
|
| 1885 |
+
"page_idx": 13
|
| 1886 |
+
},
|
| 1887 |
+
{
|
| 1888 |
+
"type": "image",
|
| 1889 |
+
"img_path": "images/5fb4c0b4afc0a3f1a56e00ab48428f673c8a69bd8eea8b47eb91ccc2480c19a4.jpg",
|
| 1890 |
+
"image_caption": [
|
| 1891 |
+
"Fig. 9: t-SNE visualization of original and aligned features. The distributions of the aligned two features overlap more compared to the original features. This indicates the effectiveness of FAF module to reduce the inter-modal inconsistency."
|
| 1892 |
+
],
|
| 1893 |
+
"image_footnote": [],
|
| 1894 |
+
"bbox": [
|
| 1895 |
+
516,
|
| 1896 |
+
375,
|
| 1897 |
+
769,
|
| 1898 |
+
545
|
| 1899 |
+
],
|
| 1900 |
+
"page_idx": 13
|
| 1901 |
+
},
|
| 1902 |
+
{
|
| 1903 |
+
"type": "text",
|
| 1904 |
+
"text": "5 Conclusion",
|
| 1905 |
+
"text_level": 1,
|
| 1906 |
+
"bbox": [
|
| 1907 |
+
215,
|
| 1908 |
+
655,
|
| 1909 |
+
359,
|
| 1910 |
+
671
|
| 1911 |
+
],
|
| 1912 |
+
"page_idx": 13
|
| 1913 |
+
},
|
| 1914 |
+
{
|
| 1915 |
+
"type": "text",
|
| 1916 |
+
"text": "In this paper, we propose an event-based image deblurring method. Firstly, we rethink the limit of traditional event preprocessing method and propose a new deviation accumulation (DA) technique. With deviation, the new approach can distinguish the order of events and perceive more motion patterns. Moreover, we design an image deblurring network equipped with recurrent motion extraction (RME) module and feature alignment and fusion (FAF) module to fit our proposed event representation. The RME module can better perceive motions in different time scales, and the FAF module promotes better integration of the two modalities. Finally, extensive experimental results over several datasets demonstrate the effectiveness of the proposed algorithm.",
|
| 1917 |
+
"bbox": [
|
| 1918 |
+
212,
|
| 1919 |
+
688,
|
| 1920 |
+
787,
|
| 1921 |
+
840
|
| 1922 |
+
],
|
| 1923 |
+
"page_idx": 13
|
| 1924 |
+
},
|
| 1925 |
+
{
|
| 1926 |
+
"type": "page_number",
|
| 1927 |
+
"text": "14",
|
| 1928 |
+
"bbox": [
|
| 1929 |
+
217,
|
| 1930 |
+
114,
|
| 1931 |
+
235,
|
| 1932 |
+
126
|
| 1933 |
+
],
|
| 1934 |
+
"page_idx": 13
|
| 1935 |
+
},
|
| 1936 |
+
{
|
| 1937 |
+
"type": "header",
|
| 1938 |
+
"text": "Z. Sun et al.",
|
| 1939 |
+
"bbox": [
|
| 1940 |
+
271,
|
| 1941 |
+
114,
|
| 1942 |
+
354,
|
| 1943 |
+
126
|
| 1944 |
+
],
|
| 1945 |
+
"page_idx": 13
|
| 1946 |
+
},
|
| 1947 |
+
{
|
| 1948 |
+
"type": "text",
|
| 1949 |
+
"text": "Acknowledgements. This work was supported by National Natural Science Foundation of China (NSFC) under Grants 62225207 and 62276243.",
|
| 1950 |
+
"bbox": [
|
| 1951 |
+
215,
|
| 1952 |
+
146,
|
| 1953 |
+
784,
|
| 1954 |
+
176
|
| 1955 |
+
],
|
| 1956 |
+
"page_idx": 14
|
| 1957 |
+
},
|
| 1958 |
+
{
|
| 1959 |
+
"type": "text",
|
| 1960 |
+
"text": "References",
|
| 1961 |
+
"text_level": 1,
|
| 1962 |
+
"bbox": [
|
| 1963 |
+
215,
|
| 1964 |
+
202,
|
| 1965 |
+
321,
|
| 1966 |
+
218
|
| 1967 |
+
],
|
| 1968 |
+
"page_idx": 14
|
| 1969 |
+
},
|
| 1970 |
+
{
|
| 1971 |
+
"type": "list",
|
| 1972 |
+
"sub_type": "ref_text",
|
| 1973 |
+
"list_items": [
|
| 1974 |
+
"1. Benosman, R., Clercq, C., Lagorce, X., Ieng, S.H., Bartolozzi, C.: Event-based visual flow. IEEE transactions on neural networks and learning systems 25(2), 407-417 (2013)",
|
| 1975 |
+
"2. Chang, M., Feng, H., Xu, Z., Li, Q.: Low-light image restoration with short-and long-exposure raw pairs. IEEE Transactions on Multimedia 24, 702-714 (2021)",
|
| 1976 |
+
"3. Chen, L., Chu, X., Zhang, X., Sun, J.: Simple baselines for image restoration. In: Proceedings of the European conference on computer vision (ECCV). pp. 17-33. Springer (2022)",
|
| 1977 |
+
"4. Chen, L., Lu, X., Zhang, J., Chu, X., Chen, C.: Hinet: Half instance normalization network for image restoration. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 182-192 (2021)",
|
| 1978 |
+
"5. Cook, M., Gugelmann, L., Jug, F., Krautz, C., Steger, A.: Interacting maps for fast visual interpretation. In: The 2011 International Joint Conference on Neural Networks. pp. 770-776. IEEE (2011)",
|
| 1979 |
+
"6. Fang, Z., Wu, F., Dong, W., Li, X., Wu, J., Shi, G.: Self-supervised non-uniform kernel estimation with flow-based motion prior for blind image deblurring. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18105-18114 (2023)",
|
| 1980 |
+
"7. Gehrig, D., Rebecq, H., Gallego, G., Scaramuzza, D.: Eklt: Asynchronous photometric feature tracking using events and frames. International Journal of Computer Vision 128(3), 601-618 (2020)",
|
| 1981 |
+
"8. Ji, S.W., Lee, J., Kim, S.W., Hong, J.P., Baek, S.J., Jung, S.W., Ko, S.J.: Xydeblur: divide and conquer for single image deblurring. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17421-17430 (2022)",
|
| 1982 |
+
"9. Jiang, Z., Zhang, Y., Zou, D., Ren, J., Lv, J., Liu, Y.: Learning event-based motion deblurring. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3320-3329 (2020)",
|
| 1983 |
+
"10. Jin, M., Meishvili, G., Favaro, P.: Learning to extract a video sequence from a single motion-blurred image. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 6334-6342 (2018)",
|
| 1984 |
+
"1. Kim, H., Handa, A., Benosman, R., Ieng, S.H., Davison, A.J.: Simultaneous mosaicing and tracking with an event camera. J. Solid State Circ 43, 566-576 (2008)",
|
| 1985 |
+
"2. Kim, T., Lee, J., Wang, L., Yoon, K.J.: Event-guided deblurring of unknown exposure time videos. In: Proceedings of the European conference on computer vision (ECCV). pp. 519-538. Springer (2022)",
|
| 1986 |
+
"3. Krishnan, D., Tay, T., Fergus, R.: Blind deconvolution using a normalized sparsity measure. In: CVPR 2011. pp. 233-240. IEEE (2011)",
|
| 1987 |
+
"4. Lagorce, X., Orchard, G., Galluppi, F., Shi, B.E., Benosman, R.B.: Hots: a hierarchy of event-based time-surfaces for pattern recognition. IEEE transactions on pattern analysis and machine intelligence 39(7), 1346-1359 (2016)",
|
| 1988 |
+
"5. Levin, A., Weiss, Y., Durand, F., Freeman, W.T.: Understanding and evaluating blind deconvolution algorithms. In: 2009 IEEE conference on computer vision and pattern recognition. pp. 1964-1971. IEEE (2009)"
|
| 1989 |
+
],
|
| 1990 |
+
"bbox": [
|
| 1991 |
+
225,
|
| 1992 |
+
237,
|
| 1993 |
+
784,
|
| 1994 |
+
839
|
| 1995 |
+
],
|
| 1996 |
+
"page_idx": 14
|
| 1997 |
+
},
|
| 1998 |
+
{
|
| 1999 |
+
"type": "header",
|
| 2000 |
+
"text": "Motion Aware Event Representation-driven Image Deblurring",
|
| 2001 |
+
"bbox": [
|
| 2002 |
+
321,
|
| 2003 |
+
114,
|
| 2004 |
+
730,
|
| 2005 |
+
128
|
| 2006 |
+
],
|
| 2007 |
+
"page_idx": 14
|
| 2008 |
+
},
|
| 2009 |
+
{
|
| 2010 |
+
"type": "page_number",
|
| 2011 |
+
"text": "15",
|
| 2012 |
+
"bbox": [
|
| 2013 |
+
767,
|
| 2014 |
+
116,
|
| 2015 |
+
784,
|
| 2016 |
+
126
|
| 2017 |
+
],
|
| 2018 |
+
"page_idx": 14
|
| 2019 |
+
},
|
| 2020 |
+
{
|
| 2021 |
+
"type": "list",
|
| 2022 |
+
"sub_type": "ref_text",
|
| 2023 |
+
"list_items": [
|
| 2024 |
+
"16. Li, D., Shi, X., Zhang, Y., Cheung, K.C., See, S., Wang, X., Qin, H., Li, H.: A simple baseline for video restoration with grouped spatial-temporal shift. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9822-9832 (2023)",
|
| 2025 |
+
"17. Li, D., Zhang, Y., Cheung, K.C., Wang, X., Qin, H., Li, H.: Learning degradation representations for image deblurring. In: Proceedings of the European conference on computer vision (ECCV). pp. 736-753. Springer (2022)",
|
| 2026 |
+
"18. Li, Y., Yu, A.W., Meng, T., Caine, B., Ngiam, J., Peng, D., Shen, J., Lu, Y., Zhou, D., Le, Q.V., et al.: Deepfusion: Lidar-camera deep fusion for multi-modal 3d object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17182-17191 (2022)",
|
| 2027 |
+
"19. Liang, M., Yang, B., Wang, S., Urtasun, R.: Deep continuous fusion for multi-sensor 3d object detection. In: Proceedings of the European conference on computer vision (ECCV). pp. 641-656 (2018)",
|
| 2028 |
+
"20. Liang, X., Qian, Y., Guo, Q., Cheng, H., Liang, J.: Af: An association-based fusion method for multi-modal classification. IEEE Transactions on Pattern Analysis and Machine Intelligence 44(12), 9236-9254 (2021)",
|
| 2029 |
+
"21. Liu, M., Delbruck, T.: Adaptive time-slice block-matching optical flow algorithm for dynamic vision sensors. BMVC (2018)",
|
| 2030 |
+
"22. Manderscheid, J., Sironi, A., Bourdis, N., Migliore, D., Lepetit, V.: Speed invariant time surface for learning to detect corner points with event-based cameras. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10245-10254 (2019)",
|
| 2031 |
+
"23. Maqueda, A.I., Loquercio, A., Gallego, G., García, N., Scaramuzza, D.: Event-based vision meets deep learning on steering prediction for self-driving cars. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 5419-5427 (2018)",
|
| 2032 |
+
"24. Nah, S., Hyun Kim, T., Mu Lee, K.: Deep multi-scale convolutional neural network for dynamic scene deblurring. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3883-3891 (2017)",
|
| 2033 |
+
"25. Pan, L., Scheerlinck, C., Yu, X., Hartley, R., Liu, M., Dai, Y.: Bringing a blurry frame alive at high frame-rate with an event camera. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6820-6829 (2019)",
|
| 2034 |
+
"26. Paredes-Vallés, F., Scheper, K.Y., De Croon, G.C.: Unsupervised learning of a hierarchical spiking neural network for optical flow estimation: From events to global motion perception. IEEE transactions on pattern analysis and machine intelligence 42(8), 2051-2064 (2019)",
|
| 2035 |
+
"27. Rebecq, H., Gehrig, D., Scaramuzza, D.: Esim: an open event camera simulator. In: Conference on robot learning. pp. 969-982. PMLR (2018)",
|
| 2036 |
+
"28. Ren, D., Shang, W., Yang, Y., Zuo, W.: Aggregating long-term sharp features via hybrid transformers for video deblurring. arXiv preprint arXiv:2309.07054 (2023)",
|
| 2037 |
+
"29. Ren, S., Du, Y., Lv, J., Han, G., He, S.: Learning from the master: Distilling cross-modal advanced knowledge for lip reading. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13325-13333 (2021)",
|
| 2038 |
+
"30. Shang, W., Ren, D., Zou, D., Ren, J.S., Luo, P., Zuo, W.: Bringing events into video deblurring with non-consecutively blurry frames. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4531-4540 (2021)",
|
| 2039 |
+
"31. Song, C., Bajaj, C., Huang, Q.: Deblursr: Event-based motion deblurring under the spiking representation. arXiv preprint arXiv:2303.08977 (2023)"
|
| 2040 |
+
],
|
| 2041 |
+
"bbox": [
|
| 2042 |
+
215,
|
| 2043 |
+
146,
|
| 2044 |
+
785,
|
| 2045 |
+
839
|
| 2046 |
+
],
|
| 2047 |
+
"page_idx": 15
|
| 2048 |
+
},
|
| 2049 |
+
{
|
| 2050 |
+
"type": "page_number",
|
| 2051 |
+
"text": "16",
|
| 2052 |
+
"bbox": [
|
| 2053 |
+
217,
|
| 2054 |
+
114,
|
| 2055 |
+
235,
|
| 2056 |
+
126
|
| 2057 |
+
],
|
| 2058 |
+
"page_idx": 15
|
| 2059 |
+
},
|
| 2060 |
+
{
|
| 2061 |
+
"type": "header",
|
| 2062 |
+
"text": "Z. Sun et al.",
|
| 2063 |
+
"bbox": [
|
| 2064 |
+
271,
|
| 2065 |
+
114,
|
| 2066 |
+
354,
|
| 2067 |
+
126
|
| 2068 |
+
],
|
| 2069 |
+
"page_idx": 15
|
| 2070 |
+
},
|
| 2071 |
+
{
|
| 2072 |
+
"type": "list",
|
| 2073 |
+
"sub_type": "ref_text",
|
| 2074 |
+
"list_items": [
|
| 2075 |
+
"32. Su, S., Delbracio, M., Wang, J., Sapiro, G., Heidrich, W., Wang, O.: Deep video deblurring for hand-held cameras. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1279-1288 (2017)",
|
| 2076 |
+
"33. Sun, L., Sakaridis, C., Liang, J., Jiang, Q., Yang, K., Sun, P., Ye, Y., Wang, K., Gool, L.V.: Event-based fusion for motion deblurring with cross-modal attention. In: Proceedings of the European conference on computer vision (ECCV). pp. 412-428. Springer (2022)",
|
| 2077 |
+
"34. Sun, L., Cho, S., Wang, J., Hays, J.: Edge-based blur kernel estimation using patch priors. In: IEEE international conference on computational photography (ICCP). pp. 1-8. IEEE (2013)",
|
| 2078 |
+
"35. Tao, X., Gao, H., Shen, X., Wang, J., Jia, J.: Scale-recurrent network for deep image deblurring. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 8174-8182 (2018)",
|
| 2079 |
+
"36. Tian, Y., Krishnan, D., Isola, P.: Contrastive multiview coding. In: Proceedings of the European conference on computer vision (ECCV). pp. 776-794. Springer (2020)",
|
| 2080 |
+
"37. Tulyakov, S., Gehrig, D., Georgoulis, S., Erbach, J., Gehrig, M., Li, Y., Scaramuzza, D.: Time lens: Event-based video frame interpolation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 16155-16164 (2021)",
|
| 2081 |
+
"38. Vitoria, P., Georgoulis, S., Tulyakov, S., Bochicchio, A., Erbach, J., Li, Y.: Event-based image deblurring with dynamic motion awareness. In: Proceedings of the European conference on computer vision (ECCV). pp. 95-112. Springer (2022)",
|
| 2082 |
+
"39. Wang, B., He, J., Yu, L., Xia, G.S., Yang, W.: Event enhanced high-quality image recovery. In: Proceedings of the European conference on computer vision (ECCV). pp. 155-171. Springer (2020)",
|
| 2083 |
+
"40. Wang, L., Ho, Y.S., Yoon, K.J., et al.: Event-based high dynamic range image and very high frame rate video generation using conditional generative adversarial networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10081-10090 (2019)",
|
| 2084 |
+
"41. Weikersdorfer, D., Conradt, J.: Event-based particle filtering for robot self-localization. In: 2012 IEEE International Conference on Robotics and Biomimetics (ROBIO). pp. 866-870. IEEE (2012)",
|
| 2085 |
+
"42. Xia, W., Li, X., Deng, A., Xiong, H., Dou, D., Hu, D.: Robust cross-modal knowledge distillation for unconstrained videos. arXiv preprint arXiv:2304.07775 (2023)",
|
| 2086 |
+
"43. Xu, F., Yu, L., Wang, B., Yang, W., Xia, G.S., Jia, X., Qiao, Z., Liu, J.: Motion deblurring with real events. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 2583-2592 (2021)",
|
| 2087 |
+
"44. Xu, L., Jia, J.: Two-phase kernel estimation for robust motion deblurring. In: Proceedings of the European conference on computer vision (ECCV). pp. 157-170. Springer (2010)",
|
| 2088 |
+
"45. Yang, D., Yamac, M.: Motion aware double attention network for dynamic scene deblurring. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1113-1123 (2022)",
|
| 2089 |
+
"46. Yuan, L., Sun, J., Quan, L., Shum, H.Y.: Image deblurring with blurred/noisy image pairs. In: ACM SIGGRAPH 2007 papers, pp. 1-es (2007)",
|
| 2090 |
+
"47. Zamir, S.W., Arora, A., Khan, S., Hayat, M., Khan, F.S., Yang, M.H.: Restormer: Efficient transformer for high-resolution image restoration. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5728-5739 (2022)"
|
| 2091 |
+
],
|
| 2092 |
+
"bbox": [
|
| 2093 |
+
212,
|
| 2094 |
+
146,
|
| 2095 |
+
787,
|
| 2096 |
+
839
|
| 2097 |
+
],
|
| 2098 |
+
"page_idx": 16
|
| 2099 |
+
},
|
| 2100 |
+
{
|
| 2101 |
+
"type": "header",
|
| 2102 |
+
"text": "Motion Aware Event Representation-driven Image Deblurring",
|
| 2103 |
+
"bbox": [
|
| 2104 |
+
321,
|
| 2105 |
+
114,
|
| 2106 |
+
730,
|
| 2107 |
+
128
|
| 2108 |
+
],
|
| 2109 |
+
"page_idx": 16
|
| 2110 |
+
},
|
| 2111 |
+
{
|
| 2112 |
+
"type": "page_number",
|
| 2113 |
+
"text": "17",
|
| 2114 |
+
"bbox": [
|
| 2115 |
+
767,
|
| 2116 |
+
116,
|
| 2117 |
+
784,
|
| 2118 |
+
126
|
| 2119 |
+
],
|
| 2120 |
+
"page_idx": 16
|
| 2121 |
+
},
|
| 2122 |
+
{
|
| 2123 |
+
"type": "list",
|
| 2124 |
+
"sub_type": "ref_text",
|
| 2125 |
+
"list_items": [
|
| 2126 |
+
"48. Zhang, S., Zhen, A., Stevenson, R.L.: Deep motion blur removal using noisy/blurry image pairs. Journal of Electronic Imaging 30(3), 033022-033022 (2021)",
|
| 2127 |
+
"49. Zhang, X., Yu, L.: Unifying motion deblurring and frame interpolation with events. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17765-17774 (2022)",
|
| 2128 |
+
"50. Zhang, X., Yu, L., Yang, W., Liu, J., Xia, G.S.: Generalizing event-based motion deblurring in real-world scenarios. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 10734-10744 (2023)",
|
| 2129 |
+
"51. Zhou, Y., Gallego, G., Rebecq, H., Kneip, L., Li, H., Scaramuzza, D.: Semi-dense 3d reconstruction with a stereo event camera. In: Proceedings of the European conference on computer vision (ECCV). pp. 235-251 (2018)"
|
| 2130 |
+
],
|
| 2131 |
+
"bbox": [
|
| 2132 |
+
215,
|
| 2133 |
+
146,
|
| 2134 |
+
785,
|
| 2135 |
+
300
|
| 2136 |
+
],
|
| 2137 |
+
"page_idx": 17
|
| 2138 |
+
},
|
| 2139 |
+
{
|
| 2140 |
+
"type": "page_number",
|
| 2141 |
+
"text": "18",
|
| 2142 |
+
"bbox": [
|
| 2143 |
+
217,
|
| 2144 |
+
114,
|
| 2145 |
+
235,
|
| 2146 |
+
126
|
| 2147 |
+
],
|
| 2148 |
+
"page_idx": 17
|
| 2149 |
+
},
|
| 2150 |
+
{
|
| 2151 |
+
"type": "header",
|
| 2152 |
+
"text": "Z. Sun et al.",
|
| 2153 |
+
"bbox": [
|
| 2154 |
+
271,
|
| 2155 |
+
114,
|
| 2156 |
+
354,
|
| 2157 |
+
126
|
| 2158 |
+
],
|
| 2159 |
+
"page_idx": 17
|
| 2160 |
+
}
|
| 2161 |
+
]
|
2024/Motion Aware Event Representation-driven Image Deblurring/a9105a02-b537-456a-9778-80b745c423ab_model.json
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
2024/Motion Aware Event Representation-driven Image Deblurring/a9105a02-b537-456a-9778-80b745c423ab_origin.pdf
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:1c1fb906670531866b90c9b0e4be252cf62060042ee9805762d0c2c32d8cd87d
|
| 3 |
+
size 2934585
|
2024/Motion Aware Event Representation-driven Image Deblurring/full.md
ADDED
|
@@ -0,0 +1,389 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Motion Aware Event Representation-driven Image Deblurring
|
| 2 |
+
|
| 3 |
+
Zhijing Sun, Xueyang Fu, Longzhuo Huang, Aiping Liu, and Zheng-Jun Zha*
|
| 4 |
+
|
| 5 |
+
School of Information Science and Technology and MoE Key Laboratory of Brain-inspired Intelligent Perception and Cognition, University of Science and Technology of China, Hefei, 230026, China {sunzhijing, hlz_0}@mail.ustc.edu.cn {xyfu, aipingl, zhazj}@ustc.edu.cn
|
| 6 |
+
|
| 7 |
+
Abstract. Traditional image deblurring struggles with high-quality reconstruction due to limited motion data from single blurred images. Excitingly, the high-temporal resolution of event cameras records motion more precisely in a different modality, transforming image deblurring. However, many event camera-based methods, which only care about the final value of the polarity accumulation, ignore the influence of the absolute intensity change where events generate so fall short in perceiving motion patterns and effectively aiding image reconstruction. To overcome this, in this work, we propose a new event preprocessing technique that accumulates the deviation from the initial moment each time the event is updated. This process can distinguish the order of events to improve the perception of object motion patterns. To complement our proposed event representation, we create a recurrent module designed to meticulously extract motion features across local and global time scales. To further facilitate the event feature and image feature integration, which assists in image reconstruction, we develop a bi-directional feature alignment and fusion module. This module works to lessen inter-modal inconsistencies. Our approach has been thoroughly tested through rigorous experiments carried out on several datasets with different distributions. These trials have delivered promising results, with our method achieving top-tier performance in both quantitative and qualitative assessments. Code is available at https://github.com/ZhijingS/DA_event_deblur.
|
| 8 |
+
|
| 9 |
+
Keywords: Image deblurring $\cdot$ Event camera $\cdot$ Event representation
|
| 10 |
+
|
| 11 |
+
# 1 Introduction
|
| 12 |
+
|
| 13 |
+
Frame-based cameras often produce blurred images due to factors like camera shake or object motion, given their configured exposure time. Such blur, a common form of image degradation, occurs across various scenes. While photographers typically aim for clear and focused photos, clear inputs are equally critical for numerous computer vision tasks. These tasks include super-resolution, target
|
| 14 |
+
|
| 15 |
+

|
| 16 |
+
(a) GT (PSNR |SSIM)
|
| 17 |
+
|
| 18 |
+

|
| 19 |
+
(b) Blurry (28.62dB |0.819)
|
| 20 |
+
|
| 21 |
+

|
| 22 |
+
(c) HINet [4] (29.14dB |0.837)
|
| 23 |
+
Fig. 1: A deblurring example on HS-ERGB [37] dataset via recent state-of-the-art methods. Compared to other methods, our approach yields a more distinct result owing to its enhanced motion perception capability.
|
| 24 |
+
|
| 25 |
+

|
| 26 |
+
(d) NAFNet [3] (30.41dB |0.855)
|
| 27 |
+
|
| 28 |
+

|
| 29 |
+
(e) EFNet [33] (30.76dB |0.856)
|
| 30 |
+
|
| 31 |
+

|
| 32 |
+
(f) Ours (31.01dB [0.858]
|
| 33 |
+
|
| 34 |
+
detection, and semantic segmentation, to name a few. Hence, image deblurring serves as a fundamental task in the realm of low-level computer vision, reinforcing the image quality and supporting the proper function of various algorithms.
|
| 35 |
+
|
| 36 |
+
Traditional image-based techniques often attempt to derive a clear image from a single blurred one. They do this by leveraging natural image priors or making assumptions about blur operations [6, 8, 13, 15, 34, 44]. However, in capturing a single blurred image, much motion information is lost. This lack of motion information during exposure creates undesirable artifacts in the deblurred results, especially in complex motion scenarios, as illustrated in Figures 1c and 1d. To counter this challenge, recent methodologies use more temporal information, such as differently exposed frames [2, 46, 48] and neighboring frames [16, 28, 32]. However, these approaches have their limitations. For instance, the use of differently exposed frames is constrained by the low temporal resolution of frame-based cameras, improving the reconstruction effect primarily in slow-motion scenes. The method that uses adjacent frames to estimate the current frame's motion information can lead to accumulative errors. In summary, traditional deblurring techniques encounter hurdles due to the inability of frame-based cameras to capture motion information during exposure.
|
| 37 |
+
|
| 38 |
+
As opposed to frame-based cameras, event cameras, inspired by biological systems, offer a promising solution to this issue. These cameras' microsecond-level temporal resolution $(\mu \mathrm{s})$ and asynchronous output architecture allow the recording of fast motions. This capability enables event streams to store pixel intensity changes and provide accurate motion information to deblur the corresponding frame. By using event frames as additional input, some methods attempt to guide the deblurring process through these event cameras [9,25,39,43,49]. One method [45] subdivides the event stream into ternary 2D data and generates a mask to localize high-blur regions, thereby enhancing the network's ability to distinguish different blur levels. Another method [33] symmetrically accumulates event polarities and fuses event and image features using an attention mechanism, enabling the integration of motion information into image reconstruction. However, these methods primarily employ the event stream in a manner akin to polarity accumulation. In essence, they consider only the final accumulation result of relative intensity change between subsequent moments, neglecting the ab
|
| 39 |
+
|
| 40 |
+
solute intensity change caused by historical events. Consequently, this approach fails to differentiate between various motion patterns, injecting unnecessary ambiguity into the deblurring process, and leading to edge artifacts as displayed in Figure 1e.
|
| 41 |
+
|
| 42 |
+
Given the limited perception of motion during exposure offered by polarity accumulation, we posit that the deviation from the initial moment provides more precise motion information for image deblurring. Each time the event generates, the absolute light intensity changes. The deviation from the initial moment enables us to consider the impact of changes in absolute light intensity, even when the absolute light intensity at the initial moment is unknown. Thus, the deviation, to a degree, implies the order of arrival of the events, offering different information for varying motion patterns. As a result, we propose the Deviation Accumulation (DA) method to enhance motion perception. Specifically, every time a new event arrives, our method accumulates the deviation relative to the initial moment, thus encompassing more comprehensive motion information. Furthermore, we find that longer time scales for event accumulation can capture global motion patterns, while shorter scales precisely depict local fast-motion patterns. To better perceive motion at both global and local time scales, we have designed a Recurrent Motion Extraction (RME) module that works with corresponding event representations. Moreover, in line with multimodal learning [18-20, 29, 36, 42], we present an effective Feature Alignment and Fusion (FAF) module. This module serves to address the issue that simple fusion methods struggle with the disparate nature of event and frame modalities. As a result, these two modalities align, ensuring better information interaction. We integrate these modules into a single image deblurring network. This network extracts the motion features within the exposure time from the event stream and combines them with the texture information supplied by the image to complete the motion blur removal, as illustrated in Figure 1f.
|
| 43 |
+
|
| 44 |
+
The key contributions of this paper are as follows:
|
| 45 |
+
|
| 46 |
+
- We introduce an event preprocessing method designed specifically for the deblurring problem. The deviation accumulation is capable of responding distinctively to various motion patterns, thereby adding accurate motion information for image deblurring tasks.
|
| 47 |
+
- We develop a network equipped with recurrent motion extraction modules, designed to better extract motion features on local and global time scales. This design enables the network to handle multiple composite motion scenarios effectively.
|
| 48 |
+
- We propose a feature alignment and fusion module to mitigate the impacts of inter-modal inconsistencies and ensure that motion information more effectively guides the image deblurring process.
|
| 49 |
+
|
| 50 |
+
Our experimental results indicate that our method achieves state-of-the-art performance on synthetic, semi-synthetic, and real-world datasets, including a 0.61dB improvement in PSNR on the GoPro dataset compared to previous studies [33]. Moreover, our method produces more distinct subjective results, as demonstrated in Figure 1.
|
| 51 |
+
|
| 52 |
+
# 2 Related Work
|
| 53 |
+
|
| 54 |
+
# 2.1 Event-based Motion Deblurring
|
| 55 |
+
|
| 56 |
+
The utilization of event information to assist in deblurring tasks has become increasingly common in recent research [9, 12, 25, 31, 39, 43, 49]. The high temporal resolution and low latency of event cameras not only capture edge positions but also embed moving temporal information within the event stream. This information is crucial to reconstruct the latent sharp image. Many methods have been attempted to explore the relationship between event information, blurred images, and sharp images. One such method [25] includes the Event-based Double Integral (EDI) model, which leverages the principles of event generation to formulate an optimization problem that establishes a correlation between a blurry image and a latent sharp image. In recent studies, the event stream is typically converted into event frames [30, 38], which simplifies the processing of event data within traditional CNN frameworks. [9] combines the motion feature extracted from event data with the image feature, using them as the input for the decoder. [50] employs a Multi-Layer Perceptron (MLP) to predict the fused feature value from event and image features. [33] integrates them through a cross-modal attention method, which generates a query matrix from the image features, but derives key and value matrices from the event features.
|
| 57 |
+
|
| 58 |
+
Confronting the prevalent issue of integrating image and event features is the problem of inter-modal inconsistencies, which traditional fusion techniques cannot easily mitigate. To address this, our approach seeks to align these two modalities, optimizing the use of event data. Through this alignment, our goal is to minimize the inconsistencies spawned when merging differing modalities, thus amplifying the efficacy of the deblurring process.
|
| 59 |
+
|
| 60 |
+
# 2.2 Event Representation
|
| 61 |
+
|
| 62 |
+
Event data, characterized by its spatial sparsity and strong temporal continuity due to its asynchronous generation, necessitates a custom-designed event representation to deliver the required information. Methods [11, 26, 41] that process events individually maximize the use of high temporal resolution, as suggested by [26], which proposed an adaptive spiking neuron model to manage the fluctuating event input. The effectiveness of these algorithms, however, can be significantly hampered by the noise within the event due to independent processing. A Time Surface [1, 14, 22, 51] is created using the timestamp of the last event at each pixel, which prominently reveals the rich temporal information contained within the event. Furthermore, the event frame [5, 7, 21, 23, 45] transforms the original 4D event data into a 2D frame by accumulating events on the number or the polarity pixel-wise, facilitating extraction with conventional image processing techniques. In [45], the authors utilize ternary data for their event frames to allocate high blur region of the image.
|
| 63 |
+
|
| 64 |
+
However, existing event frame approaches result in abundant light intensity change information being lost in the squeezing process, making it impossible to
|
| 65 |
+
|
| 66 |
+
distinguish the order of event arrivals and lacking the ability to detect multiple motion patterns. Some techniques attempt to recover some of this lost data by dividing the exposure time into smaller segments, though this essentially equates to sampling the original data. Consequently, we design a motion-aware event representation specifically tailored for motion deblurring.
|
| 67 |
+
|
| 68 |
+
# 3 Method
|
| 69 |
+
|
| 70 |
+
# 3.1 Overall Framework
|
| 71 |
+
|
| 72 |
+
The overall framework of our proposed method is depicted in Figure 2. We adopt a classical encoder-decoder architecture for our approach. Initially, the blurred image is inputted into the image branch, where we use the baseline conv block mentioned in [3] as our fundamental block, to extract relevant features. Simultaneously, the corresponding event frames generated by our deviation accumulation are fed into the Recurrent Motion Extraction (RME) module, which facilitates the extraction of features covering motion in local and global time scales. Following each block of the two branches, the Feature Alignment and Fusion (FAF) module is employed to eliminate the inconsistencies between the two modalities. By performing down-sampling operations, we are able to obtain both shallow and semantic information. In the decoder, skip connections are utilized to incorporate the extracted features for image reconstruction. The architecture of the conv block and the decoder is shown in the supplementary materials. The entire training process is conducted in an end-to-end manner, with the PSNR Loss function [4] employed to optimize the network's parameters.
|
| 73 |
+
|
| 74 |
+
# 3.2 Deviation Accumulation Event Representation
|
| 75 |
+
|
| 76 |
+
Event cameras are bio-inspired sensors which asynchronously output signals. They trigger an event while the logarithm of the intensity change exceeds the preset threshold $c$ at any pixel, which is formulated as
|
| 77 |
+
|
| 78 |
+
$$
|
| 79 |
+
p = \left\{ \begin{array}{l l} + 1 & \text {i f} \log \left(I _ {x y} (t)\right) - \log \left(I _ {x y} (t - \Delta t)\right) \geq c, \\ - 1 & \text {i f} \log \left(I _ {x y} (t)\right) - \log \left(I _ {x y} (t - \Delta t)\right) \leq - c, \end{array} \right. \tag {1}
|
| 80 |
+
$$
|
| 81 |
+
|
| 82 |
+
where $p$ is the event's polarity, $I_{xy}(t)$ and $I_{xy}(t - \Delta t)$ represent the intensity at the pixel coordinate $(x,y)$ at time $t$ and $t - \Delta t$ , respectively. An event is a quaternion value $(x,y,t,p)$ where $x,y$ denotes the pixel coordinate, $t$ represents the time of change, the polarity $p \in \{+1, -1\}$ indicates the increase and decrease. For a given blurry image $B$ , it can be formulated as the average of latent intensity images $I(t)$ during the exposure time [25]:
|
| 83 |
+
|
| 84 |
+
$$
|
| 85 |
+
B = \frac {1}{T} \int_ {0} ^ {T} I (t) d t, \tag {2}
|
| 86 |
+
$$
|
| 87 |
+
|
| 88 |
+

|
| 89 |
+
Fig. 2: Overview of the architecture of our proposed method for image deblurring. The blurred image and related event frames within the exposure time are fed into the corresponding network branches respectively. The core module of the event branch: (a) Recurrent Motion Extraction (RME) module processes event frames from longer time scales to shorter, merging old features into new features through weighted sum, which results in perceiving various motion patterns under different time ranges. The image branch extracts the features with conv block and fuses them with event features using (b) Feature Alignment and Fusion (FAF) module. FAF module eliminates the inter-modal inconsistencies in a bi-directional way ensuring better information interaction.
|
| 90 |
+
|
| 91 |
+
where $T$ is the exposure time. For simplicity of understanding, we now only consider the pixel at $(x,y)$ then combining Equations (1) and (2), we have
|
| 92 |
+
|
| 93 |
+
$$
|
| 94 |
+
\begin{array}{l} B _ {x y} = \frac {1}{T} \int_ {0} ^ {T} \exp (\log (I _ {x y} (t))) d t (3) \\ = \frac {1}{T} \int_ {0} ^ {T} e x p (\log (I _ {x y} (0)) + c \int_ {0} ^ {t} p (s) d s) d t. (4) \\ \end{array}
|
| 95 |
+
$$
|
| 96 |
+
|
| 97 |
+
The internal part of the exponential operation is the logarithmic value of the light intensity of the pixel at time $t$ , while the external part of the exponential operation is the integration and averaging of each moment within the exposure time. Based on the event data, we can discretize Equation (4):
|
| 98 |
+
|
| 99 |
+
$$
|
| 100 |
+
\begin{array}{l} B _ {x y} = \frac {1}{N} \sum_ {t ^ {\prime} \in \mathcal {T}} \exp \left(\log \left(I _ {x y} (0)\right) + c \sum_ {t = 0} ^ {t ^ {\prime}} p _ {t}\right) (5) \\ = \frac {I _ {x y} (0)}{N} \sum_ {t ^ {\prime} \in \mathcal {T}} e x p \left(c \sum_ {t = 0} ^ {t ^ {\prime}} p _ {t}\right), (6) \\ \end{array}
|
| 101 |
+
$$
|
| 102 |
+
|
| 103 |
+
where $\mathcal{T} = \{t_1, t_2, \dots, t_N\}$ indicates the set of time events occurs, $N$ is the number of events, and $p_t$ means the polarity of the event at time $t$ .
|
| 104 |
+
|
| 105 |
+
The aforementioned derivation prompts us to consider not only the polarity of the event but also the impact of the alteration in the absolute value after
|
| 106 |
+
|
| 107 |
+
# Algorithm 1 Deviation Accumulation (DA)
|
| 108 |
+
|
| 109 |
+
Input: taking pixel at $(x,y)$ for an example, the events stream $(x,y,t_i,p_i)$ in which $i\in [1,N]$ , deviation from the initial moment $D$ , accumulation of deviation $R$ , Output: the value of DA event representation $V$ .
|
| 110 |
+
|
| 111 |
+
1: $D = 0,R = 0$
|
| 112 |
+
2: Counting the number of events $N$ ;
|
| 113 |
+
3: for $(i = 1$ to $N)$ do
|
| 114 |
+
4: $D = D + p_{i}$
|
| 115 |
+
5: $R = R + D$
|
| 116 |
+
6: end for
|
| 117 |
+
7: $V = R / N$
|
| 118 |
+
|
| 119 |
+
each accumulation when considering the problem of deblurring. However, since event data cannot provide $I_{xy}(0)$ , the absolute light intensity is not accessible. But we can indirectly capture this information through the deviation from the initial moment. Therefore, we design a new event representation method for the deblurring problem presented in Algorithm 1. First, we accumulate all the events in the exposure time, so we get a 2D matrix named frequency matrix representing the number of the event occurs at each pixel position. We deal with the events considering their polarities subsequently, recording the deviation at each pixel by an additional tensor. Upon the arrival of the subsequent event, we update the deviation according to polarity and then add the deviation to the corresponding position. In the end, by dividing the cumulative matrix by the frequency matrix, we obtain the average deviation change over the exposure period as the new event representation named Deviation Accumulation (DA).
|
| 120 |
+
|
| 121 |
+
Traditional polarity accumulation methods directly accumulating event polarities lead to identical outputs whenever the number of positive and negative events within a given time frame $([0, T])$ remains unchanged. Nevertheless, by considering the deviation brought by historical events, our method can generate distinct values for events arriving in different orders, thereby improving the ability to perceive motion patterns and providing more precise motion information for the deblurring procedure. For example, as illustrated in Figure 3, (a) shows event signals generated by a motion pattern which represents that the light intensity at that pixel has risen twice and then fallen twice while (d) shows another pattern. Facing these two patterns, (b) and (e) depict the changing process of the outputs of traditional method and our method after each event is accumulated. After processing all four events, traditional polarity accumulation gets the same output of 0 for both patterns, but our DA representation yields 1 and 0.5 respectively which contains richer motion information. More examples of motion patterns our method can distinguish are shown in supplementary materials.
|
| 122 |
+
|
| 123 |
+
Overall, the phenomenon that polarity accumulation methods represent some different motion patterns as the same value introduces ambiguity into the deblurring process. Our method has a better perception of the motion patterns and offers a more distinct motion edge than conventional approaches from the visual effect as shown in Figure 3 (c) and (f). Inspired by [33], directly segmenting
|
| 124 |
+
|
| 125 |
+

|
| 126 |
+
|
| 127 |
+

|
| 128 |
+
|
| 129 |
+

|
| 130 |
+
Fig. 3: (a) and (d) are event signals of two different motion patterns. (b) shows the Polarity Accumulation (PA) can't distinguish (a) and (d) as it gets the same output of 0 after processing all events while (e) shows our Deviation Accumulation (DA) can. DA perceives more motion information than PA as shown in (c) and (f).
|
| 131 |
+
|
| 132 |
+

|
| 133 |
+
|
| 134 |
+

|
| 135 |
+
|
| 136 |
+

|
| 137 |
+
|
| 138 |
+
the entire exposure time averagely is not suitable for motion deblurring, which focuses on capturing rapid motion but falls short in adequately modeling the continuous movement throughout the exposure. In this paper, we symmetrically accumulate events along the $T / 2$ axis using the DA method. Taking $[0,T / 2]$ as an example, we accumulate events on different time scales. The longest accumulating range is $T / 2$ , followed by $T / 3$ , and the shortest is $T / 6$ . A detailed schematic is provided in the supplementary materials.
|
| 139 |
+
|
| 140 |
+
# 3.3 Recurrent Motion Extraction Module
|
| 141 |
+
|
| 142 |
+
The goal of the Recurrent Motion Extraction (RME) module is to cater to both global slow motion and local fast motion in different time scales. Based on our observations of the motion process and event representations, the focus of the event representations varies in different time scales. In shorter time intervals, the global slow motion is not significant in the event representation due to the smaller number of events triggered. On the contrary slow motion can be clearly recorded in longer intervals, but at the same time, the violent moving part will produce edge spreading artifact.
|
| 143 |
+
|
| 144 |
+
To enable the network to accurately model the various in-homogeneous motions, we devised an RME module that incorporates a recurrent mechanism. As shown in Figure 2(a), the event features are fed into the processing module from the longest time scales to the shortest and we use $e1$ , $e2$ , and $e3$ to denote them. Formally, the RME module can be presented as
|
| 145 |
+
|
| 146 |
+
$$
|
| 147 |
+
e _ {i} ^ {\prime} = \left\{ \begin{array}{l l} f _ {e x} \left(e _ {i}\right) & \text {i f} i = 1, \\ \left(1 - \alpha\right) \times f _ {e x} \left(e _ {i}\right) + \alpha \times f _ {r e} \left(e _ {i - 1} ^ {\prime}\right) & \text {i f} i = 2, 3, \end{array} \right. \tag {7}
|
| 148 |
+
$$
|
| 149 |
+
|
| 150 |
+
where $f_{ex}(\cdot)$ and $f_{re}(\cdot)$ are two baseline conv blocks, $f_{ex}(\cdot)$ is used to extract information from the current representation, $f_{re}(\cdot)$ is used to retain useful information from previous features, $\alpha$ is the learnable weight as a trade-off between two features. By sharing the parameters of $f_{ex}(\cdot)$ and $f_{re}(\cdot)$ for different time scales, we force the network to be able to perceive motions at varying time scales. Each input feature is symmetric on the time axis so that the priori information about the motion direction is simultaneously fed into the network.
|
| 151 |
+
|
| 152 |
+
# 3.4 Feature Alignment and Fusion Module
|
| 153 |
+
|
| 154 |
+
Although the information captured by the event camera exhibits a high correlation with the frame camera during the equivalent exposure period, there is still an inevitable gap originating from the modal dissimilarities between them. This phenomenon is frequently mentioned in multi-modal learning [18,42], but in event-guided image restoration tasks people usually [9,33,45] fuse features from two modalities directly using convolutions or attention mechanisms. Therefore, we propose an effective Feature Alignment and Fusion (FAF) module to facilitate enhanced feature fusion in a more coherent environment. Concretely, as shown in Figure 2(b), this module consists of a feature alignment module and a multi-head attention module. We use the operations of multiplication and addition successively to merge the raw information of one modality into the changing process of the other modality, which fills the gap of inter-modal inconsistency. The aligned features are fed into the attention module. In each fusion process, attention plays a bi-directional role, we not only generate the Q matrix through the frame features and the K and V matrix through the event features but also use the event features to perform look-ups on the frame features. The weights of the attention module are shared so that we can fuse the motion information under a more consistent space. Therefore, FAF module can be expressed as
|
| 155 |
+
|
| 156 |
+
$$
|
| 157 |
+
x ^ {\prime} = \operatorname {C o n v} _ {2} (x) + (\operatorname {C o n v} _ {1} (x) \times e + e), \tag {8}
|
| 158 |
+
$$
|
| 159 |
+
|
| 160 |
+
$$
|
| 161 |
+
e ^ {\prime} = \operatorname {C o n v} _ {2} (e) + (\operatorname {C o n v} _ {1} (e) \times x + x), \tag {9}
|
| 162 |
+
$$
|
| 163 |
+
|
| 164 |
+
$$
|
| 165 |
+
F = \operatorname {C o n v} _ {3} (\operatorname {C o n c a t} ([ \text {A t t e n t i o n} (Q _ {x ^ {\prime}}, K _ {e ^ {\prime}}, V _ {e ^ {\prime}}), \text {A t t e n t i o n} (Q _ {e ^ {\prime}}, K _ {x ^ {\prime}}, V _ {x ^ {\prime}}) ])), \tag {10}
|
| 166 |
+
$$
|
| 167 |
+
|
| 168 |
+
where $x'$ and $e'$ are aligned image features $x$ and event feature $e$ respectively, Attention(\cdot) denotes the multi-head attention operation, Concat(\cdot) denotes the concatenation operation along the channel dimension.
|
| 169 |
+
|
| 170 |
+
# 4 Experiments and Analysis
|
| 171 |
+
|
| 172 |
+
# 4.1 Experimental Settings
|
| 173 |
+
|
| 174 |
+
Datasets. We use three different datasets containing synthetic, semi-synthetic, and real-world frames and events to evaluate the proposed method.
|
| 175 |
+
|
| 176 |
+
GoPro: We train the network on the GoPro dataset which is the benchmark dataset for the image motion deblurring [24]. The blurry image is averaged from
|
| 177 |
+
|
| 178 |
+
Table 1: The quantitative results on GoPro, HS-ERGB, and REBlur test datasets. The best values are highlighted in bold. HINet+ and UFPNet+ are event-enhanced versions of HINet and UFPNet by concatenating our DA to their input.
|
| 179 |
+
|
| 180 |
+
<table><tr><td rowspan="2">Method</td><td rowspan="2">Input</td><td colspan="2">GoPro</td><td colspan="2">HS-ERGB</td><td colspan="2">REBlur</td><td rowspan="2">Params(M)</td></tr><tr><td>PSNR</td><td>SSIM</td><td>PSNR</td><td>SSIM</td><td>PSNR</td><td>SSIM</td></tr><tr><td>LEVS [10]</td><td>F</td><td>20.84</td><td>0.547</td><td>22.13</td><td>0.555</td><td>-</td><td>-</td><td>18.2</td></tr><tr><td>EDI [25]</td><td>F+E</td><td>29.06</td><td>0.943</td><td>23.93</td><td>0.704</td><td>36.52</td><td>0.964</td><td>0.5</td></tr><tr><td>SRN [35]</td><td>F</td><td>30.26</td><td>0.934</td><td>-</td><td>-</td><td>35.10</td><td>0.961</td><td>10.2</td></tr><tr><td>EVDI [49]</td><td>F+E</td><td>30.40</td><td>0.906</td><td>25.13</td><td>0.707</td><td>-</td><td>-</td><td>0.4</td></tr><tr><td>HINet [4]</td><td>F</td><td>32.71</td><td>0.959</td><td>27.32</td><td>0.807</td><td>35.58</td><td>0.965</td><td>88.7</td></tr><tr><td>Restormer [47]</td><td>F</td><td>32.92</td><td>0.961</td><td>27.55</td><td>0.808</td><td>35.50</td><td>0.959</td><td>26.1</td></tr><tr><td>MSDI-Net [17]</td><td>F</td><td>33.28</td><td>0.964</td><td>27.46</td><td>0.809</td><td>36.14</td><td>0.968</td><td>241.3</td></tr><tr><td>NAFNet [3]</td><td>F</td><td>33.71</td><td>0.967</td><td>27.64</td><td>0.811</td><td>36.15</td><td>0.969</td><td>67.8</td></tr><tr><td>UFPNet [6]</td><td>F</td><td>34.06</td><td>0.968</td><td>27.64</td><td>0.809</td><td>36.11</td><td>0.968</td><td>80.3</td></tr><tr><td>HINet+ [4]</td><td>F+E</td><td>34.63</td><td>0.968</td><td>27.66</td><td>0.808</td><td>37.92</td><td>0.976</td><td>88.7</td></tr><tr><td>UFPNet+ [6]</td><td>F+E</td><td>35.22</td><td>0.972</td><td>27.68</td><td>0.809</td><td>37.97</td><td>0.976</td><td>80.3</td></tr><tr><td>EFNet [33]</td><td>F+E</td><td>35.46</td><td>0.972</td><td>26.68</td><td>0.800</td><td>38.12</td><td>0.975</td><td>8.5</td></tr><tr><td>Ours</td><td>F+E</td><td>36.07</td><td>0.976</td><td>27.93</td><td>0.812</td><td>38.47</td><td>0.978</td><td>13.9</td></tr></table>
|
| 181 |
+
|
| 182 |
+
adjacent sharp frames, by doing so the dataset contains 3214 pairs of blurry and sharp images. According to the standard division, we adopt 2103 pairs for training and 1111 pairs for testing. For the event data, we use the GoPro raw event dataset provided by [33] which synthesizes events with a random threshold $c$ following the Gaussian distribution by ESIM simulator [27]. Following the DA method, the raw event data is preprocessed into 6 event frames for each image.
|
| 183 |
+
|
| 184 |
+
HS-ERGB: The HS-ERGB dataset [37] consists of sharp videos and real-world events, we use [50] released normal blur version which synthesizes blurry frames by averaging 49 interpolating images. Each blurry frame in the dataset has a known exposure time, allowing us to generate the DA representation for each frame within its respective exposure duration.
|
| 185 |
+
|
| 186 |
+
REBlur: The REBlur dataset collects a number of sequences of real-world event data, corresponding with blurry images and reliable ground-truth sharp images [33]. It contains 12 kinds of linear and nonlinear motions. Following the original setting, we use 486 pairs of blurry-sharp images with associated events for training and 983 pairs for testing.
|
| 187 |
+
|
| 188 |
+
Implementation details. We train the end-to-end network without pre-training on a single Tesla V100 GPU. We extract patches of size $256 \times 256$ for training images and corresponding event frames, and the batch size is set to 2 by default. The AdamW optimizer is employed ( $\beta_{1} = 0.9$ and $\beta_{2} = 0.99$ ) with the initial learning rate 0.0001 with the cosine annealing schedule where the $T_{max}$ is 400K iteration. For data augmentation, we perform horizontal and vertical flips, the random noise and the hot pixels are added into event frames in order to simulate the real-world situation. For HS-ERGB and REBlur datasets, We fine-tune the model trained on GoPro with the training set of them. The fine-tuning process
|
| 189 |
+
|
| 190 |
+

|
| 191 |
+
(a) Blur image in GoPro testset
|
| 192 |
+
|
| 193 |
+
(a) Blur image in GoPro testset
|
| 194 |
+
|
| 195 |
+

|
| 196 |
+
(b) HINet
|
| 197 |
+
|
| 198 |
+

|
| 199 |
+
(c) UFPNet
|
| 200 |
+
|
| 201 |
+

|
| 202 |
+
(d) EFNet
|
| 203 |
+
|
| 204 |
+

|
| 205 |
+
(e) NAFNet
|
| 206 |
+
|
| 207 |
+

|
| 208 |
+
(f) HINet+
|
| 209 |
+
|
| 210 |
+

|
| 211 |
+
(g) UFPNet+
|
| 212 |
+
|
| 213 |
+

|
| 214 |
+
(h) Ours
|
| 215 |
+
|
| 216 |
+

|
| 217 |
+
(i) GT
|
| 218 |
+
|
| 219 |
+

|
| 220 |
+
Fig. 4: Qualitative comparisons on the GoPro dataset. The notation is the same as in Table 1. Zoom in for better view.
|
| 221 |
+
(a) Blur image in REBlur testset
|
| 222 |
+
: Qualitative comparisons on the REBlur dataset. The notation is the same as Table 1. Zoom in for better view.
|
| 223 |
+
|
| 224 |
+

|
| 225 |
+
(b) HINet
|
| 226 |
+
|
| 227 |
+

|
| 228 |
+
(c) UFPNet
|
| 229 |
+
|
| 230 |
+

|
| 231 |
+
(d) EFNet
|
| 232 |
+
|
| 233 |
+

|
| 234 |
+
(e) NAFNet
|
| 235 |
+
|
| 236 |
+
consisted of 4000 iterations while keeping the other configurations consistent with the previous experiments.
|
| 237 |
+
|
| 238 |
+
# 4.2 Comparison with State-of-the-Art Methods
|
| 239 |
+
|
| 240 |
+
Table 1 shows the comparison results of our method and other SOTA methods on GoPro [24], HS-ERGB [50] and REBlur [33]. We compare our network with not only event-based methods, but also with some SOTA image-only deblurring networks enhancing with event input. The results illustrate that: (1) Our method outperforms all other methods in synthetic, semi-synthetic and real-world datasets, which indicates that with the proper utilization of events the designed approach is able to handle blur with different distributions. Specifically, the average gains of our method over the second-best one are 0.61dB, 0.25dB and 0.35dB in terms of PSNR on GoPro, HS-ERGB and REBlur datasets, respectively. (2) Through the enhancement of our proposed event representation, the HINet+ and the UFPNet+ increase 1.92dB and 1.16dB in PSNR on the benchmark GoPro dataset compared with their image-only version, respectively. Taking HINet as an example, with the same operation of [33], HINet with DA
|
| 241 |
+
|
| 242 |
+

|
| 243 |
+
(a) GT
|
| 244 |
+
|
| 245 |
+

|
| 246 |
+
(b) Blurry
|
| 247 |
+
|
| 248 |
+

|
| 249 |
+
(c) UFPNet
|
| 250 |
+
|
| 251 |
+

|
| 252 |
+
(d) NAFNet
|
| 253 |
+
|
| 254 |
+

|
| 255 |
+
(e) EFNet
|
| 256 |
+
|
| 257 |
+

|
| 258 |
+
(f) Ours
|
| 259 |
+
|
| 260 |
+

|
| 261 |
+
Fig. 6: Qualitative comparisons on the HS-ERGB dataset.
|
| 262 |
+
|
| 263 |
+

|
| 264 |
+
(a) Blurry
|
| 265 |
+
|
| 266 |
+

|
| 267 |
+
|
| 268 |
+

|
| 269 |
+
(b) SCER [33]
|
| 270 |
+
|
| 271 |
+

|
| 272 |
+
|
| 273 |
+

|
| 274 |
+
(c) DA (Ours)
|
| 275 |
+
|
| 276 |
+

|
| 277 |
+
|
| 278 |
+

|
| 279 |
+
(d) GT
|
| 280 |
+
Fig. 7: Visual comparisons on GoPro dataset with different event representation. Compared to SCER representation, our DA representation reduces edge artifacts.
|
| 281 |
+
|
| 282 |
+
event representation exceeds 0.94dB over the result in [33], which presents the fact that DA extracts more conducive information for motion deblurring.
|
| 283 |
+
|
| 284 |
+
We also show the visual comparison results to verify the effectiveness of our method in Figure 4, Figure 5 and Figure 6 with representative samples from each dataset. As illustrated in the figures, our model reconstructs better details and has minor spatial distortions compared with other methods. This indicates that our event representation perceives more motion information and the modules effectively embed it to reconstruct a high-quality image with enhanced details.
|
| 285 |
+
|
| 286 |
+
# 4.3 Ablation Studies
|
| 287 |
+
|
| 288 |
+
We conduct plenty ablation experiments on GoPro dataset to verify the effectiveness of our event representation and proposed method.
|
| 289 |
+
|
| 290 |
+
Effectiveness of Event Representation. To demonstrate the adaptability of our event representation for motion deblurring tasks, as shown in Table 2, we replace our DA representation with several other types of representations and trained the network accordingly. The results reveal certain shortcomings of the alternative representations. Firstly, the stack representation simply accumulates all events on their polarities which leaves out the temporal information. Secondly, the SBT approach [40], which subdivides the entire exposure time equally, proves to be inadequate for modeling the varying degrees of blur. Thirdly, the SCER
|
| 291 |
+
|
| 292 |
+
Table 2: Ablation study of each module effects.
|
| 293 |
+
|
| 294 |
+
<table><tr><td>Event Representation</td><td>RME</td><td>Fusion</td><td>PSNR</td><td>SSIM</td></tr><tr><td>Stack</td><td></td><td></td><td>33.24</td><td>0.953</td></tr><tr><td>SBT [40]</td><td>✓</td><td>FAF(Ours)</td><td>35.63</td><td>0.973</td></tr><tr><td>SCER [33]</td><td></td><td></td><td>35.81</td><td>0.975</td></tr><tr><td></td><td></td><td>Concat.</td><td>34.02</td><td>0.967</td></tr><tr><td></td><td></td><td>Add</td><td>34.74</td><td>0.970</td></tr><tr><td></td><td></td><td>Attention</td><td>35.75</td><td>0.974</td></tr><tr><td>ACA(Ours)</td><td>✓</td><td>Only Align Event+Attention</td><td>35.78</td><td>0.975</td></tr><tr><td></td><td></td><td>Only Align Image+Attention</td><td>35.74</td><td>0.974</td></tr><tr><td></td><td></td><td></td><td>35.42</td><td>0.973</td></tr><tr><td>ACA(Ours)</td><td>✗</td><td>FAF(Ours)</td><td>36.07</td><td>0.976</td></tr><tr><td></td><td>✓</td><td></td><td></td><td></td></tr></table>
|
| 295 |
+
|
| 296 |
+

|
| 297 |
+
(a) w/o RME
|
| 298 |
+
|
| 299 |
+

|
| 300 |
+
(b) w/o RME
|
| 301 |
+
|
| 302 |
+

|
| 303 |
+
(c) w/o RME
|
| 304 |
+
|
| 305 |
+

|
| 306 |
+
(d) w/RME
|
| 307 |
+
Fig. 8: Visual comparison results of w/ and w/o RME module. All of these scenarios involve movement in different directions and speeds, RME module can successfully capture motion features in these complex motion scenarios. Zoom in for better view.
|
| 308 |
+
|
| 309 |
+

|
| 310 |
+
(e) w/RME
|
| 311 |
+
|
| 312 |
+

|
| 313 |
+
(f) w/RME
|
| 314 |
+
|
| 315 |
+
method [33], which directly accumulates event polarities, is found to be incapable of capturing the influence of absolute intensity changes. This limitation is illustrated in Figure 3. In contrast, our DA representation not only possesses a solid physical foundation for event generation but also perceives more motion patterns. As a result, it achieves a 0.26dB improvement in performance compared to the best alternative representation. We also provide the visual comparison on the image reconstruction results in Figure 7. Our DA representation reduces motion-generated edge artifacts, resulting in a better deblurring effect.
|
| 316 |
+
|
| 317 |
+
Effectiveness of Recurrent Motion Extraction Module. To assess the RME module's contribution, we replace it with a baseline conv block and adjust the parameters for fair comparison. In this approach, we concatenate event frames of different time scales and feed them into the conv block. The findings, as presented in the last two rows of Table 2, clearly indicate that the contribution of $0.65\mathrm{dB}$ increment in PSNR stems from the network structure design, rather than an increase in parameter quantity. This shows the importance of incorpo
|
| 318 |
+
|
| 319 |
+
rating the module for effectively leveraging motion information in multi-timescales and improving performance in motion deblurring tasks. We also present the visual comparisons when removing the RME module in Figure 8, the RME module adeptly captures both fast local and slow global motion characteristics.
|
| 320 |
+
|
| 321 |
+
Effectiveness of Feature Alignment and Fusion Module. To illustrate the effectiveness of our FAF module, we compare it with several common fusion methods, including concatenation, addition, and cross-modality attention. As shown in Table 2, it is arduous to utilize the motion information contained in event feature through the direct concatenation or addition. While attention mechanisms can capture relative parts in the two features, the inherent inconsistencies within the two modalities still lead to a 0.32dB performance gap when compared to our FAF module. We also use a t-SNE visualization in Figure 9 to demonstrate the effectiveness of our feature alignment, where two originally in-congruent modal features are mixed up together after the FAF module.
|
| 322 |
+
|
| 323 |
+
Furthermore, we also verify the importance of bi-directional alignment by conducting experiments solely on aligning event or image features. Since both events and frames provide crucial information for image restoration, such as edge motion information and textual information, respectively, it is unreasonable to let one modality dominate the alignment process. As shown in Table 2, the bi-directional alignment outperforms the highest score achieved by one-way correction by 0.29dB. This demonstrates the necessity of fully utilizing the mutual information between the two modalities in order to achieve better image reconstruction results.
|
| 324 |
+
|
| 325 |
+

|
| 326 |
+
Fig. 9: t-SNE visualization of original and aligned features. The distributions of the aligned two features overlap more compared to the original features. This indicates the effectiveness of FAF module to reduce the inter-modal inconsistency.
|
| 327 |
+
|
| 328 |
+
# 5 Conclusion
|
| 329 |
+
|
| 330 |
+
In this paper, we propose an event-based image deblurring method. Firstly, we rethink the limit of traditional event preprocessing method and propose a new deviation accumulation (DA) technique. With deviation, the new approach can distinguish the order of events and perceive more motion patterns. Moreover, we design an image deblurring network equipped with recurrent motion extraction (RME) module and feature alignment and fusion (FAF) module to fit our proposed event representation. The RME module can better perceive motions in different time scales, and the FAF module promotes better integration of the two modalities. Finally, extensive experimental results over several datasets demonstrate the effectiveness of the proposed algorithm.
|
| 331 |
+
|
| 332 |
+
Acknowledgements. This work was supported by National Natural Science Foundation of China (NSFC) under Grants 62225207 and 62276243.
|
| 333 |
+
|
| 334 |
+
# References
|
| 335 |
+
|
| 336 |
+
1. Benosman, R., Clercq, C., Lagorce, X., Ieng, S.H., Bartolozzi, C.: Event-based visual flow. IEEE transactions on neural networks and learning systems 25(2), 407-417 (2013)
|
| 337 |
+
2. Chang, M., Feng, H., Xu, Z., Li, Q.: Low-light image restoration with short-and long-exposure raw pairs. IEEE Transactions on Multimedia 24, 702-714 (2021)
|
| 338 |
+
3. Chen, L., Chu, X., Zhang, X., Sun, J.: Simple baselines for image restoration. In: Proceedings of the European conference on computer vision (ECCV). pp. 17-33. Springer (2022)
|
| 339 |
+
4. Chen, L., Lu, X., Zhang, J., Chu, X., Chen, C.: Hinet: Half instance normalization network for image restoration. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 182-192 (2021)
|
| 340 |
+
5. Cook, M., Gugelmann, L., Jug, F., Krautz, C., Steger, A.: Interacting maps for fast visual interpretation. In: The 2011 International Joint Conference on Neural Networks. pp. 770-776. IEEE (2011)
|
| 341 |
+
6. Fang, Z., Wu, F., Dong, W., Li, X., Wu, J., Shi, G.: Self-supervised non-uniform kernel estimation with flow-based motion prior for blind image deblurring. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18105-18114 (2023)
|
| 342 |
+
7. Gehrig, D., Rebecq, H., Gallego, G., Scaramuzza, D.: Eklt: Asynchronous photometric feature tracking using events and frames. International Journal of Computer Vision 128(3), 601-618 (2020)
|
| 343 |
+
8. Ji, S.W., Lee, J., Kim, S.W., Hong, J.P., Baek, S.J., Jung, S.W., Ko, S.J.: Xydeblur: divide and conquer for single image deblurring. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17421-17430 (2022)
|
| 344 |
+
9. Jiang, Z., Zhang, Y., Zou, D., Ren, J., Lv, J., Liu, Y.: Learning event-based motion deblurring. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3320-3329 (2020)
|
| 345 |
+
10. Jin, M., Meishvili, G., Favaro, P.: Learning to extract a video sequence from a single motion-blurred image. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 6334-6342 (2018)
|
| 346 |
+
1. Kim, H., Handa, A., Benosman, R., Ieng, S.H., Davison, A.J.: Simultaneous mosaicing and tracking with an event camera. J. Solid State Circ 43, 566-576 (2008)
|
| 347 |
+
2. Kim, T., Lee, J., Wang, L., Yoon, K.J.: Event-guided deblurring of unknown exposure time videos. In: Proceedings of the European conference on computer vision (ECCV). pp. 519-538. Springer (2022)
|
| 348 |
+
3. Krishnan, D., Tay, T., Fergus, R.: Blind deconvolution using a normalized sparsity measure. In: CVPR 2011. pp. 233-240. IEEE (2011)
|
| 349 |
+
4. Lagorce, X., Orchard, G., Galluppi, F., Shi, B.E., Benosman, R.B.: Hots: a hierarchy of event-based time-surfaces for pattern recognition. IEEE transactions on pattern analysis and machine intelligence 39(7), 1346-1359 (2016)
|
| 350 |
+
5. Levin, A., Weiss, Y., Durand, F., Freeman, W.T.: Understanding and evaluating blind deconvolution algorithms. In: 2009 IEEE conference on computer vision and pattern recognition. pp. 1964-1971. IEEE (2009)
|
| 351 |
+
|
| 352 |
+
16. Li, D., Shi, X., Zhang, Y., Cheung, K.C., See, S., Wang, X., Qin, H., Li, H.: A simple baseline for video restoration with grouped spatial-temporal shift. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9822-9832 (2023)
|
| 353 |
+
17. Li, D., Zhang, Y., Cheung, K.C., Wang, X., Qin, H., Li, H.: Learning degradation representations for image deblurring. In: Proceedings of the European conference on computer vision (ECCV). pp. 736-753. Springer (2022)
|
| 354 |
+
18. Li, Y., Yu, A.W., Meng, T., Caine, B., Ngiam, J., Peng, D., Shen, J., Lu, Y., Zhou, D., Le, Q.V., et al.: Deepfusion: Lidar-camera deep fusion for multi-modal 3d object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17182-17191 (2022)
|
| 355 |
+
19. Liang, M., Yang, B., Wang, S., Urtasun, R.: Deep continuous fusion for multi-sensor 3d object detection. In: Proceedings of the European conference on computer vision (ECCV). pp. 641-656 (2018)
|
| 356 |
+
20. Liang, X., Qian, Y., Guo, Q., Cheng, H., Liang, J.: Af: An association-based fusion method for multi-modal classification. IEEE Transactions on Pattern Analysis and Machine Intelligence 44(12), 9236-9254 (2021)
|
| 357 |
+
21. Liu, M., Delbruck, T.: Adaptive time-slice block-matching optical flow algorithm for dynamic vision sensors. BMVC (2018)
|
| 358 |
+
22. Manderscheid, J., Sironi, A., Bourdis, N., Migliore, D., Lepetit, V.: Speed invariant time surface for learning to detect corner points with event-based cameras. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10245-10254 (2019)
|
| 359 |
+
23. Maqueda, A.I., Loquercio, A., Gallego, G., García, N., Scaramuzza, D.: Event-based vision meets deep learning on steering prediction for self-driving cars. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 5419-5427 (2018)
|
| 360 |
+
24. Nah, S., Hyun Kim, T., Mu Lee, K.: Deep multi-scale convolutional neural network for dynamic scene deblurring. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3883-3891 (2017)
|
| 361 |
+
25. Pan, L., Scheerlinck, C., Yu, X., Hartley, R., Liu, M., Dai, Y.: Bringing a blurry frame alive at high frame-rate with an event camera. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6820-6829 (2019)
|
| 362 |
+
26. Paredes-Vallés, F., Scheper, K.Y., De Croon, G.C.: Unsupervised learning of a hierarchical spiking neural network for optical flow estimation: From events to global motion perception. IEEE transactions on pattern analysis and machine intelligence 42(8), 2051-2064 (2019)
|
| 363 |
+
27. Rebecq, H., Gehrig, D., Scaramuzza, D.: Esim: an open event camera simulator. In: Conference on robot learning. pp. 969-982. PMLR (2018)
|
| 364 |
+
28. Ren, D., Shang, W., Yang, Y., Zuo, W.: Aggregating long-term sharp features via hybrid transformers for video deblurring. arXiv preprint arXiv:2309.07054 (2023)
|
| 365 |
+
29. Ren, S., Du, Y., Lv, J., Han, G., He, S.: Learning from the master: Distilling cross-modal advanced knowledge for lip reading. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13325-13333 (2021)
|
| 366 |
+
30. Shang, W., Ren, D., Zou, D., Ren, J.S., Luo, P., Zuo, W.: Bringing events into video deblurring with non-consecutively blurry frames. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4531-4540 (2021)
|
| 367 |
+
31. Song, C., Bajaj, C., Huang, Q.: Deblursr: Event-based motion deblurring under the spiking representation. arXiv preprint arXiv:2303.08977 (2023)
|
| 368 |
+
|
| 369 |
+
32. Su, S., Delbracio, M., Wang, J., Sapiro, G., Heidrich, W., Wang, O.: Deep video deblurring for hand-held cameras. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1279-1288 (2017)
|
| 370 |
+
33. Sun, L., Sakaridis, C., Liang, J., Jiang, Q., Yang, K., Sun, P., Ye, Y., Wang, K., Gool, L.V.: Event-based fusion for motion deblurring with cross-modal attention. In: Proceedings of the European conference on computer vision (ECCV). pp. 412-428. Springer (2022)
|
| 371 |
+
34. Sun, L., Cho, S., Wang, J., Hays, J.: Edge-based blur kernel estimation using patch priors. In: IEEE international conference on computational photography (ICCP). pp. 1-8. IEEE (2013)
|
| 372 |
+
35. Tao, X., Gao, H., Shen, X., Wang, J., Jia, J.: Scale-recurrent network for deep image deblurring. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 8174-8182 (2018)
|
| 373 |
+
36. Tian, Y., Krishnan, D., Isola, P.: Contrastive multiview coding. In: Proceedings of the European conference on computer vision (ECCV). pp. 776-794. Springer (2020)
|
| 374 |
+
37. Tulyakov, S., Gehrig, D., Georgoulis, S., Erbach, J., Gehrig, M., Li, Y., Scaramuzza, D.: Time lens: Event-based video frame interpolation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 16155-16164 (2021)
|
| 375 |
+
38. Vitoria, P., Georgoulis, S., Tulyakov, S., Bochicchio, A., Erbach, J., Li, Y.: Event-based image deblurring with dynamic motion awareness. In: Proceedings of the European conference on computer vision (ECCV). pp. 95-112. Springer (2022)
|
| 376 |
+
39. Wang, B., He, J., Yu, L., Xia, G.S., Yang, W.: Event enhanced high-quality image recovery. In: Proceedings of the European conference on computer vision (ECCV). pp. 155-171. Springer (2020)
|
| 377 |
+
40. Wang, L., Ho, Y.S., Yoon, K.J., et al.: Event-based high dynamic range image and very high frame rate video generation using conditional generative adversarial networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10081-10090 (2019)
|
| 378 |
+
41. Weikersdorfer, D., Conradt, J.: Event-based particle filtering for robot self-localization. In: 2012 IEEE International Conference on Robotics and Biomimetics (ROBIO). pp. 866-870. IEEE (2012)
|
| 379 |
+
42. Xia, W., Li, X., Deng, A., Xiong, H., Dou, D., Hu, D.: Robust cross-modal knowledge distillation for unconstrained videos. arXiv preprint arXiv:2304.07775 (2023)
|
| 380 |
+
43. Xu, F., Yu, L., Wang, B., Yang, W., Xia, G.S., Jia, X., Qiao, Z., Liu, J.: Motion deblurring with real events. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 2583-2592 (2021)
|
| 381 |
+
44. Xu, L., Jia, J.: Two-phase kernel estimation for robust motion deblurring. In: Proceedings of the European conference on computer vision (ECCV). pp. 157-170. Springer (2010)
|
| 382 |
+
45. Yang, D., Yamac, M.: Motion aware double attention network for dynamic scene deblurring. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1113-1123 (2022)
|
| 383 |
+
46. Yuan, L., Sun, J., Quan, L., Shum, H.Y.: Image deblurring with blurred/noisy image pairs. In: ACM SIGGRAPH 2007 papers, pp. 1-es (2007)
|
| 384 |
+
47. Zamir, S.W., Arora, A., Khan, S., Hayat, M., Khan, F.S., Yang, M.H.: Restormer: Efficient transformer for high-resolution image restoration. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5728-5739 (2022)
|
| 385 |
+
|
| 386 |
+
48. Zhang, S., Zhen, A., Stevenson, R.L.: Deep motion blur removal using noisy/blurry image pairs. Journal of Electronic Imaging 30(3), 033022-033022 (2021)
|
| 387 |
+
49. Zhang, X., Yu, L.: Unifying motion deblurring and frame interpolation with events. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17765-17774 (2022)
|
| 388 |
+
50. Zhang, X., Yu, L., Yang, W., Liu, J., Xia, G.S.: Generalizing event-based motion deblurring in real-world scenarios. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 10734-10744 (2023)
|
| 389 |
+
51. Zhou, Y., Gallego, G., Rebecq, H., Kneip, L., Li, H., Scaramuzza, D.: Semi-dense 3d reconstruction with a stereo event camera. In: Proceedings of the European conference on computer vision (ECCV). pp. 235-251 (2018)
|
2024/Motion Aware Event Representation-driven Image Deblurring/images.zip
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:5ac24081eab05afd3713baedd54f8320d0a6b03b3369bcecf44e17e4236c44e5
|
| 3 |
+
size 462060
|
2024/Motion Aware Event Representation-driven Image Deblurring/layout.json
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
2024/Motion Keyframe Interpolation for Any Human Skeleton using Point Cloud-based Human Motion Data Homogenisation/391ac41a-5023-4e3a-b209-fdaecfa62b6a_content_list.json
ADDED
|
@@ -0,0 +1,1685 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
{
|
| 3 |
+
"type": "text",
|
| 4 |
+
"text": "Motion Keyframe Interpolation for Any Human Skeleton via Temporally Consistent Point Cloud Sampling and Reconstruction",
|
| 5 |
+
"text_level": 1,
|
| 6 |
+
"bbox": [
|
| 7 |
+
225,
|
| 8 |
+
140,
|
| 9 |
+
779,
|
| 10 |
+
209
|
| 11 |
+
],
|
| 12 |
+
"page_idx": 0
|
| 13 |
+
},
|
| 14 |
+
{
|
| 15 |
+
"type": "text",
|
| 16 |
+
"text": "Clinton Mo $^{1}$ , Kun Hu $^{1,\\star}$ , Chengjiang Long $^{2}$ , Dong Yuan $^{1}$ , and Zhiyong Wang $^{1}$",
|
| 17 |
+
"bbox": [
|
| 18 |
+
215,
|
| 19 |
+
233,
|
| 20 |
+
782,
|
| 21 |
+
252
|
| 22 |
+
],
|
| 23 |
+
"page_idx": 0
|
| 24 |
+
},
|
| 25 |
+
{
|
| 26 |
+
"type": "text",
|
| 27 |
+
"text": "<sup>1</sup> The University of Sydney, Darlington NSW 2008, Australia \nclmo6615@uni.sydney.edu.au, {kun.hu, zhiyong.wang, dong.yuan}@sydney.edu.au",
|
| 28 |
+
"bbox": [
|
| 29 |
+
294,
|
| 30 |
+
261,
|
| 31 |
+
705,
|
| 32 |
+
303
|
| 33 |
+
],
|
| 34 |
+
"page_idx": 0
|
| 35 |
+
},
|
| 36 |
+
{
|
| 37 |
+
"type": "text",
|
| 38 |
+
"text": "$^{2}$ Meta Reality Labs, Burlingame, CA, USA",
|
| 39 |
+
"bbox": [
|
| 40 |
+
351,
|
| 41 |
+
303,
|
| 42 |
+
648,
|
| 43 |
+
318
|
| 44 |
+
],
|
| 45 |
+
"page_idx": 0
|
| 46 |
+
},
|
| 47 |
+
{
|
| 48 |
+
"type": "text",
|
| 49 |
+
"text": "clong1@meta.com",
|
| 50 |
+
"bbox": [
|
| 51 |
+
442,
|
| 52 |
+
319,
|
| 53 |
+
560,
|
| 54 |
+
330
|
| 55 |
+
],
|
| 56 |
+
"page_idx": 0
|
| 57 |
+
},
|
| 58 |
+
{
|
| 59 |
+
"type": "text",
|
| 60 |
+
"text": "Abstract. In the character animation field, modern supervised keyframe interpolation models have demonstrated exceptional performance in constructing natural human motions from sparse pose definitions. As supervised models, large motion datasets are necessary to facilitate the learning process; however, since motion is represented with fixed hierarchical skeletons, such datasets are incompatible for skeletons outside the datasets' native configurations. Consequently, the expected availability of a motion dataset for desired skeletons severely hinders the feasibility of learned interpolation in practice. To combat this limitation, we propose Point Cloud-based Motion Representation Learning (PC-MRL), an unsupervised approach to enabling cross-compatibility between skeletons for motion interpolation learning. PC-MRL consists of a skeleton obfuscation strategy using temporal point cloud sampling, and an unsupervised skeleton reconstruction method from point clouds. We devise a temporal point-wise K-nearest neighbors loss for unsupervised learning. Moreover, we propose First-frame Offset Quaternion (FOQ) and Rest Pose Augmentation (RPA) strategies to overcome necessary limitations of our unsupervised point cloud-to-skeletal motion process. Comprehensive experiments demonstrate the effectiveness of PC-MRL in motion interpolation for desired skeletons without supervision from native datasets.",
|
| 61 |
+
"bbox": [
|
| 62 |
+
259,
|
| 63 |
+
366,
|
| 64 |
+
743,
|
| 65 |
+
643
|
| 66 |
+
],
|
| 67 |
+
"page_idx": 0
|
| 68 |
+
},
|
| 69 |
+
{
|
| 70 |
+
"type": "text",
|
| 71 |
+
"text": "Keywords: 3D point clouds $\\cdot$ Human body motion $\\cdot$ Dataset creation",
|
| 72 |
+
"bbox": [
|
| 73 |
+
261,
|
| 74 |
+
656,
|
| 75 |
+
733,
|
| 76 |
+
672
|
| 77 |
+
],
|
| 78 |
+
"page_idx": 0
|
| 79 |
+
},
|
| 80 |
+
{
|
| 81 |
+
"type": "text",
|
| 82 |
+
"text": "1 Introduction",
|
| 83 |
+
"text_level": 1,
|
| 84 |
+
"bbox": [
|
| 85 |
+
215,
|
| 86 |
+
695,
|
| 87 |
+
375,
|
| 88 |
+
710
|
| 89 |
+
],
|
| 90 |
+
"page_idx": 0
|
| 91 |
+
},
|
| 92 |
+
{
|
| 93 |
+
"type": "text",
|
| 94 |
+
"text": "3D character animation workflows largely rely on the concepts of keyframing with interpolation, often referred to as the pose-to-pose principle [13]. By defining key poses at correct timings, algorithms can be employed to generate intermediate poses, and thereby eliminating the need for defining each frame individually [5, 42]. Although this keyframe-based workflow is less costly than frame-by-frame production, human motion is often complex, and natural depictions",
|
| 95 |
+
"bbox": [
|
| 96 |
+
212,
|
| 97 |
+
726,
|
| 98 |
+
787,
|
| 99 |
+
818
|
| 100 |
+
],
|
| 101 |
+
"page_idx": 0
|
| 102 |
+
},
|
| 103 |
+
{
|
| 104 |
+
"type": "page_footnote",
|
| 105 |
+
"text": "* Corresponding author.",
|
| 106 |
+
"bbox": [
|
| 107 |
+
218,
|
| 108 |
+
825,
|
| 109 |
+
385,
|
| 110 |
+
840
|
| 111 |
+
],
|
| 112 |
+
"page_idx": 0
|
| 113 |
+
},
|
| 114 |
+
{
|
| 115 |
+
"type": "image",
|
| 116 |
+
"img_path": "images/d2218860b0ef14c222532445ddae55cd3415e56b8bcf0d28f405c5ce4aab6c0c.jpg",
|
| 117 |
+
"image_caption": [
|
| 118 |
+
"(a) Linear interpolation"
|
| 119 |
+
],
|
| 120 |
+
"image_footnote": [],
|
| 121 |
+
"bbox": [
|
| 122 |
+
267,
|
| 123 |
+
150,
|
| 124 |
+
390,
|
| 125 |
+
257
|
| 126 |
+
],
|
| 127 |
+
"page_idx": 1
|
| 128 |
+
},
|
| 129 |
+
{
|
| 130 |
+
"type": "image",
|
| 131 |
+
"img_path": "images/e3bfa326343bc12db7c8e9e70f1524810ef7753ef624ffd77c7e09b16cfd69c3.jpg",
|
| 132 |
+
"image_caption": [
|
| 133 |
+
"(b) PC-MRL (Our method)"
|
| 134 |
+
],
|
| 135 |
+
"image_footnote": [],
|
| 136 |
+
"bbox": [
|
| 137 |
+
444,
|
| 138 |
+
150,
|
| 139 |
+
560,
|
| 140 |
+
257
|
| 141 |
+
],
|
| 142 |
+
"page_idx": 1
|
| 143 |
+
},
|
| 144 |
+
{
|
| 145 |
+
"type": "image",
|
| 146 |
+
"img_path": "images/44d2a55e7203e40d15ac1c42294ae3da0256fcc101b9cd092f1004ab6aa6faa2.jpg",
|
| 147 |
+
"image_caption": [
|
| 148 |
+
"(c) Ground truth"
|
| 149 |
+
],
|
| 150 |
+
"image_footnote": [],
|
| 151 |
+
"bbox": [
|
| 152 |
+
617,
|
| 153 |
+
150,
|
| 154 |
+
735,
|
| 155 |
+
257
|
| 156 |
+
],
|
| 157 |
+
"page_idx": 1
|
| 158 |
+
},
|
| 159 |
+
{
|
| 160 |
+
"type": "image",
|
| 161 |
+
"img_path": "images/ee25c25af7037e509bc3473cb4d65ecefc982ec85197badae4aac3be6468f142.jpg",
|
| 162 |
+
"image_caption": [
|
| 163 |
+
"(a) Hierarchical skeleton"
|
| 164 |
+
],
|
| 165 |
+
"image_footnote": [],
|
| 166 |
+
"bbox": [
|
| 167 |
+
295,
|
| 168 |
+
320,
|
| 169 |
+
359,
|
| 170 |
+
417
|
| 171 |
+
],
|
| 172 |
+
"page_idx": 1
|
| 173 |
+
},
|
| 174 |
+
{
|
| 175 |
+
"type": "image",
|
| 176 |
+
"img_path": "images/61f3a519e8e4097725bad99f609add4adf90d3642355878e3fb74294765cbaf4.jpg",
|
| 177 |
+
"image_caption": [
|
| 178 |
+
"Fig. 1: Motion interpolation of walking keyframes conducted by our approach.",
|
| 179 |
+
"(b) Skeletal motion"
|
| 180 |
+
],
|
| 181 |
+
"image_footnote": [],
|
| 182 |
+
"bbox": [
|
| 183 |
+
444,
|
| 184 |
+
318,
|
| 185 |
+
555,
|
| 186 |
+
417
|
| 187 |
+
],
|
| 188 |
+
"page_idx": 1
|
| 189 |
+
},
|
| 190 |
+
{
|
| 191 |
+
"type": "image",
|
| 192 |
+
"img_path": "images/f908a2abf1d811ef7f297dd1286a831a856acd18723f2f794f65c2e56c3db689.jpg",
|
| 193 |
+
"image_caption": [
|
| 194 |
+
"(c) Point cloud representation",
|
| 195 |
+
"Fig. 2: Our method leverages point cloud representations (c) to obfuscate the hierarchical dependencies of skeletal motion. Joint colours correspond to ordinal indices in original motion data representation. Point clouds are coloured by temporal index."
|
| 196 |
+
],
|
| 197 |
+
"image_footnote": [],
|
| 198 |
+
"bbox": [
|
| 199 |
+
617,
|
| 200 |
+
319,
|
| 201 |
+
733,
|
| 202 |
+
417
|
| 203 |
+
],
|
| 204 |
+
"page_idx": 1
|
| 205 |
+
},
|
| 206 |
+
{
|
| 207 |
+
"type": "text",
|
| 208 |
+
"text": "demand a significant number of keyframes, especially for realistic motions that adhere to physical properties and constraints. Although motion capture (MoCap) is frequently employed as an alternative for achieving realistic human motions, MoCap often nonetheless necessitates subsequent keyframing to correct unwanted motion elements and address other imperfections.",
|
| 209 |
+
"bbox": [
|
| 210 |
+
212,
|
| 211 |
+
521,
|
| 212 |
+
784,
|
| 213 |
+
597
|
| 214 |
+
],
|
| 215 |
+
"page_idx": 1
|
| 216 |
+
},
|
| 217 |
+
{
|
| 218 |
+
"type": "text",
|
| 219 |
+
"text": "In recent years, the interpolation process has been widely studied, particularly with the advent of machine learning methods for sequential data. Compared to conventional methods such as linear interpolation (LERP), machine learning methods have demonstrated the capability to derive visually natural motions from sparser keyframe sets [12, 18, 34, 37, 40], which we can observe in Fig. 1. Data-driven interpolation approaches require large motion capture datasets to learn the necessary motion features for effective synthesis of transitions between keyframes. In the field of 3D motions, this aspect significantly restricts the feasibility of learned interpolation methods in animation practice, as motion/pose data are tied to specific skeletal configurations and are not cross-compatible. That is, the dataset's native skeleton, i.e. source skeleton, is structurally different and incompatible with the desired skeleton, i.e. target skeleton.",
|
| 220 |
+
"bbox": [
|
| 221 |
+
212,
|
| 222 |
+
598,
|
| 223 |
+
785,
|
| 224 |
+
777
|
| 225 |
+
],
|
| 226 |
+
"page_idx": 1
|
| 227 |
+
},
|
| 228 |
+
{
|
| 229 |
+
"type": "text",
|
| 230 |
+
"text": "Therefore, we propose a novel point cloud-based motion representation learning (PC-MRL) approach, in which skeletal hierarchies are obfuscated by sampling into the point cloud medium, and reconstructed into skeletal motion data using an unsupervised neural network. Unlike raw pose and motion data, the",
|
| 231 |
+
"bbox": [
|
| 232 |
+
212,
|
| 233 |
+
779,
|
| 234 |
+
785,
|
| 235 |
+
840
|
| 236 |
+
],
|
| 237 |
+
"page_idx": 1
|
| 238 |
+
},
|
| 239 |
+
{
|
| 240 |
+
"type": "page_number",
|
| 241 |
+
"text": "2",
|
| 242 |
+
"bbox": [
|
| 243 |
+
217,
|
| 244 |
+
114,
|
| 245 |
+
228,
|
| 246 |
+
126
|
| 247 |
+
],
|
| 248 |
+
"page_idx": 1
|
| 249 |
+
},
|
| 250 |
+
{
|
| 251 |
+
"type": "header",
|
| 252 |
+
"text": "C. Mo et al.",
|
| 253 |
+
"bbox": [
|
| 254 |
+
271,
|
| 255 |
+
114,
|
| 256 |
+
354,
|
| 257 |
+
126
|
| 258 |
+
],
|
| 259 |
+
"page_idx": 1
|
| 260 |
+
},
|
| 261 |
+
{
|
| 262 |
+
"type": "text",
|
| 263 |
+
"text": "point cloud format is a non-hierarchical representation, and effectively obscures any skeletal configuration as displayed in Fig. 2. PC-MRL samples a set of points around the rest position of skeletons using a combination of uniform and normal distributions, and geometrically parents each point to an associated bone in order to effectively represent the associated skeletons' motion data in a temporally consistent manner. Subsequently, for the reconstruction of motion data into a given desired skeleton, we propose an unsupervised learning scheme, featuring a K-nearest neighbors (KNN) loss between cross-skeleton point clouds to optimize a transformer neural network. The network processes the point cloud-based motion representations and generates the corresponding target skeletal motion representations. By obscuring both source and predicted target motion data into the point cloud space, our strategy allows for a direct comparison, utilizing KNN to minimize the visual difference between the two point clouds.",
|
| 264 |
+
"bbox": [
|
| 265 |
+
212,
|
| 266 |
+
146,
|
| 267 |
+
787,
|
| 268 |
+
343
|
| 269 |
+
],
|
| 270 |
+
"page_idx": 2
|
| 271 |
+
},
|
| 272 |
+
{
|
| 273 |
+
"type": "text",
|
| 274 |
+
"text": "The unsupervised nature of our target skeleton reconstruction scheme creates two main disparities between the expected and predicted motion features. Firstly, the exact roll rotation axis of bones cannot be represented in the point cloud format. To this end, we introduce First-frame Offset Quaternion (FOQ) representations to incorporate relative roll values, which are obtainable from temporally consistent point clouds. This approach is agnostic to absolute roll rotations, providing a standardisation mechanism for the resulting motion data within the context of motion interpolation learning. Second, KNN-based objectives learn skeletal representations in a geometrically optimal manner, which does not always reflect the desired skeletal behaviour. This disparity is commonly observed with smaller skeletal features, such as shoulder and hip bones. As such, we augment training motions using Rest Pose Augmentation (RPA) to increase the range of skeletal behaviours that our interpolation model learns. Extensive experiments demonstrate the effectiveness of our proposed method, achieving performance targets near the level of direct dataset supervision.",
|
| 275 |
+
"bbox": [
|
| 276 |
+
212,
|
| 277 |
+
343,
|
| 278 |
+
787,
|
| 279 |
+
568
|
| 280 |
+
],
|
| 281 |
+
"page_idx": 2
|
| 282 |
+
},
|
| 283 |
+
{
|
| 284 |
+
"type": "text",
|
| 285 |
+
"text": "In summary, this paper presents the following key contributions:",
|
| 286 |
+
"bbox": [
|
| 287 |
+
238,
|
| 288 |
+
569,
|
| 289 |
+
702,
|
| 290 |
+
583
|
| 291 |
+
],
|
| 292 |
+
"page_idx": 2
|
| 293 |
+
},
|
| 294 |
+
{
|
| 295 |
+
"type": "list",
|
| 296 |
+
"sub_type": "text",
|
| 297 |
+
"list_items": [
|
| 298 |
+
"1. We propose a novel method for achieving unsupervised human motion reconstruction from point cloud data, enabling skeleton-agnostic motion interpolation learning for the first time.",
|
| 299 |
+
"2. To train motion interpolation models, we formulate first-frame offset quaternions to represent bone rotations with relative roll data, as well as a rest position augmentation strategy to address skeletal configuration variability.",
|
| 300 |
+
"3. We perform comprehensive experiments to demonstrate our method's effectiveness towards learning motion interpolation, despite the absence of directly compatible datasets in the training process."
|
| 301 |
+
],
|
| 302 |
+
"bbox": [
|
| 303 |
+
220,
|
| 304 |
+
592,
|
| 305 |
+
782,
|
| 306 |
+
724
|
| 307 |
+
],
|
| 308 |
+
"page_idx": 2
|
| 309 |
+
},
|
| 310 |
+
{
|
| 311 |
+
"type": "text",
|
| 312 |
+
"text": "2 Related works",
|
| 313 |
+
"text_level": 1,
|
| 314 |
+
"bbox": [
|
| 315 |
+
215,
|
| 316 |
+
750,
|
| 317 |
+
392,
|
| 318 |
+
763
|
| 319 |
+
],
|
| 320 |
+
"page_idx": 2
|
| 321 |
+
},
|
| 322 |
+
{
|
| 323 |
+
"type": "text",
|
| 324 |
+
"text": "We explore existing research on motion interpolation approaches, and motion data modelling methods in general. In addition, our proposed approach bears resemblance to unsupervised approaches for the motion re-targeting research problem, and as such, we will also analyse existing methods in this field.",
|
| 325 |
+
"bbox": [
|
| 326 |
+
212,
|
| 327 |
+
779,
|
| 328 |
+
787,
|
| 329 |
+
840
|
| 330 |
+
],
|
| 331 |
+
"page_idx": 2
|
| 332 |
+
},
|
| 333 |
+
{
|
| 334 |
+
"type": "header",
|
| 335 |
+
"text": "Motion Keyframe Interpolation for Any Human Skeleton",
|
| 336 |
+
"bbox": [
|
| 337 |
+
351,
|
| 338 |
+
114,
|
| 339 |
+
730,
|
| 340 |
+
128
|
| 341 |
+
],
|
| 342 |
+
"page_idx": 2
|
| 343 |
+
},
|
| 344 |
+
{
|
| 345 |
+
"type": "page_number",
|
| 346 |
+
"text": "3",
|
| 347 |
+
"bbox": [
|
| 348 |
+
774,
|
| 349 |
+
116,
|
| 350 |
+
784,
|
| 351 |
+
126
|
| 352 |
+
],
|
| 353 |
+
"page_idx": 2
|
| 354 |
+
},
|
| 355 |
+
{
|
| 356 |
+
"type": "text",
|
| 357 |
+
"text": "2.1 Data-Driven Motion Modelling Methods",
|
| 358 |
+
"text_level": 1,
|
| 359 |
+
"bbox": [
|
| 360 |
+
215,
|
| 361 |
+
146,
|
| 362 |
+
596,
|
| 363 |
+
161
|
| 364 |
+
],
|
| 365 |
+
"page_idx": 3
|
| 366 |
+
},
|
| 367 |
+
{
|
| 368 |
+
"type": "text",
|
| 369 |
+
"text": "Contemporary research in motion modelling predominantly explore learned data-driven methods, using neural networks to capture, model, and establish correlations between the motion features of various skeletal configurations [29, 35, 41, 51, 55]. Motion features alone have demonstrated considerable efficacy in cross-skeleton imitation using a variety of strategies, including task-driven objectives [19, 45], diffusion-based generation [25, 49, 52], and latent feature consistency [1-3, 7, 46]. Latent consistency in particular learns inter-skeleton correlations without the need for paired datasets, which reduces its barrier of entry.",
|
| 370 |
+
"bbox": [
|
| 371 |
+
212,
|
| 372 |
+
172,
|
| 373 |
+
785,
|
| 374 |
+
292
|
| 375 |
+
],
|
| 376 |
+
"page_idx": 3
|
| 377 |
+
},
|
| 378 |
+
{
|
| 379 |
+
"type": "text",
|
| 380 |
+
"text": "Likewise, motion interpolation strategies have evolved through the use of learning based approaches. The inherent numerical precision of motion data has traditionally posed a challenge for the adoption of the learning approaches, due to their approximate nature [21,27]. This challenge is further amplified by the temporally sparse distribution of keyframes. Recently, a number of recurrent neural network based methods [17, 18, 43, 53] and transformer-based networks [12, 34, 37] have demonstrated encouraging interpolation performance for the transition between keyframes. Additionally, various other motion modelling methods are capable for keyframe based interpolation, albeit with limited intermediate frame representation capabilities [19, 25, 45].",
|
| 381 |
+
"bbox": [
|
| 382 |
+
212,
|
| 383 |
+
292,
|
| 384 |
+
787,
|
| 385 |
+
443
|
| 386 |
+
],
|
| 387 |
+
"page_idx": 3
|
| 388 |
+
},
|
| 389 |
+
{
|
| 390 |
+
"type": "text",
|
| 391 |
+
"text": "A major constraint inherent in all data-driven motion modelling methods is their dependency on datasets featuring specific skeletal configurations. To address this, skeleton-free training strategies have been explored for elementary tasks involving 3D characters. These include vertex-based pose transfer [7,8,31,47], and point cloud-based human shape reconstruction [23,48]. Notably, in these methods, 3D mesh and point cloud coordinates serve as general 3D data representations. Concurrently, other research has significantly minimized the required volume of training data, bringing it down to single example sequences. This reduction has been achieved through the use of patch matching techniques [28] and imitation learning within physics-based simulations [30,39]. Reinforcement learning has facilitated the generation of goal-driven motion without pre-existing datasets [32,38,50]. However, these simulation-based methods have typically been either exceedingly complex to implement in practice, or yield results that are too inflexible and/or prone to errors for animation workflows.",
|
| 392 |
+
"bbox": [
|
| 393 |
+
212,
|
| 394 |
+
444,
|
| 395 |
+
787,
|
| 396 |
+
655
|
| 397 |
+
],
|
| 398 |
+
"page_idx": 3
|
| 399 |
+
},
|
| 400 |
+
{
|
| 401 |
+
"type": "text",
|
| 402 |
+
"text": "2.2 Motion Re-Targeting without Supervision",
|
| 403 |
+
"text_level": 1,
|
| 404 |
+
"bbox": [
|
| 405 |
+
215,
|
| 406 |
+
678,
|
| 407 |
+
606,
|
| 408 |
+
694
|
| 409 |
+
],
|
| 410 |
+
"page_idx": 3
|
| 411 |
+
},
|
| 412 |
+
{
|
| 413 |
+
"type": "text",
|
| 414 |
+
"text": "Conventional approaches in motion re-targeting have been a cornerstone of 3D animation for decades, primarily dedicated to converting motion capture data into 3D skeletal motions [4, 15]. Typically, raw motion capture data is optically recorded, yielding primarily spatial information. In contrast, motion data is largely rotational in nature [5], often represented in $SO(3)$ space. Therefore, a principal objective of motion re-targeting solutions has been to effectively bridge these two distinct types of representations. In [14], extensive manually adjustable constraints were introduced for flexible re-targeting between skeletons of identical topology (i.e., same hierarchy, different bone lengths and rest poses). Key",
|
| 415 |
+
"bbox": [
|
| 416 |
+
212,
|
| 417 |
+
704,
|
| 418 |
+
785,
|
| 419 |
+
839
|
| 420 |
+
],
|
| 421 |
+
"page_idx": 3
|
| 422 |
+
},
|
| 423 |
+
{
|
| 424 |
+
"type": "page_number",
|
| 425 |
+
"text": "4",
|
| 426 |
+
"bbox": [
|
| 427 |
+
217,
|
| 428 |
+
114,
|
| 429 |
+
228,
|
| 430 |
+
126
|
| 431 |
+
],
|
| 432 |
+
"page_idx": 3
|
| 433 |
+
},
|
| 434 |
+
{
|
| 435 |
+
"type": "header",
|
| 436 |
+
"text": "C. Mo et al.",
|
| 437 |
+
"bbox": [
|
| 438 |
+
271,
|
| 439 |
+
114,
|
| 440 |
+
354,
|
| 441 |
+
126
|
| 442 |
+
],
|
| 443 |
+
"page_idx": 3
|
| 444 |
+
},
|
| 445 |
+
{
|
| 446 |
+
"type": "image",
|
| 447 |
+
"img_path": "images/b2e7a16acdf66b05482e71dd89d4c86ffdc781ecfb2d07b2bd8b9030fd37b84b.jpg",
|
| 448 |
+
"image_caption": [
|
| 449 |
+
"Fig. 3: Illustration of point cloud-based motion representation learning (PC-MRL) pipeline for keyframe interpolation to any human skeleton. Our unsupervised point cloud-to-pose decoding approach (I) first adapts motion features from existing datasets to new skeletons. We then employ the adapted motion data to supervise motion interpolation models (II), using FOQ and rest pose augmentations to address remaining discrepancies between the adapted and expected target skeleton motion patterns."
|
| 450 |
+
],
|
| 451 |
+
"image_footnote": [],
|
| 452 |
+
"bbox": [
|
| 453 |
+
223,
|
| 454 |
+
152,
|
| 455 |
+
772,
|
| 456 |
+
280
|
| 457 |
+
],
|
| 458 |
+
"page_idx": 4
|
| 459 |
+
},
|
| 460 |
+
{
|
| 461 |
+
"type": "text",
|
| 462 |
+
"text": "practices in this approach, including end-effector matching, joint contacts, and inter-skeleton point correlations, are still relevant to current animation pipelines. Point correlations using Inverse Kinematics (IK) have also been proposed as an intermediate representation to facilitate the adaptation between topologically distinct skeletons [9, 36], albeit with hierarchical pairing limitations.",
|
| 463 |
+
"bbox": [
|
| 464 |
+
212,
|
| 465 |
+
414,
|
| 466 |
+
782,
|
| 467 |
+
489
|
| 468 |
+
],
|
| 469 |
+
"page_idx": 4
|
| 470 |
+
},
|
| 471 |
+
{
|
| 472 |
+
"type": "text",
|
| 473 |
+
"text": "In response, morphologically-independent motion control systems were proposed using IK joint chains and joint-wise constraints [20,26]. Such systems often expect highly standardised motion skeleton features, and are generally impractical for re-targeting motions produced for different control schemes. Additionally, the reliance on Euler angles for extensive classical re-targeting constraints poses a variety of issues regarding rotational freedom [11], as well as compatibility with animation workflows [33], which predominantly use quaternion-based systems.",
|
| 474 |
+
"bbox": [
|
| 475 |
+
212,
|
| 476 |
+
489,
|
| 477 |
+
784,
|
| 478 |
+
595
|
| 479 |
+
],
|
| 480 |
+
"page_idx": 4
|
| 481 |
+
},
|
| 482 |
+
{
|
| 483 |
+
"type": "text",
|
| 484 |
+
"text": "Despite these advancements, the challenge of arbitrary skeleton re-targeting from motion datasets still remains largely unresolved, limiting any widespread application of data-driven pipelines. Our paper aims to effectively address this pervasive limitation.",
|
| 485 |
+
"bbox": [
|
| 486 |
+
212,
|
| 487 |
+
595,
|
| 488 |
+
785,
|
| 489 |
+
656
|
| 490 |
+
],
|
| 491 |
+
"page_idx": 4
|
| 492 |
+
},
|
| 493 |
+
{
|
| 494 |
+
"type": "text",
|
| 495 |
+
"text": "3 Methodology",
|
| 496 |
+
"text_level": 1,
|
| 497 |
+
"bbox": [
|
| 498 |
+
214,
|
| 499 |
+
681,
|
| 500 |
+
380,
|
| 501 |
+
698
|
| 502 |
+
],
|
| 503 |
+
"page_idx": 4
|
| 504 |
+
},
|
| 505 |
+
{
|
| 506 |
+
"type": "text",
|
| 507 |
+
"text": "As illustrated in Fig. 3, our proposed method consists of two main components: point cloud-to-motion reconstruction and motion keyframe interpolation.",
|
| 508 |
+
"bbox": [
|
| 509 |
+
212,
|
| 510 |
+
714,
|
| 511 |
+
782,
|
| 512 |
+
744
|
| 513 |
+
],
|
| 514 |
+
"page_idx": 4
|
| 515 |
+
},
|
| 516 |
+
{
|
| 517 |
+
"type": "text",
|
| 518 |
+
"text": "3.1 Quaternion-based Motion Representations",
|
| 519 |
+
"text_level": 1,
|
| 520 |
+
"bbox": [
|
| 521 |
+
214,
|
| 522 |
+
768,
|
| 523 |
+
609,
|
| 524 |
+
782
|
| 525 |
+
],
|
| 526 |
+
"page_idx": 4
|
| 527 |
+
},
|
| 528 |
+
{
|
| 529 |
+
"type": "text",
|
| 530 |
+
"text": "We first declare a set of notations for representing skeletons. A skeleton $S$ is defined by a tree of bones $\\mathcal{B}$ . A bone $\\mathbf{v}_n \\in \\mathcal{B}$ is defined by its bone parent $\\mathbf{v}_m \\in \\mathcal{B}$ , and a head position $\\mathbf{H}_n \\in \\mathbb{R}^3$ which denotes the bone's base offset from",
|
| 531 |
+
"bbox": [
|
| 532 |
+
212,
|
| 533 |
+
794,
|
| 534 |
+
785,
|
| 535 |
+
840
|
| 536 |
+
],
|
| 537 |
+
"page_idx": 4
|
| 538 |
+
},
|
| 539 |
+
{
|
| 540 |
+
"type": "header",
|
| 541 |
+
"text": "Motion Keyframe Interpolation for Any Human Skeleton",
|
| 542 |
+
"bbox": [
|
| 543 |
+
351,
|
| 544 |
+
114,
|
| 545 |
+
730,
|
| 546 |
+
128
|
| 547 |
+
],
|
| 548 |
+
"page_idx": 4
|
| 549 |
+
},
|
| 550 |
+
{
|
| 551 |
+
"type": "page_number",
|
| 552 |
+
"text": "5",
|
| 553 |
+
"bbox": [
|
| 554 |
+
774,
|
| 555 |
+
116,
|
| 556 |
+
784,
|
| 557 |
+
126
|
| 558 |
+
],
|
| 559 |
+
"page_idx": 4
|
| 560 |
+
},
|
| 561 |
+
{
|
| 562 |
+
"type": "text",
|
| 563 |
+
"text": "its parent head $\\mathbf{H}_m$ . In addition, a tail position $\\mathbf{T}_n \\in \\mathbb{R}^3$ helps to define the end point of the bone's line segment representation, starting from its head position, of $\\mathbf{v}_n$ . The root bone $\\mathbf{v}_0$ has no parent, and thus no head position $\\mathbf{H}_0$ . The set of $\\mathbf{H}_n$ and $\\mathbf{T}_n$ for each bone $\\mathbf{v}_n \\in \\mathcal{B}$ make up the rest position of the skeleton.",
|
| 564 |
+
"bbox": [
|
| 565 |
+
212,
|
| 566 |
+
146,
|
| 567 |
+
782,
|
| 568 |
+
205
|
| 569 |
+
],
|
| 570 |
+
"page_idx": 5
|
| 571 |
+
},
|
| 572 |
+
{
|
| 573 |
+
"type": "text",
|
| 574 |
+
"text": "Next, we define the properties of motion data. At any given frame $t$ , a skeleton's pose position can be represented as global 3D positions $\\mathbf{p}_{n,t} \\in \\mathbb{R}^3$ and global quaternion rotations $\\mathbf{q}_{n,t} \\in \\mathbb{R}^4$ for each bone $\\mathbf{v}_n \\in \\mathcal{B}$ . For a given motion sequence $M$ , only the root bone is given a 3D position in pose position, i.e. $\\mathbf{p}_{0,t}$ as all other pose positions are derived using Forwards Kinematics (FK [11,24]):",
|
| 575 |
+
"bbox": [
|
| 576 |
+
212,
|
| 577 |
+
207,
|
| 578 |
+
784,
|
| 579 |
+
282
|
| 580 |
+
],
|
| 581 |
+
"page_idx": 5
|
| 582 |
+
},
|
| 583 |
+
{
|
| 584 |
+
"type": "equation",
|
| 585 |
+
"text": "\n$$\n\\mathbf {p} _ {n, t} = \\mathbf {p} _ {m, t} + Q R \\left(\\mathbf {H} _ {n}, \\mathbf {q} _ {m, t}\\right), \\tag {1}\n$$\n",
|
| 586 |
+
"text_format": "latex",
|
| 587 |
+
"bbox": [
|
| 588 |
+
398,
|
| 589 |
+
292,
|
| 590 |
+
784,
|
| 591 |
+
306
|
| 592 |
+
],
|
| 593 |
+
"page_idx": 5
|
| 594 |
+
},
|
| 595 |
+
{
|
| 596 |
+
"type": "text",
|
| 597 |
+
"text": "where $QR$ is the quaternion rotation function, and $\\mathbf{v}_m$ is the parent joint. In summary, a motion sequence $M$ of $|M|$ frames is defined by:",
|
| 598 |
+
"bbox": [
|
| 599 |
+
212,
|
| 600 |
+
314,
|
| 601 |
+
782,
|
| 602 |
+
345
|
| 603 |
+
],
|
| 604 |
+
"page_idx": 5
|
| 605 |
+
},
|
| 606 |
+
{
|
| 607 |
+
"type": "equation",
|
| 608 |
+
"text": "\n$$\nM = \\left\\{\\mathbf {p} _ {0, t} \\cup \\left\\{\\mathbf {q} _ {n, t} \\mid \\mathbf {v} _ {n} \\in \\mathcal {B} \\right\\} \\mid 0 \\leq t < | M | \\right\\} \\tag {2}\n$$\n",
|
| 609 |
+
"text_format": "latex",
|
| 610 |
+
"bbox": [
|
| 611 |
+
348,
|
| 612 |
+
354,
|
| 613 |
+
784,
|
| 614 |
+
369
|
| 615 |
+
],
|
| 616 |
+
"page_idx": 5
|
| 617 |
+
},
|
| 618 |
+
{
|
| 619 |
+
"type": "text",
|
| 620 |
+
"text": "3.2 Point Cloud Obfuscation and Skeleton Reconstruction",
|
| 621 |
+
"text_level": 1,
|
| 622 |
+
"bbox": [
|
| 623 |
+
214,
|
| 624 |
+
388,
|
| 625 |
+
707,
|
| 626 |
+
402
|
| 627 |
+
],
|
| 628 |
+
"page_idx": 5
|
| 629 |
+
},
|
| 630 |
+
{
|
| 631 |
+
"type": "text",
|
| 632 |
+
"text": "Our approach is based on the hypothesis that 3D motion sequences can be visually represented by a more universal format: 3D point cloud sequences. By geometrically sampling the volume around a skeleton, point cloud data becomes a non-hierarchical representation of its pose(s), obfuscating the original skeletal configuration. Due to this decoupling between the skeleton configuration and motion representation, a successful motion data reconstruction from point cloud data would be the first step to enabling learnable cross-skeleton motion features. Formally, for a given source skeleton $S_A$ with an associated motion sequence $M_A$ , and any humanoid skeletal configuration $S_B$ , the point cloud obfuscation and reconstruction module is to learn a function that can adequately perform $F_{S_A \\to S_B}(M_A) = \\hat{M}_B$ , where $\\hat{M}_B$ is a visually similar motion adaptation of $M_A$ on skeleton $S_B$ , which is an estimation of $M_B$ .",
|
| 633 |
+
"bbox": [
|
| 634 |
+
212,
|
| 635 |
+
410,
|
| 636 |
+
784,
|
| 637 |
+
592
|
| 638 |
+
],
|
| 639 |
+
"page_idx": 5
|
| 640 |
+
},
|
| 641 |
+
{
|
| 642 |
+
"type": "text",
|
| 643 |
+
"text": "Motion Data Obfuscation with Point Clouds To sample the point cloud $\\mathcal{X}$ , we first generate a set of points $u\\in \\mathcal{X}$ along the line segments defined by the global head and tail positions for each bone $\\mathbf{v}_n$ . In the bone-local space to $\\mathbf{v}_n$ , this would mean sampling $u$ by a uniformly distributed factor $\\alpha \\sim \\mathcal{U}_{[0,1]}$ . Specifically, $\\alpha = 0$ aligns with the head position of the bone, and $\\alpha = 1$ indicates the tail position. By further sampling $u$ via a normal distribution, our sampling strategy can capture the surrounding volume around the skeleton. Formally, to sample $u$ from its associated bone $\\mathbf{v}_n$ within a standard deviation of $\\sigma$ , we have:",
|
| 644 |
+
"bbox": [
|
| 645 |
+
212,
|
| 646 |
+
611,
|
| 647 |
+
784,
|
| 648 |
+
731
|
| 649 |
+
],
|
| 650 |
+
"page_idx": 5
|
| 651 |
+
},
|
| 652 |
+
{
|
| 653 |
+
"type": "equation",
|
| 654 |
+
"text": "\n$$\n\\mathbf {H} _ {x} \\sim \\mathcal {N} (\\alpha \\mathbf {T} _ {n}, \\sigma),\n$$\n",
|
| 655 |
+
"text_format": "latex",
|
| 656 |
+
"bbox": [
|
| 657 |
+
408,
|
| 658 |
+
753,
|
| 659 |
+
535,
|
| 660 |
+
768
|
| 661 |
+
],
|
| 662 |
+
"page_idx": 5
|
| 663 |
+
},
|
| 664 |
+
{
|
| 665 |
+
"type": "equation",
|
| 666 |
+
"text": "\n$$\n\\mathbf {p} _ {u, t} = \\mathbf {p} _ {n, t} + Q R (\\mathbf {H} _ {x}, \\mathbf {q} _ {n, t}) \\tag {3}\n$$\n",
|
| 667 |
+
"text_format": "latex",
|
| 668 |
+
"bbox": [
|
| 669 |
+
403,
|
| 670 |
+
763,
|
| 671 |
+
784,
|
| 672 |
+
787
|
| 673 |
+
],
|
| 674 |
+
"page_idx": 5
|
| 675 |
+
},
|
| 676 |
+
{
|
| 677 |
+
"type": "text",
|
| 678 |
+
"text": "Notably, this sampling process as in Eq. 3 is associated with the skeletal configuration only, by which the sampled points are temporally consistent for further processing with the motion sequence. Specifically, each point maintains a",
|
| 679 |
+
"bbox": [
|
| 680 |
+
212,
|
| 681 |
+
794,
|
| 682 |
+
782,
|
| 683 |
+
839
|
| 684 |
+
],
|
| 685 |
+
"page_idx": 5
|
| 686 |
+
},
|
| 687 |
+
{
|
| 688 |
+
"type": "page_number",
|
| 689 |
+
"text": "6",
|
| 690 |
+
"bbox": [
|
| 691 |
+
217,
|
| 692 |
+
114,
|
| 693 |
+
228,
|
| 694 |
+
126
|
| 695 |
+
],
|
| 696 |
+
"page_idx": 5
|
| 697 |
+
},
|
| 698 |
+
{
|
| 699 |
+
"type": "header",
|
| 700 |
+
"text": "C. Mo et al.",
|
| 701 |
+
"bbox": [
|
| 702 |
+
271,
|
| 703 |
+
114,
|
| 704 |
+
354,
|
| 705 |
+
127
|
| 706 |
+
],
|
| 707 |
+
"page_idx": 5
|
| 708 |
+
},
|
| 709 |
+
{
|
| 710 |
+
"type": "text",
|
| 711 |
+
"text": "constant position from its associated bone in the bone-local space. The lateral distance of each point from their parent bone segment allows point clouds to react to the bone's rotational roll axis.",
|
| 712 |
+
"bbox": [
|
| 713 |
+
212,
|
| 714 |
+
146,
|
| 715 |
+
782,
|
| 716 |
+
190
|
| 717 |
+
],
|
| 718 |
+
"page_idx": 6
|
| 719 |
+
},
|
| 720 |
+
{
|
| 721 |
+
"type": "text",
|
| 722 |
+
"text": "In a pure point cloud representation, it is possible to remove the relationships between bones. However, to clearly distinguish symmetrical body features in different skeletons, we introduce and embed body group associations for the points during our reconstruction process. Specifically, every bone is categorised by one of these body groups: [spine, left arm, right arm, left leg, right leg]. Each sampled point is assigned q group attribute in line with the associated bone. The attribute is represented by a one-hot feature vector $\\mathbf{g}_u$ . To this end, we characterize $u = \\{\\mathbf{p}_{u,t}, \\mathbf{g}_u\\} \\in \\mathcal{X}$ .",
|
| 723 |
+
"bbox": [
|
| 724 |
+
212,
|
| 725 |
+
191,
|
| 726 |
+
784,
|
| 727 |
+
313
|
| 728 |
+
],
|
| 729 |
+
"page_idx": 6
|
| 730 |
+
},
|
| 731 |
+
{
|
| 732 |
+
"type": "text",
|
| 733 |
+
"text": "Skeleton Reconstruction Given a sampled point cloud representation $\\mathcal{X}_A$ from $S_A$ and $M_A$ as input, we aim to train a motion adaptation function $F_{S_A\\to S_B}$ which is tasked with producing a visually similar skeletal motion $\\hat{M}_B$ for $S_B$ using a temporal and set-based neural network. As shown in Fig. 3, to construct $F_{S_A\\rightarrow S_B}$ , we employ a Point Transformer [54] to learn embeddings for the unordered point cloud sets frame-wisely, and a standard temporal transformer to process the sequence and decode $\\mathcal{X}_A$ into $\\hat{M}_B$ .",
|
| 734 |
+
"bbox": [
|
| 735 |
+
212,
|
| 736 |
+
338,
|
| 737 |
+
782,
|
| 738 |
+
443
|
| 739 |
+
],
|
| 740 |
+
"page_idx": 6
|
| 741 |
+
},
|
| 742 |
+
{
|
| 743 |
+
"type": "text",
|
| 744 |
+
"text": "To optimize $F_{S_A \\to S_B}$ , our objective function maximises the similarity between the sampled point clouds of both $M_A$ and $\\hat{M}_B$ . $\\mathcal{X}_A$ can be viewed as the ground truth to $\\mathcal{X}_B$ , which enables an unsupervised strategy for $F_{S_A \\to S_B}$ . In detail, we first sample a point cloud $\\mathcal{X}_B$ from $S_B$ based on $\\hat{M}_B$ derived by $F_{S_A \\to S_B}$ . Next, a temporally consistent KNN-Loss objective is devised to minimise the $\\ell_2$ distance between any given point of $\\mathcal{X}_B$ from its $K$ nearest neighbouring points in $\\mathcal{X}_A$ . The $K$ nearest neighbours are determined based on inter-point distance throughout the entire motion sequence, instead of a frame-wise setting, as shown in Fig. 4. The points with differing body groups are excluded from the neighbour search. Mathematically, we have:",
|
| 745 |
+
"bbox": [
|
| 746 |
+
212,
|
| 747 |
+
445,
|
| 748 |
+
784,
|
| 749 |
+
595
|
| 750 |
+
],
|
| 751 |
+
"page_idx": 6
|
| 752 |
+
},
|
| 753 |
+
{
|
| 754 |
+
"type": "equation",
|
| 755 |
+
"text": "\n$$\n\\delta \\left(u _ {A}, u _ {B}\\right) = \\left\\{ \\begin{array}{l l} \\sum_ {t} \\left\\| \\mathbf {p} _ {u _ {A}, t} - \\mathbf {p} _ {u _ {B}, t} \\right\\| _ {2}, & \\text {i f} \\mathbf {g} _ {u _ {A}} = \\mathbf {g} _ {u _ {B}}, \\\\ \\inf , & \\text {o t h e r w i s e}, \\end{array} \\right. \\tag {4}\n$$\n",
|
| 756 |
+
"text_format": "latex",
|
| 757 |
+
"bbox": [
|
| 758 |
+
316,
|
| 759 |
+
608,
|
| 760 |
+
784,
|
| 761 |
+
647
|
| 762 |
+
],
|
| 763 |
+
"page_idx": 6
|
| 764 |
+
},
|
| 765 |
+
{
|
| 766 |
+
"type": "equation",
|
| 767 |
+
"text": "\n$$\n\\mathcal {L} _ {\\mathrm {K N N}} \\left(\\mathcal {X} _ {A}, \\mathcal {X} _ {B}\\right) = \\sum_ {u _ {A} \\in \\mathcal {X} _ {A}, u _ {B} \\in N _ {k} \\left(u _ {A}, \\mathcal {X} _ {B}\\right)} \\delta \\left(u _ {A}, u _ {B}\\right), \\tag {5}\n$$\n",
|
| 768 |
+
"text_format": "latex",
|
| 769 |
+
"bbox": [
|
| 770 |
+
316,
|
| 771 |
+
651,
|
| 772 |
+
784,
|
| 773 |
+
685
|
| 774 |
+
],
|
| 775 |
+
"page_idx": 6
|
| 776 |
+
},
|
| 777 |
+
{
|
| 778 |
+
"type": "text",
|
| 779 |
+
"text": "where $N_{k}(u_{A},\\mathcal{X}_{B})$ is the set containing points that are the k-nearest neighbors of $u_{A}$ in $\\mathcal{X}_B$ , based on $\\delta (u_A,u_B)$ distance.",
|
| 780 |
+
"bbox": [
|
| 781 |
+
212,
|
| 782 |
+
696,
|
| 783 |
+
782,
|
| 784 |
+
727
|
| 785 |
+
],
|
| 786 |
+
"page_idx": 6
|
| 787 |
+
},
|
| 788 |
+
{
|
| 789 |
+
"type": "text",
|
| 790 |
+
"text": "In addition to the KNN-Loss, we introduce an optional end-effector loss in the skeleton space for the hand, foot, and head bones, which provides direct positional guidance for matched end-effector pairs $E_{AB}$ :",
|
| 791 |
+
"bbox": [
|
| 792 |
+
212,
|
| 793 |
+
728,
|
| 794 |
+
782,
|
| 795 |
+
773
|
| 796 |
+
],
|
| 797 |
+
"page_idx": 6
|
| 798 |
+
},
|
| 799 |
+
{
|
| 800 |
+
"type": "equation",
|
| 801 |
+
"text": "\n$$\n\\mathcal {L} _ {\\text {e n d}} = \\sum_ {(\\mathbf {v} _ {A}, \\mathbf {v} _ {B}) \\in E _ {A B}} \\frac {1}{| M _ {A} |} \\sum_ {t} \\left\\| \\mathbf {p} _ {\\mathbf {v} _ {A}, t} - \\mathbf {p} _ {\\mathbf {v} _ {B}, t} \\right\\| _ {2}. \\tag {6}\n$$\n",
|
| 802 |
+
"text_format": "latex",
|
| 803 |
+
"bbox": [
|
| 804 |
+
334,
|
| 805 |
+
785,
|
| 806 |
+
784,
|
| 807 |
+
821
|
| 808 |
+
],
|
| 809 |
+
"page_idx": 6
|
| 810 |
+
},
|
| 811 |
+
{
|
| 812 |
+
"type": "header",
|
| 813 |
+
"text": "Motion Keyframe Interpolation for Any Human Skeleton",
|
| 814 |
+
"bbox": [
|
| 815 |
+
351,
|
| 816 |
+
114,
|
| 817 |
+
730,
|
| 818 |
+
128
|
| 819 |
+
],
|
| 820 |
+
"page_idx": 6
|
| 821 |
+
},
|
| 822 |
+
{
|
| 823 |
+
"type": "page_number",
|
| 824 |
+
"text": "7",
|
| 825 |
+
"bbox": [
|
| 826 |
+
774,
|
| 827 |
+
116,
|
| 828 |
+
784,
|
| 829 |
+
126
|
| 830 |
+
],
|
| 831 |
+
"page_idx": 6
|
| 832 |
+
},
|
| 833 |
+
{
|
| 834 |
+
"type": "image",
|
| 835 |
+
"img_path": "images/3a14d93f1d2874a63ea21ea43eb030712e6913a3b9a1c82718ed631ffec26b07.jpg",
|
| 836 |
+
"image_caption": [
|
| 837 |
+
"Fig. 4: Visualisation of our temporally consistent KNN objective, processed independently per body group. For each point from $\\mathcal{X}_A$ (a), the total $\\ell_2$ distance throughout the sequence is measured between each point (b) from $\\mathcal{X}_B$ . The $\\ell_2$ distances from the K-nearest points are summed to produce KNN-Loss (c)."
|
| 838 |
+
],
|
| 839 |
+
"image_footnote": [],
|
| 840 |
+
"bbox": [
|
| 841 |
+
225,
|
| 842 |
+
146,
|
| 843 |
+
776,
|
| 844 |
+
280
|
| 845 |
+
],
|
| 846 |
+
"page_idx": 7
|
| 847 |
+
},
|
| 848 |
+
{
|
| 849 |
+
"type": "text",
|
| 850 |
+
"text": "3.3 First-Frame Offset Quaternions",
|
| 851 |
+
"text_level": 1,
|
| 852 |
+
"bbox": [
|
| 853 |
+
215,
|
| 854 |
+
378,
|
| 855 |
+
521,
|
| 856 |
+
393
|
| 857 |
+
],
|
| 858 |
+
"page_idx": 7
|
| 859 |
+
},
|
| 860 |
+
{
|
| 861 |
+
"type": "text",
|
| 862 |
+
"text": "Point clouds lack absolute roll axis information for each bone. As such, we introduce a First-frame Offset Quaternion (FOQ) transform to incorporate the available relative roll data for motion modelling. We formulate all quaternions of a motion sequence relative to their values from the initial frame. In detail, for any quaternion sequence $Q = \\{\\mathbf{q}_0, \\mathbf{q}_1, \\mathbf{q}_2, \\ldots\\}$ , we can simply obtain its FOQ transform as $\\{\\mathbf{q}_0 \\times \\mathbf{q}_0^{-1}, \\mathbf{q}_1 \\times \\mathbf{q}_0^{-1}, \\mathbf{q}_2 \\times \\mathbf{q}_0^{-1}, \\ldots\\}$ , where $\\mathbf{q}_0^{-1}$ is the quaternion conjugate of $\\mathbf{q}_0$ . An FOQ sequence can easily be converted back to a quaternion sequence as long as any $\\mathbf{q}_t \\in Q$ is known.",
|
| 863 |
+
"bbox": [
|
| 864 |
+
212,
|
| 865 |
+
405,
|
| 866 |
+
787,
|
| 867 |
+
526
|
| 868 |
+
],
|
| 869 |
+
"page_idx": 7
|
| 870 |
+
},
|
| 871 |
+
{
|
| 872 |
+
"type": "text",
|
| 873 |
+
"text": "Crucially, FOQ is a roll-invariant rotation representation, that is, FOQ remains identical regardless of a bone's absolute roll position in the original data. Trivially, to adjust the rotational roll of a bone $\\mathbf{v}_n$ , we perform the quaternion multiplication for each frame: $\\mathbf{q}_{n,t} \\times \\mathbf{r}$ , where $\\mathbf{r}$ is a roll quaternion. As a quaternion on the roll axis of $\\mathbf{v}_n$ , $\\mathbf{r}$ is constrained within $\\mathbf{r} = \\alpha + \\beta (x\\mathbf{i} + y\\mathbf{j} + z\\mathbf{k})$ , where $x,y,z$ are the 3D XYZ values of $\\mathbf{T}_n$ , and $\\alpha, \\beta$ are scalars that control the roll magnitude. Note that $\\mathbf{r}$ is constant throughout the sequence. Therefore, we can deduce the roll-invariant nature of any constant $\\mathbf{r}$ :",
|
| 874 |
+
"bbox": [
|
| 875 |
+
212,
|
| 876 |
+
527,
|
| 877 |
+
787,
|
| 878 |
+
647
|
| 879 |
+
],
|
| 880 |
+
"page_idx": 7
|
| 881 |
+
},
|
| 882 |
+
{
|
| 883 |
+
"type": "equation",
|
| 884 |
+
"text": "\n$$\n\\begin{array}{l} \\left(\\mathbf {q} _ {n, t} \\times \\mathbf {r}\\right) \\times \\left(\\mathbf {q} _ {0, t} \\times \\mathbf {r}\\right) ^ {- 1} = \\mathbf {q} _ {n, t} \\times \\mathbf {r} \\times \\mathbf {r} ^ {- 1} \\times \\mathbf {q} _ {0, t} ^ {- 1} \\\\ = \\mathbf {q} _ {n, t} \\mathbf {q} _ {0, t} ^ {- 1} \\\\ \\end{array}\n$$\n",
|
| 885 |
+
"text_format": "latex",
|
| 886 |
+
"bbox": [
|
| 887 |
+
326,
|
| 888 |
+
657,
|
| 889 |
+
671,
|
| 890 |
+
699
|
| 891 |
+
],
|
| 892 |
+
"page_idx": 7
|
| 893 |
+
},
|
| 894 |
+
{
|
| 895 |
+
"type": "text",
|
| 896 |
+
"text": "3.4 Motion Interpolation Transformer and Rest Pose Augmentation",
|
| 897 |
+
"text_level": 1,
|
| 898 |
+
"bbox": [
|
| 899 |
+
214,
|
| 900 |
+
720,
|
| 901 |
+
785,
|
| 902 |
+
738
|
| 903 |
+
],
|
| 904 |
+
"page_idx": 7
|
| 905 |
+
},
|
| 906 |
+
{
|
| 907 |
+
"type": "text",
|
| 908 |
+
"text": "To leverage our point cloud-based representation learning for cross-skeleton interpolation, a transformer model - CITL [34] is adapted, allowing for an efficient interpolation pipeline with existing datasets to alternative skeletons. We introduce two key modifications to CITL and its training strategy, focusing on enhancing motion quality. First, quaternion predictions are substituted with FOQ predictions. This is inspired by the RNN-based method TGcomplete [18], which predicts",
|
| 909 |
+
"bbox": [
|
| 910 |
+
212,
|
| 911 |
+
750,
|
| 912 |
+
787,
|
| 913 |
+
842
|
| 914 |
+
],
|
| 915 |
+
"page_idx": 7
|
| 916 |
+
},
|
| 917 |
+
{
|
| 918 |
+
"type": "page_number",
|
| 919 |
+
"text": "8",
|
| 920 |
+
"bbox": [
|
| 921 |
+
217,
|
| 922 |
+
114,
|
| 923 |
+
228,
|
| 924 |
+
126
|
| 925 |
+
],
|
| 926 |
+
"page_idx": 7
|
| 927 |
+
},
|
| 928 |
+
{
|
| 929 |
+
"type": "header",
|
| 930 |
+
"text": "C. Mo et al.",
|
| 931 |
+
"bbox": [
|
| 932 |
+
271,
|
| 933 |
+
114,
|
| 934 |
+
354,
|
| 935 |
+
126
|
| 936 |
+
],
|
| 937 |
+
"page_idx": 7
|
| 938 |
+
},
|
| 939 |
+
{
|
| 940 |
+
"type": "code",
|
| 941 |
+
"sub_type": "algorithm",
|
| 942 |
+
"code_caption": [
|
| 943 |
+
"Algorithm 1 IK target alignment with maximum real quaternion component"
|
| 944 |
+
],
|
| 945 |
+
"code_body": "1: Input: $\\mathbf{p_{IK}} = 3D$ IK target unit vector \n2: $\\mathbf{T} =$ resting tail position of associated bone \n3: Output: $\\mathbf{q}_{\\mathrm{max}} =$ output quaternion, to be applied on original bone quaternion \n4: $\\eta \\leftarrow \\frac{\\mathbf{T}}{|\\mathbf{T}|}$ $\\triangleright$ 3D unit coordinates $[\\eta_x,\\eta_y,\\eta_z]$ \n5: $\\omega \\leftarrow \\frac{\\mathbf{p_{IK}}}{|\\mathbf{p_{IK}}|}$ $\\triangleright$ 3D unit coordinates $[\\omega_x,\\omega_y,\\omega_z]$ \n6: $\\mathbf{q}_{\\mathrm{min}} \\gets \\frac{0 + (\\eta_x + \\omega_x)\\mathbf{i} + (\\eta_y + \\omega_y)\\mathbf{j} + (\\eta_z + \\omega_z)\\mathbf{k}}{2}$ \n7: $\\mathbf{q}_{\\mathrm{max}} \\gets \\mathbf{q}_{\\mathrm{min}} \\times (0 + \\eta_x\\mathbf{i} + \\eta_y\\mathbf{j} + \\eta_z\\mathbf{k})$",
|
| 946 |
+
"bbox": [
|
| 947 |
+
218,
|
| 948 |
+
164,
|
| 949 |
+
787,
|
| 950 |
+
273
|
| 951 |
+
],
|
| 952 |
+
"page_idx": 8
|
| 953 |
+
},
|
| 954 |
+
{
|
| 955 |
+
"type": "text",
|
| 956 |
+
"text": "quaternion offsets from the preceding frame rather than raw quaternions, thereby achieving greater generalisability and superior quality in interpolation.",
|
| 957 |
+
"bbox": [
|
| 958 |
+
212,
|
| 959 |
+
301,
|
| 960 |
+
782,
|
| 961 |
+
332
|
| 962 |
+
],
|
| 963 |
+
"page_idx": 8
|
| 964 |
+
},
|
| 965 |
+
{
|
| 966 |
+
"type": "text",
|
| 967 |
+
"text": "For our model's CITL-based architecture, we substitute the original sinusoidal positional encoding scheme with learned relative positions [44]. While the original implementation relies on the continual nature of sinusoidal functions, we observe that relative positional embeddings with zero initialisation converges upon similarly continuous behaviour, and significantly lowers the complexity of positional attention. The lessened focus on positional relations allows the model to synthesise deeper behaviours within the pose space, improving overall generalisability. As a side effect, the model also accepts arbitrary length inputs beyond the training data length.",
|
| 968 |
+
"bbox": [
|
| 969 |
+
212,
|
| 970 |
+
332,
|
| 971 |
+
784,
|
| 972 |
+
467
|
| 973 |
+
],
|
| 974 |
+
"page_idx": 8
|
| 975 |
+
},
|
| 976 |
+
{
|
| 977 |
+
"type": "text",
|
| 978 |
+
"text": "Notably, the intended rest pose of a skeletal configuration is often geometrically sub-optimal for point cloud matching, and can manifest in multiple forms. To address this, we devise a rest position augmentation strategy, leveraging a per-bone Inverse Kinematics (IK) system. RPA broadens the range of intended rest poses that the learned interpolation model can accommodate. In detail, during each phase of the Forward Kinematics (FK) process, we augment the bone tails with a random additional offset. We re-align the augmented global tail position as closely as possible to its original location using a quaternion rotation with the maximum possible real component. Algorithm 1 describes the method by which this quaternion can be derived. Since the real component describes, in cosine, the amount of rotation around an axis, this minimizes numerical disturbance from RPA while altering the visual representation to the desired extent.",
|
| 979 |
+
"bbox": [
|
| 980 |
+
212,
|
| 981 |
+
468,
|
| 982 |
+
785,
|
| 983 |
+
650
|
| 984 |
+
],
|
| 985 |
+
"page_idx": 8
|
| 986 |
+
},
|
| 987 |
+
{
|
| 988 |
+
"type": "text",
|
| 989 |
+
"text": "4 Experimental Results",
|
| 990 |
+
"text_level": 1,
|
| 991 |
+
"bbox": [
|
| 992 |
+
215,
|
| 993 |
+
686,
|
| 994 |
+
460,
|
| 995 |
+
704
|
| 996 |
+
],
|
| 997 |
+
"page_idx": 8
|
| 998 |
+
},
|
| 999 |
+
{
|
| 1000 |
+
"type": "text",
|
| 1001 |
+
"text": "4.1 Datasets",
|
| 1002 |
+
"text_level": 1,
|
| 1003 |
+
"bbox": [
|
| 1004 |
+
215,
|
| 1005 |
+
715,
|
| 1006 |
+
333,
|
| 1007 |
+
729
|
| 1008 |
+
],
|
| 1009 |
+
"page_idx": 8
|
| 1010 |
+
},
|
| 1011 |
+
{
|
| 1012 |
+
"type": "text",
|
| 1013 |
+
"text": "To thoroughly evaluate the effectiveness of the proposed method, we conducted extensive experiments on the following widely used motion capture datasets:",
|
| 1014 |
+
"bbox": [
|
| 1015 |
+
212,
|
| 1016 |
+
739,
|
| 1017 |
+
782,
|
| 1018 |
+
771
|
| 1019 |
+
],
|
| 1020 |
+
"page_idx": 8
|
| 1021 |
+
},
|
| 1022 |
+
{
|
| 1023 |
+
"type": "text",
|
| 1024 |
+
"text": "- LaFAN1 [18] contains long, high quality motions in controllable video game character styles. The dataset was recorded on a large motion capture stage, and manually cleaned to production standards. The native skeleton of LaFAN1 contains 6 spinal bones, and 4 bones for each limb.",
|
| 1025 |
+
"bbox": [
|
| 1026 |
+
223,
|
| 1027 |
+
779,
|
| 1028 |
+
792,
|
| 1029 |
+
839
|
| 1030 |
+
],
|
| 1031 |
+
"page_idx": 8
|
| 1032 |
+
},
|
| 1033 |
+
{
|
| 1034 |
+
"type": "header",
|
| 1035 |
+
"text": "Motion Keyframe Interpolation for Any Human Skeleton",
|
| 1036 |
+
"bbox": [
|
| 1037 |
+
351,
|
| 1038 |
+
114,
|
| 1039 |
+
730,
|
| 1040 |
+
128
|
| 1041 |
+
],
|
| 1042 |
+
"page_idx": 8
|
| 1043 |
+
},
|
| 1044 |
+
{
|
| 1045 |
+
"type": "page_number",
|
| 1046 |
+
"text": "9",
|
| 1047 |
+
"bbox": [
|
| 1048 |
+
774,
|
| 1049 |
+
116,
|
| 1050 |
+
784,
|
| 1051 |
+
126
|
| 1052 |
+
],
|
| 1053 |
+
"page_idx": 8
|
| 1054 |
+
},
|
| 1055 |
+
{
|
| 1056 |
+
"type": "text",
|
| 1057 |
+
"text": "- Human3.6M [6,22] comprises of 3.6 million motion frames. The motions captured in Human3.6M tend to be less dynamic, as they were recorded on a relatively compact $4 \\times 3 \\mathrm{~m}$ stage. Its native skeleton includes 5 spinal bones, 4 bones for each limb, and an additional finger bone for each hand.",
|
| 1058 |
+
"bbox": [
|
| 1059 |
+
223,
|
| 1060 |
+
145,
|
| 1061 |
+
785,
|
| 1062 |
+
205
|
| 1063 |
+
],
|
| 1064 |
+
"page_idx": 9
|
| 1065 |
+
},
|
| 1066 |
+
{
|
| 1067 |
+
"type": "text",
|
| 1068 |
+
"text": "- CMU Mocap [10] is a diverse motion dataset encompassing 113 categories. It features a mix of static and dynamic motions, captured on a $3 \\times 8 \\mathrm{~m}$ stage. The skeleton is identical to that of Human3.6M, with the exception of the pelvis, which is divided into two hip bones and a lower back bone.",
|
| 1069 |
+
"bbox": [
|
| 1070 |
+
225,
|
| 1071 |
+
205,
|
| 1072 |
+
785,
|
| 1073 |
+
267
|
| 1074 |
+
],
|
| 1075 |
+
"page_idx": 9
|
| 1076 |
+
},
|
| 1077 |
+
{
|
| 1078 |
+
"type": "text",
|
| 1079 |
+
"text": "For PC-MRL, we designated the CMU MoCap skeleton as $S_B$ . Its native dataset is exclusively utilized for testing purposes, while LaFAN1 and Human3.6M are used simultaneously for training. $^3$",
|
| 1080 |
+
"bbox": [
|
| 1081 |
+
215,
|
| 1082 |
+
271,
|
| 1083 |
+
785,
|
| 1084 |
+
318
|
| 1085 |
+
],
|
| 1086 |
+
"page_idx": 9
|
| 1087 |
+
},
|
| 1088 |
+
{
|
| 1089 |
+
"type": "text",
|
| 1090 |
+
"text": "4.2 Implementation Details",
|
| 1091 |
+
"text_level": 1,
|
| 1092 |
+
"bbox": [
|
| 1093 |
+
215,
|
| 1094 |
+
335,
|
| 1095 |
+
455,
|
| 1096 |
+
351
|
| 1097 |
+
],
|
| 1098 |
+
"page_idx": 9
|
| 1099 |
+
},
|
| 1100 |
+
{
|
| 1101 |
+
"type": "text",
|
| 1102 |
+
"text": "In our experiments, point cloud obfuscation adopted a 256-point setting for the sampling process with $\\sigma = 0.05$ and each point was characterized by a 64-element vector. To maintain consistency, we ensure that all bone offsets and motion positions are defined in metre units. For skeletal reconstruction, 4 Point Transformer [54] layers was utilized to construct the neural network. Within each layer, the feature size was doubled, and the size of the point set was downsampled by a factor of 4 through a farthest-first traversal approach, eventually resulting in a singular feature vector. In the final step, two linear layers with ReLU activations were utilized to transform the resulted feature vector in a frame-wise manner, producing the output skeletal representation. During training, we set $K = 8$ for the KNN loss to optimize the network. Additionally, the use of unit quaternion values in practice relies on modulus division for constraint, a process that can potentially lead to exploding gradients. To address this, we introduced an objective function $\\mathcal{L}_q$ that measures the $\\ell_2$ norms' difference between the raw quaternion output and the expected unit quaternion. All experiments were trained and evaluated on a single NVIDIA GeForce RTX 3090 GPU.",
|
| 1103 |
+
"bbox": [
|
| 1104 |
+
212,
|
| 1105 |
+
357,
|
| 1106 |
+
787,
|
| 1107 |
+
599
|
| 1108 |
+
],
|
| 1109 |
+
"page_idx": 9
|
| 1110 |
+
},
|
| 1111 |
+
{
|
| 1112 |
+
"type": "text",
|
| 1113 |
+
"text": "4.3 Motion Interpolation Comparison against Supervised Methods",
|
| 1114 |
+
"text_level": 1,
|
| 1115 |
+
"bbox": [
|
| 1116 |
+
215,
|
| 1117 |
+
617,
|
| 1118 |
+
779,
|
| 1119 |
+
633
|
| 1120 |
+
],
|
| 1121 |
+
"page_idx": 9
|
| 1122 |
+
},
|
| 1123 |
+
{
|
| 1124 |
+
"type": "text",
|
| 1125 |
+
"text": "To demonstrate the efficacy of our method, we produce and compare our method against the conventional LERP method, as well as three state-of-the-art interpolation models trained on the original CMU MoCap dataset. Specifically, we measure the performance of the RNN-based $\\mathrm{TG}_{\\mathrm{complete}}$ model [18], the BERT-based motion interpolation adaptation [12], and a transformer-based encoder-decoder approach [34]. We additionally provide results on a variant of our method trained without our RPA strategy. The experimental configuration for each state-of-the-art model is identical to their original implementations. To measure interpolation accuracy, we employ the standard $\\ell_2$ positional distance (L2P), $\\ell_2$ quaternion difference (L2Q), and NPSS [16] for visual similarity evaluation [18].",
|
| 1126 |
+
"bbox": [
|
| 1127 |
+
212,
|
| 1128 |
+
638,
|
| 1129 |
+
787,
|
| 1130 |
+
790
|
| 1131 |
+
],
|
| 1132 |
+
"page_idx": 9
|
| 1133 |
+
},
|
| 1134 |
+
{
|
| 1135 |
+
"type": "page_number",
|
| 1136 |
+
"text": "10",
|
| 1137 |
+
"bbox": [
|
| 1138 |
+
217,
|
| 1139 |
+
114,
|
| 1140 |
+
235,
|
| 1141 |
+
126
|
| 1142 |
+
],
|
| 1143 |
+
"page_idx": 9
|
| 1144 |
+
},
|
| 1145 |
+
{
|
| 1146 |
+
"type": "header",
|
| 1147 |
+
"text": "C. Mo et al.",
|
| 1148 |
+
"bbox": [
|
| 1149 |
+
271,
|
| 1150 |
+
114,
|
| 1151 |
+
354,
|
| 1152 |
+
126
|
| 1153 |
+
],
|
| 1154 |
+
"page_idx": 9
|
| 1155 |
+
},
|
| 1156 |
+
{
|
| 1157 |
+
"type": "page_footnote",
|
| 1158 |
+
"text": "<sup>3</sup> The authors Clinton Mo, Kun Hu and Zhiyong Wang are signatories of the dataset licenses and produced all experimental results of the paper. Meta Inc. was not granted access to the datasets.",
|
| 1159 |
+
"bbox": [
|
| 1160 |
+
217,
|
| 1161 |
+
796,
|
| 1162 |
+
785,
|
| 1163 |
+
839
|
| 1164 |
+
],
|
| 1165 |
+
"page_idx": 9
|
| 1166 |
+
},
|
| 1167 |
+
{
|
| 1168 |
+
"type": "table",
|
| 1169 |
+
"img_path": "images/d6f4353838fb737cf53e19a26d670c71b8f1f903ef28e0ecfc73dc2930b5e69b.jpg",
|
| 1170 |
+
"table_caption": [],
|
| 1171 |
+
"table_footnote": [],
|
| 1172 |
+
"table_body": "<table><tr><td rowspan=\"2\">Motion category</td><td rowspan=\"2\">Method</td><td colspan=\"3\">L2P↓</td><td colspan=\"3\">L2Q↓</td><td colspan=\"3\">NPSS↓</td></tr><tr><td>5</td><td>15</td><td>30</td><td>5</td><td>15</td><td>30</td><td>5</td><td>15</td><td>30</td></tr><tr><td rowspan=\"6\">Basketball</td><td>LERP</td><td>0.0787</td><td>0.4009</td><td>0.8033</td><td>0.1588</td><td>0.5834</td><td>0.9611</td><td>0.1315</td><td>0.6120</td><td>1.8202</td></tr><tr><td>TGcomplete [18]</td><td>0.1050</td><td>0.4419</td><td>0.7104</td><td>0.1859</td><td>0.6439</td><td>0.9303</td><td>0.1600</td><td>0.6579</td><td>1.2136</td></tr><tr><td>BERT [12]</td><td>0.0889</td><td>0.3737</td><td>0.6865</td><td>0.1585</td><td>0.5560</td><td>0.8992</td><td>0.1689</td><td>0.6177</td><td>1.4587</td></tr><tr><td>CITL [34]</td><td>0.0625</td><td>0.2973</td><td>0.5708</td><td>0.1410</td><td>0.4932</td><td>0.8011</td><td>0.1047</td><td>0.4098</td><td>1.0959</td></tr><tr><td>PC-MRL w/o RPA</td><td>0.0767</td><td>0.3534</td><td>0.6715</td><td>0.1505</td><td>0.5431</td><td>0.9049</td><td>0.1364</td><td>0.5596</td><td>1.3471</td></tr><tr><td>PC-MRL (Ours)</td><td>0.0752</td><td>0.3443</td><td>0.6662</td><td>0.1561</td><td>0.5495</td><td>0.9088</td><td>0.1397</td><td>0.5572</td><td>1.3015</td></tr><tr><td rowspan=\"6\">Golf</td><td>LERP</td><td>0.0189</td><td>0.1229</td><td>0.3187</td><td>0.0455</td><td>0.1870</td><td>0.3987</td><td>0.0587</td><td>0.3979</td><td>1.0312</td></tr><tr><td>TGcomplete [18]</td><td>0.0473</td><td>0.2616</td><td>0.4546</td><td>0.0693</td><td>0.3191</td><td>0.5297</td><td>0.0954</td><td>0.5062</td><td>1.1534</td></tr><tr><td>BERT [12]</td><td>0.0276</td><td>0.1159</td><td>0.2967</td><td>0.0612</td><td>0.2082</td><td>0.4258</td><td>0.1099</td><td>0.3906</td><td>0.9255</td></tr><tr><td>CITL [34]</td><td>0.0201</td><td>0.1043</td><td>0.2589</td><td>0.0476</td><td>0.1766</td><td>0.3552</td><td>0.0636</td><td>0.2581</td><td>0.7799</td></tr><tr><td>PC-MRL w/out RPA</td><td>0.0274</td><td>0.1304</td><td>0.3622</td><td>0.0555</td><td>0.2065</td><td>0.4923</td><td>0.0911</td><td>0.3759</td><td>1.2083</td></tr><tr><td>PC-MRL (Ours)</td><td>0.0273</td><td>0.1130</td><td>0.2706</td><td>0.0598</td><td>0.2012</td><td>0.3844</td><td>0.1049</td><td>0.3716</td><td>0.8150</td></tr><tr><td rowspan=\"6\">Swimming</td><td>LERP</td><td>0.0731</td><td>0.3680</td><td>0.7374</td><td>0.1326</td><td>0.5327</td><td>0.9476</td><td>0.2147</td><td>0.9222</td><td>1.6972</td></tr><tr><td>TGcomplete [18]</td><td>0.1292</td><td>0.5796</td><td>0.9141</td><td>0.1845</td><td>0.7420</td><td>1.1199</td><td>0.2660</td><td>0.9716</td><td>1.7698</td></tr><tr><td>BERT [12]</td><td>0.1088</td><td>0.3896</td><td>0.7516</td><td>0.1819</td><td>0.5518</td><td>0.9622</td><td>0.2877</td><td>0.9074</td><td>1.6682</td></tr><tr><td>CITL [34]</td><td>0.0905</td><td>0.3895</td><td>0.7475</td><td>0.1484</td><td>0.5535</td><td>0.9533</td><td>0.2084</td><td>0.8317</td><td>1.6122</td></tr><tr><td>PC-MRL w/o RPA</td><td>0.0722</td><td>0.3754</td><td>0.7459</td><td>0.1345</td><td>0.5442</td><td>0.9594</td><td>0.1944</td><td>0.8731</td><td>1.6608</td></tr><tr><td>PC-MRL (Ours)</td><td>0.0734</td><td>0.3595</td><td>0.7139</td><td>0.1355</td><td>0.5331</td><td>0.9248</td><td>0.1956</td><td>0.8591</td><td>1.6370</td></tr><tr><td rowspan=\"6\">Walking & Running</td><td>LERP</td><td>0.0455</td><td>0.2505</td><td>0.5527</td><td>0.1120</td><td>0.3997</td><td>0.6233</td><td>0.0655</td><td>0.2682</td><td>0.9071</td></tr><tr><td>TGcomplete [18]</td><td>0.0540</td><td>0.1786</td><td>0.3103</td><td>0.1147</td><td>0.2866</td><td>0.3866</td><td>0.0895</td><td>0.2616</td><td>0.4265</td></tr><tr><td>BERT [12]</td><td>0.0552</td><td>0.1998</td><td>0.3544</td><td>0.1111</td><td>0.3101</td><td>0.4762</td><td>0.0901</td><td>0.2760</td><td>0.6458</td></tr><tr><td>CITL [34]</td><td>0.0328</td><td>0.1034</td><td>0.2427</td><td>0.0942</td><td>0.2102</td><td>0.3384</td><td>0.0526</td><td>0.1568</td><td>0.4080</td></tr><tr><td>PC-MRL w/o RPA</td><td>0.0397</td><td>0.1657</td><td>0.3417</td><td>0.1058</td><td>0.2879</td><td>0.4722</td><td>0.0640</td><td>0.2288</td><td>0.6021</td></tr><tr><td>PC-MRL (Ours)</td><td>0.0377</td><td>0.1392</td><td>0.2848</td><td>0.1064</td><td>0.2778</td><td>0.4281</td><td>0.0658</td><td>0.2129</td><td>0.5098</td></tr></table>",
|
| 1173 |
+
"bbox": [
|
| 1174 |
+
223,
|
| 1175 |
+
143,
|
| 1176 |
+
782,
|
| 1177 |
+
388
|
| 1178 |
+
],
|
| 1179 |
+
"page_idx": 10
|
| 1180 |
+
},
|
| 1181 |
+
{
|
| 1182 |
+
"type": "text",
|
| 1183 |
+
"text": "Table 1: Performance comparisons for various motion categories, measured in L2P, L2Q, and NPSS. All methods except PC-MRL are trained on the original CMU MoCap dataset. Each motion contains 128 frames with keyframes placed every 5, 15, or 30 frames. The top 2 results for each test are highlighted in purple and blue respectively.",
|
| 1184 |
+
"bbox": [
|
| 1185 |
+
215,
|
| 1186 |
+
388,
|
| 1187 |
+
787,
|
| 1188 |
+
444
|
| 1189 |
+
],
|
| 1190 |
+
"page_idx": 10
|
| 1191 |
+
},
|
| 1192 |
+
{
|
| 1193 |
+
"type": "text",
|
| 1194 |
+
"text": "The results listed in Table 1 clearly demonstrate our method's ability to approach the accuracy levels exhibited by directly supervised state-of-the-art methods. PC-MRL consistently outperforms both the LERP standard and RNN-based $\\mathrm{TG}_{\\mathrm{complete}}$ model, particularly in longer keyframe interval scenarios where the quality and depth of learned motion features is most critical and impactful. Likewise, the inclusion of RPA also tends to improve the long interval performance of PC-MRL, particularly when precise movements (i.e. golf) or deeper motion features (i.e. locomotion) are expected. Given this, we can observe that RPA definitively improves the overall consistency of the PC-MRL method in scenarios that may be unseen or less effectively represented by our point cloud-to-motion system.",
|
| 1195 |
+
"bbox": [
|
| 1196 |
+
212,
|
| 1197 |
+
489,
|
| 1198 |
+
787,
|
| 1199 |
+
655
|
| 1200 |
+
],
|
| 1201 |
+
"page_idx": 10
|
| 1202 |
+
},
|
| 1203 |
+
{
|
| 1204 |
+
"type": "text",
|
| 1205 |
+
"text": "Native supervision, i.e. with CITL, unsurprisingly produces the most accurate interpolation model, with a notable exception of swimming motions, where our PC-MRL supervision produces even stronger results. We believe the abundance of low crawl motions, a similar motion to swimming in LaFAN1, provides more meaningful supervision over the original CMU dataset, which conversely has few similar motions once swimming motions are filtered out. On the other end of the spectrum, i.e. for high data scenarios such as walking and running locomotion from Fig. 5, all models including our PC-MRL approach observe the strongest improvements for learned methods over the naive LERP method. Both cases strongly support our method's capability to supervise highly intricate motion patterns, despite the incompleteness of the point cloud representation and reliance on relative rotations and geometric optimisation assumptions.",
|
| 1206 |
+
"bbox": [
|
| 1207 |
+
212,
|
| 1208 |
+
657,
|
| 1209 |
+
787,
|
| 1210 |
+
840
|
| 1211 |
+
],
|
| 1212 |
+
"page_idx": 10
|
| 1213 |
+
},
|
| 1214 |
+
{
|
| 1215 |
+
"type": "header",
|
| 1216 |
+
"text": "Motion Keyframe Interpolation for Any Human Skeleton",
|
| 1217 |
+
"bbox": [
|
| 1218 |
+
351,
|
| 1219 |
+
114,
|
| 1220 |
+
730,
|
| 1221 |
+
128
|
| 1222 |
+
],
|
| 1223 |
+
"page_idx": 10
|
| 1224 |
+
},
|
| 1225 |
+
{
|
| 1226 |
+
"type": "page_number",
|
| 1227 |
+
"text": "11",
|
| 1228 |
+
"bbox": [
|
| 1229 |
+
767,
|
| 1230 |
+
116,
|
| 1231 |
+
782,
|
| 1232 |
+
126
|
| 1233 |
+
],
|
| 1234 |
+
"page_idx": 10
|
| 1235 |
+
},
|
| 1236 |
+
{
|
| 1237 |
+
"type": "image",
|
| 1238 |
+
"img_path": "images/5bf89b54c8e79c1caf8b33c94c9efab4976a017e1a19a022bbde137256c2059a.jpg",
|
| 1239 |
+
"image_caption": [
|
| 1240 |
+
"Fig. 5: Walking keyframe interpolation examples produced by methods from Table 1. Close-up layered comparisons against the original motion are provided on the right."
|
| 1241 |
+
],
|
| 1242 |
+
"image_footnote": [],
|
| 1243 |
+
"bbox": [
|
| 1244 |
+
243,
|
| 1245 |
+
155,
|
| 1246 |
+
759,
|
| 1247 |
+
411
|
| 1248 |
+
],
|
| 1249 |
+
"page_idx": 11
|
| 1250 |
+
},
|
| 1251 |
+
{
|
| 1252 |
+
"type": "text",
|
| 1253 |
+
"text": "4.4 Cross-Skeleton Motion Re-Targeting Experiments",
|
| 1254 |
+
"text_level": 1,
|
| 1255 |
+
"bbox": [
|
| 1256 |
+
214,
|
| 1257 |
+
488,
|
| 1258 |
+
671,
|
| 1259 |
+
505
|
| 1260 |
+
],
|
| 1261 |
+
"page_idx": 11
|
| 1262 |
+
},
|
| 1263 |
+
{
|
| 1264 |
+
"type": "text",
|
| 1265 |
+
"text": "We further conduct an experiment on motion re-targeting to directly benchmark our method against the sole existing state-of-the-art method for unpaired motion re-targeting with unrestricted skeletons, i.e. primal skeletons (PS) [1]. For direct motion re-targeting evaluation, we utilized the KNN-Loss and end-effector loss as metrics, owing to their effective measurement of visual similarity and non-sensitivity towards absolute roll axis correctness. Though absolute rolls are generally necessary for a complete motion re-targeting solution, this aspect is out of scope for our project as our main goal is motion interpolation.",
|
| 1266 |
+
"bbox": [
|
| 1267 |
+
212,
|
| 1268 |
+
513,
|
| 1269 |
+
787,
|
| 1270 |
+
635
|
| 1271 |
+
],
|
| 1272 |
+
"page_idx": 11
|
| 1273 |
+
},
|
| 1274 |
+
{
|
| 1275 |
+
"type": "text",
|
| 1276 |
+
"text": "Table 2 indicates the superior performance of our method over the state-of-the-art PS method in terms of visual similarity. At low $\\sigma$ values, the higher relative $\\mathcal{L}_{\\mathrm{KNN}}$ of the primal skeleton method underscores its geometric deviations",
|
| 1277 |
+
"bbox": [
|
| 1278 |
+
212,
|
| 1279 |
+
636,
|
| 1280 |
+
787,
|
| 1281 |
+
683
|
| 1282 |
+
],
|
| 1283 |
+
"page_idx": 11
|
| 1284 |
+
},
|
| 1285 |
+
{
|
| 1286 |
+
"type": "table",
|
| 1287 |
+
"img_path": "images/1b9559e395566ecda0d7e8684a2e77835c2ba857c8b6aa16239b6973f57981d4.jpg",
|
| 1288 |
+
"table_caption": [],
|
| 1289 |
+
"table_footnote": [],
|
| 1290 |
+
"table_body": "<table><tr><td rowspan=\"2\">Model</td><td colspan=\"5\">\\( L_{\\text{KNN}}(\\mathcal{X}_A,\\mathcal{X}_B) \\downarrow \\)</td><td rowspan=\"2\">\\( L_{\\text{end}} \\downarrow \\)</td></tr><tr><td>\\( \\sigma = 0 \\)</td><td>\\( \\sigma = 0.01 \\)</td><td>\\( \\sigma = 0.05 \\)</td><td>\\( \\sigma = 0.1 \\)</td><td>\\( \\sigma = 0.2 \\)</td></tr><tr><td>Original motion</td><td>1.0099</td><td>2.0234</td><td>5.7399</td><td>9.6636</td><td>16.9223</td><td>0</td></tr><tr><td>Primal skeleton [1]</td><td>8.1109</td><td>8.0353</td><td>9.4206</td><td>12.8291</td><td>20.6332</td><td>9.6990</td></tr><tr><td>PC-MRL (Ours)</td><td>3.8714</td><td>4.1231</td><td>6.7277</td><td>10.3078</td><td>17.5261</td><td>5.6258</td></tr></table>",
|
| 1291 |
+
"bbox": [
|
| 1292 |
+
267,
|
| 1293 |
+
707,
|
| 1294 |
+
738,
|
| 1295 |
+
768
|
| 1296 |
+
],
|
| 1297 |
+
"page_idx": 11
|
| 1298 |
+
},
|
| 1299 |
+
{
|
| 1300 |
+
"type": "text",
|
| 1301 |
+
"text": "Table 2: Average unpaired LaFAN1 to CMU motion re-targeting performance of the primal skeleton method and our point cloud method. $\\mathcal{L}_{\\mathrm{KNN}}$ is measured using 1024-point clouds at $K = 8$ . Independently sampled point clouds from the original LaFAN1 skeleton are provided as optimal $\\mathcal{L}_{\\mathrm{KNN}}$ reference. All values are scaled by $\\times 100$ .",
|
| 1302 |
+
"bbox": [
|
| 1303 |
+
212,
|
| 1304 |
+
767,
|
| 1305 |
+
787,
|
| 1306 |
+
824
|
| 1307 |
+
],
|
| 1308 |
+
"page_idx": 11
|
| 1309 |
+
},
|
| 1310 |
+
{
|
| 1311 |
+
"type": "page_number",
|
| 1312 |
+
"text": "12",
|
| 1313 |
+
"bbox": [
|
| 1314 |
+
217,
|
| 1315 |
+
114,
|
| 1316 |
+
235,
|
| 1317 |
+
126
|
| 1318 |
+
],
|
| 1319 |
+
"page_idx": 11
|
| 1320 |
+
},
|
| 1321 |
+
{
|
| 1322 |
+
"type": "header",
|
| 1323 |
+
"text": "C. Mo et al.",
|
| 1324 |
+
"bbox": [
|
| 1325 |
+
271,
|
| 1326 |
+
114,
|
| 1327 |
+
354,
|
| 1328 |
+
126
|
| 1329 |
+
],
|
| 1330 |
+
"page_idx": 11
|
| 1331 |
+
},
|
| 1332 |
+
{
|
| 1333 |
+
"type": "image",
|
| 1334 |
+
"img_path": "images/4d71934c2946edcc88669cce7457ddcefb2a49921e2cbd04202c403bdc4878ed.jpg",
|
| 1335 |
+
"image_caption": [
|
| 1336 |
+
"Fig. 6: Jumping motion re-targeting examples produced on the CMU skeleton by the primal skeleton method (middle row) and our method (bottom row). Sampled point cloud representations of each motion sequence are provided on the right."
|
| 1337 |
+
],
|
| 1338 |
+
"image_footnote": [],
|
| 1339 |
+
"bbox": [
|
| 1340 |
+
233,
|
| 1341 |
+
146,
|
| 1342 |
+
767,
|
| 1343 |
+
277
|
| 1344 |
+
],
|
| 1345 |
+
"page_idx": 12
|
| 1346 |
+
},
|
| 1347 |
+
{
|
| 1348 |
+
"type": "text",
|
| 1349 |
+
"text": "from the original motion. In contrast, our method demonstrates significantly better visual adherence to the original geometry. As $\\sigma$ increases, our method approaches near-perfect alignment with the original motion. Figure 6 visually indicates this, showing our method follows the original motion closely, while the latent consistency-based primal skeleton method struggles to generalise and decode the skeleton's internal structures without directly supervised objectives, such as the end-effector positions provided during training. In addition, Fig 7 demonstrates that, like existing re-targeting approaches, our method is able to produce adequate results on disproportionate skeletons.",
|
| 1350 |
+
"bbox": [
|
| 1351 |
+
212,
|
| 1352 |
+
375,
|
| 1353 |
+
787,
|
| 1354 |
+
512
|
| 1355 |
+
],
|
| 1356 |
+
"page_idx": 12
|
| 1357 |
+
},
|
| 1358 |
+
{
|
| 1359 |
+
"type": "text",
|
| 1360 |
+
"text": "Since our approach is largely focused on enabling non-native motion interpolation supervision, we additionally compare our PC-MRL method against a CITL model supervised by PS in Table 3. Due to the suboptimal re-targeting performance of PS, such interpolation models exhibit expectedly poor accuracy.",
|
| 1361 |
+
"bbox": [
|
| 1362 |
+
212,
|
| 1363 |
+
516,
|
| 1364 |
+
787,
|
| 1365 |
+
578
|
| 1366 |
+
],
|
| 1367 |
+
"page_idx": 12
|
| 1368 |
+
},
|
| 1369 |
+
{
|
| 1370 |
+
"type": "image",
|
| 1371 |
+
"img_path": "images/22c90d7f153bf66827f9df7e59a0f616eb3eed7b990ed10dbb5963987623cddf.jpg",
|
| 1372 |
+
"image_caption": [
|
| 1373 |
+
"Fig. 7: Re-targeted falling motion from LaFAN1 (top) on the CMU skeleton (middle) and a short custom skeleton (bottom), performed by our point cloud-to-motion model."
|
| 1374 |
+
],
|
| 1375 |
+
"image_footnote": [],
|
| 1376 |
+
"bbox": [
|
| 1377 |
+
290,
|
| 1378 |
+
616,
|
| 1379 |
+
718,
|
| 1380 |
+
762
|
| 1381 |
+
],
|
| 1382 |
+
"page_idx": 12
|
| 1383 |
+
},
|
| 1384 |
+
{
|
| 1385 |
+
"type": "header",
|
| 1386 |
+
"text": "Motion Keyframe Interpolation for Any Human Skeleton",
|
| 1387 |
+
"bbox": [
|
| 1388 |
+
351,
|
| 1389 |
+
114,
|
| 1390 |
+
730,
|
| 1391 |
+
128
|
| 1392 |
+
],
|
| 1393 |
+
"page_idx": 12
|
| 1394 |
+
},
|
| 1395 |
+
{
|
| 1396 |
+
"type": "page_number",
|
| 1397 |
+
"text": "13",
|
| 1398 |
+
"bbox": [
|
| 1399 |
+
767,
|
| 1400 |
+
114,
|
| 1401 |
+
784,
|
| 1402 |
+
126
|
| 1403 |
+
],
|
| 1404 |
+
"page_idx": 12
|
| 1405 |
+
},
|
| 1406 |
+
{
|
| 1407 |
+
"type": "table",
|
| 1408 |
+
"img_path": "images/78c7a253c6c77c9a25a2144dec57a7944e36688644f4412cc77be868fa8bb19c.jpg",
|
| 1409 |
+
"table_caption": [],
|
| 1410 |
+
"table_footnote": [
|
| 1411 |
+
"Table 3: Average L2P, L2Q, and NPSS performance of CITL trained on native and PS-generated datasets, compared to our method."
|
| 1412 |
+
],
|
| 1413 |
+
"table_body": "<table><tr><td rowspan=\"2\">Method</td><td colspan=\"3\">L2P↓</td><td colspan=\"3\">L2Q↓</td><td colspan=\"3\">NPSS↓</td></tr><tr><td>5</td><td>15</td><td>30</td><td>5</td><td>15</td><td>30</td><td>5</td><td>15</td><td>30</td></tr><tr><td>CITL - PS</td><td>0.2043</td><td>0.6332</td><td>1.1524</td><td>0.2318</td><td>0.7428</td><td>1.1442</td><td>0.1122</td><td>0.4276</td><td>1.4622</td></tr><tr><td>CITL - Native</td><td>0.0495</td><td>0.2306</td><td>0.4898</td><td>0.0979</td><td>0.3604</td><td>0.6609</td><td>0.1149</td><td>0.4684</td><td>1.1591</td></tr><tr><td>PC-MRL</td><td>0.0494</td><td>0.2337</td><td>0.4960</td><td>0.1030</td><td>0.3771</td><td>0.6813</td><td>0.1370</td><td>0.5433</td><td>1.1760</td></tr></table>",
|
| 1414 |
+
"bbox": [
|
| 1415 |
+
223,
|
| 1416 |
+
143,
|
| 1417 |
+
782,
|
| 1418 |
+
207
|
| 1419 |
+
],
|
| 1420 |
+
"page_idx": 13
|
| 1421 |
+
},
|
| 1422 |
+
{
|
| 1423 |
+
"type": "text",
|
| 1424 |
+
"text": "5 Limitations",
|
| 1425 |
+
"text_level": 1,
|
| 1426 |
+
"bbox": [
|
| 1427 |
+
215,
|
| 1428 |
+
273,
|
| 1429 |
+
362,
|
| 1430 |
+
289
|
| 1431 |
+
],
|
| 1432 |
+
"page_idx": 13
|
| 1433 |
+
},
|
| 1434 |
+
{
|
| 1435 |
+
"type": "text",
|
| 1436 |
+
"text": "Due to the inherent limitations of the point cloud representation, certain constraints on the applicability of our point cloud-to-motion method are inevitable. Primarily, the reliance of our point cloud re-targeting method on the relative nature of First-frame Offset Quaternions (FOQ) for most motion learning tasks necessitates the presence of at least one known pose. In the absence of a known pose, relative quaternions are incapable of reconstructing absolute rotational values, which are essential for motion applications. Secondly, due to our method of sampling point clouds with a consistent deviation from each bone segment, the resultant representation inevitably obscures finer details. This includes elements like fingers and facial controls, to an extent that is beyond rectification.",
|
| 1437 |
+
"bbox": [
|
| 1438 |
+
212,
|
| 1439 |
+
304,
|
| 1440 |
+
787,
|
| 1441 |
+
455
|
| 1442 |
+
],
|
| 1443 |
+
"page_idx": 13
|
| 1444 |
+
},
|
| 1445 |
+
{
|
| 1446 |
+
"type": "text",
|
| 1447 |
+
"text": "Our proposed objective function relies entirely on geometric segments, necessitating that all bones possess a non-zero length. This stipulation is deemed a reasonable prerequisite for generic human skeletons, as each bone is typically responsible for manipulating a discernible segment of the human body. Due to technical considerations, the skeletal configurations in each of our datasets included at least one bone of zero length, which we ignored in our experiments in favour of global quaternion rotations of all bones with non-zero length.",
|
| 1448 |
+
"bbox": [
|
| 1449 |
+
212,
|
| 1450 |
+
455,
|
| 1451 |
+
787,
|
| 1452 |
+
561
|
| 1453 |
+
],
|
| 1454 |
+
"page_idx": 13
|
| 1455 |
+
},
|
| 1456 |
+
{
|
| 1457 |
+
"type": "text",
|
| 1458 |
+
"text": "As a result of these limitations, we refrained from claiming state-of-the-art performance for general motion re-targeting tasks.",
|
| 1459 |
+
"bbox": [
|
| 1460 |
+
212,
|
| 1461 |
+
561,
|
| 1462 |
+
785,
|
| 1463 |
+
592
|
| 1464 |
+
],
|
| 1465 |
+
"page_idx": 13
|
| 1466 |
+
},
|
| 1467 |
+
{
|
| 1468 |
+
"type": "text",
|
| 1469 |
+
"text": "6 Conclusion",
|
| 1470 |
+
"text_level": 1,
|
| 1471 |
+
"bbox": [
|
| 1472 |
+
215,
|
| 1473 |
+
613,
|
| 1474 |
+
359,
|
| 1475 |
+
628
|
| 1476 |
+
],
|
| 1477 |
+
"page_idx": 13
|
| 1478 |
+
},
|
| 1479 |
+
{
|
| 1480 |
+
"type": "text",
|
| 1481 |
+
"text": "This paper introduces a novel learning-based motion interpolation method designed to enable cross-skeleton compatibility with existing motion datasets. Our proposed method PC-MRL employs a process of point cloud obfuscation and skeleton reconstruction. The point cloud space represents human motions in a non-hierarchical and skeleton-agnostic manner. It enables a KNN-based objective to be optimised without dataset supervision, guiding a neural network to generate high-quality motion features with any target skeleton. We address the rotational information loss of our point cloud format by presenting an offset quaternion strategy compatible with concurrent transformer-based models. Through extensive experiments, we have demonstrated the efficacy of PC-MRL in performing motion interpolation without relying on native motion data. Moreover, PC-MRL has achieved superior visual similarity metrics in the domain of motion re-targeting. We concluded our work by discussing the limitations of PC-MRL.",
|
| 1482 |
+
"bbox": [
|
| 1483 |
+
212,
|
| 1484 |
+
643,
|
| 1485 |
+
787,
|
| 1486 |
+
840
|
| 1487 |
+
],
|
| 1488 |
+
"page_idx": 13
|
| 1489 |
+
},
|
| 1490 |
+
{
|
| 1491 |
+
"type": "page_number",
|
| 1492 |
+
"text": "14",
|
| 1493 |
+
"bbox": [
|
| 1494 |
+
217,
|
| 1495 |
+
114,
|
| 1496 |
+
235,
|
| 1497 |
+
126
|
| 1498 |
+
],
|
| 1499 |
+
"page_idx": 13
|
| 1500 |
+
},
|
| 1501 |
+
{
|
| 1502 |
+
"type": "header",
|
| 1503 |
+
"text": "C. Mo et al.",
|
| 1504 |
+
"bbox": [
|
| 1505 |
+
271,
|
| 1506 |
+
114,
|
| 1507 |
+
354,
|
| 1508 |
+
126
|
| 1509 |
+
],
|
| 1510 |
+
"page_idx": 13
|
| 1511 |
+
},
|
| 1512 |
+
{
|
| 1513 |
+
"type": "text",
|
| 1514 |
+
"text": "References",
|
| 1515 |
+
"text_level": 1,
|
| 1516 |
+
"bbox": [
|
| 1517 |
+
215,
|
| 1518 |
+
143,
|
| 1519 |
+
321,
|
| 1520 |
+
159
|
| 1521 |
+
],
|
| 1522 |
+
"page_idx": 14
|
| 1523 |
+
},
|
| 1524 |
+
{
|
| 1525 |
+
"type": "list",
|
| 1526 |
+
"sub_type": "ref_text",
|
| 1527 |
+
"list_items": [
|
| 1528 |
+
"1. Aberman, K., Li, P., Lischinski, D., Sorkine-Hornung, O., Cohen-Or, D., Chen, B.: Skeleton-aware networks for deep motion retargeting 39(4), 62–1 (2020) 4, 12",
|
| 1529 |
+
"2. Aberman, K., Wu, R., Lischinski, D., Chen, B., Cohen-Or, D.: Learning character-agnostic motion for motion retargeting in 2d 38(4), 1-14 (jul 2019) 4",
|
| 1530 |
+
"3. Annabi, L., Ma, Z., et al.: Unsupervised motion retargeting for human-robot imitation. In: Companion of the 2024 ACM/IEEE International Conference on Human-Robot Interactio. ACM (2024) 4",
|
| 1531 |
+
"4. Bodenheimer, B., Rose, C., Rosenthal, S., Pella, J.: The process of motion capture: Dealing with the data. In: Computer Animation and Simulation. pp. 3-18. Springer (1997) 4",
|
| 1532 |
+
"5. Burtnyk, N., Wein, M.: Computer-generated key-frame animation. Journal of the SMPTE 80(3), 149-153 (1971) 1, 4",
|
| 1533 |
+
"6. Catalin Ionescu, Fuxin Li, C.S.: Latent structured models for human pose estimation. In: International Conference on Computer Vision (2011) 10",
|
| 1534 |
+
"7. Chen, H., Tang, H., Timofte, R., Gool, L.V., Zhao, G.: Lart: Neural correspondence learning with latent regularization transformer for 3d motion transfer. Advances in Neural Information Processing Systems 36 (2024) 4",
|
| 1535 |
+
"8. Chen, J., Li, C., Lee, G.H.: Weakly-supervised 3D pose transfer with keypoints. In: ICCV. pp. 15156-15165 (2023) 4",
|
| 1536 |
+
"9. Choi, K.J., Ko, H.S.: Online motion retargetting. The Journal of Visualization and Computer Animation 11(5), 223-235 (2000) 5",
|
| 1537 |
+
"0. CMU: Carnegie-mellon university motion capture database (2003), http://mocap.cs.cmu.edu/ 10",
|
| 1538 |
+
"1. Dam, E.B., Koch, M., Lillholm, M.: Quaternions, interpolation and animation, vol. 2. CiteSeer (1998) 5, 6",
|
| 1539 |
+
"2. Duan, Y., Lin, Y., Zou, Z., Yuan, Y., Qian, Z., Zhang, B.: A unified framework for real time motion completion. In: AAAI. vol. 36, pp. 4459-4467 (2022) 2, 4, 10, 11",
|
| 1540 |
+
"3. Frank, T., Johnston, O.: Disney Animation: The Illusion of Life. Abbeville Publishing Group, New York (1981) 1",
|
| 1541 |
+
"4. Gleicher, M.: Retargetting motion to new characters. In: Computer Graphics and Interactive Techniques. pp. 33-42 (1998) 4",
|
| 1542 |
+
"5. Gleicher, M.: Animation from observation: Motion capture and motion editing 33(4), 51-54 (1999) 4",
|
| 1543 |
+
"6. Gopalakrishnan, A., Mali, A., Kifer, D., Giles, L., Ororbia, A.G.: A neural temporal model for human motion prediction. In: CVPR. pp. 12116-12125 (2019) 10",
|
| 1544 |
+
"7. Harvey, F.G., Pal, C.: Recurrent transition networks for character locomotion. In: SIGGRAPH Asia 2018 Technical Briefs, pp. 1-4 (2018) 4",
|
| 1545 |
+
"8. Harvey, F.G., Yurick, M., Nowrouzezahrai, D., Pal, C.: Robust motion in betweening 39(4), 60-1 (2020) 2, 4, 8, 9, 10, 11",
|
| 1546 |
+
"9. He, C., Saito, J., Zachary, J., Rushmeier, H., Zhou, Y.: Nemf: Neural motion fields for kinematic animation 35, 4244-4256 (2022) 4",
|
| 1547 |
+
"20. Hecker, C., Raabe, B., Enslow, R.W., DeWeese, J., Maynard, J., van Prooijen, K.: Real-time motion retargeting to highly varied user-created morphologies 27(3), 1-11 (2008) 5",
|
| 1548 |
+
"21. Hernandez, A., Gall, J., Moreno-Noguer, F.: Human motion prediction via spatiotemporal inpainting. In: ICCV. pp. 7134-7143 (2019) 4",
|
| 1549 |
+
"22. Ionescu, C., Papava, D., Olaru, V., Sminchisescu, C.: Human3.6m: Large scale datasets and predictive methods for 3D human sensing in natural environments."
|
| 1550 |
+
],
|
| 1551 |
+
"bbox": [
|
| 1552 |
+
225,
|
| 1553 |
+
176,
|
| 1554 |
+
784,
|
| 1555 |
+
839
|
| 1556 |
+
],
|
| 1557 |
+
"page_idx": 14
|
| 1558 |
+
},
|
| 1559 |
+
{
|
| 1560 |
+
"type": "header",
|
| 1561 |
+
"text": "Motion Keyframe Interpolation for Any Human Skeleton",
|
| 1562 |
+
"bbox": [
|
| 1563 |
+
351,
|
| 1564 |
+
114,
|
| 1565 |
+
730,
|
| 1566 |
+
128
|
| 1567 |
+
],
|
| 1568 |
+
"page_idx": 14
|
| 1569 |
+
},
|
| 1570 |
+
{
|
| 1571 |
+
"type": "page_number",
|
| 1572 |
+
"text": "15",
|
| 1573 |
+
"bbox": [
|
| 1574 |
+
767,
|
| 1575 |
+
116,
|
| 1576 |
+
784,
|
| 1577 |
+
126
|
| 1578 |
+
],
|
| 1579 |
+
"page_idx": 14
|
| 1580 |
+
},
|
| 1581 |
+
{
|
| 1582 |
+
"type": "list",
|
| 1583 |
+
"sub_type": "ref_text",
|
| 1584 |
+
"list_items": [
|
| 1585 |
+
"IEEE Transactions on Pattern Analysis and Machine Intelligence 36(7), 1325-1339 (jul 2014) 10",
|
| 1586 |
+
"23. Jiang, H., Cai, J., Zheng, J.: Skeleton-aware 3D human shape reconstruction from point clouds. In: ICCV. pp. 5431-5441 (2019) 4",
|
| 1587 |
+
"24. Kantor, I.L., Solodovnikov, A.S., Shenitzer, A.: Hypercomplex numbers: an elementary introduction to algebras, vol. 302. Springer (1989) 6",
|
| 1588 |
+
"25. Karunratanakul, K., Preechakul, K., Suwajanakorn, S., Tang, S.: Guided motion diffusion for controllable human motion synthesis. In: ICCV. pp. 2151-2162 (2023) 4",
|
| 1589 |
+
"26. Kulpa, R., Multon, F., Arnaldi, B.: Morphology-independent representation of motions for interactive human-like animation. In: Eurographics (2005) 5",
|
| 1590 |
+
"27. Lehrmann, A.M., Gehler, P.V., Nowozin, S.: Efficient nonlinear markov models for human motion. In: CVPR. pp. 1314-1321 (2014) 4",
|
| 1591 |
+
"28. Li, P., Aberman, K., Zhang, Z., Hanocka, R., Sorkine-Hornung, O.: Ganimator: Neural motion synthesis from a single sequence 41(4), 1-12 (2022) 4",
|
| 1592 |
+
"29. Li, W., Wang, Z., Mai, R., Ren, P., Zhang, Q., Zhou, Y., Xu, N., Zhuang, J., Xin, B., Gao, L., et al.: Modular design automation of the morphologies, controllers, and vision systems for intelligent robots: a survey. Visual Intelligence 1(1), 2 (2023) 4",
|
| 1593 |
+
"30. Li, Z., Peng, X.B., Abbeel, P., Levine, S., Berseth, G., Sreenath, K.: Reinforcement learning for versatile, dynamic, and robust bipedal locomotion control. arXiv preprint arXiv:2401.16889 (2024) 4",
|
| 1594 |
+
"31. Liao, Z., Yang, J., Saito, J., Pons-Moll, G., Zhou, Y.: Skeleton-free pose transfer for stylized 3D characters. In: ECCV. pp. 640-656. Springer (2022) 4",
|
| 1595 |
+
"32. Makoviychuk, V., Wawrzyniak, L., Guo, Y., Lu, M., Storey, K., Macklin, M., Hoeller, D., Rudin, N., Allshire, A., Handa, A., et al.: Isaac gym: High performancegpu-based physics simulation for robot learning. arXiv preprint arXiv:2108.10470 (2021) 4",
|
| 1596 |
+
"33. Merry, B., Marais, P., Gain, J.: Animation space: A truly linear framework for character animation. ACM Transactions on Graphics (TOG) 25(4), 1400-1423 (2006) 5",
|
| 1597 |
+
"34. Mo, C.A., Hu, K., Long, C., Wang, Z.: Continuous intermediate token learning with implicit motion manifold for keyframe based motion interpolation. In: CVPR. pp. 13894-13903 (2023) 2, 4, 8, 10, 11",
|
| 1598 |
+
"35. Mo, C.A., Hu, K., Mei, S., Chen, Z., Wang, Z.: Keyframe extraction from motion capture sequences with graph based deep reinforcement learning. In: ACM International Conference on Multimedia. pp. 5194-5202 (2021) 4",
|
| 1599 |
+
"36. Monzani, J.S., Baerlocher, P., Boulic, R., Thalmann, D.: Using an intermediate skeleton and inverse kinematics for motion retargeting. In: Computer Graphics Forum. vol. 19, pp. 11-19. Wiley Online Library (2000) 5",
|
| 1600 |
+
"37. Oreshkin, B.N., Valkanas, A., Harvey, F.G., Menard, L.S., Bocquelet, F., Coates, M.J.: Motion inbetweening via deep $\\delta$ -interpolator (2022) 2, 4",
|
| 1601 |
+
"38. Peng, X.B., Abbeel, P., Levine, S., van de Panne, M.: Deepmimic: Example-guided deep reinforcement learning of physics-based character skills 37(4), 143:1-143:14 (Jul 2018) 4",
|
| 1602 |
+
"39. Peng, X.B., Ma, Z., Abbeel, P., Levine, S., Kanazawa, A.: Amp: Adversarial motion priors for stylized physics-based character control 40(4) (Jul 2021) 4",
|
| 1603 |
+
"40. Qin, J., Zheng, Y., Zhou, K.: Motion in-between via two-stage transformers 41(6), 1-16 (2022) 2",
|
| 1604 |
+
"41. Reda, D., Won, J., Ye, Y., van de Panne, M., Winkler, A.: Physics-based motion retargeting from sparse inputs. ACM Computer Graphics and Interactive Techniques 6(3), 1-19 (2023) 4"
|
| 1605 |
+
],
|
| 1606 |
+
"bbox": [
|
| 1607 |
+
215,
|
| 1608 |
+
146,
|
| 1609 |
+
785,
|
| 1610 |
+
839
|
| 1611 |
+
],
|
| 1612 |
+
"page_idx": 15
|
| 1613 |
+
},
|
| 1614 |
+
{
|
| 1615 |
+
"type": "page_number",
|
| 1616 |
+
"text": "16",
|
| 1617 |
+
"bbox": [
|
| 1618 |
+
217,
|
| 1619 |
+
114,
|
| 1620 |
+
235,
|
| 1621 |
+
126
|
| 1622 |
+
],
|
| 1623 |
+
"page_idx": 15
|
| 1624 |
+
},
|
| 1625 |
+
{
|
| 1626 |
+
"type": "header",
|
| 1627 |
+
"text": "C. Mo et al.",
|
| 1628 |
+
"bbox": [
|
| 1629 |
+
271,
|
| 1630 |
+
114,
|
| 1631 |
+
354,
|
| 1632 |
+
126
|
| 1633 |
+
],
|
| 1634 |
+
"page_idx": 15
|
| 1635 |
+
},
|
| 1636 |
+
{
|
| 1637 |
+
"type": "list",
|
| 1638 |
+
"sub_type": "ref_text",
|
| 1639 |
+
"list_items": [
|
| 1640 |
+
"42. Reeves, W.T.: Inbetweening for computer animation utilizing moving point constraints 15(3), 263-269 (1981) 1",
|
| 1641 |
+
"43. Ren, T., Yu, J., Guo, S., Ma, Y., Ouyang, Y., Zeng, Z., Zhang, Y., Qin, Y.: Diverse motion in-betweening from sparse keyframes with dual posture stitching. IEEE Transactions on Visualization and Computer Graphics (2024) 4",
|
| 1642 |
+
"44. Shaw, P., Uszkoreit, J., Vaswani, A.: Self-attention with relative position representations. In: NAACL. Association for Computational Linguistics (2018) 9",
|
| 1643 |
+
"45. Tiwari, G., Antic, D., Lenssen, J.E., Sarafianos, N., Tung, T., Pons-Moll, G.: Posendf: Modeling human pose manifolds with neural distance fields. In: ECCV. pp. 572-589. Springer (2022) 4",
|
| 1644 |
+
"46. Villegas, R., Yang, J., Ceylan, D., Lee, H.: Neural kinematic networks for unsupervised motion retargeting. In: CVPR. pp. 8639-8648 (2018) 4",
|
| 1645 |
+
"47. Wang, J., Li, X., Liu, S., De Mello, S., Gallo, O., Wang, X., Kautz, J.: Zero-shot pose transfer for unrigged stylized 3D characters. In: CVPR. pp. 8704-8714 (2023) 4",
|
| 1646 |
+
"48. Wang, K., Xie, J., Zhang, G., Liu, L., Yang, J.: Sequential 3D human pose and shape estimation from point clouds. In: CVPR. pp. 7275-7284 (2020) 4",
|
| 1647 |
+
"49. Yuan, Y., Song, J., Iqbal, U., Vahdat, A., Kautz, J.: Physdiff: Physics-guided human motion diffusion model. In: ICCV. pp. 16010-16021 (2023) 4",
|
| 1648 |
+
"50. Yuan, Y., Wei, S.E., Simon, T., Kitani, K., Saragih, J.: Simpoe: Simulated character control for 3D human pose estimation. In: CVPR. pp. 7159-7169 (2021) 4",
|
| 1649 |
+
"51. Zhang, J., Weng, J., Kang, D., Zhao, F., Huang, S., Zhe, X., Bao, L., Shan, Y., Wang, J., Tu, Z.: Skinned motion retargeting with residual perception of motion semantics & geometry. In: CVPR. pp. 13864-13872 (2023) 4",
|
| 1650 |
+
"52. Zhang, M., Li, H., Cai, Z., Ren, J., Yang, L., Liu, Z.: Finemogen: Fine-grained spatio-temporal motion generation and editing. Advances in Neural Information Processing Systems 36 (2024) 4",
|
| 1651 |
+
"53. Zhang, X., van de Panne, M.: Data-driven autocompletion for keyframe animation. In: ACM SIGGRAPH Conference on Motion, Interaction and Games. pp. 1-11 (2018) 4",
|
| 1652 |
+
"54. Zhao, H., Jiang, L., Jia, J., Torr, P.H., Koltun, V.: Point transformer. In: ICCV. pp. 16259-16268 (2021) 7, 10",
|
| 1653 |
+
"55. Zhu, W., Yang, Z., Di, Z., Wu, W., Wang, Y., Loy, C.C.: Mocanet: Motion retargeting in-the-wild via canonicalization networks. In: AAAI. vol. 36, pp. 3617-3625 (2022) 4"
|
| 1654 |
+
],
|
| 1655 |
+
"bbox": [
|
| 1656 |
+
215,
|
| 1657 |
+
147,
|
| 1658 |
+
784,
|
| 1659 |
+
631
|
| 1660 |
+
],
|
| 1661 |
+
"page_idx": 16
|
| 1662 |
+
},
|
| 1663 |
+
{
|
| 1664 |
+
"type": "header",
|
| 1665 |
+
"text": "Motion Keyframe Interpolation for Any Human Skeleton",
|
| 1666 |
+
"bbox": [
|
| 1667 |
+
351,
|
| 1668 |
+
114,
|
| 1669 |
+
730,
|
| 1670 |
+
128
|
| 1671 |
+
],
|
| 1672 |
+
"page_idx": 16
|
| 1673 |
+
},
|
| 1674 |
+
{
|
| 1675 |
+
"type": "page_number",
|
| 1676 |
+
"text": "17",
|
| 1677 |
+
"bbox": [
|
| 1678 |
+
767,
|
| 1679 |
+
116,
|
| 1680 |
+
784,
|
| 1681 |
+
126
|
| 1682 |
+
],
|
| 1683 |
+
"page_idx": 16
|
| 1684 |
+
}
|
| 1685 |
+
]
|
2024/Motion Keyframe Interpolation for Any Human Skeleton using Point Cloud-based Human Motion Data Homogenisation/391ac41a-5023-4e3a-b209-fdaecfa62b6a_model.json
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
2024/Motion Keyframe Interpolation for Any Human Skeleton using Point Cloud-based Human Motion Data Homogenisation/391ac41a-5023-4e3a-b209-fdaecfa62b6a_origin.pdf
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:40ab048744f6ef182f5ad14bedba5bcaaac72728af59e52c5d9bb5d176049b72
|
| 3 |
+
size 11324956
|
2024/Motion Keyframe Interpolation for Any Human Skeleton using Point Cloud-based Human Motion Data Homogenisation/full.md
ADDED
|
@@ -0,0 +1,303 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Motion Keyframe Interpolation for Any Human Skeleton via Temporally Consistent Point Cloud Sampling and Reconstruction
|
| 2 |
+
|
| 3 |
+
Clinton Mo $^{1}$ , Kun Hu $^{1,\star}$ , Chengjiang Long $^{2}$ , Dong Yuan $^{1}$ , and Zhiyong Wang $^{1}$
|
| 4 |
+
|
| 5 |
+
<sup>1</sup> The University of Sydney, Darlington NSW 2008, Australia
|
| 6 |
+
clmo6615@uni.sydney.edu.au, {kun.hu, zhiyong.wang, dong.yuan}@sydney.edu.au
|
| 7 |
+
|
| 8 |
+
$^{2}$ Meta Reality Labs, Burlingame, CA, USA
|
| 9 |
+
|
| 10 |
+
clong1@meta.com
|
| 11 |
+
|
| 12 |
+
Abstract. In the character animation field, modern supervised keyframe interpolation models have demonstrated exceptional performance in constructing natural human motions from sparse pose definitions. As supervised models, large motion datasets are necessary to facilitate the learning process; however, since motion is represented with fixed hierarchical skeletons, such datasets are incompatible for skeletons outside the datasets' native configurations. Consequently, the expected availability of a motion dataset for desired skeletons severely hinders the feasibility of learned interpolation in practice. To combat this limitation, we propose Point Cloud-based Motion Representation Learning (PC-MRL), an unsupervised approach to enabling cross-compatibility between skeletons for motion interpolation learning. PC-MRL consists of a skeleton obfuscation strategy using temporal point cloud sampling, and an unsupervised skeleton reconstruction method from point clouds. We devise a temporal point-wise K-nearest neighbors loss for unsupervised learning. Moreover, we propose First-frame Offset Quaternion (FOQ) and Rest Pose Augmentation (RPA) strategies to overcome necessary limitations of our unsupervised point cloud-to-skeletal motion process. Comprehensive experiments demonstrate the effectiveness of PC-MRL in motion interpolation for desired skeletons without supervision from native datasets.
|
| 13 |
+
|
| 14 |
+
Keywords: 3D point clouds $\cdot$ Human body motion $\cdot$ Dataset creation
|
| 15 |
+
|
| 16 |
+
# 1 Introduction
|
| 17 |
+
|
| 18 |
+
3D character animation workflows largely rely on the concepts of keyframing with interpolation, often referred to as the pose-to-pose principle [13]. By defining key poses at correct timings, algorithms can be employed to generate intermediate poses, and thereby eliminating the need for defining each frame individually [5, 42]. Although this keyframe-based workflow is less costly than frame-by-frame production, human motion is often complex, and natural depictions
|
| 19 |
+
|
| 20 |
+

|
| 21 |
+
(a) Linear interpolation
|
| 22 |
+
|
| 23 |
+

|
| 24 |
+
(b) PC-MRL (Our method)
|
| 25 |
+
|
| 26 |
+

|
| 27 |
+
(c) Ground truth
|
| 28 |
+
|
| 29 |
+

|
| 30 |
+
(a) Hierarchical skeleton
|
| 31 |
+
|
| 32 |
+

|
| 33 |
+
Fig. 1: Motion interpolation of walking keyframes conducted by our approach.
|
| 34 |
+
(b) Skeletal motion
|
| 35 |
+
|
| 36 |
+

|
| 37 |
+
(c) Point cloud representation
|
| 38 |
+
Fig. 2: Our method leverages point cloud representations (c) to obfuscate the hierarchical dependencies of skeletal motion. Joint colours correspond to ordinal indices in original motion data representation. Point clouds are coloured by temporal index.
|
| 39 |
+
|
| 40 |
+
demand a significant number of keyframes, especially for realistic motions that adhere to physical properties and constraints. Although motion capture (MoCap) is frequently employed as an alternative for achieving realistic human motions, MoCap often nonetheless necessitates subsequent keyframing to correct unwanted motion elements and address other imperfections.
|
| 41 |
+
|
| 42 |
+
In recent years, the interpolation process has been widely studied, particularly with the advent of machine learning methods for sequential data. Compared to conventional methods such as linear interpolation (LERP), machine learning methods have demonstrated the capability to derive visually natural motions from sparser keyframe sets [12, 18, 34, 37, 40], which we can observe in Fig. 1. Data-driven interpolation approaches require large motion capture datasets to learn the necessary motion features for effective synthesis of transitions between keyframes. In the field of 3D motions, this aspect significantly restricts the feasibility of learned interpolation methods in animation practice, as motion/pose data are tied to specific skeletal configurations and are not cross-compatible. That is, the dataset's native skeleton, i.e. source skeleton, is structurally different and incompatible with the desired skeleton, i.e. target skeleton.
|
| 43 |
+
|
| 44 |
+
Therefore, we propose a novel point cloud-based motion representation learning (PC-MRL) approach, in which skeletal hierarchies are obfuscated by sampling into the point cloud medium, and reconstructed into skeletal motion data using an unsupervised neural network. Unlike raw pose and motion data, the
|
| 45 |
+
|
| 46 |
+
point cloud format is a non-hierarchical representation, and effectively obscures any skeletal configuration as displayed in Fig. 2. PC-MRL samples a set of points around the rest position of skeletons using a combination of uniform and normal distributions, and geometrically parents each point to an associated bone in order to effectively represent the associated skeletons' motion data in a temporally consistent manner. Subsequently, for the reconstruction of motion data into a given desired skeleton, we propose an unsupervised learning scheme, featuring a K-nearest neighbors (KNN) loss between cross-skeleton point clouds to optimize a transformer neural network. The network processes the point cloud-based motion representations and generates the corresponding target skeletal motion representations. By obscuring both source and predicted target motion data into the point cloud space, our strategy allows for a direct comparison, utilizing KNN to minimize the visual difference between the two point clouds.
|
| 47 |
+
|
| 48 |
+
The unsupervised nature of our target skeleton reconstruction scheme creates two main disparities between the expected and predicted motion features. Firstly, the exact roll rotation axis of bones cannot be represented in the point cloud format. To this end, we introduce First-frame Offset Quaternion (FOQ) representations to incorporate relative roll values, which are obtainable from temporally consistent point clouds. This approach is agnostic to absolute roll rotations, providing a standardisation mechanism for the resulting motion data within the context of motion interpolation learning. Second, KNN-based objectives learn skeletal representations in a geometrically optimal manner, which does not always reflect the desired skeletal behaviour. This disparity is commonly observed with smaller skeletal features, such as shoulder and hip bones. As such, we augment training motions using Rest Pose Augmentation (RPA) to increase the range of skeletal behaviours that our interpolation model learns. Extensive experiments demonstrate the effectiveness of our proposed method, achieving performance targets near the level of direct dataset supervision.
|
| 49 |
+
|
| 50 |
+
In summary, this paper presents the following key contributions:
|
| 51 |
+
|
| 52 |
+
1. We propose a novel method for achieving unsupervised human motion reconstruction from point cloud data, enabling skeleton-agnostic motion interpolation learning for the first time.
|
| 53 |
+
2. To train motion interpolation models, we formulate first-frame offset quaternions to represent bone rotations with relative roll data, as well as a rest position augmentation strategy to address skeletal configuration variability.
|
| 54 |
+
3. We perform comprehensive experiments to demonstrate our method's effectiveness towards learning motion interpolation, despite the absence of directly compatible datasets in the training process.
|
| 55 |
+
|
| 56 |
+
# 2 Related works
|
| 57 |
+
|
| 58 |
+
We explore existing research on motion interpolation approaches, and motion data modelling methods in general. In addition, our proposed approach bears resemblance to unsupervised approaches for the motion re-targeting research problem, and as such, we will also analyse existing methods in this field.
|
| 59 |
+
|
| 60 |
+
# 2.1 Data-Driven Motion Modelling Methods
|
| 61 |
+
|
| 62 |
+
Contemporary research in motion modelling predominantly explore learned data-driven methods, using neural networks to capture, model, and establish correlations between the motion features of various skeletal configurations [29, 35, 41, 51, 55]. Motion features alone have demonstrated considerable efficacy in cross-skeleton imitation using a variety of strategies, including task-driven objectives [19, 45], diffusion-based generation [25, 49, 52], and latent feature consistency [1-3, 7, 46]. Latent consistency in particular learns inter-skeleton correlations without the need for paired datasets, which reduces its barrier of entry.
|
| 63 |
+
|
| 64 |
+
Likewise, motion interpolation strategies have evolved through the use of learning based approaches. The inherent numerical precision of motion data has traditionally posed a challenge for the adoption of the learning approaches, due to their approximate nature [21,27]. This challenge is further amplified by the temporally sparse distribution of keyframes. Recently, a number of recurrent neural network based methods [17, 18, 43, 53] and transformer-based networks [12, 34, 37] have demonstrated encouraging interpolation performance for the transition between keyframes. Additionally, various other motion modelling methods are capable for keyframe based interpolation, albeit with limited intermediate frame representation capabilities [19, 25, 45].
|
| 65 |
+
|
| 66 |
+
A major constraint inherent in all data-driven motion modelling methods is their dependency on datasets featuring specific skeletal configurations. To address this, skeleton-free training strategies have been explored for elementary tasks involving 3D characters. These include vertex-based pose transfer [7,8,31,47], and point cloud-based human shape reconstruction [23,48]. Notably, in these methods, 3D mesh and point cloud coordinates serve as general 3D data representations. Concurrently, other research has significantly minimized the required volume of training data, bringing it down to single example sequences. This reduction has been achieved through the use of patch matching techniques [28] and imitation learning within physics-based simulations [30,39]. Reinforcement learning has facilitated the generation of goal-driven motion without pre-existing datasets [32,38,50]. However, these simulation-based methods have typically been either exceedingly complex to implement in practice, or yield results that are too inflexible and/or prone to errors for animation workflows.
|
| 67 |
+
|
| 68 |
+
# 2.2 Motion Re-Targeting without Supervision
|
| 69 |
+
|
| 70 |
+
Conventional approaches in motion re-targeting have been a cornerstone of 3D animation for decades, primarily dedicated to converting motion capture data into 3D skeletal motions [4, 15]. Typically, raw motion capture data is optically recorded, yielding primarily spatial information. In contrast, motion data is largely rotational in nature [5], often represented in $SO(3)$ space. Therefore, a principal objective of motion re-targeting solutions has been to effectively bridge these two distinct types of representations. In [14], extensive manually adjustable constraints were introduced for flexible re-targeting between skeletons of identical topology (i.e., same hierarchy, different bone lengths and rest poses). Key
|
| 71 |
+
|
| 72 |
+

|
| 73 |
+
Fig. 3: Illustration of point cloud-based motion representation learning (PC-MRL) pipeline for keyframe interpolation to any human skeleton. Our unsupervised point cloud-to-pose decoding approach (I) first adapts motion features from existing datasets to new skeletons. We then employ the adapted motion data to supervise motion interpolation models (II), using FOQ and rest pose augmentations to address remaining discrepancies between the adapted and expected target skeleton motion patterns.
|
| 74 |
+
|
| 75 |
+
practices in this approach, including end-effector matching, joint contacts, and inter-skeleton point correlations, are still relevant to current animation pipelines. Point correlations using Inverse Kinematics (IK) have also been proposed as an intermediate representation to facilitate the adaptation between topologically distinct skeletons [9, 36], albeit with hierarchical pairing limitations.
|
| 76 |
+
|
| 77 |
+
In response, morphologically-independent motion control systems were proposed using IK joint chains and joint-wise constraints [20,26]. Such systems often expect highly standardised motion skeleton features, and are generally impractical for re-targeting motions produced for different control schemes. Additionally, the reliance on Euler angles for extensive classical re-targeting constraints poses a variety of issues regarding rotational freedom [11], as well as compatibility with animation workflows [33], which predominantly use quaternion-based systems.
|
| 78 |
+
|
| 79 |
+
Despite these advancements, the challenge of arbitrary skeleton re-targeting from motion datasets still remains largely unresolved, limiting any widespread application of data-driven pipelines. Our paper aims to effectively address this pervasive limitation.
|
| 80 |
+
|
| 81 |
+
# 3 Methodology
|
| 82 |
+
|
| 83 |
+
As illustrated in Fig. 3, our proposed method consists of two main components: point cloud-to-motion reconstruction and motion keyframe interpolation.
|
| 84 |
+
|
| 85 |
+
# 3.1 Quaternion-based Motion Representations
|
| 86 |
+
|
| 87 |
+
We first declare a set of notations for representing skeletons. A skeleton $S$ is defined by a tree of bones $\mathcal{B}$ . A bone $\mathbf{v}_n \in \mathcal{B}$ is defined by its bone parent $\mathbf{v}_m \in \mathcal{B}$ , and a head position $\mathbf{H}_n \in \mathbb{R}^3$ which denotes the bone's base offset from
|
| 88 |
+
|
| 89 |
+
its parent head $\mathbf{H}_m$ . In addition, a tail position $\mathbf{T}_n \in \mathbb{R}^3$ helps to define the end point of the bone's line segment representation, starting from its head position, of $\mathbf{v}_n$ . The root bone $\mathbf{v}_0$ has no parent, and thus no head position $\mathbf{H}_0$ . The set of $\mathbf{H}_n$ and $\mathbf{T}_n$ for each bone $\mathbf{v}_n \in \mathcal{B}$ make up the rest position of the skeleton.
|
| 90 |
+
|
| 91 |
+
Next, we define the properties of motion data. At any given frame $t$ , a skeleton's pose position can be represented as global 3D positions $\mathbf{p}_{n,t} \in \mathbb{R}^3$ and global quaternion rotations $\mathbf{q}_{n,t} \in \mathbb{R}^4$ for each bone $\mathbf{v}_n \in \mathcal{B}$ . For a given motion sequence $M$ , only the root bone is given a 3D position in pose position, i.e. $\mathbf{p}_{0,t}$ as all other pose positions are derived using Forwards Kinematics (FK [11,24]):
|
| 92 |
+
|
| 93 |
+
$$
|
| 94 |
+
\mathbf {p} _ {n, t} = \mathbf {p} _ {m, t} + Q R \left(\mathbf {H} _ {n}, \mathbf {q} _ {m, t}\right), \tag {1}
|
| 95 |
+
$$
|
| 96 |
+
|
| 97 |
+
where $QR$ is the quaternion rotation function, and $\mathbf{v}_m$ is the parent joint. In summary, a motion sequence $M$ of $|M|$ frames is defined by:
|
| 98 |
+
|
| 99 |
+
$$
|
| 100 |
+
M = \left\{\mathbf {p} _ {0, t} \cup \left\{\mathbf {q} _ {n, t} \mid \mathbf {v} _ {n} \in \mathcal {B} \right\} \mid 0 \leq t < | M | \right\} \tag {2}
|
| 101 |
+
$$
|
| 102 |
+
|
| 103 |
+
# 3.2 Point Cloud Obfuscation and Skeleton Reconstruction
|
| 104 |
+
|
| 105 |
+
Our approach is based on the hypothesis that 3D motion sequences can be visually represented by a more universal format: 3D point cloud sequences. By geometrically sampling the volume around a skeleton, point cloud data becomes a non-hierarchical representation of its pose(s), obfuscating the original skeletal configuration. Due to this decoupling between the skeleton configuration and motion representation, a successful motion data reconstruction from point cloud data would be the first step to enabling learnable cross-skeleton motion features. Formally, for a given source skeleton $S_A$ with an associated motion sequence $M_A$ , and any humanoid skeletal configuration $S_B$ , the point cloud obfuscation and reconstruction module is to learn a function that can adequately perform $F_{S_A \to S_B}(M_A) = \hat{M}_B$ , where $\hat{M}_B$ is a visually similar motion adaptation of $M_A$ on skeleton $S_B$ , which is an estimation of $M_B$ .
|
| 106 |
+
|
| 107 |
+
Motion Data Obfuscation with Point Clouds To sample the point cloud $\mathcal{X}$ , we first generate a set of points $u\in \mathcal{X}$ along the line segments defined by the global head and tail positions for each bone $\mathbf{v}_n$ . In the bone-local space to $\mathbf{v}_n$ , this would mean sampling $u$ by a uniformly distributed factor $\alpha \sim \mathcal{U}_{[0,1]}$ . Specifically, $\alpha = 0$ aligns with the head position of the bone, and $\alpha = 1$ indicates the tail position. By further sampling $u$ via a normal distribution, our sampling strategy can capture the surrounding volume around the skeleton. Formally, to sample $u$ from its associated bone $\mathbf{v}_n$ within a standard deviation of $\sigma$ , we have:
|
| 108 |
+
|
| 109 |
+
$$
|
| 110 |
+
\mathbf {H} _ {x} \sim \mathcal {N} (\alpha \mathbf {T} _ {n}, \sigma),
|
| 111 |
+
$$
|
| 112 |
+
|
| 113 |
+
$$
|
| 114 |
+
\mathbf {p} _ {u, t} = \mathbf {p} _ {n, t} + Q R (\mathbf {H} _ {x}, \mathbf {q} _ {n, t}) \tag {3}
|
| 115 |
+
$$
|
| 116 |
+
|
| 117 |
+
Notably, this sampling process as in Eq. 3 is associated with the skeletal configuration only, by which the sampled points are temporally consistent for further processing with the motion sequence. Specifically, each point maintains a
|
| 118 |
+
|
| 119 |
+
constant position from its associated bone in the bone-local space. The lateral distance of each point from their parent bone segment allows point clouds to react to the bone's rotational roll axis.
|
| 120 |
+
|
| 121 |
+
In a pure point cloud representation, it is possible to remove the relationships between bones. However, to clearly distinguish symmetrical body features in different skeletons, we introduce and embed body group associations for the points during our reconstruction process. Specifically, every bone is categorised by one of these body groups: [spine, left arm, right arm, left leg, right leg]. Each sampled point is assigned q group attribute in line with the associated bone. The attribute is represented by a one-hot feature vector $\mathbf{g}_u$ . To this end, we characterize $u = \{\mathbf{p}_{u,t}, \mathbf{g}_u\} \in \mathcal{X}$ .
|
| 122 |
+
|
| 123 |
+
Skeleton Reconstruction Given a sampled point cloud representation $\mathcal{X}_A$ from $S_A$ and $M_A$ as input, we aim to train a motion adaptation function $F_{S_A\to S_B}$ which is tasked with producing a visually similar skeletal motion $\hat{M}_B$ for $S_B$ using a temporal and set-based neural network. As shown in Fig. 3, to construct $F_{S_A\rightarrow S_B}$ , we employ a Point Transformer [54] to learn embeddings for the unordered point cloud sets frame-wisely, and a standard temporal transformer to process the sequence and decode $\mathcal{X}_A$ into $\hat{M}_B$ .
|
| 124 |
+
|
| 125 |
+
To optimize $F_{S_A \to S_B}$ , our objective function maximises the similarity between the sampled point clouds of both $M_A$ and $\hat{M}_B$ . $\mathcal{X}_A$ can be viewed as the ground truth to $\mathcal{X}_B$ , which enables an unsupervised strategy for $F_{S_A \to S_B}$ . In detail, we first sample a point cloud $\mathcal{X}_B$ from $S_B$ based on $\hat{M}_B$ derived by $F_{S_A \to S_B}$ . Next, a temporally consistent KNN-Loss objective is devised to minimise the $\ell_2$ distance between any given point of $\mathcal{X}_B$ from its $K$ nearest neighbouring points in $\mathcal{X}_A$ . The $K$ nearest neighbours are determined based on inter-point distance throughout the entire motion sequence, instead of a frame-wise setting, as shown in Fig. 4. The points with differing body groups are excluded from the neighbour search. Mathematically, we have:
|
| 126 |
+
|
| 127 |
+
$$
|
| 128 |
+
\delta \left(u _ {A}, u _ {B}\right) = \left\{ \begin{array}{l l} \sum_ {t} \left\| \mathbf {p} _ {u _ {A}, t} - \mathbf {p} _ {u _ {B}, t} \right\| _ {2}, & \text {i f} \mathbf {g} _ {u _ {A}} = \mathbf {g} _ {u _ {B}}, \\ \inf , & \text {o t h e r w i s e}, \end{array} \right. \tag {4}
|
| 129 |
+
$$
|
| 130 |
+
|
| 131 |
+
$$
|
| 132 |
+
\mathcal {L} _ {\mathrm {K N N}} \left(\mathcal {X} _ {A}, \mathcal {X} _ {B}\right) = \sum_ {u _ {A} \in \mathcal {X} _ {A}, u _ {B} \in N _ {k} \left(u _ {A}, \mathcal {X} _ {B}\right)} \delta \left(u _ {A}, u _ {B}\right), \tag {5}
|
| 133 |
+
$$
|
| 134 |
+
|
| 135 |
+
where $N_{k}(u_{A},\mathcal{X}_{B})$ is the set containing points that are the k-nearest neighbors of $u_{A}$ in $\mathcal{X}_B$ , based on $\delta (u_A,u_B)$ distance.
|
| 136 |
+
|
| 137 |
+
In addition to the KNN-Loss, we introduce an optional end-effector loss in the skeleton space for the hand, foot, and head bones, which provides direct positional guidance for matched end-effector pairs $E_{AB}$ :
|
| 138 |
+
|
| 139 |
+
$$
|
| 140 |
+
\mathcal {L} _ {\text {e n d}} = \sum_ {(\mathbf {v} _ {A}, \mathbf {v} _ {B}) \in E _ {A B}} \frac {1}{| M _ {A} |} \sum_ {t} \left\| \mathbf {p} _ {\mathbf {v} _ {A}, t} - \mathbf {p} _ {\mathbf {v} _ {B}, t} \right\| _ {2}. \tag {6}
|
| 141 |
+
$$
|
| 142 |
+
|
| 143 |
+

|
| 144 |
+
Fig. 4: Visualisation of our temporally consistent KNN objective, processed independently per body group. For each point from $\mathcal{X}_A$ (a), the total $\ell_2$ distance throughout the sequence is measured between each point (b) from $\mathcal{X}_B$ . The $\ell_2$ distances from the K-nearest points are summed to produce KNN-Loss (c).
|
| 145 |
+
|
| 146 |
+
# 3.3 First-Frame Offset Quaternions
|
| 147 |
+
|
| 148 |
+
Point clouds lack absolute roll axis information for each bone. As such, we introduce a First-frame Offset Quaternion (FOQ) transform to incorporate the available relative roll data for motion modelling. We formulate all quaternions of a motion sequence relative to their values from the initial frame. In detail, for any quaternion sequence $Q = \{\mathbf{q}_0, \mathbf{q}_1, \mathbf{q}_2, \ldots\}$ , we can simply obtain its FOQ transform as $\{\mathbf{q}_0 \times \mathbf{q}_0^{-1}, \mathbf{q}_1 \times \mathbf{q}_0^{-1}, \mathbf{q}_2 \times \mathbf{q}_0^{-1}, \ldots\}$ , where $\mathbf{q}_0^{-1}$ is the quaternion conjugate of $\mathbf{q}_0$ . An FOQ sequence can easily be converted back to a quaternion sequence as long as any $\mathbf{q}_t \in Q$ is known.
|
| 149 |
+
|
| 150 |
+
Crucially, FOQ is a roll-invariant rotation representation, that is, FOQ remains identical regardless of a bone's absolute roll position in the original data. Trivially, to adjust the rotational roll of a bone $\mathbf{v}_n$ , we perform the quaternion multiplication for each frame: $\mathbf{q}_{n,t} \times \mathbf{r}$ , where $\mathbf{r}$ is a roll quaternion. As a quaternion on the roll axis of $\mathbf{v}_n$ , $\mathbf{r}$ is constrained within $\mathbf{r} = \alpha + \beta (x\mathbf{i} + y\mathbf{j} + z\mathbf{k})$ , where $x,y,z$ are the 3D XYZ values of $\mathbf{T}_n$ , and $\alpha, \beta$ are scalars that control the roll magnitude. Note that $\mathbf{r}$ is constant throughout the sequence. Therefore, we can deduce the roll-invariant nature of any constant $\mathbf{r}$ :
|
| 151 |
+
|
| 152 |
+
$$
|
| 153 |
+
\begin{array}{l} \left(\mathbf {q} _ {n, t} \times \mathbf {r}\right) \times \left(\mathbf {q} _ {0, t} \times \mathbf {r}\right) ^ {- 1} = \mathbf {q} _ {n, t} \times \mathbf {r} \times \mathbf {r} ^ {- 1} \times \mathbf {q} _ {0, t} ^ {- 1} \\ = \mathbf {q} _ {n, t} \mathbf {q} _ {0, t} ^ {- 1} \\ \end{array}
|
| 154 |
+
$$
|
| 155 |
+
|
| 156 |
+
# 3.4 Motion Interpolation Transformer and Rest Pose Augmentation
|
| 157 |
+
|
| 158 |
+
To leverage our point cloud-based representation learning for cross-skeleton interpolation, a transformer model - CITL [34] is adapted, allowing for an efficient interpolation pipeline with existing datasets to alternative skeletons. We introduce two key modifications to CITL and its training strategy, focusing on enhancing motion quality. First, quaternion predictions are substituted with FOQ predictions. This is inspired by the RNN-based method TGcomplete [18], which predicts
|
| 159 |
+
|
| 160 |
+
Algorithm 1 IK target alignment with maximum real quaternion component
|
| 161 |
+
1: Input: $\mathbf{p_{IK}} = 3D$ IK target unit vector
|
| 162 |
+
2: $\mathbf{T} =$ resting tail position of associated bone
|
| 163 |
+
3: Output: $\mathbf{q}_{\mathrm{max}} =$ output quaternion, to be applied on original bone quaternion
|
| 164 |
+
4: $\eta \leftarrow \frac{\mathbf{T}}{|\mathbf{T}|}$ $\triangleright$ 3D unit coordinates $[\eta_x,\eta_y,\eta_z]$
|
| 165 |
+
5: $\omega \leftarrow \frac{\mathbf{p_{IK}}}{|\mathbf{p_{IK}}|}$ $\triangleright$ 3D unit coordinates $[\omega_x,\omega_y,\omega_z]$
|
| 166 |
+
6: $\mathbf{q}_{\mathrm{min}} \gets \frac{0 + (\eta_x + \omega_x)\mathbf{i} + (\eta_y + \omega_y)\mathbf{j} + (\eta_z + \omega_z)\mathbf{k}}{2}$
|
| 167 |
+
7: $\mathbf{q}_{\mathrm{max}} \gets \mathbf{q}_{\mathrm{min}} \times (0 + \eta_x\mathbf{i} + \eta_y\mathbf{j} + \eta_z\mathbf{k})$
|
| 168 |
+
|
| 169 |
+
quaternion offsets from the preceding frame rather than raw quaternions, thereby achieving greater generalisability and superior quality in interpolation.
|
| 170 |
+
|
| 171 |
+
For our model's CITL-based architecture, we substitute the original sinusoidal positional encoding scheme with learned relative positions [44]. While the original implementation relies on the continual nature of sinusoidal functions, we observe that relative positional embeddings with zero initialisation converges upon similarly continuous behaviour, and significantly lowers the complexity of positional attention. The lessened focus on positional relations allows the model to synthesise deeper behaviours within the pose space, improving overall generalisability. As a side effect, the model also accepts arbitrary length inputs beyond the training data length.
|
| 172 |
+
|
| 173 |
+
Notably, the intended rest pose of a skeletal configuration is often geometrically sub-optimal for point cloud matching, and can manifest in multiple forms. To address this, we devise a rest position augmentation strategy, leveraging a per-bone Inverse Kinematics (IK) system. RPA broadens the range of intended rest poses that the learned interpolation model can accommodate. In detail, during each phase of the Forward Kinematics (FK) process, we augment the bone tails with a random additional offset. We re-align the augmented global tail position as closely as possible to its original location using a quaternion rotation with the maximum possible real component. Algorithm 1 describes the method by which this quaternion can be derived. Since the real component describes, in cosine, the amount of rotation around an axis, this minimizes numerical disturbance from RPA while altering the visual representation to the desired extent.
|
| 174 |
+
|
| 175 |
+
# 4 Experimental Results
|
| 176 |
+
|
| 177 |
+
# 4.1 Datasets
|
| 178 |
+
|
| 179 |
+
To thoroughly evaluate the effectiveness of the proposed method, we conducted extensive experiments on the following widely used motion capture datasets:
|
| 180 |
+
|
| 181 |
+
- LaFAN1 [18] contains long, high quality motions in controllable video game character styles. The dataset was recorded on a large motion capture stage, and manually cleaned to production standards. The native skeleton of LaFAN1 contains 6 spinal bones, and 4 bones for each limb.
|
| 182 |
+
|
| 183 |
+
- Human3.6M [6,22] comprises of 3.6 million motion frames. The motions captured in Human3.6M tend to be less dynamic, as they were recorded on a relatively compact $4 \times 3 \mathrm{~m}$ stage. Its native skeleton includes 5 spinal bones, 4 bones for each limb, and an additional finger bone for each hand.
|
| 184 |
+
|
| 185 |
+
- CMU Mocap [10] is a diverse motion dataset encompassing 113 categories. It features a mix of static and dynamic motions, captured on a $3 \times 8 \mathrm{~m}$ stage. The skeleton is identical to that of Human3.6M, with the exception of the pelvis, which is divided into two hip bones and a lower back bone.
|
| 186 |
+
|
| 187 |
+
For PC-MRL, we designated the CMU MoCap skeleton as $S_B$ . Its native dataset is exclusively utilized for testing purposes, while LaFAN1 and Human3.6M are used simultaneously for training. $^3$
|
| 188 |
+
|
| 189 |
+
# 4.2 Implementation Details
|
| 190 |
+
|
| 191 |
+
In our experiments, point cloud obfuscation adopted a 256-point setting for the sampling process with $\sigma = 0.05$ and each point was characterized by a 64-element vector. To maintain consistency, we ensure that all bone offsets and motion positions are defined in metre units. For skeletal reconstruction, 4 Point Transformer [54] layers was utilized to construct the neural network. Within each layer, the feature size was doubled, and the size of the point set was downsampled by a factor of 4 through a farthest-first traversal approach, eventually resulting in a singular feature vector. In the final step, two linear layers with ReLU activations were utilized to transform the resulted feature vector in a frame-wise manner, producing the output skeletal representation. During training, we set $K = 8$ for the KNN loss to optimize the network. Additionally, the use of unit quaternion values in practice relies on modulus division for constraint, a process that can potentially lead to exploding gradients. To address this, we introduced an objective function $\mathcal{L}_q$ that measures the $\ell_2$ norms' difference between the raw quaternion output and the expected unit quaternion. All experiments were trained and evaluated on a single NVIDIA GeForce RTX 3090 GPU.
|
| 192 |
+
|
| 193 |
+
# 4.3 Motion Interpolation Comparison against Supervised Methods
|
| 194 |
+
|
| 195 |
+
To demonstrate the efficacy of our method, we produce and compare our method against the conventional LERP method, as well as three state-of-the-art interpolation models trained on the original CMU MoCap dataset. Specifically, we measure the performance of the RNN-based $\mathrm{TG}_{\mathrm{complete}}$ model [18], the BERT-based motion interpolation adaptation [12], and a transformer-based encoder-decoder approach [34]. We additionally provide results on a variant of our method trained without our RPA strategy. The experimental configuration for each state-of-the-art model is identical to their original implementations. To measure interpolation accuracy, we employ the standard $\ell_2$ positional distance (L2P), $\ell_2$ quaternion difference (L2Q), and NPSS [16] for visual similarity evaluation [18].
|
| 196 |
+
|
| 197 |
+
<table><tr><td rowspan="2">Motion category</td><td rowspan="2">Method</td><td colspan="3">L2P↓</td><td colspan="3">L2Q↓</td><td colspan="3">NPSS↓</td></tr><tr><td>5</td><td>15</td><td>30</td><td>5</td><td>15</td><td>30</td><td>5</td><td>15</td><td>30</td></tr><tr><td rowspan="6">Basketball</td><td>LERP</td><td>0.0787</td><td>0.4009</td><td>0.8033</td><td>0.1588</td><td>0.5834</td><td>0.9611</td><td>0.1315</td><td>0.6120</td><td>1.8202</td></tr><tr><td>TGcomplete [18]</td><td>0.1050</td><td>0.4419</td><td>0.7104</td><td>0.1859</td><td>0.6439</td><td>0.9303</td><td>0.1600</td><td>0.6579</td><td>1.2136</td></tr><tr><td>BERT [12]</td><td>0.0889</td><td>0.3737</td><td>0.6865</td><td>0.1585</td><td>0.5560</td><td>0.8992</td><td>0.1689</td><td>0.6177</td><td>1.4587</td></tr><tr><td>CITL [34]</td><td>0.0625</td><td>0.2973</td><td>0.5708</td><td>0.1410</td><td>0.4932</td><td>0.8011</td><td>0.1047</td><td>0.4098</td><td>1.0959</td></tr><tr><td>PC-MRL w/o RPA</td><td>0.0767</td><td>0.3534</td><td>0.6715</td><td>0.1505</td><td>0.5431</td><td>0.9049</td><td>0.1364</td><td>0.5596</td><td>1.3471</td></tr><tr><td>PC-MRL (Ours)</td><td>0.0752</td><td>0.3443</td><td>0.6662</td><td>0.1561</td><td>0.5495</td><td>0.9088</td><td>0.1397</td><td>0.5572</td><td>1.3015</td></tr><tr><td rowspan="6">Golf</td><td>LERP</td><td>0.0189</td><td>0.1229</td><td>0.3187</td><td>0.0455</td><td>0.1870</td><td>0.3987</td><td>0.0587</td><td>0.3979</td><td>1.0312</td></tr><tr><td>TGcomplete [18]</td><td>0.0473</td><td>0.2616</td><td>0.4546</td><td>0.0693</td><td>0.3191</td><td>0.5297</td><td>0.0954</td><td>0.5062</td><td>1.1534</td></tr><tr><td>BERT [12]</td><td>0.0276</td><td>0.1159</td><td>0.2967</td><td>0.0612</td><td>0.2082</td><td>0.4258</td><td>0.1099</td><td>0.3906</td><td>0.9255</td></tr><tr><td>CITL [34]</td><td>0.0201</td><td>0.1043</td><td>0.2589</td><td>0.0476</td><td>0.1766</td><td>0.3552</td><td>0.0636</td><td>0.2581</td><td>0.7799</td></tr><tr><td>PC-MRL w/out RPA</td><td>0.0274</td><td>0.1304</td><td>0.3622</td><td>0.0555</td><td>0.2065</td><td>0.4923</td><td>0.0911</td><td>0.3759</td><td>1.2083</td></tr><tr><td>PC-MRL (Ours)</td><td>0.0273</td><td>0.1130</td><td>0.2706</td><td>0.0598</td><td>0.2012</td><td>0.3844</td><td>0.1049</td><td>0.3716</td><td>0.8150</td></tr><tr><td rowspan="6">Swimming</td><td>LERP</td><td>0.0731</td><td>0.3680</td><td>0.7374</td><td>0.1326</td><td>0.5327</td><td>0.9476</td><td>0.2147</td><td>0.9222</td><td>1.6972</td></tr><tr><td>TGcomplete [18]</td><td>0.1292</td><td>0.5796</td><td>0.9141</td><td>0.1845</td><td>0.7420</td><td>1.1199</td><td>0.2660</td><td>0.9716</td><td>1.7698</td></tr><tr><td>BERT [12]</td><td>0.1088</td><td>0.3896</td><td>0.7516</td><td>0.1819</td><td>0.5518</td><td>0.9622</td><td>0.2877</td><td>0.9074</td><td>1.6682</td></tr><tr><td>CITL [34]</td><td>0.0905</td><td>0.3895</td><td>0.7475</td><td>0.1484</td><td>0.5535</td><td>0.9533</td><td>0.2084</td><td>0.8317</td><td>1.6122</td></tr><tr><td>PC-MRL w/o RPA</td><td>0.0722</td><td>0.3754</td><td>0.7459</td><td>0.1345</td><td>0.5442</td><td>0.9594</td><td>0.1944</td><td>0.8731</td><td>1.6608</td></tr><tr><td>PC-MRL (Ours)</td><td>0.0734</td><td>0.3595</td><td>0.7139</td><td>0.1355</td><td>0.5331</td><td>0.9248</td><td>0.1956</td><td>0.8591</td><td>1.6370</td></tr><tr><td rowspan="6">Walking & Running</td><td>LERP</td><td>0.0455</td><td>0.2505</td><td>0.5527</td><td>0.1120</td><td>0.3997</td><td>0.6233</td><td>0.0655</td><td>0.2682</td><td>0.9071</td></tr><tr><td>TGcomplete [18]</td><td>0.0540</td><td>0.1786</td><td>0.3103</td><td>0.1147</td><td>0.2866</td><td>0.3866</td><td>0.0895</td><td>0.2616</td><td>0.4265</td></tr><tr><td>BERT [12]</td><td>0.0552</td><td>0.1998</td><td>0.3544</td><td>0.1111</td><td>0.3101</td><td>0.4762</td><td>0.0901</td><td>0.2760</td><td>0.6458</td></tr><tr><td>CITL [34]</td><td>0.0328</td><td>0.1034</td><td>0.2427</td><td>0.0942</td><td>0.2102</td><td>0.3384</td><td>0.0526</td><td>0.1568</td><td>0.4080</td></tr><tr><td>PC-MRL w/o RPA</td><td>0.0397</td><td>0.1657</td><td>0.3417</td><td>0.1058</td><td>0.2879</td><td>0.4722</td><td>0.0640</td><td>0.2288</td><td>0.6021</td></tr><tr><td>PC-MRL (Ours)</td><td>0.0377</td><td>0.1392</td><td>0.2848</td><td>0.1064</td><td>0.2778</td><td>0.4281</td><td>0.0658</td><td>0.2129</td><td>0.5098</td></tr></table>
|
| 198 |
+
|
| 199 |
+
Table 1: Performance comparisons for various motion categories, measured in L2P, L2Q, and NPSS. All methods except PC-MRL are trained on the original CMU MoCap dataset. Each motion contains 128 frames with keyframes placed every 5, 15, or 30 frames. The top 2 results for each test are highlighted in purple and blue respectively.
|
| 200 |
+
|
| 201 |
+
The results listed in Table 1 clearly demonstrate our method's ability to approach the accuracy levels exhibited by directly supervised state-of-the-art methods. PC-MRL consistently outperforms both the LERP standard and RNN-based $\mathrm{TG}_{\mathrm{complete}}$ model, particularly in longer keyframe interval scenarios where the quality and depth of learned motion features is most critical and impactful. Likewise, the inclusion of RPA also tends to improve the long interval performance of PC-MRL, particularly when precise movements (i.e. golf) or deeper motion features (i.e. locomotion) are expected. Given this, we can observe that RPA definitively improves the overall consistency of the PC-MRL method in scenarios that may be unseen or less effectively represented by our point cloud-to-motion system.
|
| 202 |
+
|
| 203 |
+
Native supervision, i.e. with CITL, unsurprisingly produces the most accurate interpolation model, with a notable exception of swimming motions, where our PC-MRL supervision produces even stronger results. We believe the abundance of low crawl motions, a similar motion to swimming in LaFAN1, provides more meaningful supervision over the original CMU dataset, which conversely has few similar motions once swimming motions are filtered out. On the other end of the spectrum, i.e. for high data scenarios such as walking and running locomotion from Fig. 5, all models including our PC-MRL approach observe the strongest improvements for learned methods over the naive LERP method. Both cases strongly support our method's capability to supervise highly intricate motion patterns, despite the incompleteness of the point cloud representation and reliance on relative rotations and geometric optimisation assumptions.
|
| 204 |
+
|
| 205 |
+

|
| 206 |
+
Fig. 5: Walking keyframe interpolation examples produced by methods from Table 1. Close-up layered comparisons against the original motion are provided on the right.
|
| 207 |
+
|
| 208 |
+
# 4.4 Cross-Skeleton Motion Re-Targeting Experiments
|
| 209 |
+
|
| 210 |
+
We further conduct an experiment on motion re-targeting to directly benchmark our method against the sole existing state-of-the-art method for unpaired motion re-targeting with unrestricted skeletons, i.e. primal skeletons (PS) [1]. For direct motion re-targeting evaluation, we utilized the KNN-Loss and end-effector loss as metrics, owing to their effective measurement of visual similarity and non-sensitivity towards absolute roll axis correctness. Though absolute rolls are generally necessary for a complete motion re-targeting solution, this aspect is out of scope for our project as our main goal is motion interpolation.
|
| 211 |
+
|
| 212 |
+
Table 2 indicates the superior performance of our method over the state-of-the-art PS method in terms of visual similarity. At low $\sigma$ values, the higher relative $\mathcal{L}_{\mathrm{KNN}}$ of the primal skeleton method underscores its geometric deviations
|
| 213 |
+
|
| 214 |
+
<table><tr><td rowspan="2">Model</td><td colspan="5">\( L_{\text{KNN}}(\mathcal{X}_A,\mathcal{X}_B) \downarrow \)</td><td rowspan="2">\( L_{\text{end}} \downarrow \)</td></tr><tr><td>\( \sigma = 0 \)</td><td>\( \sigma = 0.01 \)</td><td>\( \sigma = 0.05 \)</td><td>\( \sigma = 0.1 \)</td><td>\( \sigma = 0.2 \)</td></tr><tr><td>Original motion</td><td>1.0099</td><td>2.0234</td><td>5.7399</td><td>9.6636</td><td>16.9223</td><td>0</td></tr><tr><td>Primal skeleton [1]</td><td>8.1109</td><td>8.0353</td><td>9.4206</td><td>12.8291</td><td>20.6332</td><td>9.6990</td></tr><tr><td>PC-MRL (Ours)</td><td>3.8714</td><td>4.1231</td><td>6.7277</td><td>10.3078</td><td>17.5261</td><td>5.6258</td></tr></table>
|
| 215 |
+
|
| 216 |
+
Table 2: Average unpaired LaFAN1 to CMU motion re-targeting performance of the primal skeleton method and our point cloud method. $\mathcal{L}_{\mathrm{KNN}}$ is measured using 1024-point clouds at $K = 8$ . Independently sampled point clouds from the original LaFAN1 skeleton are provided as optimal $\mathcal{L}_{\mathrm{KNN}}$ reference. All values are scaled by $\times 100$ .
|
| 217 |
+
|
| 218 |
+

|
| 219 |
+
Fig. 6: Jumping motion re-targeting examples produced on the CMU skeleton by the primal skeleton method (middle row) and our method (bottom row). Sampled point cloud representations of each motion sequence are provided on the right.
|
| 220 |
+
|
| 221 |
+
from the original motion. In contrast, our method demonstrates significantly better visual adherence to the original geometry. As $\sigma$ increases, our method approaches near-perfect alignment with the original motion. Figure 6 visually indicates this, showing our method follows the original motion closely, while the latent consistency-based primal skeleton method struggles to generalise and decode the skeleton's internal structures without directly supervised objectives, such as the end-effector positions provided during training. In addition, Fig 7 demonstrates that, like existing re-targeting approaches, our method is able to produce adequate results on disproportionate skeletons.
|
| 222 |
+
|
| 223 |
+
Since our approach is largely focused on enabling non-native motion interpolation supervision, we additionally compare our PC-MRL method against a CITL model supervised by PS in Table 3. Due to the suboptimal re-targeting performance of PS, such interpolation models exhibit expectedly poor accuracy.
|
| 224 |
+
|
| 225 |
+

|
| 226 |
+
Fig. 7: Re-targeted falling motion from LaFAN1 (top) on the CMU skeleton (middle) and a short custom skeleton (bottom), performed by our point cloud-to-motion model.
|
| 227 |
+
|
| 228 |
+
<table><tr><td rowspan="2">Method</td><td colspan="3">L2P↓</td><td colspan="3">L2Q↓</td><td colspan="3">NPSS↓</td></tr><tr><td>5</td><td>15</td><td>30</td><td>5</td><td>15</td><td>30</td><td>5</td><td>15</td><td>30</td></tr><tr><td>CITL - PS</td><td>0.2043</td><td>0.6332</td><td>1.1524</td><td>0.2318</td><td>0.7428</td><td>1.1442</td><td>0.1122</td><td>0.4276</td><td>1.4622</td></tr><tr><td>CITL - Native</td><td>0.0495</td><td>0.2306</td><td>0.4898</td><td>0.0979</td><td>0.3604</td><td>0.6609</td><td>0.1149</td><td>0.4684</td><td>1.1591</td></tr><tr><td>PC-MRL</td><td>0.0494</td><td>0.2337</td><td>0.4960</td><td>0.1030</td><td>0.3771</td><td>0.6813</td><td>0.1370</td><td>0.5433</td><td>1.1760</td></tr></table>
|
| 229 |
+
|
| 230 |
+
Table 3: Average L2P, L2Q, and NPSS performance of CITL trained on native and PS-generated datasets, compared to our method.
|
| 231 |
+
|
| 232 |
+
# 5 Limitations
|
| 233 |
+
|
| 234 |
+
Due to the inherent limitations of the point cloud representation, certain constraints on the applicability of our point cloud-to-motion method are inevitable. Primarily, the reliance of our point cloud re-targeting method on the relative nature of First-frame Offset Quaternions (FOQ) for most motion learning tasks necessitates the presence of at least one known pose. In the absence of a known pose, relative quaternions are incapable of reconstructing absolute rotational values, which are essential for motion applications. Secondly, due to our method of sampling point clouds with a consistent deviation from each bone segment, the resultant representation inevitably obscures finer details. This includes elements like fingers and facial controls, to an extent that is beyond rectification.
|
| 235 |
+
|
| 236 |
+
Our proposed objective function relies entirely on geometric segments, necessitating that all bones possess a non-zero length. This stipulation is deemed a reasonable prerequisite for generic human skeletons, as each bone is typically responsible for manipulating a discernible segment of the human body. Due to technical considerations, the skeletal configurations in each of our datasets included at least one bone of zero length, which we ignored in our experiments in favour of global quaternion rotations of all bones with non-zero length.
|
| 237 |
+
|
| 238 |
+
As a result of these limitations, we refrained from claiming state-of-the-art performance for general motion re-targeting tasks.
|
| 239 |
+
|
| 240 |
+
# 6 Conclusion
|
| 241 |
+
|
| 242 |
+
This paper introduces a novel learning-based motion interpolation method designed to enable cross-skeleton compatibility with existing motion datasets. Our proposed method PC-MRL employs a process of point cloud obfuscation and skeleton reconstruction. The point cloud space represents human motions in a non-hierarchical and skeleton-agnostic manner. It enables a KNN-based objective to be optimised without dataset supervision, guiding a neural network to generate high-quality motion features with any target skeleton. We address the rotational information loss of our point cloud format by presenting an offset quaternion strategy compatible with concurrent transformer-based models. Through extensive experiments, we have demonstrated the efficacy of PC-MRL in performing motion interpolation without relying on native motion data. Moreover, PC-MRL has achieved superior visual similarity metrics in the domain of motion re-targeting. We concluded our work by discussing the limitations of PC-MRL.
|
| 243 |
+
|
| 244 |
+
# References
|
| 245 |
+
|
| 246 |
+
1. Aberman, K., Li, P., Lischinski, D., Sorkine-Hornung, O., Cohen-Or, D., Chen, B.: Skeleton-aware networks for deep motion retargeting 39(4), 62–1 (2020) 4, 12
|
| 247 |
+
2. Aberman, K., Wu, R., Lischinski, D., Chen, B., Cohen-Or, D.: Learning character-agnostic motion for motion retargeting in 2d 38(4), 1-14 (jul 2019) 4
|
| 248 |
+
3. Annabi, L., Ma, Z., et al.: Unsupervised motion retargeting for human-robot imitation. In: Companion of the 2024 ACM/IEEE International Conference on Human-Robot Interactio. ACM (2024) 4
|
| 249 |
+
4. Bodenheimer, B., Rose, C., Rosenthal, S., Pella, J.: The process of motion capture: Dealing with the data. In: Computer Animation and Simulation. pp. 3-18. Springer (1997) 4
|
| 250 |
+
5. Burtnyk, N., Wein, M.: Computer-generated key-frame animation. Journal of the SMPTE 80(3), 149-153 (1971) 1, 4
|
| 251 |
+
6. Catalin Ionescu, Fuxin Li, C.S.: Latent structured models for human pose estimation. In: International Conference on Computer Vision (2011) 10
|
| 252 |
+
7. Chen, H., Tang, H., Timofte, R., Gool, L.V., Zhao, G.: Lart: Neural correspondence learning with latent regularization transformer for 3d motion transfer. Advances in Neural Information Processing Systems 36 (2024) 4
|
| 253 |
+
8. Chen, J., Li, C., Lee, G.H.: Weakly-supervised 3D pose transfer with keypoints. In: ICCV. pp. 15156-15165 (2023) 4
|
| 254 |
+
9. Choi, K.J., Ko, H.S.: Online motion retargetting. The Journal of Visualization and Computer Animation 11(5), 223-235 (2000) 5
|
| 255 |
+
0. CMU: Carnegie-mellon university motion capture database (2003), http://mocap.cs.cmu.edu/ 10
|
| 256 |
+
1. Dam, E.B., Koch, M., Lillholm, M.: Quaternions, interpolation and animation, vol. 2. CiteSeer (1998) 5, 6
|
| 257 |
+
2. Duan, Y., Lin, Y., Zou, Z., Yuan, Y., Qian, Z., Zhang, B.: A unified framework for real time motion completion. In: AAAI. vol. 36, pp. 4459-4467 (2022) 2, 4, 10, 11
|
| 258 |
+
3. Frank, T., Johnston, O.: Disney Animation: The Illusion of Life. Abbeville Publishing Group, New York (1981) 1
|
| 259 |
+
4. Gleicher, M.: Retargetting motion to new characters. In: Computer Graphics and Interactive Techniques. pp. 33-42 (1998) 4
|
| 260 |
+
5. Gleicher, M.: Animation from observation: Motion capture and motion editing 33(4), 51-54 (1999) 4
|
| 261 |
+
6. Gopalakrishnan, A., Mali, A., Kifer, D., Giles, L., Ororbia, A.G.: A neural temporal model for human motion prediction. In: CVPR. pp. 12116-12125 (2019) 10
|
| 262 |
+
7. Harvey, F.G., Pal, C.: Recurrent transition networks for character locomotion. In: SIGGRAPH Asia 2018 Technical Briefs, pp. 1-4 (2018) 4
|
| 263 |
+
8. Harvey, F.G., Yurick, M., Nowrouzezahrai, D., Pal, C.: Robust motion in betweening 39(4), 60-1 (2020) 2, 4, 8, 9, 10, 11
|
| 264 |
+
9. He, C., Saito, J., Zachary, J., Rushmeier, H., Zhou, Y.: Nemf: Neural motion fields for kinematic animation 35, 4244-4256 (2022) 4
|
| 265 |
+
20. Hecker, C., Raabe, B., Enslow, R.W., DeWeese, J., Maynard, J., van Prooijen, K.: Real-time motion retargeting to highly varied user-created morphologies 27(3), 1-11 (2008) 5
|
| 266 |
+
21. Hernandez, A., Gall, J., Moreno-Noguer, F.: Human motion prediction via spatiotemporal inpainting. In: ICCV. pp. 7134-7143 (2019) 4
|
| 267 |
+
22. Ionescu, C., Papava, D., Olaru, V., Sminchisescu, C.: Human3.6m: Large scale datasets and predictive methods for 3D human sensing in natural environments.
|
| 268 |
+
|
| 269 |
+
IEEE Transactions on Pattern Analysis and Machine Intelligence 36(7), 1325-1339 (jul 2014) 10
|
| 270 |
+
23. Jiang, H., Cai, J., Zheng, J.: Skeleton-aware 3D human shape reconstruction from point clouds. In: ICCV. pp. 5431-5441 (2019) 4
|
| 271 |
+
24. Kantor, I.L., Solodovnikov, A.S., Shenitzer, A.: Hypercomplex numbers: an elementary introduction to algebras, vol. 302. Springer (1989) 6
|
| 272 |
+
25. Karunratanakul, K., Preechakul, K., Suwajanakorn, S., Tang, S.: Guided motion diffusion for controllable human motion synthesis. In: ICCV. pp. 2151-2162 (2023) 4
|
| 273 |
+
26. Kulpa, R., Multon, F., Arnaldi, B.: Morphology-independent representation of motions for interactive human-like animation. In: Eurographics (2005) 5
|
| 274 |
+
27. Lehrmann, A.M., Gehler, P.V., Nowozin, S.: Efficient nonlinear markov models for human motion. In: CVPR. pp. 1314-1321 (2014) 4
|
| 275 |
+
28. Li, P., Aberman, K., Zhang, Z., Hanocka, R., Sorkine-Hornung, O.: Ganimator: Neural motion synthesis from a single sequence 41(4), 1-12 (2022) 4
|
| 276 |
+
29. Li, W., Wang, Z., Mai, R., Ren, P., Zhang, Q., Zhou, Y., Xu, N., Zhuang, J., Xin, B., Gao, L., et al.: Modular design automation of the morphologies, controllers, and vision systems for intelligent robots: a survey. Visual Intelligence 1(1), 2 (2023) 4
|
| 277 |
+
30. Li, Z., Peng, X.B., Abbeel, P., Levine, S., Berseth, G., Sreenath, K.: Reinforcement learning for versatile, dynamic, and robust bipedal locomotion control. arXiv preprint arXiv:2401.16889 (2024) 4
|
| 278 |
+
31. Liao, Z., Yang, J., Saito, J., Pons-Moll, G., Zhou, Y.: Skeleton-free pose transfer for stylized 3D characters. In: ECCV. pp. 640-656. Springer (2022) 4
|
| 279 |
+
32. Makoviychuk, V., Wawrzyniak, L., Guo, Y., Lu, M., Storey, K., Macklin, M., Hoeller, D., Rudin, N., Allshire, A., Handa, A., et al.: Isaac gym: High performancegpu-based physics simulation for robot learning. arXiv preprint arXiv:2108.10470 (2021) 4
|
| 280 |
+
33. Merry, B., Marais, P., Gain, J.: Animation space: A truly linear framework for character animation. ACM Transactions on Graphics (TOG) 25(4), 1400-1423 (2006) 5
|
| 281 |
+
34. Mo, C.A., Hu, K., Long, C., Wang, Z.: Continuous intermediate token learning with implicit motion manifold for keyframe based motion interpolation. In: CVPR. pp. 13894-13903 (2023) 2, 4, 8, 10, 11
|
| 282 |
+
35. Mo, C.A., Hu, K., Mei, S., Chen, Z., Wang, Z.: Keyframe extraction from motion capture sequences with graph based deep reinforcement learning. In: ACM International Conference on Multimedia. pp. 5194-5202 (2021) 4
|
| 283 |
+
36. Monzani, J.S., Baerlocher, P., Boulic, R., Thalmann, D.: Using an intermediate skeleton and inverse kinematics for motion retargeting. In: Computer Graphics Forum. vol. 19, pp. 11-19. Wiley Online Library (2000) 5
|
| 284 |
+
37. Oreshkin, B.N., Valkanas, A., Harvey, F.G., Menard, L.S., Bocquelet, F., Coates, M.J.: Motion inbetweening via deep $\delta$ -interpolator (2022) 2, 4
|
| 285 |
+
38. Peng, X.B., Abbeel, P., Levine, S., van de Panne, M.: Deepmimic: Example-guided deep reinforcement learning of physics-based character skills 37(4), 143:1-143:14 (Jul 2018) 4
|
| 286 |
+
39. Peng, X.B., Ma, Z., Abbeel, P., Levine, S., Kanazawa, A.: Amp: Adversarial motion priors for stylized physics-based character control 40(4) (Jul 2021) 4
|
| 287 |
+
40. Qin, J., Zheng, Y., Zhou, K.: Motion in-between via two-stage transformers 41(6), 1-16 (2022) 2
|
| 288 |
+
41. Reda, D., Won, J., Ye, Y., van de Panne, M., Winkler, A.: Physics-based motion retargeting from sparse inputs. ACM Computer Graphics and Interactive Techniques 6(3), 1-19 (2023) 4
|
| 289 |
+
|
| 290 |
+
42. Reeves, W.T.: Inbetweening for computer animation utilizing moving point constraints 15(3), 263-269 (1981) 1
|
| 291 |
+
43. Ren, T., Yu, J., Guo, S., Ma, Y., Ouyang, Y., Zeng, Z., Zhang, Y., Qin, Y.: Diverse motion in-betweening from sparse keyframes with dual posture stitching. IEEE Transactions on Visualization and Computer Graphics (2024) 4
|
| 292 |
+
44. Shaw, P., Uszkoreit, J., Vaswani, A.: Self-attention with relative position representations. In: NAACL. Association for Computational Linguistics (2018) 9
|
| 293 |
+
45. Tiwari, G., Antic, D., Lenssen, J.E., Sarafianos, N., Tung, T., Pons-Moll, G.: Posendf: Modeling human pose manifolds with neural distance fields. In: ECCV. pp. 572-589. Springer (2022) 4
|
| 294 |
+
46. Villegas, R., Yang, J., Ceylan, D., Lee, H.: Neural kinematic networks for unsupervised motion retargeting. In: CVPR. pp. 8639-8648 (2018) 4
|
| 295 |
+
47. Wang, J., Li, X., Liu, S., De Mello, S., Gallo, O., Wang, X., Kautz, J.: Zero-shot pose transfer for unrigged stylized 3D characters. In: CVPR. pp. 8704-8714 (2023) 4
|
| 296 |
+
48. Wang, K., Xie, J., Zhang, G., Liu, L., Yang, J.: Sequential 3D human pose and shape estimation from point clouds. In: CVPR. pp. 7275-7284 (2020) 4
|
| 297 |
+
49. Yuan, Y., Song, J., Iqbal, U., Vahdat, A., Kautz, J.: Physdiff: Physics-guided human motion diffusion model. In: ICCV. pp. 16010-16021 (2023) 4
|
| 298 |
+
50. Yuan, Y., Wei, S.E., Simon, T., Kitani, K., Saragih, J.: Simpoe: Simulated character control for 3D human pose estimation. In: CVPR. pp. 7159-7169 (2021) 4
|
| 299 |
+
51. Zhang, J., Weng, J., Kang, D., Zhao, F., Huang, S., Zhe, X., Bao, L., Shan, Y., Wang, J., Tu, Z.: Skinned motion retargeting with residual perception of motion semantics & geometry. In: CVPR. pp. 13864-13872 (2023) 4
|
| 300 |
+
52. Zhang, M., Li, H., Cai, Z., Ren, J., Yang, L., Liu, Z.: Finemogen: Fine-grained spatio-temporal motion generation and editing. Advances in Neural Information Processing Systems 36 (2024) 4
|
| 301 |
+
53. Zhang, X., van de Panne, M.: Data-driven autocompletion for keyframe animation. In: ACM SIGGRAPH Conference on Motion, Interaction and Games. pp. 1-11 (2018) 4
|
| 302 |
+
54. Zhao, H., Jiang, L., Jia, J., Torr, P.H., Koltun, V.: Point transformer. In: ICCV. pp. 16259-16268 (2021) 7, 10
|
| 303 |
+
55. Zhu, W., Yang, Z., Di, Z., Wu, W., Wang, Y., Loy, C.C.: Mocanet: Motion retargeting in-the-wild via canonicalization networks. In: AAAI. vol. 36, pp. 3617-3625 (2022) 4
|
2024/Motion Keyframe Interpolation for Any Human Skeleton using Point Cloud-based Human Motion Data Homogenisation/images.zip
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:f33e8b8893dee3a93be5dac335e00446af47660fff3b3700a27aa5b6c6e85a33
|
| 3 |
+
size 565290
|
2024/Motion Keyframe Interpolation for Any Human Skeleton using Point Cloud-based Human Motion Data Homogenisation/layout.json
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
2024/Motion Mamba_ Efficient and Long Sequence Motion Generation/1b586349-9221-4bad-a56d-e274ec79c852_content_list.json
ADDED
|
@@ -0,0 +1,1595 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
{
|
| 3 |
+
"type": "text",
|
| 4 |
+
"text": "Motion Mamba: Efficient and Long Sequence Motion Generation",
|
| 5 |
+
"text_level": 1,
|
| 6 |
+
"bbox": [
|
| 7 |
+
243,
|
| 8 |
+
140,
|
| 9 |
+
759,
|
| 10 |
+
184
|
| 11 |
+
],
|
| 12 |
+
"page_idx": 0
|
| 13 |
+
},
|
| 14 |
+
{
|
| 15 |
+
"type": "text",
|
| 16 |
+
"text": "Zeyu Zhang $^{12*}$ , Akide Liu $^{1*}$ , Ian Reid $^{3}$ , Richard Hartley $^{2}$ , Bohan Zhuang $^{1}$ , and Hao Tang $^{4\\boxtimes}$",
|
| 17 |
+
"bbox": [
|
| 18 |
+
238,
|
| 19 |
+
212,
|
| 20 |
+
764,
|
| 21 |
+
244
|
| 22 |
+
],
|
| 23 |
+
"page_idx": 0
|
| 24 |
+
},
|
| 25 |
+
{
|
| 26 |
+
"type": "text",
|
| 27 |
+
"text": "<sup>1</sup> Monash University",
|
| 28 |
+
"bbox": [
|
| 29 |
+
431,
|
| 30 |
+
253,
|
| 31 |
+
573,
|
| 32 |
+
268
|
| 33 |
+
],
|
| 34 |
+
"page_idx": 0
|
| 35 |
+
},
|
| 36 |
+
{
|
| 37 |
+
"type": "text",
|
| 38 |
+
"text": "2 The Australian National University",
|
| 39 |
+
"bbox": [
|
| 40 |
+
375,
|
| 41 |
+
268,
|
| 42 |
+
627,
|
| 43 |
+
282
|
| 44 |
+
],
|
| 45 |
+
"page_idx": 0
|
| 46 |
+
},
|
| 47 |
+
{
|
| 48 |
+
"type": "text",
|
| 49 |
+
"text": "<sup>3</sup> Mohamed bin Zayed University of Artificial Intelligence",
|
| 50 |
+
"bbox": [
|
| 51 |
+
308,
|
| 52 |
+
282,
|
| 53 |
+
692,
|
| 54 |
+
297
|
| 55 |
+
],
|
| 56 |
+
"page_idx": 0
|
| 57 |
+
},
|
| 58 |
+
{
|
| 59 |
+
"type": "text",
|
| 60 |
+
"text": "$^{4}$ National Key Laboratory for Multimedia Information Processing, School of Computer Science, Peking University",
|
| 61 |
+
"bbox": [
|
| 62 |
+
276,
|
| 63 |
+
296,
|
| 64 |
+
725,
|
| 65 |
+
325
|
| 66 |
+
],
|
| 67 |
+
"page_idx": 0
|
| 68 |
+
},
|
| 69 |
+
{
|
| 70 |
+
"type": "text",
|
| 71 |
+
"text": "https://steve-zeyu-zhang.github.io/MotionMamba",
|
| 72 |
+
"bbox": [
|
| 73 |
+
334,
|
| 74 |
+
335,
|
| 75 |
+
666,
|
| 76 |
+
348
|
| 77 |
+
],
|
| 78 |
+
"page_idx": 0
|
| 79 |
+
},
|
| 80 |
+
{
|
| 81 |
+
"type": "image",
|
| 82 |
+
"img_path": "images/9d50515d2181f0c7d935e2bbb103cb5293cf413d46914332a1dd381edc6ed7a8.jpg",
|
| 83 |
+
"image_caption": [
|
| 84 |
+
"Fig.1: Motion Mamba has achieved significantly superior performance on long sequence modeling and motion generation efficiency compared with other well-designed state-of-the-art methods such as MLD [6], MotionDiffuse [54], and MDM [49]."
|
| 85 |
+
],
|
| 86 |
+
"image_footnote": [],
|
| 87 |
+
"bbox": [
|
| 88 |
+
222,
|
| 89 |
+
362,
|
| 90 |
+
419,
|
| 91 |
+
513
|
| 92 |
+
],
|
| 93 |
+
"page_idx": 0
|
| 94 |
+
},
|
| 95 |
+
{
|
| 96 |
+
"type": "image",
|
| 97 |
+
"img_path": "images/48cf11f3798a6aedad48113acb76dca9df21fbaad9e72b652f24ec0f0b7e17e3.jpg",
|
| 98 |
+
"image_caption": [],
|
| 99 |
+
"image_footnote": [],
|
| 100 |
+
"bbox": [
|
| 101 |
+
421,
|
| 102 |
+
361,
|
| 103 |
+
614,
|
| 104 |
+
513
|
| 105 |
+
],
|
| 106 |
+
"page_idx": 0
|
| 107 |
+
},
|
| 108 |
+
{
|
| 109 |
+
"type": "image",
|
| 110 |
+
"img_path": "images/210741e7ade2cfcbcf250c0c7992f4b4a689a716569803d67cdd28922bb7d8d0.jpg",
|
| 111 |
+
"image_caption": [],
|
| 112 |
+
"image_footnote": [],
|
| 113 |
+
"bbox": [
|
| 114 |
+
627,
|
| 115 |
+
361,
|
| 116 |
+
787,
|
| 117 |
+
513
|
| 118 |
+
],
|
| 119 |
+
"page_idx": 0
|
| 120 |
+
},
|
| 121 |
+
{
|
| 122 |
+
"type": "text",
|
| 123 |
+
"text": "Abstract. Human motion generation stands as a significant pursuit in generative computer vision, while achieving long-sequence and efficient motion generation remains challenging. Recent advancements in state space models (SSMs), notably Mamba, have showcased considerable promise in long sequence modeling with an efficient hardware-aware design, which appears to be a significant direction upon building motion generation model. Nevertheless, adapting SSMs to motion generation faces hurdles since the lack of a specialized design architecture to model motion sequence. To address these challenges, we propose Motion Mamba, a simple yet efficient approach that presents the pioneering motion generation model utilized SSMs. Specifically, we design a Hierarchical Temporal Mamba (HTM) block to process temporal data by ensemble varying numbers of isolated SSM modules across a symmetric U-Net architecture aimed at preserving motion consistency between frames. We also design a Bidirectional Spatial Mamba (BSM) block",
|
| 124 |
+
"bbox": [
|
| 125 |
+
259,
|
| 126 |
+
583,
|
| 127 |
+
743,
|
| 128 |
+
792
|
| 129 |
+
],
|
| 130 |
+
"page_idx": 0
|
| 131 |
+
},
|
| 132 |
+
{
|
| 133 |
+
"type": "page_footnote",
|
| 134 |
+
"text": "*Equal contribution.",
|
| 135 |
+
"bbox": [
|
| 136 |
+
232,
|
| 137 |
+
804,
|
| 138 |
+
354,
|
| 139 |
+
816
|
| 140 |
+
],
|
| 141 |
+
"page_idx": 0
|
| 142 |
+
},
|
| 143 |
+
{
|
| 144 |
+
"type": "page_footnote",
|
| 145 |
+
"text": "†Work done while being a research assistant at Monash University.",
|
| 146 |
+
"bbox": [
|
| 147 |
+
233,
|
| 148 |
+
816,
|
| 149 |
+
617,
|
| 150 |
+
827
|
| 151 |
+
],
|
| 152 |
+
"page_idx": 0
|
| 153 |
+
},
|
| 154 |
+
{
|
| 155 |
+
"type": "page_footnote",
|
| 156 |
+
"text": "Corresponding author: bjdxtanghao@gmail.com",
|
| 157 |
+
"bbox": [
|
| 158 |
+
233,
|
| 159 |
+
827,
|
| 160 |
+
517,
|
| 161 |
+
839
|
| 162 |
+
],
|
| 163 |
+
"page_idx": 0
|
| 164 |
+
},
|
| 165 |
+
{
|
| 166 |
+
"type": "text",
|
| 167 |
+
"text": "to bidirectionally process latent poses, to enhance accurate motion generation within a temporal frame. Our proposed method achieves up to $50\\%$ FID improvement and up to 4 times faster on the HumanML3D and KIT-ML datasets compared to the previous best diffusion-based method, which demonstrates strong capabilities of high-quality long sequence motion modeling and real-time human motion generation.",
|
| 168 |
+
"bbox": [
|
| 169 |
+
259,
|
| 170 |
+
147,
|
| 171 |
+
743,
|
| 172 |
+
231
|
| 173 |
+
],
|
| 174 |
+
"page_idx": 1
|
| 175 |
+
},
|
| 176 |
+
{
|
| 177 |
+
"type": "text",
|
| 178 |
+
"text": "Keywords: Human Motion Generation $\\cdot$ Selective State Space Models $\\cdot$ Latent Diffusion Models",
|
| 179 |
+
"bbox": [
|
| 180 |
+
261,
|
| 181 |
+
244,
|
| 182 |
+
743,
|
| 183 |
+
272
|
| 184 |
+
],
|
| 185 |
+
"page_idx": 1
|
| 186 |
+
},
|
| 187 |
+
{
|
| 188 |
+
"type": "text",
|
| 189 |
+
"text": "1 Introduction",
|
| 190 |
+
"text_level": 1,
|
| 191 |
+
"bbox": [
|
| 192 |
+
215,
|
| 193 |
+
303,
|
| 194 |
+
375,
|
| 195 |
+
319
|
| 196 |
+
],
|
| 197 |
+
"page_idx": 1
|
| 198 |
+
},
|
| 199 |
+
{
|
| 200 |
+
"type": "text",
|
| 201 |
+
"text": "Human motion generation involves creating high-quality 3D human motion based on given conditions, such as a text prompt, which generates Euler angles for each joint in a human skeleton. To emulate human motion effectively, virtual characters must respond to the conditional context, exhibit natural movement, and perform motion accurately. Recent motion generation models are categorized into four main approaches: autoencoder-based [1,10,17,36,48,56], utilizing transformers for latent space compression and motion synthesis; GAN-based [4,20,30], using discriminators to enhance the realism of generated motions; autoregressive models [24], treating motion sequences as languages with specialized codebooks; and diffusion-based [6,49,54], employing denoising steps for motion generation. Challenges vary across methods, with autoencoder models struggling to generate accurate motions from detailed descriptions due to textual information compression, GAN-based models facing training difficulties, especially in conditional tasks, and diffusion-based models relying on complex transformer-based architectures results in inefficient of motion prediction.",
|
| 202 |
+
"bbox": [
|
| 203 |
+
212,
|
| 204 |
+
339,
|
| 205 |
+
787,
|
| 206 |
+
566
|
| 207 |
+
],
|
| 208 |
+
"page_idx": 1
|
| 209 |
+
},
|
| 210 |
+
{
|
| 211 |
+
"type": "text",
|
| 212 |
+
"text": "Although diffusion-based models excel at generating motion with robust performance and often exhibit superior diversity, they encounter two limitations. 1) Convolutional or transformer-based diffusion methods exhibit limitations in generating long-range motion sequences. Previous transformer-based methodologies [6,49,54] have focused on modeling long-range dependencies and acquiring comprehensive global context information. Despite these advances, they are frequently associated with a substantial increase in computational requirements. Furthermore, transformer architectures are not intrinsically designed for temporal sequential modeling, which poses an inherent limitation. 2) The efficiency of inference in transformer-based diffusion methods is constrained. Although prior research has attempted to leverage the Variational Autoencoder for denoising operations in the latent space [6], the inference speed remains adversely affected by the attention mechanism's quadratic scaling, leading to inefficient motion generation. Consequently, exploring a new architectural paradigm that accommodates long-range dependencies and maintains a linear computational complexity is crucial for sustaining motion generation tasks.",
|
| 213 |
+
"bbox": [
|
| 214 |
+
212,
|
| 215 |
+
566,
|
| 216 |
+
787,
|
| 217 |
+
808
|
| 218 |
+
],
|
| 219 |
+
"page_idx": 1
|
| 220 |
+
},
|
| 221 |
+
{
|
| 222 |
+
"type": "text",
|
| 223 |
+
"text": "Recent advances have sparked renewed interest in state space models (SSMs) [14, 15], a field that originated from the foundational classic state space model",
|
| 224 |
+
"bbox": [
|
| 225 |
+
215,
|
| 226 |
+
809,
|
| 227 |
+
785,
|
| 228 |
+
840
|
| 229 |
+
],
|
| 230 |
+
"page_idx": 1
|
| 231 |
+
},
|
| 232 |
+
{
|
| 233 |
+
"type": "page_number",
|
| 234 |
+
"text": "2",
|
| 235 |
+
"bbox": [
|
| 236 |
+
217,
|
| 237 |
+
114,
|
| 238 |
+
228,
|
| 239 |
+
126
|
| 240 |
+
],
|
| 241 |
+
"page_idx": 1
|
| 242 |
+
},
|
| 243 |
+
{
|
| 244 |
+
"type": "header",
|
| 245 |
+
"text": "Zeyu and Akide et al.",
|
| 246 |
+
"bbox": [
|
| 247 |
+
271,
|
| 248 |
+
114,
|
| 249 |
+
416,
|
| 250 |
+
128
|
| 251 |
+
],
|
| 252 |
+
"page_idx": 1
|
| 253 |
+
},
|
| 254 |
+
{
|
| 255 |
+
"type": "text",
|
| 256 |
+
"text": "[25]. Modern versions of SSMs stand out due to their ability to effectively capture long-range dependencies, a capability greatly improved by the introduction of parallel training techniques. This evolution has led to various methodologies based on SSM, notably the linear state space layers (LSSL) [15], the structured state-space sequence model (S4) [14], the diagonal state space (DSS) [19], and S4D [13]. These methods have been carefully designed to handle sequential data across various tasks and modalities, paying special attention to modeling long-range dependencies. Their efficacy in managing long sequences is attributed to the implementation of convolutional computations [15] and near-linear computational strategies, such as mamba [12], marking a significant stride in sequentially oriented tasks, including large language model decoding and motion sequence generation.",
|
| 257 |
+
"bbox": [
|
| 258 |
+
212,
|
| 259 |
+
145,
|
| 260 |
+
787,
|
| 261 |
+
327
|
| 262 |
+
],
|
| 263 |
+
"page_idx": 2
|
| 264 |
+
},
|
| 265 |
+
{
|
| 266 |
+
"type": "text",
|
| 267 |
+
"text": "Adapting selective state space modules for motion generation tasks presents notable challenges, primarily due to the lack of specialized design in SSMs for capturing the sensitive motion details required for temporal representation and the complexities involved in aggregating latent space. In response to these challenges, we have meticulously developed a motion generation architecture, specifically tailored to address the intricacies of long-term sequence generation, while optimizing for computational efficiency with near-linear-time complexity. This innovation is embodied in the Motion Mamba model, a simple yet potent approach to motion generation. The Motion Mamba framework pioneers a diffusion-based generative system, incorporating two key components oriented toward SSM as shown in Figure. 2: (1) a Hierarchical Temporal Mamba (HTM) block: This component is ingeniously crafted to arrange motion frames in sequential order, using hierarchically adjusted scanning. It is adept at identifying temporal dependencies at various depths, thereby facilitating a thorough comprehension of the dynamics inherent in motion sequences. (2) a Bidirectional Spatial Mamba (BSM) block: This block is designed to unravel the structured latent skeleton by evaluating data from both forward and reverse directions. Its primary goal is to safeguard the continuity of information flow, significantly bolstering the model's capacity for precise motion generation through the retention of dense informational exchange.",
|
| 268 |
+
"bbox": [
|
| 269 |
+
212,
|
| 270 |
+
330,
|
| 271 |
+
787,
|
| 272 |
+
633
|
| 273 |
+
],
|
| 274 |
+
"page_idx": 2
|
| 275 |
+
},
|
| 276 |
+
{
|
| 277 |
+
"type": "text",
|
| 278 |
+
"text": "The Motion Mamba introduces a new approach to motion generation that strikes an exceptional trade-off between accuracy and efficiency, shown in Fig. 1. Our experimental results underscore the significant improvements brought about by Motion Mamba, showcasing a remarkable improvement in the Fréchet Inception Distance (FID), with a reduction of up to $50\\%$ from the prior state-of-the-art metric of 0.473 to an impressive 0.281 on the HumanML3D dataset [17]. Furthermore, we emphasize our framework's unparalleled inference speed, which is four times faster than previous methods, achieving an average inference time of only 0.058 seconds per sequence compared to the 0.217 seconds required by the MLD [6] method per sequence. These outcomes unequivocally establish Motion Mamba's state-of-the-art performance, concurrently ensuring fast inference speeds for conditional human motion generation tasks.",
|
| 279 |
+
"bbox": [
|
| 280 |
+
212,
|
| 281 |
+
638,
|
| 282 |
+
787,
|
| 283 |
+
820
|
| 284 |
+
],
|
| 285 |
+
"page_idx": 2
|
| 286 |
+
},
|
| 287 |
+
{
|
| 288 |
+
"type": "text",
|
| 289 |
+
"text": "Our contributions to the field of motion generation can be summarized as:",
|
| 290 |
+
"bbox": [
|
| 291 |
+
238,
|
| 292 |
+
824,
|
| 293 |
+
774,
|
| 294 |
+
839
|
| 295 |
+
],
|
| 296 |
+
"page_idx": 2
|
| 297 |
+
},
|
| 298 |
+
{
|
| 299 |
+
"type": "header",
|
| 300 |
+
"text": "Motion Mamba",
|
| 301 |
+
"bbox": [
|
| 302 |
+
625,
|
| 303 |
+
114,
|
| 304 |
+
730,
|
| 305 |
+
126
|
| 306 |
+
],
|
| 307 |
+
"page_idx": 2
|
| 308 |
+
},
|
| 309 |
+
{
|
| 310 |
+
"type": "page_number",
|
| 311 |
+
"text": "3",
|
| 312 |
+
"bbox": [
|
| 313 |
+
774,
|
| 314 |
+
116,
|
| 315 |
+
784,
|
| 316 |
+
126
|
| 317 |
+
],
|
| 318 |
+
"page_idx": 2
|
| 319 |
+
},
|
| 320 |
+
{
|
| 321 |
+
"type": "list",
|
| 322 |
+
"sub_type": "text",
|
| 323 |
+
"list_items": [
|
| 324 |
+
"1. We introduce a simple yet effective framework, named Motion Mamba, which is a pioneering method integrates a selective scanning mechanism into motion generation tasks.",
|
| 325 |
+
"2. Motion Mamba is comprised of two modules: Hierarchical Temporal Mamba (HTM) and Bidirectional Spatial Mamba (BSM), which are designed for temporal and spatial modeling, respectively. HTM blocks are tasked with processing temporal motion data, aiming to enhance motion consistency across frames. BSM blocks are engineered to bidirectionally capture the channel-wise flow of hidden information within the latent pose representations.",
|
| 326 |
+
"3. Motion Mamba framework demonstrated exceptional performance on text-to-motion generation task, through experimental validation on the HumanML3D [17] and KIT-ML [38] datasets. Our methodology achieved state-of-the-art generation quality and significantly improved long-squence modeling, meanwhile optimizing inference speed."
|
| 327 |
+
],
|
| 328 |
+
"bbox": [
|
| 329 |
+
215,
|
| 330 |
+
146,
|
| 331 |
+
785,
|
| 332 |
+
357
|
| 333 |
+
],
|
| 334 |
+
"page_idx": 3
|
| 335 |
+
},
|
| 336 |
+
{
|
| 337 |
+
"type": "text",
|
| 338 |
+
"text": "2 Related Works",
|
| 339 |
+
"text_level": 1,
|
| 340 |
+
"bbox": [
|
| 341 |
+
215,
|
| 342 |
+
383,
|
| 343 |
+
395,
|
| 344 |
+
398
|
| 345 |
+
],
|
| 346 |
+
"page_idx": 3
|
| 347 |
+
},
|
| 348 |
+
{
|
| 349 |
+
"type": "text",
|
| 350 |
+
"text": "Human Motion Generation. Generating human motion is a significant application of computer vision, essential for various applications like 3D modelling and robot manipulation. Recently, the predominant method of achieving human motion generation, known as the Text-to-Motion task, involves learning a common latent space for both language and motion.",
|
| 351 |
+
"bbox": [
|
| 352 |
+
212,
|
| 353 |
+
416,
|
| 354 |
+
785,
|
| 355 |
+
491
|
| 356 |
+
],
|
| 357 |
+
"page_idx": 3
|
| 358 |
+
},
|
| 359 |
+
{
|
| 360 |
+
"type": "text",
|
| 361 |
+
"text": "DVGAN [30] creates the GAN [11] discriminator by densely validating at each time-scale and perturbing the discriminator input for translation invariance, enabling motion generation and completion. ERD-QV [20] enhances latent representations through two additive modifiers: a time-to-arrival embedding applied universally and an additive scheduled target noise vector used during extended transitions. It further improves transition quality by incorporating a GAN framework with two discriminators operating at different timescales. HP-GAN [4], trained with a modified version of the improved WGAN-GP [16], utilizes a custom loss function designed for human motion prediction. It learns a probability density function of future human poses conditioned on previous poses.",
|
| 362 |
+
"bbox": [
|
| 363 |
+
212,
|
| 364 |
+
492,
|
| 365 |
+
785,
|
| 366 |
+
643
|
| 367 |
+
],
|
| 368 |
+
"page_idx": 3
|
| 369 |
+
},
|
| 370 |
+
{
|
| 371 |
+
"type": "text",
|
| 372 |
+
"text": "Autoencoders [27, 43] are notable generative models known for their ability to represent data robustly by compressing high-dimensional data into a latent space, which is widely adopted in human motion generation. JL2P [1] uses RNN-based autoencoders [23] to learn a combined representation of language and pose. It restricts the direct mapping from text to motion to a one-to-one relationship. MotionCLIP [48] uses Transformer-based Autoencoders [51] to reconstruct motion while ensuring alignment with the corresponding text label in the CLIP [40] space. This alignment effectively integrates the semantic knowledge from CLIP into the human motion manifold. TEMOS [36] and T2M [17] combine a Transformer-based VAE [26] with a text encoder to generate distribution parameters that work within the VAE latent space. AttT2M [56] and TM2D [10] incorporate a body-part spatio-temporal encoder into VQ-VAE [50] for enhanced learning of a discrete latent space with increased expressiveness.",
|
| 373 |
+
"bbox": [
|
| 374 |
+
212,
|
| 375 |
+
643,
|
| 376 |
+
785,
|
| 377 |
+
839
|
| 378 |
+
],
|
| 379 |
+
"page_idx": 3
|
| 380 |
+
},
|
| 381 |
+
{
|
| 382 |
+
"type": "page_number",
|
| 383 |
+
"text": "4",
|
| 384 |
+
"bbox": [
|
| 385 |
+
217,
|
| 386 |
+
114,
|
| 387 |
+
228,
|
| 388 |
+
126
|
| 389 |
+
],
|
| 390 |
+
"page_idx": 3
|
| 391 |
+
},
|
| 392 |
+
{
|
| 393 |
+
"type": "header",
|
| 394 |
+
"text": "Zeyu and Akide et al.",
|
| 395 |
+
"bbox": [
|
| 396 |
+
271,
|
| 397 |
+
114,
|
| 398 |
+
416,
|
| 399 |
+
128
|
| 400 |
+
],
|
| 401 |
+
"page_idx": 3
|
| 402 |
+
},
|
| 403 |
+
{
|
| 404 |
+
"type": "text",
|
| 405 |
+
"text": "Diffusion models [7, 22, 41, 47] have recently surpassed GANs and VAEs in generating 2D images. Developing a motion generation model based on diffusion models is obviously an attractive direction. MotionDiffuse [54] introduces the inaugural framework for text-driven motion generation based on diffusion models. It showcases several desirable properties, including probabilistic mapping, realistic synthesis, and multi-level manipulation. MDM [49] utilizes a classifier-free Transformer-based diffusion model for the human motion domain to predict sample rather than noise in each diffusion step. MLD [6] performs a diffusion process in latent motion space, rather than using a diffusion model to establish connections between raw motion sequences and conditional inputs.",
|
| 406 |
+
"bbox": [
|
| 407 |
+
218,
|
| 408 |
+
145,
|
| 409 |
+
785,
|
| 410 |
+
296
|
| 411 |
+
],
|
| 412 |
+
"page_idx": 4
|
| 413 |
+
},
|
| 414 |
+
{
|
| 415 |
+
"type": "text",
|
| 416 |
+
"text": "State Space Models. Recently, state space sequence models (SSMs) [14, 15], drawing inspiration from classical state-space models [25], have emerged as a promising architecture for sequence modeling. Mamba [12] introduces a selective SSM architecture, integrating time-varying parameters into the SSM framework, and proposes a hardware-aware algorithm to facilitate highly efficient training and inference processes. Some research works leverage SSM in computer vision to process 2D data. The 2D SSM [3] introduces an SSM block at the beginning of each transformer block [8, 51]. This approach aims to achieve efficient parameterization, accelerated computation, and a suitable normalization scheme. SGConvNeXt [29] presents a structured global convolution method inspired by ConvNeXt [32], incorporating multi-scale sub-kernels to achieve both parameterization efficiency and effective long sequence modeling. ConvSSM [46] integrates the tensor modeling principles of ConvLSTM [45] with SSMs, elucidating the utilization of parallel scans in convolutional recurrences. This approach enables subquadratic parallelization and rapid autoregressive generation. Vim [57] introduces a bidirectional SSM block [52] for efficient and versatile visual representation learning, achieving performance comparable to established ViT [8] methods. VMamba [31] introduces a Cross-Scan Module (CSM) designed to traverse the spatial domain and transform any non-causal visual image into ordered patch sequences. This approach achieves linear complexity while preserving global receptive fields. There have also been attempts to utilize SSMs to handle higher-dimensional data. Mamba-ND [28] explores various combinations of SSM and different scan directions within the SSM block to adapt Mamba [12] to higher-dimensional tasks. Recent efforts have sought to replace the traditional transformer-based U-Net within the diffusion denoiser with the SSM block, with the aim of enhancing image generation efficiency. DiffuSSM [53] adeptly manages higher resolutions without relying on global compression, thus maintaining detailed image representation throughout the diffusion process.",
|
| 417 |
+
"bbox": [
|
| 418 |
+
217,
|
| 419 |
+
297,
|
| 420 |
+
785,
|
| 421 |
+
719
|
| 422 |
+
],
|
| 423 |
+
"page_idx": 4
|
| 424 |
+
},
|
| 425 |
+
{
|
| 426 |
+
"type": "text",
|
| 427 |
+
"text": "3 The Proposed Method",
|
| 428 |
+
"text_level": 1,
|
| 429 |
+
"bbox": [
|
| 430 |
+
217,
|
| 431 |
+
744,
|
| 432 |
+
470,
|
| 433 |
+
762
|
| 434 |
+
],
|
| 435 |
+
"page_idx": 4
|
| 436 |
+
},
|
| 437 |
+
{
|
| 438 |
+
"type": "text",
|
| 439 |
+
"text": "In this section, we delineate the architecture and operational principles of the Motion Mamba framework, designed for generating human motion over long ranges efficiently from textual descriptions. Initially, we discuss the foundational concepts underpinning our approach, including the Mamba Model [12] and the",
|
| 440 |
+
"bbox": [
|
| 441 |
+
217,
|
| 442 |
+
779,
|
| 443 |
+
785,
|
| 444 |
+
839
|
| 445 |
+
],
|
| 446 |
+
"page_idx": 4
|
| 447 |
+
},
|
| 448 |
+
{
|
| 449 |
+
"type": "header",
|
| 450 |
+
"text": "Motion Mamba",
|
| 451 |
+
"bbox": [
|
| 452 |
+
625,
|
| 453 |
+
114,
|
| 454 |
+
730,
|
| 455 |
+
126
|
| 456 |
+
],
|
| 457 |
+
"page_idx": 4
|
| 458 |
+
},
|
| 459 |
+
{
|
| 460 |
+
"type": "page_number",
|
| 461 |
+
"text": "5",
|
| 462 |
+
"bbox": [
|
| 463 |
+
774,
|
| 464 |
+
116,
|
| 465 |
+
784,
|
| 466 |
+
126
|
| 467 |
+
],
|
| 468 |
+
"page_idx": 4
|
| 469 |
+
},
|
| 470 |
+
{
|
| 471 |
+
"type": "text",
|
| 472 |
+
"text": "latent diffusion model [6]. Following this, we detail our uniquely crafted architecture that leverages the Mamba model to enhance motion generation efficiency. This architecture comprises two principal components: the Hierarchical Temporal Mamba (HTM) block, which addresses temporal aspects, and the Bidirectional Spatial Mamba (BSM) block, focusing on spatial dynamics.",
|
| 473 |
+
"bbox": [
|
| 474 |
+
212,
|
| 475 |
+
146,
|
| 476 |
+
787,
|
| 477 |
+
223
|
| 478 |
+
],
|
| 479 |
+
"page_idx": 5
|
| 480 |
+
},
|
| 481 |
+
{
|
| 482 |
+
"type": "text",
|
| 483 |
+
"text": "3.1 Preliminaries",
|
| 484 |
+
"text_level": 1,
|
| 485 |
+
"bbox": [
|
| 486 |
+
215,
|
| 487 |
+
244,
|
| 488 |
+
372,
|
| 489 |
+
258
|
| 490 |
+
],
|
| 491 |
+
"page_idx": 5
|
| 492 |
+
},
|
| 493 |
+
{
|
| 494 |
+
"type": "text",
|
| 495 |
+
"text": "Selective Structured State Space Sequence Model. SSMs particularly through the contributions of structured state space sequence models (S4) and Mamba, have demonstrated exceptional proficiency in handling long sequences. These models operationalize the mapping of a 1-D function or sequence, $x(t) \\in \\mathbb{R} \\mapsto y(t) \\in \\mathbb{R}$ , through a hidden state $h(t) \\in \\mathbb{R}^N$ , employing $\\mathbf{A} \\in \\mathbb{R}^{N \\times N}$ as the evolution parameters, $\\mathbf{B} \\in \\mathbb{R}^{N \\times 1}$ and $\\mathbf{C} \\in \\mathbb{R}^{1 \\times N}$ as the projection parameters, respectively.",
|
| 496 |
+
"bbox": [
|
| 497 |
+
212,
|
| 498 |
+
270,
|
| 499 |
+
787,
|
| 500 |
+
375
|
| 501 |
+
],
|
| 502 |
+
"page_idx": 5
|
| 503 |
+
},
|
| 504 |
+
{
|
| 505 |
+
"type": "text",
|
| 506 |
+
"text": "The discretized system can then be expressed as follows, incorporating a step size $\\Delta$ :",
|
| 507 |
+
"bbox": [
|
| 508 |
+
212,
|
| 509 |
+
376,
|
| 510 |
+
787,
|
| 511 |
+
404
|
| 512 |
+
],
|
| 513 |
+
"page_idx": 5
|
| 514 |
+
},
|
| 515 |
+
{
|
| 516 |
+
"type": "equation",
|
| 517 |
+
"text": "\n$$\n\\begin{array}{c} h _ {t} = \\overline {{\\mathbf {A}}} h _ {t - 1} + \\overline {{\\mathbf {B}}} x _ {t}, \\\\ \\mathbf {C} ^ {\\prime} \\end{array} \\tag {1}\n$$\n",
|
| 518 |
+
"text_format": "latex",
|
| 519 |
+
"bbox": [
|
| 520 |
+
433,
|
| 521 |
+
402,
|
| 522 |
+
784,
|
| 523 |
+
426
|
| 524 |
+
],
|
| 525 |
+
"page_idx": 5
|
| 526 |
+
},
|
| 527 |
+
{
|
| 528 |
+
"type": "equation",
|
| 529 |
+
"text": "\n$$\ny _ {t} = \\mathbf {C} h _ {t}.\n$$\n",
|
| 530 |
+
"text_format": "latex",
|
| 531 |
+
"bbox": [
|
| 532 |
+
434,
|
| 533 |
+
422,
|
| 534 |
+
504,
|
| 535 |
+
436
|
| 536 |
+
],
|
| 537 |
+
"page_idx": 5
|
| 538 |
+
},
|
| 539 |
+
{
|
| 540 |
+
"type": "text",
|
| 541 |
+
"text": "This adaptation facilitates the computation of output through global convolution, leveraging a structured convolutional kernel $\\overline{\\mathbf{K}}$ , which encompasses the entire length $M$ of the input sequence $\\mathbf{x}$ :",
|
| 542 |
+
"bbox": [
|
| 543 |
+
212,
|
| 544 |
+
444,
|
| 545 |
+
787,
|
| 546 |
+
491
|
| 547 |
+
],
|
| 548 |
+
"page_idx": 5
|
| 549 |
+
},
|
| 550 |
+
{
|
| 551 |
+
"type": "equation",
|
| 552 |
+
"text": "\n$$\n\\overline {{\\mathbf {K}}} = \\left(\\underset {-} {\\mathbf {C}} \\overline {{\\mathbf {B}}}, \\mathbf {C} \\overline {{\\mathbf {A}}} \\mathbf {B}, \\dots , \\mathbf {C} \\overline {{\\mathbf {A}}} ^ {M - 1} \\overline {{\\mathbf {B}}}\\right), \\tag {2}\n$$\n",
|
| 553 |
+
"text_format": "latex",
|
| 554 |
+
"bbox": [
|
| 555 |
+
382,
|
| 556 |
+
501,
|
| 557 |
+
784,
|
| 558 |
+
527
|
| 559 |
+
],
|
| 560 |
+
"page_idx": 5
|
| 561 |
+
},
|
| 562 |
+
{
|
| 563 |
+
"type": "equation",
|
| 564 |
+
"text": "\n$$\n\\mathbf {y} = \\mathbf {x} * \\overline {{\\mathbf {K}}}.\n$$\n",
|
| 565 |
+
"text_format": "latex",
|
| 566 |
+
"bbox": [
|
| 567 |
+
387,
|
| 568 |
+
525,
|
| 569 |
+
464,
|
| 570 |
+
541
|
| 571 |
+
],
|
| 572 |
+
"page_idx": 5
|
| 573 |
+
},
|
| 574 |
+
{
|
| 575 |
+
"type": "text",
|
| 576 |
+
"text": "Selective models like Mamba introduce time-varying parameters, deviating from the linear time invariance (LTI) assumption and complicating parallel computation. However, hardware-aware optimizations, such as associative scans, have been developed to address these computational challenges, highlighting the ongoing evolution and application of SSMs in modeling complex temporal dynamics.",
|
| 577 |
+
"bbox": [
|
| 578 |
+
212,
|
| 579 |
+
553,
|
| 580 |
+
787,
|
| 581 |
+
628
|
| 582 |
+
],
|
| 583 |
+
"page_idx": 5
|
| 584 |
+
},
|
| 585 |
+
{
|
| 586 |
+
"type": "text",
|
| 587 |
+
"text": "Latent Motion Diffusion Model. Diffusion probabilistic models offer a significant advancement in motion generation by gradually reducing noise from a Gaussian distribution to a target data distribution $p(x)$ through a T-length learned Markov process [7,22,41,44,47,49,54], giving $\\{\\pmb{x}_t\\}_{t=1}^T$ . In the motion generation, we define our trainable diffusion models with a denoiser $\\epsilon_{\\theta}(x_t,t)$ which anneal the random noise to motion sequence $\\{\\hat{x}_t^{1:N}\\}_{t=1}^T$ iteratively. To address the inefficiencies of applying diffusion models directly to raw motion sequences, we employ a low-dimensional motion latent space for the diffusion process. Given an input condition $c$ , such as a descriptive sentence $\\pmb{w}^{1:N} = \\{w^i\\}_{i=1}^N$ , an action label $a$ from a predefined set $\\mathcal{A}$ , or an empty condition $c = \\emptyset$ , and the motion representation that combines 3D joint rotations, positions, velocities, and foot contact as proposed in [17]. The frozen CLIP [40] text encoder $\\tau_{\\theta}^w$ has been employed to obtain projected text embedding $\\tau_{\\theta}^w(w^{1:N}) \\in \\mathbb{R}^{1 \\times d}$ , thereby conditional denoiser comprised in term of $\\epsilon_{\\theta}(z_t,t,\\tau_{\\theta}(c))$ . The latent diffusion model $\\epsilon_{\\theta}(x_t,t)$ aimed",
|
| 588 |
+
"bbox": [
|
| 589 |
+
212,
|
| 590 |
+
628,
|
| 591 |
+
787,
|
| 592 |
+
843
|
| 593 |
+
],
|
| 594 |
+
"page_idx": 5
|
| 595 |
+
},
|
| 596 |
+
{
|
| 597 |
+
"type": "page_number",
|
| 598 |
+
"text": "6",
|
| 599 |
+
"bbox": [
|
| 600 |
+
217,
|
| 601 |
+
114,
|
| 602 |
+
228,
|
| 603 |
+
126
|
| 604 |
+
],
|
| 605 |
+
"page_idx": 5
|
| 606 |
+
},
|
| 607 |
+
{
|
| 608 |
+
"type": "header",
|
| 609 |
+
"text": "Zeyu and Akide et al.",
|
| 610 |
+
"bbox": [
|
| 611 |
+
271,
|
| 612 |
+
114,
|
| 613 |
+
418,
|
| 614 |
+
128
|
| 615 |
+
],
|
| 616 |
+
"page_idx": 5
|
| 617 |
+
},
|
| 618 |
+
{
|
| 619 |
+
"type": "image",
|
| 620 |
+
"img_path": "images/c47e35af41944d8524c810593fb0260c0e75549c04159ef1f62bdcc5c0d563b4.jpg",
|
| 621 |
+
"image_caption": [
|
| 622 |
+
"Fig. 2: This figure illustrates the architecture of the proposed Motion Mamba model. Each of encoder and decoder blocks consists of a Hierarchical Temporal Mamba block (HTM) and a Bidirectional Spatial Mamba (BSM) block, which possess hierarchical scan and bidirectional scan within SSM layers respectively. This symmetric distribution of scans ensure a balanced and coherence framework across the encoder-decoder architecture."
|
| 623 |
+
],
|
| 624 |
+
"image_footnote": [],
|
| 625 |
+
"bbox": [
|
| 626 |
+
212,
|
| 627 |
+
143,
|
| 628 |
+
787,
|
| 629 |
+
359
|
| 630 |
+
],
|
| 631 |
+
"page_idx": 6
|
| 632 |
+
},
|
| 633 |
+
{
|
| 634 |
+
"type": "text",
|
| 635 |
+
"text": "to generate the human motion sequence in terms of $\\hat{x}^{1:L} = \\{\\hat{x}^i\\}_{i=1}^L$ , where L denotes the sequence length or number of frames [34,35,37,55]. After that we reused the motion Variational AutoEncoder (VAE) $\\mathcal{V} = \\{\\mathcal{E},\\mathcal{D}\\}$ proposed in MLD [6] to manipulate the motion sequence in latent space $z = \\mathcal{E}(x^{1:L})$ , and decompress the intermediate representation to motion sequence by $\\hat{x}^{1:L} = \\mathcal{D}(z) = \\mathcal{D}\\mathcal{E}(x^{1:L})$ [26, 42, 51]. Finally, our latent diffusion model is trained with an objective focusing on minimization of MSE between true and predicted noise in the latent space, facilitating efficient and high-quality motion generation [2, 22].",
|
| 636 |
+
"bbox": [
|
| 637 |
+
212,
|
| 638 |
+
468,
|
| 639 |
+
787,
|
| 640 |
+
592
|
| 641 |
+
],
|
| 642 |
+
"page_idx": 6
|
| 643 |
+
},
|
| 644 |
+
{
|
| 645 |
+
"type": "text",
|
| 646 |
+
"text": "3.2 Motion Mamba",
|
| 647 |
+
"text_level": 1,
|
| 648 |
+
"bbox": [
|
| 649 |
+
215,
|
| 650 |
+
611,
|
| 651 |
+
392,
|
| 652 |
+
625
|
| 653 |
+
],
|
| 654 |
+
"page_idx": 6
|
| 655 |
+
},
|
| 656 |
+
{
|
| 657 |
+
"type": "text",
|
| 658 |
+
"text": "The architecture of the proposed Motion Mamba framework is illustrated in Figure. 2. At its core, Motion Mamba utilizes a denoising U-Net architecture, which is distinguished for its effectiveness in modeling the continuous, temporal sequences of motion frames. This effectiveness is attributed to the inherent long-sequence modeling capacity of the Mamba model. The denoiser, denoted by $\\epsilon_{\\theta}$ , comprises $N$ blocks including encoder $E_{1..N}$ and decoder $D_{1..N}$ . Additionally, the architecture is enhanced with a transformer-based attention mixer block $M$ , designed to augment the model's ability to capture complex temporal dynamics.",
|
| 659 |
+
"bbox": [
|
| 660 |
+
212,
|
| 661 |
+
636,
|
| 662 |
+
787,
|
| 663 |
+
758
|
| 664 |
+
],
|
| 665 |
+
"page_idx": 6
|
| 666 |
+
},
|
| 667 |
+
{
|
| 668 |
+
"type": "equation",
|
| 669 |
+
"text": "\n$$\n\\epsilon_ {\\theta} (x) \\equiv \\{E _ {1 \\dots N}, M, D _ {1 \\dots N} \\}. \\tag {3}\n$$\n",
|
| 670 |
+
"text_format": "latex",
|
| 671 |
+
"bbox": [
|
| 672 |
+
401,
|
| 673 |
+
771,
|
| 674 |
+
785,
|
| 675 |
+
789
|
| 676 |
+
],
|
| 677 |
+
"page_idx": 6
|
| 678 |
+
},
|
| 679 |
+
{
|
| 680 |
+
"type": "text",
|
| 681 |
+
"text": "The encoder blocks are represented as $E_{1..N}$ , arranged sequentially, and the decoder blocks as $D_{1..N}$ , configured in reverse order to facilitate effective bottom-up and top-down information flow. Given that selective operations have signif",
|
| 682 |
+
"bbox": [
|
| 683 |
+
212,
|
| 684 |
+
794,
|
| 685 |
+
787,
|
| 686 |
+
842
|
| 687 |
+
],
|
| 688 |
+
"page_idx": 6
|
| 689 |
+
},
|
| 690 |
+
{
|
| 691 |
+
"type": "header",
|
| 692 |
+
"text": "Motion Mamba",
|
| 693 |
+
"bbox": [
|
| 694 |
+
625,
|
| 695 |
+
114,
|
| 696 |
+
730,
|
| 697 |
+
126
|
| 698 |
+
],
|
| 699 |
+
"page_idx": 6
|
| 700 |
+
},
|
| 701 |
+
{
|
| 702 |
+
"type": "page_number",
|
| 703 |
+
"text": "7",
|
| 704 |
+
"bbox": [
|
| 705 |
+
774,
|
| 706 |
+
114,
|
| 707 |
+
784,
|
| 708 |
+
126
|
| 709 |
+
],
|
| 710 |
+
"page_idx": 6
|
| 711 |
+
},
|
| 712 |
+
{
|
| 713 |
+
"type": "code",
|
| 714 |
+
"sub_type": "algorithm",
|
| 715 |
+
"code_caption": [
|
| 716 |
+
"Algorithm 1 Hierarchical Temporal Mamba (HTM) Block."
|
| 717 |
+
],
|
| 718 |
+
"code_body": "Require: compressed latent representations $z:(T,B,C)$ \nEnsure: transformed representations $z_{\\mathrm{HTM}}:(T,B,E)$ \n1:/\\* linear projection layer \\*/ \n2: $x,z:(T,B,E)\\gets \\mathrm{Linear}(z)$ \n3:/\\* set of scans and memory matrices \\*/ \n4: $K = \\{S_{2N_n - 1},S_{2N_{n - 1} - 1},\\ldots ,S_1\\}$ \n5: Memory matrices: $\\{A_1,\\dots ,A_k\\}$ \n6: for each scan $S_{i}$ in $K$ do \n7: $x_o^{\\prime}:(T,B,E)\\gets \\mathrm{Conv1D}(x)$ \n8: $B_{o},C_{o},\\varDelta_{o}\\gets \\mathrm{Linear}(x_{o}^{\\prime})$ \n9: Transform $\\overline{A}_o$ and $\\overline{B}_o$ using $\\varDelta_{o}$ \n10: $O_{i}\\gets \\mathrm{SSMA}_{i,x}(x_{o}^{\\prime})$ \n11: end for \n12:/\\* aggregation of outputs \\*/ \n13: $z_{\\mathrm{HTM}}:(T,B,E)\\gets \\mathrm{Linear(Aggregate(\\{O_1,\\ldots,O_k\\}))}$ \n14: Return: $z_{\\mathrm{HTM}}$",
|
| 719 |
+
"bbox": [
|
| 720 |
+
215,
|
| 721 |
+
166,
|
| 722 |
+
602,
|
| 723 |
+
391
|
| 724 |
+
],
|
| 725 |
+
"page_idx": 7
|
| 726 |
+
},
|
| 727 |
+
{
|
| 728 |
+
"type": "text",
|
| 729 |
+
"text": "icantly lower computational complexity compared to attention-based methods, we have increased the number of scans to achieve higher quality generations. Concurrently, it is imperative to maintain a balance between the model's parameters and its efficiency. Thereby, a novel aspect of our model is the introduction of a hierarchical scan strategy, characterized by a sequence of scan numbers as,",
|
| 730 |
+
"bbox": [
|
| 731 |
+
212,
|
| 732 |
+
421,
|
| 733 |
+
787,
|
| 734 |
+
497
|
| 735 |
+
],
|
| 736 |
+
"page_idx": 7
|
| 737 |
+
},
|
| 738 |
+
{
|
| 739 |
+
"type": "equation",
|
| 740 |
+
"text": "\n$$\nK = \\left\\{S _ {2 N - 1}, S _ {2 (N - 1) - 1}, \\dots , S _ {1} \\right\\}. \\tag {4}\n$$\n",
|
| 741 |
+
"text_format": "latex",
|
| 742 |
+
"bbox": [
|
| 743 |
+
379,
|
| 744 |
+
507,
|
| 745 |
+
785,
|
| 746 |
+
525
|
| 747 |
+
],
|
| 748 |
+
"page_idx": 7
|
| 749 |
+
},
|
| 750 |
+
{
|
| 751 |
+
"type": "text",
|
| 752 |
+
"text": "This sequence specifies the number of scans allocated to each layer, in descending order of complexity. For instance, the uppermost encoder layer, $E_{1}$ , and the lowermost decoder layer, $D_{N}$ , are allocated $S_{2N - 1}$ scans, indicating the highest scanning complexity. Conversely, the lowest encoder layer, $E_{N}$ , and the uppermost decoder layer, $D_{1}$ , are assigned $S_{1}$ scans, reflecting the lowest level of scanning complexity.",
|
| 753 |
+
"bbox": [
|
| 754 |
+
212,
|
| 755 |
+
534,
|
| 756 |
+
787,
|
| 757 |
+
626
|
| 758 |
+
],
|
| 759 |
+
"page_idx": 7
|
| 760 |
+
},
|
| 761 |
+
{
|
| 762 |
+
"type": "equation",
|
| 763 |
+
"text": "\n$$\nE _ {i} (S) = \\left\\{ \\begin{array}{l l} S _ {2 N - 1} & \\text {f o r} i = 1 \\\\ S _ {2 (N - i) - 1} & \\text {f o r} i = 2, \\dots , N - 1 \\\\ S _ {1} & \\text {f o r} i = N \\end{array} \\right. \\tag {5}\n$$\n",
|
| 764 |
+
"text_format": "latex",
|
| 765 |
+
"bbox": [
|
| 766 |
+
344,
|
| 767 |
+
635,
|
| 768 |
+
785,
|
| 769 |
+
691
|
| 770 |
+
],
|
| 771 |
+
"page_idx": 7
|
| 772 |
+
},
|
| 773 |
+
{
|
| 774 |
+
"type": "equation",
|
| 775 |
+
"text": "\n$$\nD _ {j} (S) = \\left\\{ \\begin{array}{l l} S _ {2 N - 1} & \\text {f o r} j = N \\\\ S _ {2 (N - j) - 1} & \\text {f o r} j = N - 1, \\dots , 2 \\\\ S _ {1} & \\text {f o r} j = 1 \\end{array} \\right. \\tag {6}\n$$\n",
|
| 776 |
+
"text_format": "latex",
|
| 777 |
+
"bbox": [
|
| 778 |
+
341,
|
| 779 |
+
703,
|
| 780 |
+
785,
|
| 781 |
+
760
|
| 782 |
+
],
|
| 783 |
+
"page_idx": 7
|
| 784 |
+
},
|
| 785 |
+
{
|
| 786 |
+
"type": "text",
|
| 787 |
+
"text": "This hierarchical scanning approach ensures that processing capabilities are evenly distributed throughout the encoder-decoder architecture., facilitating a detailed and nuanced analysis of temporal sequences. Within this structured framework, each denoiser is equipped with a specialized Hierarchical Temporal Mamba (HTM) block, which serves to augment the model's ability to process",
|
| 788 |
+
"bbox": [
|
| 789 |
+
212,
|
| 790 |
+
763,
|
| 791 |
+
787,
|
| 792 |
+
842
|
| 793 |
+
],
|
| 794 |
+
"page_idx": 7
|
| 795 |
+
},
|
| 796 |
+
{
|
| 797 |
+
"type": "page_number",
|
| 798 |
+
"text": "8",
|
| 799 |
+
"bbox": [
|
| 800 |
+
217,
|
| 801 |
+
114,
|
| 802 |
+
228,
|
| 803 |
+
126
|
| 804 |
+
],
|
| 805 |
+
"page_idx": 7
|
| 806 |
+
},
|
| 807 |
+
{
|
| 808 |
+
"type": "header",
|
| 809 |
+
"text": "Zeyu and Akide et al.",
|
| 810 |
+
"bbox": [
|
| 811 |
+
271,
|
| 812 |
+
114,
|
| 813 |
+
418,
|
| 814 |
+
128
|
| 815 |
+
],
|
| 816 |
+
"page_idx": 7
|
| 817 |
+
},
|
| 818 |
+
{
|
| 819 |
+
"type": "code",
|
| 820 |
+
"sub_type": "algorithm",
|
| 821 |
+
"code_caption": [
|
| 822 |
+
"Algorithm 2 Bidirectional Spatial Mamba (BSM) Block."
|
| 823 |
+
],
|
| 824 |
+
"code_body": "Require: compressed latent representations $z:(T,B,C)$ \nEnsure: transformed representations $z_{\\mathrm{BSM}}:(C,B,E)$ \n1:/\\* dimension rearrangement \\*/ \n2: $z^{\\prime}:(C,B,T)\\gets$ Rearrange(z) \n3:/\\* linear projection after normalization \\*/ \n4: $z^{\\prime}:(C,B,T)\\gets$ Norm(z) \n5:x,z:(C,B,E) $\\leftarrow$ Linear(z') \n6: for o in {forward, backward} do \n7: $x_o':(C,B,E)\\gets$ Conv1D(x) \n8: $B_{o},C_{o},\\varDelta_{o}\\gets$ Linear $(x_o^{\\prime})$ \n9: Transform $\\overline{A}_o$ and $\\overline{B}_o$ using $\\varDelta_{o}$ \n10: $y_{o}\\gets$ SSM(Ao,Bo,Co) \n11: end for \n12:/\\* gating and summing outputs \\*/ \n13: $z_{\\mathrm{BSM}}:(C,B,T)\\gets$ GateAndSum(yforward,ybackward,z) \n14: $z_{\\mathrm{BSM}}^{\\prime}:(T,B,C)\\gets$ Rearrange(z) \n15: Return: $z_{\\mathrm{BSM}}^{\\prime}$",
|
| 825 |
+
"bbox": [
|
| 826 |
+
215,
|
| 827 |
+
166,
|
| 828 |
+
619,
|
| 829 |
+
404
|
| 830 |
+
],
|
| 831 |
+
"page_idx": 8
|
| 832 |
+
},
|
| 833 |
+
{
|
| 834 |
+
"type": "text",
|
| 835 |
+
"text": "temporal information effectively. Additionally, the proposed Motion Mamba incorporates an attention-based mixer block denoted as $M$ , strategically integrated to enhance conditional fusion.",
|
| 836 |
+
"bbox": [
|
| 837 |
+
212,
|
| 838 |
+
436,
|
| 839 |
+
784,
|
| 840 |
+
479
|
| 841 |
+
],
|
| 842 |
+
"page_idx": 8
|
| 843 |
+
},
|
| 844 |
+
{
|
| 845 |
+
"type": "text",
|
| 846 |
+
"text": "Hierarchical Temporal Mamba (HTM) block processes compressed latent representations, denoted as $z$ , with the dimensions $(T,B,C)$ , of which procedure shown in Algorithm 1. Here, $T$ signifies the temporal dimension, as specified in the Variational AutoEncoder (VAE) framework. Initially, the input $z$ is subjected to a linear projection layer, producing transformed representations $x$ and $z$ with dimension $E$ . Our analysis revealed an increased density of motion within the lower-level feature spaces. Consequently, we developed a hierarchical scanning methodology that is executed at various depths of the network. This approach not only accommodates the diverse motion densities, but also significantly reduces computational overhead. This step utilizes a hierarchically structured set of scans, $K = \\{S_{2N_n - 1}, S_{2N_{n - 1} - 1}, \\ldots, S_1\\}$ , in conjunction with a corresponding series of memory matrices $\\{A_1, \\ldots, A_k\\}$ . Each sub-SSM scan first applies a 1-D convolution to $x$ , resulting in $x_o'$ . $x_o'$ is then linearly projected to derive $B_o$ , $C_o$ , and $\\Delta_o$ . These projections $B_o$ , $C_o$ use $\\Delta_o$ to effect transformations in $\\overline{A}_o$ and $\\overline{B}_o$ , respectively. After executing a sequence of SSM scans $\\{SSM_{A_1,x}, SSM_{A_2,x}, \\ldots, SSM_{A_k,x}\\}$ , a set of outputs $\\{O_1, \\ldots, O_k\\}$ is compiled. This collection is subsequently aggregated via a linear projection to obtain the final output of the HTM block.",
|
| 847 |
+
"bbox": [
|
| 848 |
+
212,
|
| 849 |
+
482,
|
| 850 |
+
787,
|
| 851 |
+
753
|
| 852 |
+
],
|
| 853 |
+
"page_idx": 8
|
| 854 |
+
},
|
| 855 |
+
{
|
| 856 |
+
"type": "text",
|
| 857 |
+
"text": "Bidirectional Spatial Mamba (BSM) block focuses on enhancing latent representation learning through a novel approach of dimension rearrangement and bidirectional scanning, of which the process is shown in Algorithm 2. Initially, it alters the original input dimensions from $(T,B,C)$ to $(C,B,T)$ , effectively swap",
|
| 858 |
+
"bbox": [
|
| 859 |
+
212,
|
| 860 |
+
779,
|
| 861 |
+
787,
|
| 862 |
+
840
|
| 863 |
+
],
|
| 864 |
+
"page_idx": 8
|
| 865 |
+
},
|
| 866 |
+
{
|
| 867 |
+
"type": "header",
|
| 868 |
+
"text": "Motion Mamba",
|
| 869 |
+
"bbox": [
|
| 870 |
+
625,
|
| 871 |
+
114,
|
| 872 |
+
730,
|
| 873 |
+
126
|
| 874 |
+
],
|
| 875 |
+
"page_idx": 8
|
| 876 |
+
},
|
| 877 |
+
{
|
| 878 |
+
"type": "page_number",
|
| 879 |
+
"text": "9",
|
| 880 |
+
"bbox": [
|
| 881 |
+
774,
|
| 882 |
+
114,
|
| 883 |
+
784,
|
| 884 |
+
126
|
| 885 |
+
],
|
| 886 |
+
"page_idx": 8
|
| 887 |
+
},
|
| 888 |
+
{
|
| 889 |
+
"type": "text",
|
| 890 |
+
"text": "ping the temporal and channel dimensions. After this rearrangement, the input, now denoted $z'$ , undergoes a linear projection after normalization, resulting in dimensions $x$ and $z$ of size $E$ . The process involves bidirectional scanning of the latent channel dimension, where $\\mathbf{x}$ is subjected to a 1-D convolution, yielding $\\mathbf{x}_o'$ for both forward and backward directions. Each $\\mathbf{x}_o'$ is then linearly projected to obtain $B_o$ , $C_o$ , and $\\Delta_o$ , which are utilized to transform $\\overline{A}_o$ and $\\overline{B}_o$ , respectively. The final output token sequence, $\\mathbf{z}_1$ , is computed by gating and summing the forward $y_{\\mathrm{forward}}$ and backward $y_{\\mathrm{backward}}$ output with $\\mathbf{z}$ . This component is engineered to decode the structured latent skeleton by analyzing data from both forward and reverse viewpoints. Its main objective is to ensure the seamless continuity of information flow, thereby substantially enhancing the model's ability to generate accurate motion. This is achieved through the maintenance of a dense informational exchange, which is critical for the model's performance.",
|
| 891 |
+
"bbox": [
|
| 892 |
+
212,
|
| 893 |
+
146,
|
| 894 |
+
787,
|
| 895 |
+
343
|
| 896 |
+
],
|
| 897 |
+
"page_idx": 9
|
| 898 |
+
},
|
| 899 |
+
{
|
| 900 |
+
"type": "text",
|
| 901 |
+
"text": "4 Experiments",
|
| 902 |
+
"text_level": 1,
|
| 903 |
+
"bbox": [
|
| 904 |
+
215,
|
| 905 |
+
362,
|
| 906 |
+
375,
|
| 907 |
+
380
|
| 908 |
+
],
|
| 909 |
+
"page_idx": 9
|
| 910 |
+
},
|
| 911 |
+
{
|
| 912 |
+
"type": "text",
|
| 913 |
+
"text": "4.1 Datasets",
|
| 914 |
+
"text_level": 1,
|
| 915 |
+
"bbox": [
|
| 916 |
+
215,
|
| 917 |
+
390,
|
| 918 |
+
334,
|
| 919 |
+
402
|
| 920 |
+
],
|
| 921 |
+
"page_idx": 9
|
| 922 |
+
},
|
| 923 |
+
{
|
| 924 |
+
"type": "text",
|
| 925 |
+
"text": "We evaluate our proposed Motion Mamba on two prominent Text-to-Motion synthesis benchmarks as follows:",
|
| 926 |
+
"bbox": [
|
| 927 |
+
212,
|
| 928 |
+
412,
|
| 929 |
+
785,
|
| 930 |
+
440
|
| 931 |
+
],
|
| 932 |
+
"page_idx": 9
|
| 933 |
+
},
|
| 934 |
+
{
|
| 935 |
+
"type": "text",
|
| 936 |
+
"text": "HumanML3D. The HumanML3D [17] dataset aggregates 14,616 motions sourced from the AMASS [33] and HumanAct12 [18] datasets, with each motion accompanied by three textual descriptions, culminating in 44,970 scripts. This dataset spans a wide range of actions such as exercising, dancing, and acrobatics, presenting a rich motion-language corpus.",
|
| 937 |
+
"bbox": [
|
| 938 |
+
212,
|
| 939 |
+
441,
|
| 940 |
+
797,
|
| 941 |
+
516
|
| 942 |
+
],
|
| 943 |
+
"page_idx": 9
|
| 944 |
+
},
|
| 945 |
+
{
|
| 946 |
+
"type": "text",
|
| 947 |
+
"text": "KIT-ML. The KIT-ML dataset [38] is comprised of 3,911 motions paired with 6,278 textual descriptions, serving as a compact yet effective benchmark for evaluation. For both datasets, the pose representation adopted is derived from T2M [17], ensuring consistency in motion representation across evaluations.",
|
| 948 |
+
"bbox": [
|
| 949 |
+
212,
|
| 950 |
+
517,
|
| 951 |
+
785,
|
| 952 |
+
579
|
| 953 |
+
],
|
| 954 |
+
"page_idx": 9
|
| 955 |
+
},
|
| 956 |
+
{
|
| 957 |
+
"type": "text",
|
| 958 |
+
"text": "4.2 Evaluation Metrics",
|
| 959 |
+
"text_level": 1,
|
| 960 |
+
"bbox": [
|
| 961 |
+
215,
|
| 962 |
+
595,
|
| 963 |
+
419,
|
| 964 |
+
611
|
| 965 |
+
],
|
| 966 |
+
"page_idx": 9
|
| 967 |
+
},
|
| 968 |
+
{
|
| 969 |
+
"type": "text",
|
| 970 |
+
"text": "We adapt the standard evaluation metrics on following aspects throughout our experiments, including: Generation Quality. We implement a Fréchet inception distance (FID) [21] to quantify the realism and diversity of motion generated by models. Moreover, we use multi-modal distance (MM Dist) to measure the distance between motions and texts and assess motion-text alignment. Diversity. We use the diversity metric to measure motion diversity, which calculates variance in features extracted from the motions. Additionally, we employ multimodality (MModality) to assess diversity within generated motions sharing the same text description.",
|
| 971 |
+
"bbox": [
|
| 972 |
+
212,
|
| 973 |
+
618,
|
| 974 |
+
787,
|
| 975 |
+
753
|
| 976 |
+
],
|
| 977 |
+
"page_idx": 9
|
| 978 |
+
},
|
| 979 |
+
{
|
| 980 |
+
"type": "text",
|
| 981 |
+
"text": "4.3 Comparative Studies",
|
| 982 |
+
"text_level": 1,
|
| 983 |
+
"bbox": [
|
| 984 |
+
215,
|
| 985 |
+
772,
|
| 986 |
+
434,
|
| 987 |
+
787
|
| 988 |
+
],
|
| 989 |
+
"page_idx": 9
|
| 990 |
+
},
|
| 991 |
+
{
|
| 992 |
+
"type": "text",
|
| 993 |
+
"text": "We evaluate our method against the state-of-the-art methods on the HumanML3D [17] and KIT-ML [38] datasets. We train our Motion Mamba with HTM arrangement strategy MM ( $\\{S_{2N_n - 1},\\ldots ,S_1\\}$ ), BSM bidirectional block strategy on the",
|
| 994 |
+
"bbox": [
|
| 995 |
+
212,
|
| 996 |
+
794,
|
| 997 |
+
794,
|
| 998 |
+
842
|
| 999 |
+
],
|
| 1000 |
+
"page_idx": 9
|
| 1001 |
+
},
|
| 1002 |
+
{
|
| 1003 |
+
"type": "page_number",
|
| 1004 |
+
"text": "10",
|
| 1005 |
+
"bbox": [
|
| 1006 |
+
217,
|
| 1007 |
+
114,
|
| 1008 |
+
235,
|
| 1009 |
+
126
|
| 1010 |
+
],
|
| 1011 |
+
"page_idx": 9
|
| 1012 |
+
},
|
| 1013 |
+
{
|
| 1014 |
+
"type": "header",
|
| 1015 |
+
"text": "Zeyu and Akide et al.",
|
| 1016 |
+
"bbox": [
|
| 1017 |
+
271,
|
| 1018 |
+
114,
|
| 1019 |
+
416,
|
| 1020 |
+
128
|
| 1021 |
+
],
|
| 1022 |
+
"page_idx": 9
|
| 1023 |
+
},
|
| 1024 |
+
{
|
| 1025 |
+
"type": "table",
|
| 1026 |
+
"img_path": "images/f22668e60db39236e4993bff3b2a318b4fa3c2d281126a45986137457738618f.jpg",
|
| 1027 |
+
"table_caption": [
|
| 1028 |
+
"Table 1: Comparison of text-conditional motion synthesis on HumanML3D [17]. These metrics are evaluated by the motion encoder from [17]. Empty MModality indicates the non-diverse generation methods. We employ real motion as a reference and sort all methods by descending FIDs. The right arrow $\\rightarrow$ means that the closer to the real motion, the better. Bold and underline indicate the best and second best result."
|
| 1029 |
+
],
|
| 1030 |
+
"table_footnote": [],
|
| 1031 |
+
"table_body": "<table><tr><td rowspan=\"2\">Method</td><td colspan=\"3\">R Precision ↑</td><td rowspan=\"2\">FID↓</td><td rowspan=\"2\">MM Dist↓</td><td rowspan=\"2\" colspan=\"2\">Diversity→MModality↑</td></tr><tr><td>Top 1</td><td>Top 2</td><td>Top 3</td></tr><tr><td>Real</td><td>0.511±.003</td><td>0.703±.003</td><td>0.797±.002</td><td>0.002±.000</td><td>2.974±.008</td><td>9.503±.065</td><td>-</td></tr><tr><td>Seq2Seq [39]</td><td>0.180±.002</td><td>0.300±.002</td><td>0.396±.002</td><td>11.75±.035</td><td>5.529±.007</td><td>6.223±.061</td><td>-</td></tr><tr><td>LJ2P [1]</td><td>0.246±.001</td><td>0.387±.002</td><td>0.486±.002</td><td>11.02±.046</td><td>5.296±.008</td><td>7.676±.058</td><td>-</td></tr><tr><td>T2G [5]</td><td>0.165±.001</td><td>0.267±.002</td><td>0.345±.002</td><td>7.664±.030</td><td>6.030±.008</td><td>6.409±.071</td><td>-</td></tr><tr><td>Hier [9]</td><td>0.301±.002</td><td>0.425±.002</td><td>0.552±.004</td><td>6.532±.024</td><td>5.012±.018</td><td>8.332±.042</td><td>-</td></tr><tr><td>TEMOS [37]</td><td>0.424±.002</td><td>0.612±.002</td><td>0.722±.002</td><td>3.734±.028</td><td>3.703±.008</td><td>8.973±.071</td><td>0.368±.018</td></tr><tr><td>T2M [17]</td><td>0.457±.002</td><td>0.639±.003</td><td>0.740±.003</td><td>1.067±.002</td><td>3.340±.008</td><td>9.188±.002</td><td>2.090±.083</td></tr><tr><td>MDM [49]</td><td>0.320±.005</td><td>0.498±.004</td><td>0.611±.007</td><td>0.544±.044</td><td>5.566±.027</td><td>9.559±.086</td><td>2.799±.072</td></tr><tr><td>MotionDiffuse [54]</td><td>0.491±.001</td><td>0.681±.001</td><td>0.782±.001</td><td>0.630±.001</td><td>3.113±.001</td><td>9.410±.049</td><td>1.553±.042</td></tr><tr><td>MLD [6]</td><td>0.481±.003</td><td>0.673±.003</td><td>0.772±.002</td><td>0.473±.013</td><td>3.196±.010</td><td>9.724±.082</td><td>2.413±.079</td></tr><tr><td>Motion Mamba (Ours)</td><td>0.502±.003</td><td>0.693±.002</td><td>0.792±.002</td><td>0.281±.009</td><td>3.060±.058</td><td>9.871±.084</td><td>2.294±.058</td></tr></table>",
|
| 1032 |
+
"bbox": [
|
| 1033 |
+
215,
|
| 1034 |
+
213,
|
| 1035 |
+
785,
|
| 1036 |
+
380
|
| 1037 |
+
],
|
| 1038 |
+
"page_idx": 10
|
| 1039 |
+
},
|
| 1040 |
+
{
|
| 1041 |
+
"type": "table",
|
| 1042 |
+
"img_path": "images/d4a6902799c96d2278dd1a098881e5ed26a9fd074019736e497d0d9b158b0591.jpg",
|
| 1043 |
+
"table_caption": [
|
| 1044 |
+
"Table 2: We involve KIT-ML [38] dataset and evaluate the SOTA methods on the text-to-motion task."
|
| 1045 |
+
],
|
| 1046 |
+
"table_footnote": [],
|
| 1047 |
+
"table_body": "<table><tr><td rowspan=\"2\">Method</td><td colspan=\"3\">R Precision ↑</td><td rowspan=\"2\">FID↓</td><td rowspan=\"2\">MM Dist↓</td><td rowspan=\"2\" colspan=\"2\">Diversity→MModality↑</td></tr><tr><td>Top 1</td><td>Top 2</td><td>Top 3</td></tr><tr><td>Real</td><td>0.424±.005</td><td>0.649±.006</td><td>0.779±.006</td><td>0.031±.004</td><td>2.788±.012</td><td>11.08±.097</td><td>-</td></tr><tr><td>Seq2Seq [39]</td><td>0.103±.003</td><td>0.178±.005</td><td>0.241±.006</td><td>24.86±.348</td><td>7.960±.031</td><td>6.744±.106</td><td>-</td></tr><tr><td>T2G [5]</td><td>0.156±.004</td><td>0.255±.004</td><td>0.338±.005</td><td>12.12±.183</td><td>6.964±.029</td><td>9.334±.079</td><td>-</td></tr><tr><td>LJ2P [1]</td><td>0.221±.005</td><td>0.373±.004</td><td>0.483±.005</td><td>6.545±.072</td><td>5.147±.030</td><td>9.073±.100</td><td>-</td></tr><tr><td>Hier [9]</td><td>0.255±.006</td><td>0.432±.007</td><td>0.531±.007</td><td>5.203±.107</td><td>4.986±.027</td><td>9.563±.072</td><td>2.090±.083</td></tr><tr><td>TEMOS [37]</td><td>0.353±.006</td><td>0.561±.007</td><td>0.687±.005</td><td>3.717±.051</td><td>3.417±.019</td><td>10.84±.100</td><td>0.532±.034</td></tr><tr><td>T2M [17]</td><td>0.370±.005</td><td>0.569±.007</td><td>0.693±.007</td><td>2.770±.109</td><td>3.401±.008</td><td>10.91±.119</td><td>1.482±.065</td></tr><tr><td>MDM [49]</td><td>0.164±.004</td><td>0.291±.004</td><td>0.396±.004</td><td>0.497±.021</td><td>9.191±.022</td><td>10.85±.109</td><td>1.907±.214</td></tr><tr><td>MotionDiffuse [54]</td><td>0.417±.004</td><td>0.621±.004</td><td>0.739±.004</td><td>1.954±.062</td><td>2.958±.005</td><td>11.10±.143</td><td>0.730±.013</td></tr><tr><td>MLD [6]</td><td>0.390±.008</td><td>0.609±.008</td><td>0.734±.007</td><td>0.404±.027</td><td>3.204±.027</td><td>10.80±.117</td><td>2.192±.071</td></tr><tr><td>Motion Mamba (Ours)</td><td>0.419±.006</td><td>0.645±.005</td><td>0.765±.006</td><td>0.307±.041</td><td>3.021±.025</td><td>11.02±.098</td><td>1.678±.064</td></tr></table>",
|
| 1048 |
+
"bbox": [
|
| 1049 |
+
215,
|
| 1050 |
+
414,
|
| 1051 |
+
785,
|
| 1052 |
+
583
|
| 1053 |
+
],
|
| 1054 |
+
"page_idx": 10
|
| 1055 |
+
},
|
| 1056 |
+
{
|
| 1057 |
+
"type": "text",
|
| 1058 |
+
"text": "latent dimension $= 2$ with 11 layers. We evaluate our model and previous works with suggested metrics in HumanML3D [17] and calculate $95\\%$ confidence interval by repeat evaluation 20 times. The results for the HumanML3D dataset are presented in Table 1. Our model outperforms other methods significantly across various evaluation metrics, including FID, R precision, multi-modal distance, and diversity. For instance, our Motion Mamba outperforms previous best diffusion based motion generation MLD by $40.5\\%$ in terms of FID, and up to $10\\%$ improvement on R Precision, we aslo obatined best MModality by 3.060. The results for the KIT-ML dataset are presented in Table 2. We have also outperformed other well-established methods in FID and multi-modal distance.",
|
| 1059 |
+
"bbox": [
|
| 1060 |
+
212,
|
| 1061 |
+
595,
|
| 1062 |
+
787,
|
| 1063 |
+
748
|
| 1064 |
+
],
|
| 1065 |
+
"page_idx": 10
|
| 1066 |
+
},
|
| 1067 |
+
{
|
| 1068 |
+
"type": "text",
|
| 1069 |
+
"text": "4.4 Ablation Studies",
|
| 1070 |
+
"text_level": 1,
|
| 1071 |
+
"bbox": [
|
| 1072 |
+
215,
|
| 1073 |
+
768,
|
| 1074 |
+
401,
|
| 1075 |
+
782
|
| 1076 |
+
],
|
| 1077 |
+
"page_idx": 10
|
| 1078 |
+
},
|
| 1079 |
+
{
|
| 1080 |
+
"type": "text",
|
| 1081 |
+
"text": "We concluded the ablation studies including long sequence evaluation, hierarchical design with HTM, bidirectional design in the BSM, number of latent dimensions, and number of layers of our proposed motion mamba in Table 4.",
|
| 1082 |
+
"bbox": [
|
| 1083 |
+
212,
|
| 1084 |
+
792,
|
| 1085 |
+
785,
|
| 1086 |
+
838
|
| 1087 |
+
],
|
| 1088 |
+
"page_idx": 10
|
| 1089 |
+
},
|
| 1090 |
+
{
|
| 1091 |
+
"type": "header",
|
| 1092 |
+
"text": "Motion Mamba",
|
| 1093 |
+
"bbox": [
|
| 1094 |
+
625,
|
| 1095 |
+
114,
|
| 1096 |
+
730,
|
| 1097 |
+
126
|
| 1098 |
+
],
|
| 1099 |
+
"page_idx": 10
|
| 1100 |
+
},
|
| 1101 |
+
{
|
| 1102 |
+
"type": "page_number",
|
| 1103 |
+
"text": "11",
|
| 1104 |
+
"bbox": [
|
| 1105 |
+
767,
|
| 1106 |
+
114,
|
| 1107 |
+
782,
|
| 1108 |
+
126
|
| 1109 |
+
],
|
| 1110 |
+
"page_idx": 10
|
| 1111 |
+
},
|
| 1112 |
+
{
|
| 1113 |
+
"type": "table",
|
| 1114 |
+
"img_path": "images/f10dee507e46e0ced2ba07214fcf83530aab51e42cf950b0ed7dd0baa6212667.jpg",
|
| 1115 |
+
"table_caption": [
|
| 1116 |
+
"Table 3: In order to evaluate the models' capability in long sequence motion generation, we compared our method with an existing approach on the recently introduced HumanML3D-LS dataset. This dataset comprises motion sequences longer than 190 frames from the original evaluation set. Our model demonstrates superior performance compared to other methods."
|
| 1117 |
+
],
|
| 1118 |
+
"table_footnote": [],
|
| 1119 |
+
"table_body": "<table><tr><td rowspan=\"2\">Method</td><td colspan=\"3\">R Precision ↑</td><td rowspan=\"2\">FID↓</td><td rowspan=\"2\">MM Dist↓</td><td rowspan=\"2\">Diversity→</td><td rowspan=\"2\">MModality↑</td></tr><tr><td>Top 1</td><td>Top 2</td><td>Top 3</td></tr><tr><td>Real</td><td>0.437±.003</td><td>0.622±.004</td><td>0.721±.004</td><td>0.004±.000</td><td>3.343±.015</td><td>8.423±.090</td><td>-</td></tr><tr><td>MDM [49]</td><td>0.368±.005</td><td>0.553±.006</td><td>0.672±.005</td><td>0.802±.044</td><td>3.860±.025</td><td>8.817±.068</td><td>-</td></tr><tr><td>MotionDiffuse [54]</td><td>0.367±.004</td><td>0.521±.004</td><td>0.623±.004</td><td>2.460±.062</td><td>3.789±.005</td><td>8.707±.143</td><td>1.602±.013</td></tr><tr><td>MLD [6]</td><td>0.403±.005</td><td>0.584±.005</td><td>0.690±.005</td><td>0.952±.020</td><td>3.580±.016</td><td>9.050±.085</td><td>2.711±.104</td></tr><tr><td>Motion Mamba (Ours)</td><td>0.417±.003</td><td>0.606±.003</td><td>0.713±.004</td><td>0.668±.019</td><td>3.435±.015</td><td>9.021±.070</td><td>2.373±.084</td></tr></table>",
|
| 1120 |
+
"bbox": [
|
| 1121 |
+
215,
|
| 1122 |
+
212,
|
| 1123 |
+
785,
|
| 1124 |
+
316
|
| 1125 |
+
],
|
| 1126 |
+
"page_idx": 11
|
| 1127 |
+
},
|
| 1128 |
+
{
|
| 1129 |
+
"type": "text",
|
| 1130 |
+
"text": "Long Sequence Motion Generation. The HumanML3D [17] dataset exhibits a long-tailed and right-skewed distribution with a significant proportion of long-sequence human motions, as shown in Figure 3. We suggest previous study overlooked the challenges in the long-sequence generation problem. Thus, we introduce a new dataset variant, HumanML3D-LS, comprising motion sequences longer than 190 frames extracted from the original test set. This addition allows us to showcase our capability in generating long-sequence motions. Subsequently, we eval",
|
| 1131 |
+
"bbox": [
|
| 1132 |
+
212,
|
| 1133 |
+
335,
|
| 1134 |
+
544,
|
| 1135 |
+
532
|
| 1136 |
+
],
|
| 1137 |
+
"page_idx": 11
|
| 1138 |
+
},
|
| 1139 |
+
{
|
| 1140 |
+
"type": "image",
|
| 1141 |
+
"img_path": "images/aae2391766f3828f2e11babdcf89aea649a1a4ec6a3b1b4196a8a1944d3acb76.jpg",
|
| 1142 |
+
"image_caption": [
|
| 1143 |
+
"Fig. 3: The figure shows a long tail distribution of the HumanML3D [17], which has a significant proportion of long-sequence human motions."
|
| 1144 |
+
],
|
| 1145 |
+
"image_footnote": [],
|
| 1146 |
+
"bbox": [
|
| 1147 |
+
555,
|
| 1148 |
+
334,
|
| 1149 |
+
787,
|
| 1150 |
+
443
|
| 1151 |
+
],
|
| 1152 |
+
"page_idx": 11
|
| 1153 |
+
},
|
| 1154 |
+
{
|
| 1155 |
+
"type": "text",
|
| 1156 |
+
"text": "uate the performance of our method on HumanML3D-LS and compare it with other diffusion-based motion generation approaches. The comparative results are presented in Table 3. Motion Mamba by leverage the benefits on long-range dependency modeling make it well suitable for long sequence motion generation.",
|
| 1157 |
+
"bbox": [
|
| 1158 |
+
212,
|
| 1159 |
+
532,
|
| 1160 |
+
787,
|
| 1161 |
+
594
|
| 1162 |
+
],
|
| 1163 |
+
"page_idx": 11
|
| 1164 |
+
},
|
| 1165 |
+
{
|
| 1166 |
+
"type": "text",
|
| 1167 |
+
"text": "Hierarchical Design with HTM. In our ablation studies, we observed a slight improvement upon reversing the scan order from a lower to a higher level, specifically transitioning from MM $\\{S_1,\\ldots ,S_N\\}$ to MM $\\{S_N,\\dots ,S_1\\}$ . This enhancement suggests a correlation with the increase in temporal motion density within the lower-level feature spaces. Furthermore, to achieve the optimal result, we introduce the hierarchical design to arrange the scanning frequency, resulting in the sequence MM $\\{S_{2N_n - 1},\\ldots ,S_1\\}$ . This expansion in the number of scans led to a performance increase. We attribute this enhancement to the observation that individual selective scan operations significantly reduce the parameter count, especially when compared to the parameter-intensive constructs of self-attention and feedforward network blocks prevalent in transformer architectures.",
|
| 1168 |
+
"bbox": [
|
| 1169 |
+
212,
|
| 1170 |
+
595,
|
| 1171 |
+
787,
|
| 1172 |
+
762
|
| 1173 |
+
],
|
| 1174 |
+
"page_idx": 11
|
| 1175 |
+
},
|
| 1176 |
+
{
|
| 1177 |
+
"type": "text",
|
| 1178 |
+
"text": "Bidirectional Design in BSM. We developed three distinct variations of latent scanning mechanisms, differentiated by their scanning directions. In the context of motion generation tasks, we posit that the flow of hidden information within the structured latent skeleton holds significance, an aspect previously underexplored. Our ablation study reveals that a single scan across the la",
|
| 1179 |
+
"bbox": [
|
| 1180 |
+
212,
|
| 1181 |
+
763,
|
| 1182 |
+
787,
|
| 1183 |
+
840
|
| 1184 |
+
],
|
| 1185 |
+
"page_idx": 11
|
| 1186 |
+
},
|
| 1187 |
+
{
|
| 1188 |
+
"type": "page_number",
|
| 1189 |
+
"text": "12",
|
| 1190 |
+
"bbox": [
|
| 1191 |
+
217,
|
| 1192 |
+
114,
|
| 1193 |
+
235,
|
| 1194 |
+
126
|
| 1195 |
+
],
|
| 1196 |
+
"page_idx": 11
|
| 1197 |
+
},
|
| 1198 |
+
{
|
| 1199 |
+
"type": "header",
|
| 1200 |
+
"text": "Zeyu and Akide et al.",
|
| 1201 |
+
"bbox": [
|
| 1202 |
+
271,
|
| 1203 |
+
114,
|
| 1204 |
+
416,
|
| 1205 |
+
128
|
| 1206 |
+
],
|
| 1207 |
+
"page_idx": 11
|
| 1208 |
+
},
|
| 1209 |
+
{
|
| 1210 |
+
"type": "table",
|
| 1211 |
+
"img_path": "images/91cc7db9915584cb2721c037bc8a8073a2e1c0bf79422ad9a14143816aef9ff8.jpg",
|
| 1212 |
+
"table_caption": [
|
| 1213 |
+
"Table 4: Evaluation of text-based motion synthesis on HumanML3D [17]: we use metrics in Table 1 and provides real reference, we evaluate the various HTM and BSM design choices, the dimension of the latent input, the different number of layer of Motion Mamba model."
|
| 1214 |
+
],
|
| 1215 |
+
"table_footnote": [],
|
| 1216 |
+
"table_body": "<table><tr><td>Models</td><td>R Precision Top 3↑</td><td>FID↓</td><td>MM Dist.↓</td><td>Diversity→</td><td>MModality↑</td></tr><tr><td>Real</td><td>0.797±.002</td><td>0.002±.000</td><td>2.974±.008</td><td>9.503±.065</td><td>-</td></tr><tr><td>MM ( {S1, ..., SN})</td><td>0.673±.003</td><td>1.278±.012</td><td>3.802±.041</td><td>8.678±.096</td><td>3.127±.024</td></tr><tr><td>MM ( {SN, ..., S1})</td><td>0.738±.002</td><td>0.962±.011</td><td>3.433±.003</td><td>9.180±.071</td><td>2.723±.033</td></tr><tr><td>MM ( {S1, ..., S2Nn-1})</td><td>0.698±.002</td><td>0.856±.008</td><td>3.624±.037</td><td>9.229±.067</td><td>2.826±.017</td></tr><tr><td>MM ( {S2Nn-1, ..., S1})</td><td>0.792±.002</td><td>0.281±.009</td><td>3.060±.058</td><td>9.871±.084</td><td>2.294±.058</td></tr><tr><td>MM (SingleScan)</td><td>0.736±.003</td><td>1.063±.010</td><td>3.443±.026</td><td>9.180±.067</td><td>2.676±.041</td></tr><tr><td>MM (BiScan, layer)</td><td>0.735±.004</td><td>0.789±.007</td><td>3.408±.034</td><td>9.374±.059</td><td>2.591±.046</td></tr><tr><td>MM (BiScan, block)</td><td>0.792±.002</td><td>0.281±.009</td><td>3.060±.058</td><td>9.871±.084</td><td>2.294±.058</td></tr><tr><td>MM (Dim, 1)</td><td>0.706±.003</td><td>0.652±.011</td><td>3.541±.072</td><td>9.141±.082</td><td>2.612±.055</td></tr><tr><td>MM (Dim, 2)</td><td>0.792±.002</td><td>0.281±.009</td><td>3.060±.058</td><td>9.871±.084</td><td>2.294±.058</td></tr><tr><td>MM (Dim, 5)</td><td>0.741±.008</td><td>0.728±.009</td><td>3.307±.027</td><td>9.427±.099</td><td>2.314±.062</td></tr><tr><td>MM (Dim, 7)</td><td>0.738±.004</td><td>0.599±.007</td><td>3.359±.068</td><td>9.166±.075</td><td>2.488±.037</td></tr><tr><td>MM (Dim, 10)</td><td>0.715±.003</td><td>0.628±.008</td><td>3.548±.043</td><td>9.200±.075</td><td>2.884±.096</td></tr><tr><td>MM (9 layers)</td><td>0.755±.002</td><td>1.080±.012</td><td>3.309±.057</td><td>9.721±.081</td><td>2.974±.039</td></tr><tr><td>MM (11 layers)</td><td>0.792±.002</td><td>0.281±.009</td><td>3.060±.058</td><td>9.871±.084</td><td>2.294±.058</td></tr><tr><td>MM (27 layers)</td><td>0.750±.003</td><td>0.975±.008</td><td>3.336±.096</td><td>9.249±.071</td><td>2.821±.063</td></tr><tr><td>MM (37 layers)</td><td>0.754±.005</td><td>0.809±.010</td><td>3.338±.061</td><td>9.355±.062</td><td>2.741±.077</td></tr></table>",
|
| 1217 |
+
"bbox": [
|
| 1218 |
+
267,
|
| 1219 |
+
199,
|
| 1220 |
+
728,
|
| 1221 |
+
444
|
| 1222 |
+
],
|
| 1223 |
+
"page_idx": 12
|
| 1224 |
+
},
|
| 1225 |
+
{
|
| 1226 |
+
"type": "text",
|
| 1227 |
+
"text": "tent dimension yields minimal improvement. Subsequently, we investigated both layer-based and block-based bidirectional scans. Our findings indicate that the block-based bidirectional scan achieves optimal performance. This suggests that spatial information flows are encoded within the latent spaces and that bidirectional scanning facilitates the exchange of this information, thereby enhancing the efficacy of motion generation tasks.",
|
| 1228 |
+
"bbox": [
|
| 1229 |
+
212,
|
| 1230 |
+
460,
|
| 1231 |
+
787,
|
| 1232 |
+
551
|
| 1233 |
+
],
|
| 1234 |
+
"page_idx": 12
|
| 1235 |
+
},
|
| 1236 |
+
{
|
| 1237 |
+
"type": "text",
|
| 1238 |
+
"text": "Architecture Design for Motion Mamba. The proposed Motion Mamba which is grounded in a standardized motion latent diffusion system. We delved into the interplay between dimensional aspects and the module's capacity (measured by the number of layers) to ascertain their impact on system performance. Experimental results demonstrate that the Motion Mamba achieves superior performance at a latent dimension of 2, diverging from prior works where the optimal dimension was identified as 1. We attribute this discrepancy to our HTM, which necessitates multiple scans correlating with the sequence length, thus implicating dimensionality as a pivotal factor. A marginal increase in dimensionality enabled us to attain peak performance, simultaneously enhancing efficiency compared to models with a dimensionality of 10. Furthermore, we conducted experiments to determine the optimal layer count for Motion Mamba, inspired by the design of its selective scanning mechanism. Notably, a single Mamba layer achieves a parameter reduction of approximately $75\\%$ compared to a conventional transformer encoder block. By increasing the number of layers, we aim to uncover the relationship between model capacity and its performance. Our findings reveal that, through the integration of our specially designed HTM and BSM (Bidirectional Scanning Module) blocks, the Motion Mamba reaches its optimal performance with 11 layers. This represents a slight increase over the MLD [6] baseline. How-",
|
| 1239 |
+
"bbox": [
|
| 1240 |
+
212,
|
| 1241 |
+
551,
|
| 1242 |
+
787,
|
| 1243 |
+
843
|
| 1244 |
+
],
|
| 1245 |
+
"page_idx": 12
|
| 1246 |
+
},
|
| 1247 |
+
{
|
| 1248 |
+
"type": "header",
|
| 1249 |
+
"text": "Motion Mamba",
|
| 1250 |
+
"bbox": [
|
| 1251 |
+
625,
|
| 1252 |
+
114,
|
| 1253 |
+
730,
|
| 1254 |
+
126
|
| 1255 |
+
],
|
| 1256 |
+
"page_idx": 12
|
| 1257 |
+
},
|
| 1258 |
+
{
|
| 1259 |
+
"type": "page_number",
|
| 1260 |
+
"text": "13",
|
| 1261 |
+
"bbox": [
|
| 1262 |
+
767,
|
| 1263 |
+
114,
|
| 1264 |
+
784,
|
| 1265 |
+
126
|
| 1266 |
+
],
|
| 1267 |
+
"page_idx": 12
|
| 1268 |
+
},
|
| 1269 |
+
{
|
| 1270 |
+
"type": "text",
|
| 1271 |
+
"text": "ever, due to the reduced parameter count in each layer, Motion Mamba exhibits significantly greater efficiency than previous methodologies.",
|
| 1272 |
+
"bbox": [
|
| 1273 |
+
212,
|
| 1274 |
+
146,
|
| 1275 |
+
784,
|
| 1276 |
+
176
|
| 1277 |
+
],
|
| 1278 |
+
"page_idx": 13
|
| 1279 |
+
},
|
| 1280 |
+
{
|
| 1281 |
+
"type": "text",
|
| 1282 |
+
"text": "4.5 Inference Time",
|
| 1283 |
+
"text_level": 1,
|
| 1284 |
+
"bbox": [
|
| 1285 |
+
215,
|
| 1286 |
+
199,
|
| 1287 |
+
387,
|
| 1288 |
+
213
|
| 1289 |
+
],
|
| 1290 |
+
"page_idx": 13
|
| 1291 |
+
},
|
| 1292 |
+
{
|
| 1293 |
+
"type": "text",
|
| 1294 |
+
"text": "Inference time remains a significant challenge for diffusion-based methods. To address this, we enhance the inference speed by incorporating the efficient Mamba block within a lightweight architecture. Compared to the previous strong baseline, such as the MLD model cited in [6], which reports an average inference time of 0.217 seconds, our Motion Mamba model achieves a notable reduction in computational overhead, as shown in Figure 4. Specifically, it requires four times less computational effort, thereby facilitating faster and real-time inference speeds.",
|
| 1295 |
+
"bbox": [
|
| 1296 |
+
212,
|
| 1297 |
+
224,
|
| 1298 |
+
517,
|
| 1299 |
+
436
|
| 1300 |
+
],
|
| 1301 |
+
"page_idx": 13
|
| 1302 |
+
},
|
| 1303 |
+
{
|
| 1304 |
+
"type": "image",
|
| 1305 |
+
"img_path": "images/3f16a76fbbc0786783ac633696339787bd4ab38a49ca1a6d7338b047d5fb89d2.jpg",
|
| 1306 |
+
"image_caption": [
|
| 1307 |
+
"Fig.4: The figure shows the average inference time per sentence (AIT) vs FID, our proposed motion mamba obtained 0.058s AIT and 0.281 FID over all outperform previous methods. We evaluate all methods on a single V100 GPU."
|
| 1308 |
+
],
|
| 1309 |
+
"image_footnote": [],
|
| 1310 |
+
"bbox": [
|
| 1311 |
+
537,
|
| 1312 |
+
220,
|
| 1313 |
+
764,
|
| 1314 |
+
353
|
| 1315 |
+
],
|
| 1316 |
+
"page_idx": 13
|
| 1317 |
+
},
|
| 1318 |
+
{
|
| 1319 |
+
"type": "text",
|
| 1320 |
+
"text": "5 Discussion and Conclusion",
|
| 1321 |
+
"text_level": 1,
|
| 1322 |
+
"bbox": [
|
| 1323 |
+
215,
|
| 1324 |
+
458,
|
| 1325 |
+
509,
|
| 1326 |
+
474
|
| 1327 |
+
],
|
| 1328 |
+
"page_idx": 13
|
| 1329 |
+
},
|
| 1330 |
+
{
|
| 1331 |
+
"type": "text",
|
| 1332 |
+
"text": "In this study, we introduced Motion Mamba, a novel framework designed for efficient and extended sequence motion generation. Our approach represents the inaugural integration of the Mamba model within the domain of motion generation, featuring significant advancements including the implementation of Hierarchical Temporal Mamba (HTM) blocks. These blocks are specifically engineered to enhance temporal alignment through hierarchically organized selective scanning. Furthermore, Bidirectional Spatial Mamba (BSM) blocks have been developed to amplify the exchange of information flow within latent spaces, thereby augmenting the model's ability to bidirectionally capture skeleton-level density features with greater precision. Compared to previous diffusion-based motion generation methodologies that predominantly utilize transformer blocks, our Motion Mamba framework achieves SOTA performance, evidencing an improvement of up to $50\\%$ in FID scores and a quadrupled improvement in inference speed. Through comprehensive experimentation across a variety of human motion generation tasks, the effectiveness and efficiency of our proposed Motion Mamba model have been robustly demonstrated, marking a significant leap forward in the field of human motion generation.",
|
| 1333 |
+
"bbox": [
|
| 1334 |
+
212,
|
| 1335 |
+
489,
|
| 1336 |
+
787,
|
| 1337 |
+
750
|
| 1338 |
+
],
|
| 1339 |
+
"page_idx": 13
|
| 1340 |
+
},
|
| 1341 |
+
{
|
| 1342 |
+
"type": "text",
|
| 1343 |
+
"text": "Acknowledgements",
|
| 1344 |
+
"text_level": 1,
|
| 1345 |
+
"bbox": [
|
| 1346 |
+
215,
|
| 1347 |
+
771,
|
| 1348 |
+
403,
|
| 1349 |
+
787
|
| 1350 |
+
],
|
| 1351 |
+
"page_idx": 13
|
| 1352 |
+
},
|
| 1353 |
+
{
|
| 1354 |
+
"type": "text",
|
| 1355 |
+
"text": "This work is partially supported by the Fundamental Research Funds for the Central Universities, Peking University.",
|
| 1356 |
+
"bbox": [
|
| 1357 |
+
212,
|
| 1358 |
+
803,
|
| 1359 |
+
785,
|
| 1360 |
+
834
|
| 1361 |
+
],
|
| 1362 |
+
"page_idx": 13
|
| 1363 |
+
},
|
| 1364 |
+
{
|
| 1365 |
+
"type": "page_number",
|
| 1366 |
+
"text": "14",
|
| 1367 |
+
"bbox": [
|
| 1368 |
+
217,
|
| 1369 |
+
114,
|
| 1370 |
+
235,
|
| 1371 |
+
126
|
| 1372 |
+
],
|
| 1373 |
+
"page_idx": 13
|
| 1374 |
+
},
|
| 1375 |
+
{
|
| 1376 |
+
"type": "header",
|
| 1377 |
+
"text": "Zeyu and Akide et al.",
|
| 1378 |
+
"bbox": [
|
| 1379 |
+
271,
|
| 1380 |
+
114,
|
| 1381 |
+
416,
|
| 1382 |
+
128
|
| 1383 |
+
],
|
| 1384 |
+
"page_idx": 13
|
| 1385 |
+
},
|
| 1386 |
+
{
|
| 1387 |
+
"type": "text",
|
| 1388 |
+
"text": "References",
|
| 1389 |
+
"text_level": 1,
|
| 1390 |
+
"bbox": [
|
| 1391 |
+
215,
|
| 1392 |
+
143,
|
| 1393 |
+
321,
|
| 1394 |
+
159
|
| 1395 |
+
],
|
| 1396 |
+
"page_idx": 14
|
| 1397 |
+
},
|
| 1398 |
+
{
|
| 1399 |
+
"type": "list",
|
| 1400 |
+
"sub_type": "ref_text",
|
| 1401 |
+
"list_items": [
|
| 1402 |
+
"1. Ahuja, C., Morency, L.P.: Language2pose: Natural language grounded pose forecasting. In: 2019 International Conference on 3D Vision (3DV). pp. 719-728. IEEE (2019)",
|
| 1403 |
+
"2. Bao, F., Li, C., Cao, Y., Zhu, J.: All are worth words: a vit backbone for score-based diffusion models. arXiv preprint arXiv:2209.12152 (2022)",
|
| 1404 |
+
"3. Baron, E., Zimerman, I., Wolf, L.: 2-d ssm: A general spatial layer for visual transformers. arXiv preprint arXiv:2306.06635 (2023)",
|
| 1405 |
+
"4. Barsoum, E., Kender, J., Liu, Z.: Hp-gan: Probabilistic 3d human motion prediction via gan. In: Proceedings of the IEEE conference on computer vision and pattern recognition workshops. pp. 1418-1427 (2018)",
|
| 1406 |
+
"5. Bhattacharya, U., Rewkowski, N., Banerjee, A., Guhan, P., Bera, A., Manocha, D.: Text2gestures: A transformer-based network for generating emotive body gestures for virtual agents. In: 2021 IEEE virtual reality and 3D user interfaces (VR). pp. 1-10. IEEE (2021)",
|
| 1407 |
+
"6. Chen, X., Jiang, B., Liu, W., Huang, Z., Fu, B., Chen, T., Yu, G.: Executing your commands via motion diffusion in latent space. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18000-18010 (2023)",
|
| 1408 |
+
"7. Dhariwal, P., Nichol, A.: Diffusion models beat gans on image synthesis. Advances in Neural Information Processing Systems 34, 8780-8794 (2021)",
|
| 1409 |
+
"8. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. In: International Conference on Learning Representations (2020)",
|
| 1410 |
+
"9. Ghosh, A., Cheema, N., Oguz, C., Theobalt, C., Slusallek, P.: Synthesis of compositional animations from textual descriptions. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 1396-1406 (2021)",
|
| 1411 |
+
"0. Gong, K., Lian, D., Chang, H., Guo, C., Jiang, Z., Zuo, X., Mi, M.B., Wang, X.: Tm2d: Bimodality driven 3d dance generation via music-text integration. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 9942-9952 (2023)",
|
| 1412 |
+
"1. Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., Bengio, Y.: Generative adversarial nets. Advances in neural information processing systems 27 (2014)",
|
| 1413 |
+
"2. Gu, A., Dao, T.: Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752 (2023)",
|
| 1414 |
+
"3. Gu, A., Goel, K., Gupta, A., Ré, C.: On the parameterization and initialization of diagonal state space models. Advances in Neural Information Processing Systems 35, 35971-35983 (2022)",
|
| 1415 |
+
"4. Gu, A., Goel, K., Ré, C.: Efficiently modeling long sequences with structured state spaces. arXiv preprint arXiv:2111.00396 (2021)",
|
| 1416 |
+
"5. Gu, A., Johnson, I., Goel, K., Saab, K., Dao, T., Rudra, A., Ré, C.: Combining recurrent, convolutional, and continuous-time models with linear state space layers. Advances in neural information processing systems 34, 572-585 (2021)",
|
| 1417 |
+
"6. Gulrajani, I., Ahmed, F., Arjovsky, M., Dumoulin, V., Courville, A.C.: Improved training of wasserstein gans. Advances in neural information processing systems 30 (2017)",
|
| 1418 |
+
"7. Guo, C., Zou, S., Zuo, X., Wang, S., Ji, W., Li, X., Cheng, L.: Generating diverse and natural 3d human motions from text. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5152-5161 (2022)"
|
| 1419 |
+
],
|
| 1420 |
+
"bbox": [
|
| 1421 |
+
225,
|
| 1422 |
+
172,
|
| 1423 |
+
785,
|
| 1424 |
+
839
|
| 1425 |
+
],
|
| 1426 |
+
"page_idx": 14
|
| 1427 |
+
},
|
| 1428 |
+
{
|
| 1429 |
+
"type": "header",
|
| 1430 |
+
"text": "Motion Mamba",
|
| 1431 |
+
"bbox": [
|
| 1432 |
+
625,
|
| 1433 |
+
114,
|
| 1434 |
+
730,
|
| 1435 |
+
126
|
| 1436 |
+
],
|
| 1437 |
+
"page_idx": 14
|
| 1438 |
+
},
|
| 1439 |
+
{
|
| 1440 |
+
"type": "page_number",
|
| 1441 |
+
"text": "15",
|
| 1442 |
+
"bbox": [
|
| 1443 |
+
767,
|
| 1444 |
+
114,
|
| 1445 |
+
785,
|
| 1446 |
+
126
|
| 1447 |
+
],
|
| 1448 |
+
"page_idx": 14
|
| 1449 |
+
},
|
| 1450 |
+
{
|
| 1451 |
+
"type": "list",
|
| 1452 |
+
"sub_type": "ref_text",
|
| 1453 |
+
"list_items": [
|
| 1454 |
+
"18. Guo, C., Zuo, X., Wang, S., Zou, S., Sun, Q., Deng, A., Gong, M., Cheng, L.: Action2motion: Conditioned generation of 3d human motions. In: Proceedings of the 28th ACM International Conference on Multimedia. pp. 2021-2029 (2020)",
|
| 1455 |
+
"19. Gupta, A., Gu, A., Berant, J.: Diagonal state spaces are as effective as structured state spaces. Advances in Neural Information Processing Systems 35, 22982-22994 (2022)",
|
| 1456 |
+
"20. Harvey, F.G., Yurick, M., Nowrouzezahrai, D., Pal, C.: Robust motion in-between. ACM Transactions on Graphics (TOG) 39(4), 60-1 (2020)",
|
| 1457 |
+
"21. Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., Hochreiter, S.: Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems 30 (2017)",
|
| 1458 |
+
"22. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems 33, 6840-6851 (2020)",
|
| 1459 |
+
"23. Hopfield, J.J.: Neural networks and physical systems with emergent collective computational abilities. Proceedings of the national academy of sciences 79(8), 2554-2558 (1982)",
|
| 1460 |
+
"24. Jiang, B., Chen, X., Liu, W., Yu, J., Yu, G., Chen, T.: Motiongpt: Human motion as a foreign language. Advances in Neural Information Processing Systems 36 (2024)",
|
| 1461 |
+
"25. Kalman, R.E.: A New Approach to Linear Filtering and Prediction Problems. Journal of Basic Engineering 82(1), 35-45 (1960)",
|
| 1462 |
+
"26. Kingma, D.P., Welling, M.: Auto-encoding variational bayes. stat 1050, 1 (2014)",
|
| 1463 |
+
"27. Kramer, M.A.: Nonlinear principal component analysis using autoassociative neural networks. AIChE journal 37(2), 233-243 (1991)",
|
| 1464 |
+
"28. Li, S., Singh, H., Grover, A.: Mamba-nd: Selective state space modeling for multidimensional data. arXiv preprint arXiv:2402.05892 (2024)",
|
| 1465 |
+
"29. Li, Y., Cai, T., Zhang, Y., Chen, D., Dey, D.: What makes convolutional models great on long sequence modeling? In: The Eleventh International Conference on Learning Representations (2022)",
|
| 1466 |
+
"30. Lin, X., Amer, M.R.: Human motion modeling using dvgans. arXiv preprint arXiv:1804.10652 (2018)",
|
| 1467 |
+
"31. Liu, Y., Tian, Y., Zhao, Y., Yu, H., Xie, L., Wang, Y., Ye, Q., Liu, Y.: Vmamba: Visual state space model. arXiv preprint arXiv:2401.10166 (2024)",
|
| 1468 |
+
"32. Liu, Z., Mao, H., Wu, C.Y., Feichtenhofer, C., Darrell, T., Xie, S.: A convnet for the 2020s. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 11976-11986 (2022)",
|
| 1469 |
+
"33. Mahmood, N., Ghorbani, N., Troje, N.F., Pons-Moll, G., Black, M.J.: Amass: Archive of motion capture as surface shapes. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5442-5451 (2019)",
|
| 1470 |
+
"34. Pavlakos, G., Choutas, V., Ghorbani, N., Bolkart, T., Osman, A.A.A., Tzionas, D., Black, M.J.: Expressive body capture: 3d hands, face, and body from a single image. In: Proceedings IEEE Conf. on Computer Vision and Pattern Recognition (CVPR) (2019)",
|
| 1471 |
+
"35. Petrovich, M., Black, M.J., Varol, G.: Action-conditioned 3D human motion synthesis with transformer VAE. In: International Conference on Computer Vision (ICCV) (2021)",
|
| 1472 |
+
"36. Petrovich, M., Black, M.J., Varol, G.: Temos: Generating diverse human motions from textual descriptions. In: European Conference on Computer Vision. pp. 480-497. Springer (2022)"
|
| 1473 |
+
],
|
| 1474 |
+
"bbox": [
|
| 1475 |
+
215,
|
| 1476 |
+
146,
|
| 1477 |
+
785,
|
| 1478 |
+
839
|
| 1479 |
+
],
|
| 1480 |
+
"page_idx": 15
|
| 1481 |
+
},
|
| 1482 |
+
{
|
| 1483 |
+
"type": "page_number",
|
| 1484 |
+
"text": "16",
|
| 1485 |
+
"bbox": [
|
| 1486 |
+
217,
|
| 1487 |
+
114,
|
| 1488 |
+
235,
|
| 1489 |
+
126
|
| 1490 |
+
],
|
| 1491 |
+
"page_idx": 15
|
| 1492 |
+
},
|
| 1493 |
+
{
|
| 1494 |
+
"type": "header",
|
| 1495 |
+
"text": "Zeyu and Akide et al.",
|
| 1496 |
+
"bbox": [
|
| 1497 |
+
271,
|
| 1498 |
+
114,
|
| 1499 |
+
416,
|
| 1500 |
+
127
|
| 1501 |
+
],
|
| 1502 |
+
"page_idx": 15
|
| 1503 |
+
},
|
| 1504 |
+
{
|
| 1505 |
+
"type": "list",
|
| 1506 |
+
"sub_type": "ref_text",
|
| 1507 |
+
"list_items": [
|
| 1508 |
+
"37. Petrovich, M., Black, M.J., Varol, G.: TEMOS: Generating diverse human motions from textual descriptions. In: European Conference on Computer Vision (ECCV) (2022)",
|
| 1509 |
+
"38. Plappert, M., Mandery, C., Asfour, T.: The kit motion-language dataset. Big data 4(4), 236-252 (2016)",
|
| 1510 |
+
"39. Plappert, M., Mandery, C., Asfour, T.: Learning a bidirectional mapping between human whole-body motion and natural language using deep recurrent neural networks. Robotics and Autonomous Systems 109, 13-26 (2018)",
|
| 1511 |
+
"40. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021)",
|
| 1512 |
+
"41. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10684-10695 (2022)",
|
| 1513 |
+
"42. Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomedical image segmentation. In: International Conference on Medical image computing and computer-assisted intervention. pp. 234-241. Springer (2015)",
|
| 1514 |
+
"43. Rumelhart, D.E., Hinton, G.E., Williams, R.J.: Learning internal representations by error propagation. Parallel Distributed Processing pp. 318-362 (1986)",
|
| 1515 |
+
"44. Sahara, C., Ho, J., Chan, W., Salimans, T., Fleet, D.J., Norouzi, M.: Image superresolution via iterative refinement. IEEE Transactions on Pattern Analysis and Machine Intelligence (2022)",
|
| 1516 |
+
"45. Shi, X., Chen, Z., Wang, H., Yeung, D.Y., Wong, W.K., Woo, W.c.: Convolutional LSTM network: A machine learning approach for precipitation nowcasting. Advances in neural information processing systems 28 (2015)",
|
| 1517 |
+
"46. Smith, J., De Mello, S., Kautz, J., Linderman, S., Byeon, W.: Convolutional state space models for long-range spatiotemporal modeling. Advances in Neural Information Processing Systems 36 (2024)",
|
| 1518 |
+
"47. Sohl-Dickstein, J., Weiss, E., Maheswaranathan, N., Ganguli, S.: Deep unsupervised learning using nonequilibrium thermodynamics. In: International conference on machine learning. pp. 2256-2265. PMLR (2015)",
|
| 1519 |
+
"48. Tevet, G., Gordon, B., Hertz, A., Bermano, A.H., Cohen-Or, D.: Motionclip: Exposing human motion generation to clip space. In: European Conference on Computer Vision. pp. 358-374. Springer (2022)",
|
| 1520 |
+
"49. Tevet, G., Raab, S., Gordon, B., Shafir, Y., Cohen-or, D., Bermano, A.H.: Human motion diffusion model. In: The Eleventh International Conference on Learning Representations (2022)",
|
| 1521 |
+
"50. Van Den Oord, A., Vinyals, O., et al.: Neural discrete representation learning. Advances in neural information processing systems 30 (2017)",
|
| 1522 |
+
"51. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. Advances in neural information processing systems 30 (2017)",
|
| 1523 |
+
"52. Wang, J., Yan, J.N., Gu, A., Rush, A.M.: Pretraining without attention. arXiv preprint arXiv:2212.10544 (2022)",
|
| 1524 |
+
"53. Yan, J.N., Gu, J., Rush, A.M.: Diffusion models without attention. arXiv preprint arXiv:2311.18257 (2023)",
|
| 1525 |
+
"54. Zhang, M., Cai, Z., Pan, L., Hong, F., Guo, X., Yang, L., Liu, Z.: Motiondiffuse: Text-driven human motion generation with diffusion model. IEEE Transactions on Pattern Analysis and Machine Intelligence (2024)"
|
| 1526 |
+
],
|
| 1527 |
+
"bbox": [
|
| 1528 |
+
212,
|
| 1529 |
+
146,
|
| 1530 |
+
787,
|
| 1531 |
+
840
|
| 1532 |
+
],
|
| 1533 |
+
"page_idx": 16
|
| 1534 |
+
},
|
| 1535 |
+
{
|
| 1536 |
+
"type": "header",
|
| 1537 |
+
"text": "Motion Mamba",
|
| 1538 |
+
"bbox": [
|
| 1539 |
+
625,
|
| 1540 |
+
114,
|
| 1541 |
+
730,
|
| 1542 |
+
126
|
| 1543 |
+
],
|
| 1544 |
+
"page_idx": 16
|
| 1545 |
+
},
|
| 1546 |
+
{
|
| 1547 |
+
"type": "page_number",
|
| 1548 |
+
"text": "17",
|
| 1549 |
+
"bbox": [
|
| 1550 |
+
767,
|
| 1551 |
+
114,
|
| 1552 |
+
785,
|
| 1553 |
+
126
|
| 1554 |
+
],
|
| 1555 |
+
"page_idx": 16
|
| 1556 |
+
},
|
| 1557 |
+
{
|
| 1558 |
+
"type": "list",
|
| 1559 |
+
"sub_type": "ref_text",
|
| 1560 |
+
"list_items": [
|
| 1561 |
+
"55. Zhang, Y., Black, M.J., Tang, S.: We are more than our joints: Predicting how 3d bodies move. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3372-3382 (2021)",
|
| 1562 |
+
"56. Zhong, C., Hu, L., Zhang, Z., Xia, S.: Attt2m: Text-driven human motion generation with multi-perspective attention mechanism. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 509-519 (2023)",
|
| 1563 |
+
"57. Zhu, L., Liao, B., Zhang, Q., Wang, X., Liu, W., Wang, X.: Vision mamba: Efficient visual representation learning with bidirectional state space model. arXiv preprint arXiv:2401.09417 (2024)"
|
| 1564 |
+
],
|
| 1565 |
+
"bbox": [
|
| 1566 |
+
215,
|
| 1567 |
+
146,
|
| 1568 |
+
784,
|
| 1569 |
+
272
|
| 1570 |
+
],
|
| 1571 |
+
"page_idx": 17
|
| 1572 |
+
},
|
| 1573 |
+
{
|
| 1574 |
+
"type": "page_number",
|
| 1575 |
+
"text": "18",
|
| 1576 |
+
"bbox": [
|
| 1577 |
+
217,
|
| 1578 |
+
114,
|
| 1579 |
+
235,
|
| 1580 |
+
126
|
| 1581 |
+
],
|
| 1582 |
+
"page_idx": 17
|
| 1583 |
+
},
|
| 1584 |
+
{
|
| 1585 |
+
"type": "header",
|
| 1586 |
+
"text": "Zeyu and Akide et al.",
|
| 1587 |
+
"bbox": [
|
| 1588 |
+
271,
|
| 1589 |
+
114,
|
| 1590 |
+
416,
|
| 1591 |
+
128
|
| 1592 |
+
],
|
| 1593 |
+
"page_idx": 17
|
| 1594 |
+
}
|
| 1595 |
+
]
|
2024/Motion Mamba_ Efficient and Long Sequence Motion Generation/1b586349-9221-4bad-a56d-e274ec79c852_model.json
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
2024/Motion Mamba_ Efficient and Long Sequence Motion Generation/1b586349-9221-4bad-a56d-e274ec79c852_origin.pdf
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:c0a72d2370bc1dd6ec9376c87822020048d65fa2bde9842164183544c9e0a152
|
| 3 |
+
size 3543768
|
2024/Motion Mamba_ Efficient and Long Sequence Motion Generation/full.md
ADDED
|
@@ -0,0 +1,305 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Motion Mamba: Efficient and Long Sequence Motion Generation
|
| 2 |
+
|
| 3 |
+
Zeyu Zhang $^{12*}$ , Akide Liu $^{1*}$ , Ian Reid $^{3}$ , Richard Hartley $^{2}$ , Bohan Zhuang $^{1}$ , and Hao Tang $^{4\boxtimes}$
|
| 4 |
+
|
| 5 |
+
<sup>1</sup> Monash University
|
| 6 |
+
|
| 7 |
+
2 The Australian National University
|
| 8 |
+
|
| 9 |
+
<sup>3</sup> Mohamed bin Zayed University of Artificial Intelligence
|
| 10 |
+
|
| 11 |
+
$^{4}$ National Key Laboratory for Multimedia Information Processing, School of Computer Science, Peking University
|
| 12 |
+
|
| 13 |
+
https://steve-zeyu-zhang.github.io/MotionMamba
|
| 14 |
+
|
| 15 |
+

|
| 16 |
+
Fig.1: Motion Mamba has achieved significantly superior performance on long sequence modeling and motion generation efficiency compared with other well-designed state-of-the-art methods such as MLD [6], MotionDiffuse [54], and MDM [49].
|
| 17 |
+
|
| 18 |
+

|
| 19 |
+
|
| 20 |
+

|
| 21 |
+
|
| 22 |
+
Abstract. Human motion generation stands as a significant pursuit in generative computer vision, while achieving long-sequence and efficient motion generation remains challenging. Recent advancements in state space models (SSMs), notably Mamba, have showcased considerable promise in long sequence modeling with an efficient hardware-aware design, which appears to be a significant direction upon building motion generation model. Nevertheless, adapting SSMs to motion generation faces hurdles since the lack of a specialized design architecture to model motion sequence. To address these challenges, we propose Motion Mamba, a simple yet efficient approach that presents the pioneering motion generation model utilized SSMs. Specifically, we design a Hierarchical Temporal Mamba (HTM) block to process temporal data by ensemble varying numbers of isolated SSM modules across a symmetric U-Net architecture aimed at preserving motion consistency between frames. We also design a Bidirectional Spatial Mamba (BSM) block
|
| 23 |
+
|
| 24 |
+
to bidirectionally process latent poses, to enhance accurate motion generation within a temporal frame. Our proposed method achieves up to $50\%$ FID improvement and up to 4 times faster on the HumanML3D and KIT-ML datasets compared to the previous best diffusion-based method, which demonstrates strong capabilities of high-quality long sequence motion modeling and real-time human motion generation.
|
| 25 |
+
|
| 26 |
+
Keywords: Human Motion Generation $\cdot$ Selective State Space Models $\cdot$ Latent Diffusion Models
|
| 27 |
+
|
| 28 |
+
# 1 Introduction
|
| 29 |
+
|
| 30 |
+
Human motion generation involves creating high-quality 3D human motion based on given conditions, such as a text prompt, which generates Euler angles for each joint in a human skeleton. To emulate human motion effectively, virtual characters must respond to the conditional context, exhibit natural movement, and perform motion accurately. Recent motion generation models are categorized into four main approaches: autoencoder-based [1,10,17,36,48,56], utilizing transformers for latent space compression and motion synthesis; GAN-based [4,20,30], using discriminators to enhance the realism of generated motions; autoregressive models [24], treating motion sequences as languages with specialized codebooks; and diffusion-based [6,49,54], employing denoising steps for motion generation. Challenges vary across methods, with autoencoder models struggling to generate accurate motions from detailed descriptions due to textual information compression, GAN-based models facing training difficulties, especially in conditional tasks, and diffusion-based models relying on complex transformer-based architectures results in inefficient of motion prediction.
|
| 31 |
+
|
| 32 |
+
Although diffusion-based models excel at generating motion with robust performance and often exhibit superior diversity, they encounter two limitations. 1) Convolutional or transformer-based diffusion methods exhibit limitations in generating long-range motion sequences. Previous transformer-based methodologies [6,49,54] have focused on modeling long-range dependencies and acquiring comprehensive global context information. Despite these advances, they are frequently associated with a substantial increase in computational requirements. Furthermore, transformer architectures are not intrinsically designed for temporal sequential modeling, which poses an inherent limitation. 2) The efficiency of inference in transformer-based diffusion methods is constrained. Although prior research has attempted to leverage the Variational Autoencoder for denoising operations in the latent space [6], the inference speed remains adversely affected by the attention mechanism's quadratic scaling, leading to inefficient motion generation. Consequently, exploring a new architectural paradigm that accommodates long-range dependencies and maintains a linear computational complexity is crucial for sustaining motion generation tasks.
|
| 33 |
+
|
| 34 |
+
Recent advances have sparked renewed interest in state space models (SSMs) [14, 15], a field that originated from the foundational classic state space model
|
| 35 |
+
|
| 36 |
+
[25]. Modern versions of SSMs stand out due to their ability to effectively capture long-range dependencies, a capability greatly improved by the introduction of parallel training techniques. This evolution has led to various methodologies based on SSM, notably the linear state space layers (LSSL) [15], the structured state-space sequence model (S4) [14], the diagonal state space (DSS) [19], and S4D [13]. These methods have been carefully designed to handle sequential data across various tasks and modalities, paying special attention to modeling long-range dependencies. Their efficacy in managing long sequences is attributed to the implementation of convolutional computations [15] and near-linear computational strategies, such as mamba [12], marking a significant stride in sequentially oriented tasks, including large language model decoding and motion sequence generation.
|
| 37 |
+
|
| 38 |
+
Adapting selective state space modules for motion generation tasks presents notable challenges, primarily due to the lack of specialized design in SSMs for capturing the sensitive motion details required for temporal representation and the complexities involved in aggregating latent space. In response to these challenges, we have meticulously developed a motion generation architecture, specifically tailored to address the intricacies of long-term sequence generation, while optimizing for computational efficiency with near-linear-time complexity. This innovation is embodied in the Motion Mamba model, a simple yet potent approach to motion generation. The Motion Mamba framework pioneers a diffusion-based generative system, incorporating two key components oriented toward SSM as shown in Figure. 2: (1) a Hierarchical Temporal Mamba (HTM) block: This component is ingeniously crafted to arrange motion frames in sequential order, using hierarchically adjusted scanning. It is adept at identifying temporal dependencies at various depths, thereby facilitating a thorough comprehension of the dynamics inherent in motion sequences. (2) a Bidirectional Spatial Mamba (BSM) block: This block is designed to unravel the structured latent skeleton by evaluating data from both forward and reverse directions. Its primary goal is to safeguard the continuity of information flow, significantly bolstering the model's capacity for precise motion generation through the retention of dense informational exchange.
|
| 39 |
+
|
| 40 |
+
The Motion Mamba introduces a new approach to motion generation that strikes an exceptional trade-off between accuracy and efficiency, shown in Fig. 1. Our experimental results underscore the significant improvements brought about by Motion Mamba, showcasing a remarkable improvement in the Fréchet Inception Distance (FID), with a reduction of up to $50\%$ from the prior state-of-the-art metric of 0.473 to an impressive 0.281 on the HumanML3D dataset [17]. Furthermore, we emphasize our framework's unparalleled inference speed, which is four times faster than previous methods, achieving an average inference time of only 0.058 seconds per sequence compared to the 0.217 seconds required by the MLD [6] method per sequence. These outcomes unequivocally establish Motion Mamba's state-of-the-art performance, concurrently ensuring fast inference speeds for conditional human motion generation tasks.
|
| 41 |
+
|
| 42 |
+
Our contributions to the field of motion generation can be summarized as:
|
| 43 |
+
|
| 44 |
+
1. We introduce a simple yet effective framework, named Motion Mamba, which is a pioneering method integrates a selective scanning mechanism into motion generation tasks.
|
| 45 |
+
2. Motion Mamba is comprised of two modules: Hierarchical Temporal Mamba (HTM) and Bidirectional Spatial Mamba (BSM), which are designed for temporal and spatial modeling, respectively. HTM blocks are tasked with processing temporal motion data, aiming to enhance motion consistency across frames. BSM blocks are engineered to bidirectionally capture the channel-wise flow of hidden information within the latent pose representations.
|
| 46 |
+
3. Motion Mamba framework demonstrated exceptional performance on text-to-motion generation task, through experimental validation on the HumanML3D [17] and KIT-ML [38] datasets. Our methodology achieved state-of-the-art generation quality and significantly improved long-squence modeling, meanwhile optimizing inference speed.
|
| 47 |
+
|
| 48 |
+
# 2 Related Works
|
| 49 |
+
|
| 50 |
+
Human Motion Generation. Generating human motion is a significant application of computer vision, essential for various applications like 3D modelling and robot manipulation. Recently, the predominant method of achieving human motion generation, known as the Text-to-Motion task, involves learning a common latent space for both language and motion.
|
| 51 |
+
|
| 52 |
+
DVGAN [30] creates the GAN [11] discriminator by densely validating at each time-scale and perturbing the discriminator input for translation invariance, enabling motion generation and completion. ERD-QV [20] enhances latent representations through two additive modifiers: a time-to-arrival embedding applied universally and an additive scheduled target noise vector used during extended transitions. It further improves transition quality by incorporating a GAN framework with two discriminators operating at different timescales. HP-GAN [4], trained with a modified version of the improved WGAN-GP [16], utilizes a custom loss function designed for human motion prediction. It learns a probability density function of future human poses conditioned on previous poses.
|
| 53 |
+
|
| 54 |
+
Autoencoders [27, 43] are notable generative models known for their ability to represent data robustly by compressing high-dimensional data into a latent space, which is widely adopted in human motion generation. JL2P [1] uses RNN-based autoencoders [23] to learn a combined representation of language and pose. It restricts the direct mapping from text to motion to a one-to-one relationship. MotionCLIP [48] uses Transformer-based Autoencoders [51] to reconstruct motion while ensuring alignment with the corresponding text label in the CLIP [40] space. This alignment effectively integrates the semantic knowledge from CLIP into the human motion manifold. TEMOS [36] and T2M [17] combine a Transformer-based VAE [26] with a text encoder to generate distribution parameters that work within the VAE latent space. AttT2M [56] and TM2D [10] incorporate a body-part spatio-temporal encoder into VQ-VAE [50] for enhanced learning of a discrete latent space with increased expressiveness.
|
| 55 |
+
|
| 56 |
+
Diffusion models [7, 22, 41, 47] have recently surpassed GANs and VAEs in generating 2D images. Developing a motion generation model based on diffusion models is obviously an attractive direction. MotionDiffuse [54] introduces the inaugural framework for text-driven motion generation based on diffusion models. It showcases several desirable properties, including probabilistic mapping, realistic synthesis, and multi-level manipulation. MDM [49] utilizes a classifier-free Transformer-based diffusion model for the human motion domain to predict sample rather than noise in each diffusion step. MLD [6] performs a diffusion process in latent motion space, rather than using a diffusion model to establish connections between raw motion sequences and conditional inputs.
|
| 57 |
+
|
| 58 |
+
State Space Models. Recently, state space sequence models (SSMs) [14, 15], drawing inspiration from classical state-space models [25], have emerged as a promising architecture for sequence modeling. Mamba [12] introduces a selective SSM architecture, integrating time-varying parameters into the SSM framework, and proposes a hardware-aware algorithm to facilitate highly efficient training and inference processes. Some research works leverage SSM in computer vision to process 2D data. The 2D SSM [3] introduces an SSM block at the beginning of each transformer block [8, 51]. This approach aims to achieve efficient parameterization, accelerated computation, and a suitable normalization scheme. SGConvNeXt [29] presents a structured global convolution method inspired by ConvNeXt [32], incorporating multi-scale sub-kernels to achieve both parameterization efficiency and effective long sequence modeling. ConvSSM [46] integrates the tensor modeling principles of ConvLSTM [45] with SSMs, elucidating the utilization of parallel scans in convolutional recurrences. This approach enables subquadratic parallelization and rapid autoregressive generation. Vim [57] introduces a bidirectional SSM block [52] for efficient and versatile visual representation learning, achieving performance comparable to established ViT [8] methods. VMamba [31] introduces a Cross-Scan Module (CSM) designed to traverse the spatial domain and transform any non-causal visual image into ordered patch sequences. This approach achieves linear complexity while preserving global receptive fields. There have also been attempts to utilize SSMs to handle higher-dimensional data. Mamba-ND [28] explores various combinations of SSM and different scan directions within the SSM block to adapt Mamba [12] to higher-dimensional tasks. Recent efforts have sought to replace the traditional transformer-based U-Net within the diffusion denoiser with the SSM block, with the aim of enhancing image generation efficiency. DiffuSSM [53] adeptly manages higher resolutions without relying on global compression, thus maintaining detailed image representation throughout the diffusion process.
|
| 59 |
+
|
| 60 |
+
# 3 The Proposed Method
|
| 61 |
+
|
| 62 |
+
In this section, we delineate the architecture and operational principles of the Motion Mamba framework, designed for generating human motion over long ranges efficiently from textual descriptions. Initially, we discuss the foundational concepts underpinning our approach, including the Mamba Model [12] and the
|
| 63 |
+
|
| 64 |
+
latent diffusion model [6]. Following this, we detail our uniquely crafted architecture that leverages the Mamba model to enhance motion generation efficiency. This architecture comprises two principal components: the Hierarchical Temporal Mamba (HTM) block, which addresses temporal aspects, and the Bidirectional Spatial Mamba (BSM) block, focusing on spatial dynamics.
|
| 65 |
+
|
| 66 |
+
# 3.1 Preliminaries
|
| 67 |
+
|
| 68 |
+
Selective Structured State Space Sequence Model. SSMs particularly through the contributions of structured state space sequence models (S4) and Mamba, have demonstrated exceptional proficiency in handling long sequences. These models operationalize the mapping of a 1-D function or sequence, $x(t) \in \mathbb{R} \mapsto y(t) \in \mathbb{R}$ , through a hidden state $h(t) \in \mathbb{R}^N$ , employing $\mathbf{A} \in \mathbb{R}^{N \times N}$ as the evolution parameters, $\mathbf{B} \in \mathbb{R}^{N \times 1}$ and $\mathbf{C} \in \mathbb{R}^{1 \times N}$ as the projection parameters, respectively.
|
| 69 |
+
|
| 70 |
+
The discretized system can then be expressed as follows, incorporating a step size $\Delta$ :
|
| 71 |
+
|
| 72 |
+
$$
|
| 73 |
+
\begin{array}{c} h _ {t} = \overline {{\mathbf {A}}} h _ {t - 1} + \overline {{\mathbf {B}}} x _ {t}, \\ \mathbf {C} ^ {\prime} \end{array} \tag {1}
|
| 74 |
+
$$
|
| 75 |
+
|
| 76 |
+
$$
|
| 77 |
+
y _ {t} = \mathbf {C} h _ {t}.
|
| 78 |
+
$$
|
| 79 |
+
|
| 80 |
+
This adaptation facilitates the computation of output through global convolution, leveraging a structured convolutional kernel $\overline{\mathbf{K}}$ , which encompasses the entire length $M$ of the input sequence $\mathbf{x}$ :
|
| 81 |
+
|
| 82 |
+
$$
|
| 83 |
+
\overline {{\mathbf {K}}} = \left(\underset {-} {\mathbf {C}} \overline {{\mathbf {B}}}, \mathbf {C} \overline {{\mathbf {A}}} \mathbf {B}, \dots , \mathbf {C} \overline {{\mathbf {A}}} ^ {M - 1} \overline {{\mathbf {B}}}\right), \tag {2}
|
| 84 |
+
$$
|
| 85 |
+
|
| 86 |
+
$$
|
| 87 |
+
\mathbf {y} = \mathbf {x} * \overline {{\mathbf {K}}}.
|
| 88 |
+
$$
|
| 89 |
+
|
| 90 |
+
Selective models like Mamba introduce time-varying parameters, deviating from the linear time invariance (LTI) assumption and complicating parallel computation. However, hardware-aware optimizations, such as associative scans, have been developed to address these computational challenges, highlighting the ongoing evolution and application of SSMs in modeling complex temporal dynamics.
|
| 91 |
+
|
| 92 |
+
Latent Motion Diffusion Model. Diffusion probabilistic models offer a significant advancement in motion generation by gradually reducing noise from a Gaussian distribution to a target data distribution $p(x)$ through a T-length learned Markov process [7,22,41,44,47,49,54], giving $\{\pmb{x}_t\}_{t=1}^T$ . In the motion generation, we define our trainable diffusion models with a denoiser $\epsilon_{\theta}(x_t,t)$ which anneal the random noise to motion sequence $\{\hat{x}_t^{1:N}\}_{t=1}^T$ iteratively. To address the inefficiencies of applying diffusion models directly to raw motion sequences, we employ a low-dimensional motion latent space for the diffusion process. Given an input condition $c$ , such as a descriptive sentence $\pmb{w}^{1:N} = \{w^i\}_{i=1}^N$ , an action label $a$ from a predefined set $\mathcal{A}$ , or an empty condition $c = \emptyset$ , and the motion representation that combines 3D joint rotations, positions, velocities, and foot contact as proposed in [17]. The frozen CLIP [40] text encoder $\tau_{\theta}^w$ has been employed to obtain projected text embedding $\tau_{\theta}^w(w^{1:N}) \in \mathbb{R}^{1 \times d}$ , thereby conditional denoiser comprised in term of $\epsilon_{\theta}(z_t,t,\tau_{\theta}(c))$ . The latent diffusion model $\epsilon_{\theta}(x_t,t)$ aimed
|
| 93 |
+
|
| 94 |
+

|
| 95 |
+
Fig. 2: This figure illustrates the architecture of the proposed Motion Mamba model. Each of encoder and decoder blocks consists of a Hierarchical Temporal Mamba block (HTM) and a Bidirectional Spatial Mamba (BSM) block, which possess hierarchical scan and bidirectional scan within SSM layers respectively. This symmetric distribution of scans ensure a balanced and coherence framework across the encoder-decoder architecture.
|
| 96 |
+
|
| 97 |
+
to generate the human motion sequence in terms of $\hat{x}^{1:L} = \{\hat{x}^i\}_{i=1}^L$ , where L denotes the sequence length or number of frames [34,35,37,55]. After that we reused the motion Variational AutoEncoder (VAE) $\mathcal{V} = \{\mathcal{E},\mathcal{D}\}$ proposed in MLD [6] to manipulate the motion sequence in latent space $z = \mathcal{E}(x^{1:L})$ , and decompress the intermediate representation to motion sequence by $\hat{x}^{1:L} = \mathcal{D}(z) = \mathcal{D}\mathcal{E}(x^{1:L})$ [26, 42, 51]. Finally, our latent diffusion model is trained with an objective focusing on minimization of MSE between true and predicted noise in the latent space, facilitating efficient and high-quality motion generation [2, 22].
|
| 98 |
+
|
| 99 |
+
# 3.2 Motion Mamba
|
| 100 |
+
|
| 101 |
+
The architecture of the proposed Motion Mamba framework is illustrated in Figure. 2. At its core, Motion Mamba utilizes a denoising U-Net architecture, which is distinguished for its effectiveness in modeling the continuous, temporal sequences of motion frames. This effectiveness is attributed to the inherent long-sequence modeling capacity of the Mamba model. The denoiser, denoted by $\epsilon_{\theta}$ , comprises $N$ blocks including encoder $E_{1..N}$ and decoder $D_{1..N}$ . Additionally, the architecture is enhanced with a transformer-based attention mixer block $M$ , designed to augment the model's ability to capture complex temporal dynamics.
|
| 102 |
+
|
| 103 |
+
$$
|
| 104 |
+
\epsilon_ {\theta} (x) \equiv \{E _ {1 \dots N}, M, D _ {1 \dots N} \}. \tag {3}
|
| 105 |
+
$$
|
| 106 |
+
|
| 107 |
+
The encoder blocks are represented as $E_{1..N}$ , arranged sequentially, and the decoder blocks as $D_{1..N}$ , configured in reverse order to facilitate effective bottom-up and top-down information flow. Given that selective operations have signif
|
| 108 |
+
|
| 109 |
+
Algorithm 1 Hierarchical Temporal Mamba (HTM) Block.
|
| 110 |
+
Require: compressed latent representations $z:(T,B,C)$
|
| 111 |
+
Ensure: transformed representations $z_{\mathrm{HTM}}:(T,B,E)$
|
| 112 |
+
1:/\* linear projection layer \*/
|
| 113 |
+
2: $x,z:(T,B,E)\gets \mathrm{Linear}(z)$
|
| 114 |
+
3:/\* set of scans and memory matrices \*/
|
| 115 |
+
4: $K = \{S_{2N_n - 1},S_{2N_{n - 1} - 1},\ldots ,S_1\}$
|
| 116 |
+
5: Memory matrices: $\{A_1,\dots ,A_k\}$
|
| 117 |
+
6: for each scan $S_{i}$ in $K$ do
|
| 118 |
+
7: $x_o^{\prime}:(T,B,E)\gets \mathrm{Conv1D}(x)$
|
| 119 |
+
8: $B_{o},C_{o},\varDelta_{o}\gets \mathrm{Linear}(x_{o}^{\prime})$
|
| 120 |
+
9: Transform $\overline{A}_o$ and $\overline{B}_o$ using $\varDelta_{o}$
|
| 121 |
+
10: $O_{i}\gets \mathrm{SSMA}_{i,x}(x_{o}^{\prime})$
|
| 122 |
+
11: end for
|
| 123 |
+
12:/\* aggregation of outputs \*/
|
| 124 |
+
13: $z_{\mathrm{HTM}}:(T,B,E)\gets \mathrm{Linear(Aggregate(\{O_1,\ldots,O_k\}))}$
|
| 125 |
+
14: Return: $z_{\mathrm{HTM}}$
|
| 126 |
+
|
| 127 |
+
icantly lower computational complexity compared to attention-based methods, we have increased the number of scans to achieve higher quality generations. Concurrently, it is imperative to maintain a balance between the model's parameters and its efficiency. Thereby, a novel aspect of our model is the introduction of a hierarchical scan strategy, characterized by a sequence of scan numbers as,
|
| 128 |
+
|
| 129 |
+
$$
|
| 130 |
+
K = \left\{S _ {2 N - 1}, S _ {2 (N - 1) - 1}, \dots , S _ {1} \right\}. \tag {4}
|
| 131 |
+
$$
|
| 132 |
+
|
| 133 |
+
This sequence specifies the number of scans allocated to each layer, in descending order of complexity. For instance, the uppermost encoder layer, $E_{1}$ , and the lowermost decoder layer, $D_{N}$ , are allocated $S_{2N - 1}$ scans, indicating the highest scanning complexity. Conversely, the lowest encoder layer, $E_{N}$ , and the uppermost decoder layer, $D_{1}$ , are assigned $S_{1}$ scans, reflecting the lowest level of scanning complexity.
|
| 134 |
+
|
| 135 |
+
$$
|
| 136 |
+
E _ {i} (S) = \left\{ \begin{array}{l l} S _ {2 N - 1} & \text {f o r} i = 1 \\ S _ {2 (N - i) - 1} & \text {f o r} i = 2, \dots , N - 1 \\ S _ {1} & \text {f o r} i = N \end{array} \right. \tag {5}
|
| 137 |
+
$$
|
| 138 |
+
|
| 139 |
+
$$
|
| 140 |
+
D _ {j} (S) = \left\{ \begin{array}{l l} S _ {2 N - 1} & \text {f o r} j = N \\ S _ {2 (N - j) - 1} & \text {f o r} j = N - 1, \dots , 2 \\ S _ {1} & \text {f o r} j = 1 \end{array} \right. \tag {6}
|
| 141 |
+
$$
|
| 142 |
+
|
| 143 |
+
This hierarchical scanning approach ensures that processing capabilities are evenly distributed throughout the encoder-decoder architecture., facilitating a detailed and nuanced analysis of temporal sequences. Within this structured framework, each denoiser is equipped with a specialized Hierarchical Temporal Mamba (HTM) block, which serves to augment the model's ability to process
|
| 144 |
+
|
| 145 |
+
Algorithm 2 Bidirectional Spatial Mamba (BSM) Block.
|
| 146 |
+
Require: compressed latent representations $z:(T,B,C)$
|
| 147 |
+
Ensure: transformed representations $z_{\mathrm{BSM}}:(C,B,E)$
|
| 148 |
+
1:/\* dimension rearrangement \*/
|
| 149 |
+
2: $z^{\prime}:(C,B,T)\gets$ Rearrange(z)
|
| 150 |
+
3:/\* linear projection after normalization \*/
|
| 151 |
+
4: $z^{\prime}:(C,B,T)\gets$ Norm(z)
|
| 152 |
+
5:x,z:(C,B,E) $\leftarrow$ Linear(z')
|
| 153 |
+
6: for o in {forward, backward} do
|
| 154 |
+
7: $x_o':(C,B,E)\gets$ Conv1D(x)
|
| 155 |
+
8: $B_{o},C_{o},\varDelta_{o}\gets$ Linear $(x_o^{\prime})$
|
| 156 |
+
9: Transform $\overline{A}_o$ and $\overline{B}_o$ using $\varDelta_{o}$
|
| 157 |
+
10: $y_{o}\gets$ SSM(Ao,Bo,Co)
|
| 158 |
+
11: end for
|
| 159 |
+
12:/\* gating and summing outputs \*/
|
| 160 |
+
13: $z_{\mathrm{BSM}}:(C,B,T)\gets$ GateAndSum(yforward,ybackward,z)
|
| 161 |
+
14: $z_{\mathrm{BSM}}^{\prime}:(T,B,C)\gets$ Rearrange(z)
|
| 162 |
+
15: Return: $z_{\mathrm{BSM}}^{\prime}$
|
| 163 |
+
|
| 164 |
+
temporal information effectively. Additionally, the proposed Motion Mamba incorporates an attention-based mixer block denoted as $M$ , strategically integrated to enhance conditional fusion.
|
| 165 |
+
|
| 166 |
+
Hierarchical Temporal Mamba (HTM) block processes compressed latent representations, denoted as $z$ , with the dimensions $(T,B,C)$ , of which procedure shown in Algorithm 1. Here, $T$ signifies the temporal dimension, as specified in the Variational AutoEncoder (VAE) framework. Initially, the input $z$ is subjected to a linear projection layer, producing transformed representations $x$ and $z$ with dimension $E$ . Our analysis revealed an increased density of motion within the lower-level feature spaces. Consequently, we developed a hierarchical scanning methodology that is executed at various depths of the network. This approach not only accommodates the diverse motion densities, but also significantly reduces computational overhead. This step utilizes a hierarchically structured set of scans, $K = \{S_{2N_n - 1}, S_{2N_{n - 1} - 1}, \ldots, S_1\}$ , in conjunction with a corresponding series of memory matrices $\{A_1, \ldots, A_k\}$ . Each sub-SSM scan first applies a 1-D convolution to $x$ , resulting in $x_o'$ . $x_o'$ is then linearly projected to derive $B_o$ , $C_o$ , and $\Delta_o$ . These projections $B_o$ , $C_o$ use $\Delta_o$ to effect transformations in $\overline{A}_o$ and $\overline{B}_o$ , respectively. After executing a sequence of SSM scans $\{SSM_{A_1,x}, SSM_{A_2,x}, \ldots, SSM_{A_k,x}\}$ , a set of outputs $\{O_1, \ldots, O_k\}$ is compiled. This collection is subsequently aggregated via a linear projection to obtain the final output of the HTM block.
|
| 167 |
+
|
| 168 |
+
Bidirectional Spatial Mamba (BSM) block focuses on enhancing latent representation learning through a novel approach of dimension rearrangement and bidirectional scanning, of which the process is shown in Algorithm 2. Initially, it alters the original input dimensions from $(T,B,C)$ to $(C,B,T)$ , effectively swap
|
| 169 |
+
|
| 170 |
+
ping the temporal and channel dimensions. After this rearrangement, the input, now denoted $z'$ , undergoes a linear projection after normalization, resulting in dimensions $x$ and $z$ of size $E$ . The process involves bidirectional scanning of the latent channel dimension, where $\mathbf{x}$ is subjected to a 1-D convolution, yielding $\mathbf{x}_o'$ for both forward and backward directions. Each $\mathbf{x}_o'$ is then linearly projected to obtain $B_o$ , $C_o$ , and $\Delta_o$ , which are utilized to transform $\overline{A}_o$ and $\overline{B}_o$ , respectively. The final output token sequence, $\mathbf{z}_1$ , is computed by gating and summing the forward $y_{\mathrm{forward}}$ and backward $y_{\mathrm{backward}}$ output with $\mathbf{z}$ . This component is engineered to decode the structured latent skeleton by analyzing data from both forward and reverse viewpoints. Its main objective is to ensure the seamless continuity of information flow, thereby substantially enhancing the model's ability to generate accurate motion. This is achieved through the maintenance of a dense informational exchange, which is critical for the model's performance.
|
| 171 |
+
|
| 172 |
+
# 4 Experiments
|
| 173 |
+
|
| 174 |
+
# 4.1 Datasets
|
| 175 |
+
|
| 176 |
+
We evaluate our proposed Motion Mamba on two prominent Text-to-Motion synthesis benchmarks as follows:
|
| 177 |
+
|
| 178 |
+
HumanML3D. The HumanML3D [17] dataset aggregates 14,616 motions sourced from the AMASS [33] and HumanAct12 [18] datasets, with each motion accompanied by three textual descriptions, culminating in 44,970 scripts. This dataset spans a wide range of actions such as exercising, dancing, and acrobatics, presenting a rich motion-language corpus.
|
| 179 |
+
|
| 180 |
+
KIT-ML. The KIT-ML dataset [38] is comprised of 3,911 motions paired with 6,278 textual descriptions, serving as a compact yet effective benchmark for evaluation. For both datasets, the pose representation adopted is derived from T2M [17], ensuring consistency in motion representation across evaluations.
|
| 181 |
+
|
| 182 |
+
# 4.2 Evaluation Metrics
|
| 183 |
+
|
| 184 |
+
We adapt the standard evaluation metrics on following aspects throughout our experiments, including: Generation Quality. We implement a Fréchet inception distance (FID) [21] to quantify the realism and diversity of motion generated by models. Moreover, we use multi-modal distance (MM Dist) to measure the distance between motions and texts and assess motion-text alignment. Diversity. We use the diversity metric to measure motion diversity, which calculates variance in features extracted from the motions. Additionally, we employ multimodality (MModality) to assess diversity within generated motions sharing the same text description.
|
| 185 |
+
|
| 186 |
+
# 4.3 Comparative Studies
|
| 187 |
+
|
| 188 |
+
We evaluate our method against the state-of-the-art methods on the HumanML3D [17] and KIT-ML [38] datasets. We train our Motion Mamba with HTM arrangement strategy MM ( $\{S_{2N_n - 1},\ldots ,S_1\}$ ), BSM bidirectional block strategy on the
|
| 189 |
+
|
| 190 |
+
Table 1: Comparison of text-conditional motion synthesis on HumanML3D [17]. These metrics are evaluated by the motion encoder from [17]. Empty MModality indicates the non-diverse generation methods. We employ real motion as a reference and sort all methods by descending FIDs. The right arrow $\rightarrow$ means that the closer to the real motion, the better. Bold and underline indicate the best and second best result.
|
| 191 |
+
|
| 192 |
+
<table><tr><td rowspan="2">Method</td><td colspan="3">R Precision ↑</td><td rowspan="2">FID↓</td><td rowspan="2">MM Dist↓</td><td rowspan="2" colspan="2">Diversity→MModality↑</td></tr><tr><td>Top 1</td><td>Top 2</td><td>Top 3</td></tr><tr><td>Real</td><td>0.511±.003</td><td>0.703±.003</td><td>0.797±.002</td><td>0.002±.000</td><td>2.974±.008</td><td>9.503±.065</td><td>-</td></tr><tr><td>Seq2Seq [39]</td><td>0.180±.002</td><td>0.300±.002</td><td>0.396±.002</td><td>11.75±.035</td><td>5.529±.007</td><td>6.223±.061</td><td>-</td></tr><tr><td>LJ2P [1]</td><td>0.246±.001</td><td>0.387±.002</td><td>0.486±.002</td><td>11.02±.046</td><td>5.296±.008</td><td>7.676±.058</td><td>-</td></tr><tr><td>T2G [5]</td><td>0.165±.001</td><td>0.267±.002</td><td>0.345±.002</td><td>7.664±.030</td><td>6.030±.008</td><td>6.409±.071</td><td>-</td></tr><tr><td>Hier [9]</td><td>0.301±.002</td><td>0.425±.002</td><td>0.552±.004</td><td>6.532±.024</td><td>5.012±.018</td><td>8.332±.042</td><td>-</td></tr><tr><td>TEMOS [37]</td><td>0.424±.002</td><td>0.612±.002</td><td>0.722±.002</td><td>3.734±.028</td><td>3.703±.008</td><td>8.973±.071</td><td>0.368±.018</td></tr><tr><td>T2M [17]</td><td>0.457±.002</td><td>0.639±.003</td><td>0.740±.003</td><td>1.067±.002</td><td>3.340±.008</td><td>9.188±.002</td><td>2.090±.083</td></tr><tr><td>MDM [49]</td><td>0.320±.005</td><td>0.498±.004</td><td>0.611±.007</td><td>0.544±.044</td><td>5.566±.027</td><td>9.559±.086</td><td>2.799±.072</td></tr><tr><td>MotionDiffuse [54]</td><td>0.491±.001</td><td>0.681±.001</td><td>0.782±.001</td><td>0.630±.001</td><td>3.113±.001</td><td>9.410±.049</td><td>1.553±.042</td></tr><tr><td>MLD [6]</td><td>0.481±.003</td><td>0.673±.003</td><td>0.772±.002</td><td>0.473±.013</td><td>3.196±.010</td><td>9.724±.082</td><td>2.413±.079</td></tr><tr><td>Motion Mamba (Ours)</td><td>0.502±.003</td><td>0.693±.002</td><td>0.792±.002</td><td>0.281±.009</td><td>3.060±.058</td><td>9.871±.084</td><td>2.294±.058</td></tr></table>
|
| 193 |
+
|
| 194 |
+
Table 2: We involve KIT-ML [38] dataset and evaluate the SOTA methods on the text-to-motion task.
|
| 195 |
+
|
| 196 |
+
<table><tr><td rowspan="2">Method</td><td colspan="3">R Precision ↑</td><td rowspan="2">FID↓</td><td rowspan="2">MM Dist↓</td><td rowspan="2" colspan="2">Diversity→MModality↑</td></tr><tr><td>Top 1</td><td>Top 2</td><td>Top 3</td></tr><tr><td>Real</td><td>0.424±.005</td><td>0.649±.006</td><td>0.779±.006</td><td>0.031±.004</td><td>2.788±.012</td><td>11.08±.097</td><td>-</td></tr><tr><td>Seq2Seq [39]</td><td>0.103±.003</td><td>0.178±.005</td><td>0.241±.006</td><td>24.86±.348</td><td>7.960±.031</td><td>6.744±.106</td><td>-</td></tr><tr><td>T2G [5]</td><td>0.156±.004</td><td>0.255±.004</td><td>0.338±.005</td><td>12.12±.183</td><td>6.964±.029</td><td>9.334±.079</td><td>-</td></tr><tr><td>LJ2P [1]</td><td>0.221±.005</td><td>0.373±.004</td><td>0.483±.005</td><td>6.545±.072</td><td>5.147±.030</td><td>9.073±.100</td><td>-</td></tr><tr><td>Hier [9]</td><td>0.255±.006</td><td>0.432±.007</td><td>0.531±.007</td><td>5.203±.107</td><td>4.986±.027</td><td>9.563±.072</td><td>2.090±.083</td></tr><tr><td>TEMOS [37]</td><td>0.353±.006</td><td>0.561±.007</td><td>0.687±.005</td><td>3.717±.051</td><td>3.417±.019</td><td>10.84±.100</td><td>0.532±.034</td></tr><tr><td>T2M [17]</td><td>0.370±.005</td><td>0.569±.007</td><td>0.693±.007</td><td>2.770±.109</td><td>3.401±.008</td><td>10.91±.119</td><td>1.482±.065</td></tr><tr><td>MDM [49]</td><td>0.164±.004</td><td>0.291±.004</td><td>0.396±.004</td><td>0.497±.021</td><td>9.191±.022</td><td>10.85±.109</td><td>1.907±.214</td></tr><tr><td>MotionDiffuse [54]</td><td>0.417±.004</td><td>0.621±.004</td><td>0.739±.004</td><td>1.954±.062</td><td>2.958±.005</td><td>11.10±.143</td><td>0.730±.013</td></tr><tr><td>MLD [6]</td><td>0.390±.008</td><td>0.609±.008</td><td>0.734±.007</td><td>0.404±.027</td><td>3.204±.027</td><td>10.80±.117</td><td>2.192±.071</td></tr><tr><td>Motion Mamba (Ours)</td><td>0.419±.006</td><td>0.645±.005</td><td>0.765±.006</td><td>0.307±.041</td><td>3.021±.025</td><td>11.02±.098</td><td>1.678±.064</td></tr></table>
|
| 197 |
+
|
| 198 |
+
latent dimension $= 2$ with 11 layers. We evaluate our model and previous works with suggested metrics in HumanML3D [17] and calculate $95\%$ confidence interval by repeat evaluation 20 times. The results for the HumanML3D dataset are presented in Table 1. Our model outperforms other methods significantly across various evaluation metrics, including FID, R precision, multi-modal distance, and diversity. For instance, our Motion Mamba outperforms previous best diffusion based motion generation MLD by $40.5\%$ in terms of FID, and up to $10\%$ improvement on R Precision, we aslo obatined best MModality by 3.060. The results for the KIT-ML dataset are presented in Table 2. We have also outperformed other well-established methods in FID and multi-modal distance.
|
| 199 |
+
|
| 200 |
+
# 4.4 Ablation Studies
|
| 201 |
+
|
| 202 |
+
We concluded the ablation studies including long sequence evaluation, hierarchical design with HTM, bidirectional design in the BSM, number of latent dimensions, and number of layers of our proposed motion mamba in Table 4.
|
| 203 |
+
|
| 204 |
+
Table 3: In order to evaluate the models' capability in long sequence motion generation, we compared our method with an existing approach on the recently introduced HumanML3D-LS dataset. This dataset comprises motion sequences longer than 190 frames from the original evaluation set. Our model demonstrates superior performance compared to other methods.
|
| 205 |
+
|
| 206 |
+
<table><tr><td rowspan="2">Method</td><td colspan="3">R Precision ↑</td><td rowspan="2">FID↓</td><td rowspan="2">MM Dist↓</td><td rowspan="2">Diversity→</td><td rowspan="2">MModality↑</td></tr><tr><td>Top 1</td><td>Top 2</td><td>Top 3</td></tr><tr><td>Real</td><td>0.437±.003</td><td>0.622±.004</td><td>0.721±.004</td><td>0.004±.000</td><td>3.343±.015</td><td>8.423±.090</td><td>-</td></tr><tr><td>MDM [49]</td><td>0.368±.005</td><td>0.553±.006</td><td>0.672±.005</td><td>0.802±.044</td><td>3.860±.025</td><td>8.817±.068</td><td>-</td></tr><tr><td>MotionDiffuse [54]</td><td>0.367±.004</td><td>0.521±.004</td><td>0.623±.004</td><td>2.460±.062</td><td>3.789±.005</td><td>8.707±.143</td><td>1.602±.013</td></tr><tr><td>MLD [6]</td><td>0.403±.005</td><td>0.584±.005</td><td>0.690±.005</td><td>0.952±.020</td><td>3.580±.016</td><td>9.050±.085</td><td>2.711±.104</td></tr><tr><td>Motion Mamba (Ours)</td><td>0.417±.003</td><td>0.606±.003</td><td>0.713±.004</td><td>0.668±.019</td><td>3.435±.015</td><td>9.021±.070</td><td>2.373±.084</td></tr></table>
|
| 207 |
+
|
| 208 |
+
Long Sequence Motion Generation. The HumanML3D [17] dataset exhibits a long-tailed and right-skewed distribution with a significant proportion of long-sequence human motions, as shown in Figure 3. We suggest previous study overlooked the challenges in the long-sequence generation problem. Thus, we introduce a new dataset variant, HumanML3D-LS, comprising motion sequences longer than 190 frames extracted from the original test set. This addition allows us to showcase our capability in generating long-sequence motions. Subsequently, we eval
|
| 209 |
+
|
| 210 |
+

|
| 211 |
+
Fig. 3: The figure shows a long tail distribution of the HumanML3D [17], which has a significant proportion of long-sequence human motions.
|
| 212 |
+
|
| 213 |
+
uate the performance of our method on HumanML3D-LS and compare it with other diffusion-based motion generation approaches. The comparative results are presented in Table 3. Motion Mamba by leverage the benefits on long-range dependency modeling make it well suitable for long sequence motion generation.
|
| 214 |
+
|
| 215 |
+
Hierarchical Design with HTM. In our ablation studies, we observed a slight improvement upon reversing the scan order from a lower to a higher level, specifically transitioning from MM $\{S_1,\ldots ,S_N\}$ to MM $\{S_N,\dots ,S_1\}$ . This enhancement suggests a correlation with the increase in temporal motion density within the lower-level feature spaces. Furthermore, to achieve the optimal result, we introduce the hierarchical design to arrange the scanning frequency, resulting in the sequence MM $\{S_{2N_n - 1},\ldots ,S_1\}$ . This expansion in the number of scans led to a performance increase. We attribute this enhancement to the observation that individual selective scan operations significantly reduce the parameter count, especially when compared to the parameter-intensive constructs of self-attention and feedforward network blocks prevalent in transformer architectures.
|
| 216 |
+
|
| 217 |
+
Bidirectional Design in BSM. We developed three distinct variations of latent scanning mechanisms, differentiated by their scanning directions. In the context of motion generation tasks, we posit that the flow of hidden information within the structured latent skeleton holds significance, an aspect previously underexplored. Our ablation study reveals that a single scan across the la
|
| 218 |
+
|
| 219 |
+
Table 4: Evaluation of text-based motion synthesis on HumanML3D [17]: we use metrics in Table 1 and provides real reference, we evaluate the various HTM and BSM design choices, the dimension of the latent input, the different number of layer of Motion Mamba model.
|
| 220 |
+
|
| 221 |
+
<table><tr><td>Models</td><td>R Precision Top 3↑</td><td>FID↓</td><td>MM Dist.↓</td><td>Diversity→</td><td>MModality↑</td></tr><tr><td>Real</td><td>0.797±.002</td><td>0.002±.000</td><td>2.974±.008</td><td>9.503±.065</td><td>-</td></tr><tr><td>MM ( {S1, ..., SN})</td><td>0.673±.003</td><td>1.278±.012</td><td>3.802±.041</td><td>8.678±.096</td><td>3.127±.024</td></tr><tr><td>MM ( {SN, ..., S1})</td><td>0.738±.002</td><td>0.962±.011</td><td>3.433±.003</td><td>9.180±.071</td><td>2.723±.033</td></tr><tr><td>MM ( {S1, ..., S2Nn-1})</td><td>0.698±.002</td><td>0.856±.008</td><td>3.624±.037</td><td>9.229±.067</td><td>2.826±.017</td></tr><tr><td>MM ( {S2Nn-1, ..., S1})</td><td>0.792±.002</td><td>0.281±.009</td><td>3.060±.058</td><td>9.871±.084</td><td>2.294±.058</td></tr><tr><td>MM (SingleScan)</td><td>0.736±.003</td><td>1.063±.010</td><td>3.443±.026</td><td>9.180±.067</td><td>2.676±.041</td></tr><tr><td>MM (BiScan, layer)</td><td>0.735±.004</td><td>0.789±.007</td><td>3.408±.034</td><td>9.374±.059</td><td>2.591±.046</td></tr><tr><td>MM (BiScan, block)</td><td>0.792±.002</td><td>0.281±.009</td><td>3.060±.058</td><td>9.871±.084</td><td>2.294±.058</td></tr><tr><td>MM (Dim, 1)</td><td>0.706±.003</td><td>0.652±.011</td><td>3.541±.072</td><td>9.141±.082</td><td>2.612±.055</td></tr><tr><td>MM (Dim, 2)</td><td>0.792±.002</td><td>0.281±.009</td><td>3.060±.058</td><td>9.871±.084</td><td>2.294±.058</td></tr><tr><td>MM (Dim, 5)</td><td>0.741±.008</td><td>0.728±.009</td><td>3.307±.027</td><td>9.427±.099</td><td>2.314±.062</td></tr><tr><td>MM (Dim, 7)</td><td>0.738±.004</td><td>0.599±.007</td><td>3.359±.068</td><td>9.166±.075</td><td>2.488±.037</td></tr><tr><td>MM (Dim, 10)</td><td>0.715±.003</td><td>0.628±.008</td><td>3.548±.043</td><td>9.200��.075</td><td>2.884±.096</td></tr><tr><td>MM (9 layers)</td><td>0.755±.002</td><td>1.080±.012</td><td>3.309±.057</td><td>9.721±.081</td><td>2.974±.039</td></tr><tr><td>MM (11 layers)</td><td>0.792±.002</td><td>0.281±.009</td><td>3.060±.058</td><td>9.871±.084</td><td>2.294±.058</td></tr><tr><td>MM (27 layers)</td><td>0.750±.003</td><td>0.975±.008</td><td>3.336±.096</td><td>9.249±.071</td><td>2.821±.063</td></tr><tr><td>MM (37 layers)</td><td>0.754±.005</td><td>0.809±.010</td><td>3.338±.061</td><td>9.355±.062</td><td>2.741±.077</td></tr></table>
|
| 222 |
+
|
| 223 |
+
tent dimension yields minimal improvement. Subsequently, we investigated both layer-based and block-based bidirectional scans. Our findings indicate that the block-based bidirectional scan achieves optimal performance. This suggests that spatial information flows are encoded within the latent spaces and that bidirectional scanning facilitates the exchange of this information, thereby enhancing the efficacy of motion generation tasks.
|
| 224 |
+
|
| 225 |
+
Architecture Design for Motion Mamba. The proposed Motion Mamba which is grounded in a standardized motion latent diffusion system. We delved into the interplay between dimensional aspects and the module's capacity (measured by the number of layers) to ascertain their impact on system performance. Experimental results demonstrate that the Motion Mamba achieves superior performance at a latent dimension of 2, diverging from prior works where the optimal dimension was identified as 1. We attribute this discrepancy to our HTM, which necessitates multiple scans correlating with the sequence length, thus implicating dimensionality as a pivotal factor. A marginal increase in dimensionality enabled us to attain peak performance, simultaneously enhancing efficiency compared to models with a dimensionality of 10. Furthermore, we conducted experiments to determine the optimal layer count for Motion Mamba, inspired by the design of its selective scanning mechanism. Notably, a single Mamba layer achieves a parameter reduction of approximately $75\%$ compared to a conventional transformer encoder block. By increasing the number of layers, we aim to uncover the relationship between model capacity and its performance. Our findings reveal that, through the integration of our specially designed HTM and BSM (Bidirectional Scanning Module) blocks, the Motion Mamba reaches its optimal performance with 11 layers. This represents a slight increase over the MLD [6] baseline. How-
|
| 226 |
+
|
| 227 |
+
ever, due to the reduced parameter count in each layer, Motion Mamba exhibits significantly greater efficiency than previous methodologies.
|
| 228 |
+
|
| 229 |
+
# 4.5 Inference Time
|
| 230 |
+
|
| 231 |
+
Inference time remains a significant challenge for diffusion-based methods. To address this, we enhance the inference speed by incorporating the efficient Mamba block within a lightweight architecture. Compared to the previous strong baseline, such as the MLD model cited in [6], which reports an average inference time of 0.217 seconds, our Motion Mamba model achieves a notable reduction in computational overhead, as shown in Figure 4. Specifically, it requires four times less computational effort, thereby facilitating faster and real-time inference speeds.
|
| 232 |
+
|
| 233 |
+

|
| 234 |
+
Fig.4: The figure shows the average inference time per sentence (AIT) vs FID, our proposed motion mamba obtained 0.058s AIT and 0.281 FID over all outperform previous methods. We evaluate all methods on a single V100 GPU.
|
| 235 |
+
|
| 236 |
+
# 5 Discussion and Conclusion
|
| 237 |
+
|
| 238 |
+
In this study, we introduced Motion Mamba, a novel framework designed for efficient and extended sequence motion generation. Our approach represents the inaugural integration of the Mamba model within the domain of motion generation, featuring significant advancements including the implementation of Hierarchical Temporal Mamba (HTM) blocks. These blocks are specifically engineered to enhance temporal alignment through hierarchically organized selective scanning. Furthermore, Bidirectional Spatial Mamba (BSM) blocks have been developed to amplify the exchange of information flow within latent spaces, thereby augmenting the model's ability to bidirectionally capture skeleton-level density features with greater precision. Compared to previous diffusion-based motion generation methodologies that predominantly utilize transformer blocks, our Motion Mamba framework achieves SOTA performance, evidencing an improvement of up to $50\%$ in FID scores and a quadrupled improvement in inference speed. Through comprehensive experimentation across a variety of human motion generation tasks, the effectiveness and efficiency of our proposed Motion Mamba model have been robustly demonstrated, marking a significant leap forward in the field of human motion generation.
|
| 239 |
+
|
| 240 |
+
# Acknowledgements
|
| 241 |
+
|
| 242 |
+
This work is partially supported by the Fundamental Research Funds for the Central Universities, Peking University.
|
| 243 |
+
|
| 244 |
+
# References
|
| 245 |
+
|
| 246 |
+
1. Ahuja, C., Morency, L.P.: Language2pose: Natural language grounded pose forecasting. In: 2019 International Conference on 3D Vision (3DV). pp. 719-728. IEEE (2019)
|
| 247 |
+
2. Bao, F., Li, C., Cao, Y., Zhu, J.: All are worth words: a vit backbone for score-based diffusion models. arXiv preprint arXiv:2209.12152 (2022)
|
| 248 |
+
3. Baron, E., Zimerman, I., Wolf, L.: 2-d ssm: A general spatial layer for visual transformers. arXiv preprint arXiv:2306.06635 (2023)
|
| 249 |
+
4. Barsoum, E., Kender, J., Liu, Z.: Hp-gan: Probabilistic 3d human motion prediction via gan. In: Proceedings of the IEEE conference on computer vision and pattern recognition workshops. pp. 1418-1427 (2018)
|
| 250 |
+
5. Bhattacharya, U., Rewkowski, N., Banerjee, A., Guhan, P., Bera, A., Manocha, D.: Text2gestures: A transformer-based network for generating emotive body gestures for virtual agents. In: 2021 IEEE virtual reality and 3D user interfaces (VR). pp. 1-10. IEEE (2021)
|
| 251 |
+
6. Chen, X., Jiang, B., Liu, W., Huang, Z., Fu, B., Chen, T., Yu, G.: Executing your commands via motion diffusion in latent space. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18000-18010 (2023)
|
| 252 |
+
7. Dhariwal, P., Nichol, A.: Diffusion models beat gans on image synthesis. Advances in Neural Information Processing Systems 34, 8780-8794 (2021)
|
| 253 |
+
8. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. In: International Conference on Learning Representations (2020)
|
| 254 |
+
9. Ghosh, A., Cheema, N., Oguz, C., Theobalt, C., Slusallek, P.: Synthesis of compositional animations from textual descriptions. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 1396-1406 (2021)
|
| 255 |
+
0. Gong, K., Lian, D., Chang, H., Guo, C., Jiang, Z., Zuo, X., Mi, M.B., Wang, X.: Tm2d: Bimodality driven 3d dance generation via music-text integration. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 9942-9952 (2023)
|
| 256 |
+
1. Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., Bengio, Y.: Generative adversarial nets. Advances in neural information processing systems 27 (2014)
|
| 257 |
+
2. Gu, A., Dao, T.: Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752 (2023)
|
| 258 |
+
3. Gu, A., Goel, K., Gupta, A., Ré, C.: On the parameterization and initialization of diagonal state space models. Advances in Neural Information Processing Systems 35, 35971-35983 (2022)
|
| 259 |
+
4. Gu, A., Goel, K., Ré, C.: Efficiently modeling long sequences with structured state spaces. arXiv preprint arXiv:2111.00396 (2021)
|
| 260 |
+
5. Gu, A., Johnson, I., Goel, K., Saab, K., Dao, T., Rudra, A., Ré, C.: Combining recurrent, convolutional, and continuous-time models with linear state space layers. Advances in neural information processing systems 34, 572-585 (2021)
|
| 261 |
+
6. Gulrajani, I., Ahmed, F., Arjovsky, M., Dumoulin, V., Courville, A.C.: Improved training of wasserstein gans. Advances in neural information processing systems 30 (2017)
|
| 262 |
+
7. Guo, C., Zou, S., Zuo, X., Wang, S., Ji, W., Li, X., Cheng, L.: Generating diverse and natural 3d human motions from text. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5152-5161 (2022)
|
| 263 |
+
|
| 264 |
+
18. Guo, C., Zuo, X., Wang, S., Zou, S., Sun, Q., Deng, A., Gong, M., Cheng, L.: Action2motion: Conditioned generation of 3d human motions. In: Proceedings of the 28th ACM International Conference on Multimedia. pp. 2021-2029 (2020)
|
| 265 |
+
19. Gupta, A., Gu, A., Berant, J.: Diagonal state spaces are as effective as structured state spaces. Advances in Neural Information Processing Systems 35, 22982-22994 (2022)
|
| 266 |
+
20. Harvey, F.G., Yurick, M., Nowrouzezahrai, D., Pal, C.: Robust motion in-between. ACM Transactions on Graphics (TOG) 39(4), 60-1 (2020)
|
| 267 |
+
21. Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., Hochreiter, S.: Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems 30 (2017)
|
| 268 |
+
22. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems 33, 6840-6851 (2020)
|
| 269 |
+
23. Hopfield, J.J.: Neural networks and physical systems with emergent collective computational abilities. Proceedings of the national academy of sciences 79(8), 2554-2558 (1982)
|
| 270 |
+
24. Jiang, B., Chen, X., Liu, W., Yu, J., Yu, G., Chen, T.: Motiongpt: Human motion as a foreign language. Advances in Neural Information Processing Systems 36 (2024)
|
| 271 |
+
25. Kalman, R.E.: A New Approach to Linear Filtering and Prediction Problems. Journal of Basic Engineering 82(1), 35-45 (1960)
|
| 272 |
+
26. Kingma, D.P., Welling, M.: Auto-encoding variational bayes. stat 1050, 1 (2014)
|
| 273 |
+
27. Kramer, M.A.: Nonlinear principal component analysis using autoassociative neural networks. AIChE journal 37(2), 233-243 (1991)
|
| 274 |
+
28. Li, S., Singh, H., Grover, A.: Mamba-nd: Selective state space modeling for multidimensional data. arXiv preprint arXiv:2402.05892 (2024)
|
| 275 |
+
29. Li, Y., Cai, T., Zhang, Y., Chen, D., Dey, D.: What makes convolutional models great on long sequence modeling? In: The Eleventh International Conference on Learning Representations (2022)
|
| 276 |
+
30. Lin, X., Amer, M.R.: Human motion modeling using dvgans. arXiv preprint arXiv:1804.10652 (2018)
|
| 277 |
+
31. Liu, Y., Tian, Y., Zhao, Y., Yu, H., Xie, L., Wang, Y., Ye, Q., Liu, Y.: Vmamba: Visual state space model. arXiv preprint arXiv:2401.10166 (2024)
|
| 278 |
+
32. Liu, Z., Mao, H., Wu, C.Y., Feichtenhofer, C., Darrell, T., Xie, S.: A convnet for the 2020s. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 11976-11986 (2022)
|
| 279 |
+
33. Mahmood, N., Ghorbani, N., Troje, N.F., Pons-Moll, G., Black, M.J.: Amass: Archive of motion capture as surface shapes. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5442-5451 (2019)
|
| 280 |
+
34. Pavlakos, G., Choutas, V., Ghorbani, N., Bolkart, T., Osman, A.A.A., Tzionas, D., Black, M.J.: Expressive body capture: 3d hands, face, and body from a single image. In: Proceedings IEEE Conf. on Computer Vision and Pattern Recognition (CVPR) (2019)
|
| 281 |
+
35. Petrovich, M., Black, M.J., Varol, G.: Action-conditioned 3D human motion synthesis with transformer VAE. In: International Conference on Computer Vision (ICCV) (2021)
|
| 282 |
+
36. Petrovich, M., Black, M.J., Varol, G.: Temos: Generating diverse human motions from textual descriptions. In: European Conference on Computer Vision. pp. 480-497. Springer (2022)
|
| 283 |
+
|
| 284 |
+
37. Petrovich, M., Black, M.J., Varol, G.: TEMOS: Generating diverse human motions from textual descriptions. In: European Conference on Computer Vision (ECCV) (2022)
|
| 285 |
+
38. Plappert, M., Mandery, C., Asfour, T.: The kit motion-language dataset. Big data 4(4), 236-252 (2016)
|
| 286 |
+
39. Plappert, M., Mandery, C., Asfour, T.: Learning a bidirectional mapping between human whole-body motion and natural language using deep recurrent neural networks. Robotics and Autonomous Systems 109, 13-26 (2018)
|
| 287 |
+
40. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021)
|
| 288 |
+
41. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10684-10695 (2022)
|
| 289 |
+
42. Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomedical image segmentation. In: International Conference on Medical image computing and computer-assisted intervention. pp. 234-241. Springer (2015)
|
| 290 |
+
43. Rumelhart, D.E., Hinton, G.E., Williams, R.J.: Learning internal representations by error propagation. Parallel Distributed Processing pp. 318-362 (1986)
|
| 291 |
+
44. Sahara, C., Ho, J., Chan, W., Salimans, T., Fleet, D.J., Norouzi, M.: Image superresolution via iterative refinement. IEEE Transactions on Pattern Analysis and Machine Intelligence (2022)
|
| 292 |
+
45. Shi, X., Chen, Z., Wang, H., Yeung, D.Y., Wong, W.K., Woo, W.c.: Convolutional LSTM network: A machine learning approach for precipitation nowcasting. Advances in neural information processing systems 28 (2015)
|
| 293 |
+
46. Smith, J., De Mello, S., Kautz, J., Linderman, S., Byeon, W.: Convolutional state space models for long-range spatiotemporal modeling. Advances in Neural Information Processing Systems 36 (2024)
|
| 294 |
+
47. Sohl-Dickstein, J., Weiss, E., Maheswaranathan, N., Ganguli, S.: Deep unsupervised learning using nonequilibrium thermodynamics. In: International conference on machine learning. pp. 2256-2265. PMLR (2015)
|
| 295 |
+
48. Tevet, G., Gordon, B., Hertz, A., Bermano, A.H., Cohen-Or, D.: Motionclip: Exposing human motion generation to clip space. In: European Conference on Computer Vision. pp. 358-374. Springer (2022)
|
| 296 |
+
49. Tevet, G., Raab, S., Gordon, B., Shafir, Y., Cohen-or, D., Bermano, A.H.: Human motion diffusion model. In: The Eleventh International Conference on Learning Representations (2022)
|
| 297 |
+
50. Van Den Oord, A., Vinyals, O., et al.: Neural discrete representation learning. Advances in neural information processing systems 30 (2017)
|
| 298 |
+
51. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. Advances in neural information processing systems 30 (2017)
|
| 299 |
+
52. Wang, J., Yan, J.N., Gu, A., Rush, A.M.: Pretraining without attention. arXiv preprint arXiv:2212.10544 (2022)
|
| 300 |
+
53. Yan, J.N., Gu, J., Rush, A.M.: Diffusion models without attention. arXiv preprint arXiv:2311.18257 (2023)
|
| 301 |
+
54. Zhang, M., Cai, Z., Pan, L., Hong, F., Guo, X., Yang, L., Liu, Z.: Motiondiffuse: Text-driven human motion generation with diffusion model. IEEE Transactions on Pattern Analysis and Machine Intelligence (2024)
|
| 302 |
+
|
| 303 |
+
55. Zhang, Y., Black, M.J., Tang, S.: We are more than our joints: Predicting how 3d bodies move. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3372-3382 (2021)
|
| 304 |
+
56. Zhong, C., Hu, L., Zhang, Z., Xia, S.: Attt2m: Text-driven human motion generation with multi-perspective attention mechanism. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 509-519 (2023)
|
| 305 |
+
57. Zhu, L., Liao, B., Zhang, Q., Wang, X., Liu, W., Wang, X.: Vision mamba: Efficient visual representation learning with bidirectional state space model. arXiv preprint arXiv:2401.09417 (2024)
|
2024/Motion Mamba_ Efficient and Long Sequence Motion Generation/images.zip
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:1197c102592c1467b060b73dd3f39de20deb581b7121d862525565437c2688b7
|
| 3 |
+
size 447387
|
2024/Motion Mamba_ Efficient and Long Sequence Motion Generation/layout.json
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
2024/Motion and Structure from Event-based Normal Flow/2dddd958-d905-447a-8afe-98428ea7b185_content_list.json
ADDED
|
@@ -0,0 +1,1909 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
{
|
| 3 |
+
"type": "text",
|
| 4 |
+
"text": "Motion and Structure from Event-based Normal Flow",
|
| 5 |
+
"text_level": 1,
|
| 6 |
+
"bbox": [
|
| 7 |
+
271,
|
| 8 |
+
140,
|
| 9 |
+
730,
|
| 10 |
+
183
|
| 11 |
+
],
|
| 12 |
+
"page_idx": 0
|
| 13 |
+
},
|
| 14 |
+
{
|
| 15 |
+
"type": "text",
|
| 16 |
+
"text": "Zhongyang Ren $^{1*}$ , Bangyan Liao $^{2*}$ , Delei Kong $^{1}$ , Jinghang Li $^{1}$ , Peidong Liu $^{2}$ , Laurent Kneip $^{3}$ , Guillermo Gallego $^{4}$ , and Yi Zhou $^{1*}$",
|
| 17 |
+
"bbox": [
|
| 18 |
+
235,
|
| 19 |
+
210,
|
| 20 |
+
766,
|
| 21 |
+
243
|
| 22 |
+
],
|
| 23 |
+
"page_idx": 0
|
| 24 |
+
},
|
| 25 |
+
{
|
| 26 |
+
"type": "text",
|
| 27 |
+
"text": "$^{1}$ School of Robotics, Hunan University",
|
| 28 |
+
"bbox": [
|
| 29 |
+
367,
|
| 30 |
+
253,
|
| 31 |
+
635,
|
| 32 |
+
268
|
| 33 |
+
],
|
| 34 |
+
"page_idx": 0
|
| 35 |
+
},
|
| 36 |
+
{
|
| 37 |
+
"type": "text",
|
| 38 |
+
"text": "$^{2}$ School of Engineering, Westlake University",
|
| 39 |
+
"bbox": [
|
| 40 |
+
349,
|
| 41 |
+
268,
|
| 42 |
+
651,
|
| 43 |
+
282
|
| 44 |
+
],
|
| 45 |
+
"page_idx": 0
|
| 46 |
+
},
|
| 47 |
+
{
|
| 48 |
+
"type": "text",
|
| 49 |
+
"text": "<sup>3</sup> School of Information Science and Technology, ShanghaiTech University",
|
| 50 |
+
"bbox": [
|
| 51 |
+
254,
|
| 52 |
+
282,
|
| 53 |
+
746,
|
| 54 |
+
297
|
| 55 |
+
],
|
| 56 |
+
"page_idx": 0
|
| 57 |
+
},
|
| 58 |
+
{
|
| 59 |
+
"type": "text",
|
| 60 |
+
"text": "TU Berlin, ECDF, SCIoI Excellence Cluster and Robotics Institute Germany",
|
| 61 |
+
"bbox": [
|
| 62 |
+
238,
|
| 63 |
+
296,
|
| 64 |
+
764,
|
| 65 |
+
310
|
| 66 |
+
],
|
| 67 |
+
"page_idx": 0
|
| 68 |
+
},
|
| 69 |
+
{
|
| 70 |
+
"type": "text",
|
| 71 |
+
"text": "Abstract. Recovering the camera motion and scene geometry from visual data is a fundamental problem in computer vision. Its success in conventional (frame-based) vision is attributed to the maturity of feature extraction, data association and multi-view geometry. The emergence of asynchronous (event-based) cameras calls for new approaches that use raw event data as input to solve this fundamental problem. State-of-the-art solutions typically infer data association implicitly by iteratively reversing the event data generation process. However, the nonlinear nature of these methods limits their applicability in real-time tasks, and the constant-motion assumption leads to unstable results under agile motion. To this end, we reformulate the problem in a way that aligns better with the differential working principle of event cameras. We show that event-based normal flow can be used, via the proposed geometric error term, as an alternative to the full (optical) flow in solving a family of geometric problems that involve instantaneous first-order kinematics and scene geometry. Furthermore, we develop a fast linear solver and a continuous-time nonlinear solver on top of the proposed geometric error term. Experiments on both synthetic and real data show the superiority of our linear solver in terms of accuracy and efficiency, and its practicality as an initializer for previous nonlinear solvers. Besides, our continuous-time non-linear solver exhibits exceptional capabilities in accommodating sudden variations in motion since it does not rely on the constant-motion assumption. Our project page can be found at https://nail-hnu.github.io/EvLinearSolver/.",
|
| 72 |
+
"bbox": [
|
| 73 |
+
259,
|
| 74 |
+
347,
|
| 75 |
+
743,
|
| 76 |
+
680
|
| 77 |
+
],
|
| 78 |
+
"page_idx": 0
|
| 79 |
+
},
|
| 80 |
+
{
|
| 81 |
+
"type": "text",
|
| 82 |
+
"text": "Keywords: Event cameras $\\cdot$ Geometric model fitting $\\cdot$ Normal flow",
|
| 83 |
+
"bbox": [
|
| 84 |
+
261,
|
| 85 |
+
693,
|
| 86 |
+
718,
|
| 87 |
+
707
|
| 88 |
+
],
|
| 89 |
+
"page_idx": 0
|
| 90 |
+
},
|
| 91 |
+
{
|
| 92 |
+
"type": "text",
|
| 93 |
+
"text": "1 Introduction",
|
| 94 |
+
"text_level": 1,
|
| 95 |
+
"bbox": [
|
| 96 |
+
215,
|
| 97 |
+
755,
|
| 98 |
+
375,
|
| 99 |
+
771
|
| 100 |
+
],
|
| 101 |
+
"page_idx": 0
|
| 102 |
+
},
|
| 103 |
+
{
|
| 104 |
+
"type": "text",
|
| 105 |
+
"text": "Neuromorphic event-based cameras are bio-inspired sensors that report only brightness changes at each pixel location asynchronously. Their high temporal",
|
| 106 |
+
"bbox": [
|
| 107 |
+
212,
|
| 108 |
+
786,
|
| 109 |
+
787,
|
| 110 |
+
816
|
| 111 |
+
],
|
| 112 |
+
"page_idx": 0
|
| 113 |
+
},
|
| 114 |
+
{
|
| 115 |
+
"type": "page_footnote",
|
| 116 |
+
"text": "* equal contribution; $\\boxtimes$ corresponding author (eeyzhou@hnu.edu.cn).",
|
| 117 |
+
"bbox": [
|
| 118 |
+
230,
|
| 119 |
+
825,
|
| 120 |
+
692,
|
| 121 |
+
840
|
| 122 |
+
],
|
| 123 |
+
"page_idx": 0
|
| 124 |
+
},
|
| 125 |
+
{
|
| 126 |
+
"type": "text",
|
| 127 |
+
"text": "resolution enables them to sense at exactly the same rate as the scene dynamics, and thus, event cameras are suitable for dealing with robotic perception [33,38, 39,41,43], navigation [21,34,35,42], and control [6,9] tasks that involve agile motion. A general and fundamental problem among these tasks is to recover the camera's ego-motion and scene geometry from event data. One of the bottlenecks for solving this kind of problems at very high speeds is the lack of methods for efficient and robust data association.",
|
| 128 |
+
"bbox": [
|
| 129 |
+
212,
|
| 130 |
+
146,
|
| 131 |
+
787,
|
| 132 |
+
252
|
| 133 |
+
],
|
| 134 |
+
"page_idx": 1
|
| 135 |
+
},
|
| 136 |
+
{
|
| 137 |
+
"type": "text",
|
| 138 |
+
"text": "Unlike its standard-vision counterparts, an event camera cannot observe visual information (i.e. intensity changes) when motion is parallel to edges. Such a partial observability of event data makes optical flow estimation (as well as feature detection and matching) an ill-posed problem. To this end, many existing methods [14,17,19,22,30] for solving the problem of motion-model fitting on event data propose specific metrics that assess event-data association established in either an implicit manner or an explicitly model-driven style. The best fit can be retrieved by iteratively optimizing the data association, often expressed in terms of event alignment. However, we argue two flaws of these methods: $(i)$ The success of these non-linear solutions requires a proper initialization. While a brute-force search may be applicable, it is computationally expensive for real-time application, specially if the dimension of the parameter is high. Efficient and deterministic initialization methods are lacking and urgently needed. $(ii)$ The common assumption of constant velocity only holds within a short time interval. Nevertheless, a time window that is too narrow may contain insufficient number of events for a valid model fitting. The heuristic of using a constant number of events may violate the constant velocity assumption, especially in the case of sudden changes in speed.",
|
| 139 |
+
"bbox": [
|
| 140 |
+
212,
|
| 141 |
+
253,
|
| 142 |
+
787,
|
| 143 |
+
526
|
| 144 |
+
],
|
| 145 |
+
"page_idx": 1
|
| 146 |
+
},
|
| 147 |
+
{
|
| 148 |
+
"type": "text",
|
| 149 |
+
"text": "In this paper, we rethink the problem formulation in a way that is more consistent with the differential working principle of event cameras. Unlike existing pipelines that use events occurring within a time interval to resolve the relative motion, we attempt to recover the structure and instantaneous first-order kinematic parameters involved in the motion field equation, using as input an instantaneous observation, namely event-based normal flow. To this end, we explore effective constraints between normal-flow observations and model parameters for solving a family of motion-and-structure estimation problems, such as the estimation of depth and full optical flow, instantaneous angular velocity and linear velocity, and differential homography (see Fig. 1).",
|
| 150 |
+
"bbox": [
|
| 151 |
+
212,
|
| 152 |
+
529,
|
| 153 |
+
787,
|
| 154 |
+
680
|
| 155 |
+
],
|
| 156 |
+
"page_idx": 1
|
| 157 |
+
},
|
| 158 |
+
{
|
| 159 |
+
"type": "text",
|
| 160 |
+
"text": "The contributions of this paper can be summarized as follows:",
|
| 161 |
+
"bbox": [
|
| 162 |
+
240,
|
| 163 |
+
681,
|
| 164 |
+
687,
|
| 165 |
+
696
|
| 166 |
+
],
|
| 167 |
+
"page_idx": 1
|
| 168 |
+
},
|
| 169 |
+
{
|
| 170 |
+
"type": "list",
|
| 171 |
+
"sub_type": "text",
|
| 172 |
+
"list_items": [
|
| 173 |
+
"- A normal flow constraint with a geometric connection to event data. It overcomes the partial observability issue (i.e., errors caused by using normal flow as replacement for full optical flow) in general problems of geometric model fitting when normal flow is used as input.",
|
| 174 |
+
"- Two solvers for a family of motion-and-structure estimation problems using as input sparse event-based normal flow. Our linear solver (can be used with RANSAC) leads to closed-form and deterministic solutions that can be used as an initialization to existing nonlinear methods. Our nonlinear solver,"
|
| 175 |
+
],
|
| 176 |
+
"bbox": [
|
| 177 |
+
225,
|
| 178 |
+
715,
|
| 179 |
+
784,
|
| 180 |
+
837
|
| 181 |
+
],
|
| 182 |
+
"page_idx": 1
|
| 183 |
+
},
|
| 184 |
+
{
|
| 185 |
+
"type": "page_number",
|
| 186 |
+
"text": "2",
|
| 187 |
+
"bbox": [
|
| 188 |
+
217,
|
| 189 |
+
114,
|
| 190 |
+
228,
|
| 191 |
+
126
|
| 192 |
+
],
|
| 193 |
+
"page_idx": 1
|
| 194 |
+
},
|
| 195 |
+
{
|
| 196 |
+
"type": "header",
|
| 197 |
+
"text": "Z. Ren et al.",
|
| 198 |
+
"bbox": [
|
| 199 |
+
271,
|
| 200 |
+
114,
|
| 201 |
+
357,
|
| 202 |
+
127
|
| 203 |
+
],
|
| 204 |
+
"page_idx": 1
|
| 205 |
+
},
|
| 206 |
+
{
|
| 207 |
+
"type": "image",
|
| 208 |
+
"img_path": "images/d1f0b4c57510e255c6b64fbfb4fe1cca27618059d35bbcd12fc8f37b680f0a43.jpg",
|
| 209 |
+
"image_caption": [
|
| 210 |
+
"Fig. 1: Overview of our event-based motion estimation approach. Normal flow vectors are computed from the raw event data and are used as input to the two solvers. The result of the linear solver consists of fitted motion models $\\{\\pmb{\\theta}_1, \\pmb{\\theta}_2, \\dots\\}$ at discrete time instants, which can be used to initialize the continuous-time nonlinear solver."
|
| 211 |
+
],
|
| 212 |
+
"image_footnote": [],
|
| 213 |
+
"bbox": [
|
| 214 |
+
233,
|
| 215 |
+
143,
|
| 216 |
+
767,
|
| 217 |
+
292
|
| 218 |
+
],
|
| 219 |
+
"page_idx": 2
|
| 220 |
+
},
|
| 221 |
+
{
|
| 222 |
+
"type": "text",
|
| 223 |
+
"text": "under a continuous-time formulation, can handle sudden speed variations, and thus, is free of the constant-motion assumption.",
|
| 224 |
+
"bbox": [
|
| 225 |
+
240,
|
| 226 |
+
388,
|
| 227 |
+
782,
|
| 228 |
+
417
|
| 229 |
+
],
|
| 230 |
+
"page_idx": 2
|
| 231 |
+
},
|
| 232 |
+
{
|
| 233 |
+
"type": "text",
|
| 234 |
+
"text": "- A thorough evaluation on the two proposed solvers, comparing against state-of-the-art methods, and also including an investigation on the performance of existing non-linear methods initialized using our linear solver.",
|
| 235 |
+
"bbox": [
|
| 236 |
+
223,
|
| 237 |
+
417,
|
| 238 |
+
782,
|
| 239 |
+
463
|
| 240 |
+
],
|
| 241 |
+
"page_idx": 2
|
| 242 |
+
},
|
| 243 |
+
{
|
| 244 |
+
"type": "text",
|
| 245 |
+
"text": "Outline: The rest of the paper is organized as follows. We first discuss the related work in Sec. 2. Then a compact review of preliminary knowledge is given in Sec. 3. Our approach is presented in Sec. 4, including the normal-flow constraint, the linear models of a family of motion-and-structure estimation problems, and the proposed linear and nonlinear solvers. Experimental evaluation is conducted in Sec. 5 and conclusion are drawn in Sec. 6.",
|
| 246 |
+
"bbox": [
|
| 247 |
+
212,
|
| 248 |
+
472,
|
| 249 |
+
787,
|
| 250 |
+
564
|
| 251 |
+
],
|
| 252 |
+
"page_idx": 2
|
| 253 |
+
},
|
| 254 |
+
{
|
| 255 |
+
"type": "text",
|
| 256 |
+
"text": "2 Related Work",
|
| 257 |
+
"text_level": 1,
|
| 258 |
+
"bbox": [
|
| 259 |
+
215,
|
| 260 |
+
585,
|
| 261 |
+
387,
|
| 262 |
+
602
|
| 263 |
+
],
|
| 264 |
+
"page_idx": 2
|
| 265 |
+
},
|
| 266 |
+
{
|
| 267 |
+
"type": "text",
|
| 268 |
+
"text": "2.1 Event-based Geometric Model Fitting",
|
| 269 |
+
"text_level": 1,
|
| 270 |
+
"bbox": [
|
| 271 |
+
214,
|
| 272 |
+
618,
|
| 273 |
+
576,
|
| 274 |
+
633
|
| 275 |
+
],
|
| 276 |
+
"page_idx": 2
|
| 277 |
+
},
|
| 278 |
+
{
|
| 279 |
+
"type": "text",
|
| 280 |
+
"text": "The major body of literature on this specific topic follows an iterative scheme, which typically has its own special metric evaluating quantitatively the association between events, and solve the model fitting as a nonlinear optimization problem. Among the first works is the Contrast Maximization method [13, 14], which provides a unified framework for solving a family of motion-model fitting problems. It searches the optimal model parameters that ultimately lead to a sharp image of warped events (IWE). The goodness of fit is evaluated by the alignment of all involved events along point trajectories on the image plane, and such alignment is assessed by the strength of the contours of the IWE. The computational complexity of CMax is high because of iteratively updating of IWE, whose cost scales with the spatial resolution and the amount of event data.",
|
| 281 |
+
"bbox": [
|
| 282 |
+
212,
|
| 283 |
+
643,
|
| 284 |
+
787,
|
| 285 |
+
809
|
| 286 |
+
],
|
| 287 |
+
"page_idx": 2
|
| 288 |
+
},
|
| 289 |
+
{
|
| 290 |
+
"type": "text",
|
| 291 |
+
"text": "To lessen the dependency on IWE, Nunes et al. proposes an entropy minimization framework (EMin) [30] that replaces the dispersion metric of [14] with",
|
| 292 |
+
"bbox": [
|
| 293 |
+
214,
|
| 294 |
+
809,
|
| 295 |
+
785,
|
| 296 |
+
840
|
| 297 |
+
],
|
| 298 |
+
"page_idx": 2
|
| 299 |
+
},
|
| 300 |
+
{
|
| 301 |
+
"type": "header",
|
| 302 |
+
"text": "Motion and Structure from Event-based Normal Flow",
|
| 303 |
+
"bbox": [
|
| 304 |
+
372,
|
| 305 |
+
114,
|
| 306 |
+
732,
|
| 307 |
+
127
|
| 308 |
+
],
|
| 309 |
+
"page_idx": 2
|
| 310 |
+
},
|
| 311 |
+
{
|
| 312 |
+
"type": "page_number",
|
| 313 |
+
"text": "3",
|
| 314 |
+
"bbox": [
|
| 315 |
+
774,
|
| 316 |
+
116,
|
| 317 |
+
784,
|
| 318 |
+
126
|
| 319 |
+
],
|
| 320 |
+
"page_idx": 2
|
| 321 |
+
},
|
| 322 |
+
{
|
| 323 |
+
"type": "text",
|
| 324 |
+
"text": "a similarity measurement between events, enabling to evaluate the goodness of fit in a higher dimensional space. EMin is, still, too expensive for real-time application. Liu et al. [22] propose a two-step iterative method for the problem of 3-D rotation estimation using events as input. Given a rotation estimate, the first step determines event correspondences by looking for the nearest neighbour in the spatio-temporal domain. The second step employs a spatio-temporal registration which aligns corresponding events by minimizing the sum of squared geodesic residual. The least-square nature of the energy function in the second step leads to a computation friendly performance. Gu et al. [17] present a probabilistic approach for spatio-temporal alignment of events. The aligned events at a particular pixel are modeled as a spatio-temporal Poisson point process and the goodness of fit can be quantified using the likelihood of the data. The optimal model is retrieved when the likelihood is maximized. More recently, an event-to-map alignment scheme is proposed in [19], which adopts the time surface (TS) map as a reference for event alignment. The proposed TS loss function drives later events to get increasingly closer to pixel positions of corresponding earlier events, thus, leading to a progressively sharp TS map as iterations proceed.",
|
| 325 |
+
"bbox": [
|
| 326 |
+
212,
|
| 327 |
+
146,
|
| 328 |
+
787,
|
| 329 |
+
404
|
| 330 |
+
],
|
| 331 |
+
"page_idx": 3
|
| 332 |
+
},
|
| 333 |
+
{
|
| 334 |
+
"type": "text",
|
| 335 |
+
"text": "In summary, all the above-mentioned methods are nonlinear, and thus, a proper initialization is necessary. Although a brute-force search is widely adopted, sometimes it can be too expensive to locate the convergence basin in parameter space. Moreover, [14,30] may come across a convergence to false-positive minima, known as event collapse [36, 37]. Hence, a quick and deterministic solver (e.g., a closed-form solution) is desirable. Besides, most of these methods assume a constant-motion model, which limits their applicability to scenarios of sudden speed variations. In addition, we have noticed that a recent work [16] can jointly estimate the line parameters and the camera's linear velocities. However, this method is not a general one due to the dependence on line structures and the usage of an IMU. Thus, it is not within the scope of our discussion.",
|
| 336 |
+
"bbox": [
|
| 337 |
+
212,
|
| 338 |
+
404,
|
| 339 |
+
787,
|
| 340 |
+
570
|
| 341 |
+
],
|
| 342 |
+
"page_idx": 3
|
| 343 |
+
},
|
| 344 |
+
{
|
| 345 |
+
"type": "text",
|
| 346 |
+
"text": "2.2 Motion and Structure from Normal Flow",
|
| 347 |
+
"text_level": 1,
|
| 348 |
+
"bbox": [
|
| 349 |
+
215,
|
| 350 |
+
597,
|
| 351 |
+
602,
|
| 352 |
+
612
|
| 353 |
+
],
|
| 354 |
+
"page_idx": 3
|
| 355 |
+
},
|
| 356 |
+
{
|
| 357 |
+
"type": "text",
|
| 358 |
+
"text": "Due to the aperture problem, a local estimate of optical flow recovers only the partial component of the full flow vector along the direction of image gradient. Preliminary knowledge about normal flow can be found in the supplementary material. Recovering motion and structure from normal flow on intensity images is a well-studied topic. Its main body of literature typically proposes a qualitatively deterministic solution by which a set (even infinite) of applicable ego motions are determined via an analysis of normal flow classification [3, 11, 20]. Only recently, researches devised ways to formulate the Cheirality constraint as a differentiable term and use it in a continuous optimization approach [1] or as a loss function for training an artificial neural network [31].",
|
| 359 |
+
"bbox": [
|
| 360 |
+
212,
|
| 361 |
+
627,
|
| 362 |
+
787,
|
| 363 |
+
779
|
| 364 |
+
],
|
| 365 |
+
"page_idx": 3
|
| 366 |
+
},
|
| 367 |
+
{
|
| 368 |
+
"type": "text",
|
| 369 |
+
"text": "In event-based perception and navigation, similar methodologies as the frame-based ones are also witnessed, where normal flow computed from event data is used as input to determine the Focus of Expansion (FoE) for drones' course estimation [10] and obstacle avoidance [5].",
|
| 370 |
+
"bbox": [
|
| 371 |
+
212,
|
| 372 |
+
779,
|
| 373 |
+
787,
|
| 374 |
+
840
|
| 375 |
+
],
|
| 376 |
+
"page_idx": 3
|
| 377 |
+
},
|
| 378 |
+
{
|
| 379 |
+
"type": "page_number",
|
| 380 |
+
"text": "4",
|
| 381 |
+
"bbox": [
|
| 382 |
+
217,
|
| 383 |
+
114,
|
| 384 |
+
228,
|
| 385 |
+
126
|
| 386 |
+
],
|
| 387 |
+
"page_idx": 3
|
| 388 |
+
},
|
| 389 |
+
{
|
| 390 |
+
"type": "header",
|
| 391 |
+
"text": "Z. Ren et al.",
|
| 392 |
+
"bbox": [
|
| 393 |
+
271,
|
| 394 |
+
114,
|
| 395 |
+
357,
|
| 396 |
+
127
|
| 397 |
+
],
|
| 398 |
+
"page_idx": 3
|
| 399 |
+
},
|
| 400 |
+
{
|
| 401 |
+
"type": "text",
|
| 402 |
+
"text": "Most of these heuristic methods, while effective, are cumbersome for practical application. In contrast, our method delivers a deterministic way to retrieve motion and structure from normal flow. Additionally, it is noticed that normal flow can replace full flow in some specific problems of geometric model fitting, such as affine models [26, 27]. In this way, the geometric model fitting problem can be easily solved in closed-form. Nevertheless, these methods do not consider the observability bias in normal flows, and thus, such naive substitution may lead to inaccurate model fitting. To this end, we present a robust linear solver based on a carefully-designed geometric measurement model that can overcome the partial observability of event-based normal flow. It is worth mentioning that a recent work [24] also determines instantaneous linear velocity using as input the normal flow. Our difference is seen in the scope of applicability. The usage of normal flow is restricted in the motion flow equation of [24], while our paper extends it to a wider scope (i.e., a family of geometric model fitting problems) by combining normal flow with e.g., differential epipolar geometry, differential homography, etc.",
|
| 403 |
+
"bbox": [
|
| 404 |
+
212,
|
| 405 |
+
146,
|
| 406 |
+
787,
|
| 407 |
+
388
|
| 408 |
+
],
|
| 409 |
+
"page_idx": 4
|
| 410 |
+
},
|
| 411 |
+
{
|
| 412 |
+
"type": "text",
|
| 413 |
+
"text": "3 Preliminaries",
|
| 414 |
+
"text_level": 1,
|
| 415 |
+
"bbox": [
|
| 416 |
+
215,
|
| 417 |
+
409,
|
| 418 |
+
380,
|
| 419 |
+
425
|
| 420 |
+
],
|
| 421 |
+
"page_idx": 4
|
| 422 |
+
},
|
| 423 |
+
{
|
| 424 |
+
"type": "text",
|
| 425 |
+
"text": "To make this paper self-contained, some preliminary knowledge is revisited. Due to space limitations, here we only list essential concepts. A more detailed review of preliminary knowledge can be found in the supplementary material.",
|
| 426 |
+
"bbox": [
|
| 427 |
+
212,
|
| 428 |
+
439,
|
| 429 |
+
787,
|
| 430 |
+
486
|
| 431 |
+
],
|
| 432 |
+
"page_idx": 4
|
| 433 |
+
},
|
| 434 |
+
{
|
| 435 |
+
"type": "text",
|
| 436 |
+
"text": "Motion Field. Consider a perspective camera moving in a static environment, with instantaneous angular and linear velocities in the camera frame $\\boldsymbol{\\omega} = (\\omega_{x},\\omega_{y},\\omega_{z})^{\\top}$ and $\\pmb {\\nu} = (\\nu_{x},\\nu_{y},\\nu_{z})^{\\top}$ , respectively. The motion field at pixel $\\mathbf{x} = (x,y)^{\\top}$ produced by projecting the 3D motion at point with depth $Z(\\mathbf{x})$ is",
|
| 437 |
+
"bbox": [
|
| 438 |
+
212,
|
| 439 |
+
505,
|
| 440 |
+
787,
|
| 441 |
+
568
|
| 442 |
+
],
|
| 443 |
+
"page_idx": 4
|
| 444 |
+
},
|
| 445 |
+
{
|
| 446 |
+
"type": "equation",
|
| 447 |
+
"text": "\n$$\n\\mathbf {u} = \\frac {1}{Z (\\mathbf {x})} \\mathbf {A} (\\mathbf {x}) \\boldsymbol {\\nu} + \\mathbf {B} (\\mathbf {x}) \\boldsymbol {\\omega}, \\tag {1}\n$$\n",
|
| 448 |
+
"text_format": "latex",
|
| 449 |
+
"bbox": [
|
| 450 |
+
403,
|
| 451 |
+
575,
|
| 452 |
+
785,
|
| 453 |
+
608
|
| 454 |
+
],
|
| 455 |
+
"page_idx": 4
|
| 456 |
+
},
|
| 457 |
+
{
|
| 458 |
+
"type": "text",
|
| 459 |
+
"text": "where $\\mathbf{A}(\\mathbf{x})$ and $\\mathbf{B}(\\mathbf{x})$ are $2\\times 3$ matrices that solely depend on $\\mathbf{x}$ . Equation (1) is the motion field equation [40, Ch.8] [23]. Optical flow in rigid scenes aims at estimating the motion field from visual data, which is only possible in regions that do not suffer from the aperture problem.",
|
| 460 |
+
"bbox": [
|
| 461 |
+
212,
|
| 462 |
+
614,
|
| 463 |
+
787,
|
| 464 |
+
676
|
| 465 |
+
],
|
| 466 |
+
"page_idx": 4
|
| 467 |
+
},
|
| 468 |
+
{
|
| 469 |
+
"type": "text",
|
| 470 |
+
"text": "Brightness constancy is the assumption stating that the true motion trajectories $\\mathbf{x}(t)$ followed by objects on the image plane are those at which brightness is constant: $I(\\mathbf{x}(t), t) = \\mathrm{const}$ . Differentiating in time (total derivative) gives the equation at $(\\mathbf{x}, t)$ : $\\nabla I \\cdot \\dot{\\mathbf{x}} + \\partial_t I = 0$ , where $\\dot{\\mathbf{x}} = d\\mathbf{x} / dt$ is the velocity of image point $\\mathbf{x}(t)$ . Given $I$ , this equation provides one constraint per point $(\\mathbf{x}, t)$ to estimate $\\dot{\\mathbf{x}}$ , now called optical flow. Without additional equations, only the normal component of the optical flow can be determined [40],",
|
| 471 |
+
"bbox": [
|
| 472 |
+
212,
|
| 473 |
+
696,
|
| 474 |
+
787,
|
| 475 |
+
803
|
| 476 |
+
],
|
| 477 |
+
"page_idx": 4
|
| 478 |
+
},
|
| 479 |
+
{
|
| 480 |
+
"type": "equation",
|
| 481 |
+
"text": "\n$$\n\\mathbf {n} = - \\frac {\\partial_ {t} I}{\\| \\nabla I \\| ^ {2}} \\nabla I. \\tag {2}\n$$\n",
|
| 482 |
+
"text_format": "latex",
|
| 483 |
+
"bbox": [
|
| 484 |
+
439,
|
| 485 |
+
811,
|
| 486 |
+
785,
|
| 487 |
+
844
|
| 488 |
+
],
|
| 489 |
+
"page_idx": 4
|
| 490 |
+
},
|
| 491 |
+
{
|
| 492 |
+
"type": "header",
|
| 493 |
+
"text": "Motion and Structure from Event-based Normal Flow",
|
| 494 |
+
"bbox": [
|
| 495 |
+
372,
|
| 496 |
+
114,
|
| 497 |
+
732,
|
| 498 |
+
127
|
| 499 |
+
],
|
| 500 |
+
"page_idx": 4
|
| 501 |
+
},
|
| 502 |
+
{
|
| 503 |
+
"type": "page_number",
|
| 504 |
+
"text": "5",
|
| 505 |
+
"bbox": [
|
| 506 |
+
774,
|
| 507 |
+
116,
|
| 508 |
+
784,
|
| 509 |
+
126
|
| 510 |
+
],
|
| 511 |
+
"page_idx": 4
|
| 512 |
+
},
|
| 513 |
+
{
|
| 514 |
+
"type": "text",
|
| 515 |
+
"text": "Normal Flow. Decomposing the optical flow $\\mathbf{u} = \\mathbf{u}_{\\perp} + \\mathbf{u}_{\\parallel}$ into its normal and parallel components to the local image gradient $(\\nabla I)$ , the normal flow is $\\mathbf{n} = \\mathbf{u}_{\\perp}$ , and the dot-product with $\\mathbf{n}$ gives",
|
| 516 |
+
"bbox": [
|
| 517 |
+
215,
|
| 518 |
+
146,
|
| 519 |
+
787,
|
| 520 |
+
191
|
| 521 |
+
],
|
| 522 |
+
"page_idx": 5
|
| 523 |
+
},
|
| 524 |
+
{
|
| 525 |
+
"type": "equation",
|
| 526 |
+
"text": "\n$$\n\\mathbf {n} ^ {\\top} \\mathbf {u} = \\| \\mathbf {n} \\| ^ {2}, \\quad \\text {o r e q u i v a l e n t l y} \\quad (\\mathbf {u} - \\mathbf {n}) \\cdot \\mathbf {n} = 0. \\tag {3}\n$$\n",
|
| 527 |
+
"text_format": "latex",
|
| 528 |
+
"bbox": [
|
| 529 |
+
323,
|
| 530 |
+
202,
|
| 531 |
+
785,
|
| 532 |
+
219
|
| 533 |
+
],
|
| 534 |
+
"page_idx": 5
|
| 535 |
+
},
|
| 536 |
+
{
|
| 537 |
+
"type": "text",
|
| 538 |
+
"text": "Differential Homography. (Continuous Homography Matrix [25, Ch.5]). If all scene points lie on a plane, their coordinates are related to their image projections via a homography transformation (which encodes the information of the scene plane $\\{\\mathbf{N}, d\\}$ and the camera pose). In this case, the motion field (1) can be rewritten to adopt the form in homogeneous coordinates",
|
| 539 |
+
"bbox": [
|
| 540 |
+
214,
|
| 541 |
+
241,
|
| 542 |
+
782,
|
| 543 |
+
316
|
| 544 |
+
],
|
| 545 |
+
"page_idx": 5
|
| 546 |
+
},
|
| 547 |
+
{
|
| 548 |
+
"type": "equation",
|
| 549 |
+
"text": "\n$$\n\\hat {\\mathbf {u}} (\\mathbf {x}) = \\left(\\mathbf {1} - \\hat {\\mathbf {x}} \\mathbf {e} _ {3} ^ {\\top}\\right) \\mathbf {H} _ {d} \\hat {\\mathbf {x}}, \\tag {4}\n$$\n",
|
| 550 |
+
"text_format": "latex",
|
| 551 |
+
"bbox": [
|
| 552 |
+
416,
|
| 553 |
+
328,
|
| 554 |
+
784,
|
| 555 |
+
345
|
| 556 |
+
],
|
| 557 |
+
"page_idx": 5
|
| 558 |
+
},
|
| 559 |
+
{
|
| 560 |
+
"type": "text",
|
| 561 |
+
"text": "where $\\mathbf{H}_d\\dot{=} -\\left([ \\boldsymbol {\\omega}]_{\\times} + d^{-1}\\boldsymbol {\\nu}\\mathbf{N}^{\\top}\\right)$ denotes the differential homography matrix $([\\cdot ]_{\\times}$ refers to the cross-product, skew-symmetric matrix), $\\mathbf{1}$ is the identity matrix, $\\hat{\\mathbf{x}} = (x,y,1)^{\\top}$ , $\\hat{\\mathbf{u}} = (u_x,u_y,0)^{\\top}$ and $\\mathbf{e}_3 = (0,0,1)^{\\top}$ .",
|
| 562 |
+
"bbox": [
|
| 563 |
+
215,
|
| 564 |
+
357,
|
| 565 |
+
782,
|
| 566 |
+
404
|
| 567 |
+
],
|
| 568 |
+
"page_idx": 5
|
| 569 |
+
},
|
| 570 |
+
{
|
| 571 |
+
"type": "text",
|
| 572 |
+
"text": "Differential Epipolar Geometry. (Continuous Epipolar Geometry [25, Ch.5]). The epipolar geometry examines the distance between a potential match to the determined epipolar line, and can be expressed as $\\hat{\\mathbf{x}}^{\\prime \\top}\\mathbf{E}\\hat{\\mathbf{x}} = 0$ . The differential epipolar geometry can be derived analogously, providing a relationship between $\\mathbf{x},\\mathbf{u}$ and the camera velocities $\\pmb {\\nu},\\pmb{\\omega}$ . In homogeneous coordinates:",
|
| 573 |
+
"bbox": [
|
| 574 |
+
214,
|
| 575 |
+
425,
|
| 576 |
+
787,
|
| 577 |
+
501
|
| 578 |
+
],
|
| 579 |
+
"page_idx": 5
|
| 580 |
+
},
|
| 581 |
+
{
|
| 582 |
+
"type": "equation",
|
| 583 |
+
"text": "\n$$\n\\hat {\\mathbf {u}} (\\mathbf {x}) ^ {\\top} [ \\boldsymbol {\\nu} ] _ {\\times} \\hat {\\mathbf {x}} - \\hat {\\mathbf {x}} ^ {\\top} \\mathbf {s} \\hat {\\mathbf {x}} = 0, \\tag {5}\n$$\n",
|
| 584 |
+
"text_format": "latex",
|
| 585 |
+
"bbox": [
|
| 586 |
+
413,
|
| 587 |
+
512,
|
| 588 |
+
784,
|
| 589 |
+
529
|
| 590 |
+
],
|
| 591 |
+
"page_idx": 5
|
| 592 |
+
},
|
| 593 |
+
{
|
| 594 |
+
"type": "text",
|
| 595 |
+
"text": "where $\\mathbf{s} = \\frac{1}{2} ([\\pmb {\\nu}]_{\\times}[\\pmb {\\omega}]_{\\times} + [\\pmb {\\omega}]_{\\times}[\\pmb {\\nu}]_{\\times})$ is a $3\\times 3$ symmetric matrix.",
|
| 596 |
+
"bbox": [
|
| 597 |
+
215,
|
| 598 |
+
540,
|
| 599 |
+
671,
|
| 600 |
+
558
|
| 601 |
+
],
|
| 602 |
+
"page_idx": 5
|
| 603 |
+
},
|
| 604 |
+
{
|
| 605 |
+
"type": "text",
|
| 606 |
+
"text": "4 Methodology",
|
| 607 |
+
"text_level": 1,
|
| 608 |
+
"bbox": [
|
| 609 |
+
215,
|
| 610 |
+
579,
|
| 611 |
+
380,
|
| 612 |
+
597
|
| 613 |
+
],
|
| 614 |
+
"page_idx": 5
|
| 615 |
+
},
|
| 616 |
+
{
|
| 617 |
+
"type": "text",
|
| 618 |
+
"text": "Figure 1 summarizes our approach, grounded upon event-based normal flow. In this section, we first disclose the normal-flow constraint that can overcome the partial observability of normal flow when replacing full optical flow with normal flow in general tasks of geometric model fitting (Sec. 4.1). By combining the constraint and the theory in Sec. 3, we derive linear equations for a family of motion-and-structure estimation problems (Sec. 4.2). To solve these problems, we present a linear solver and a continuous-time nonlinear solver (Sec. 4.3).",
|
| 619 |
+
"bbox": [
|
| 620 |
+
214,
|
| 621 |
+
611,
|
| 622 |
+
787,
|
| 623 |
+
717
|
| 624 |
+
],
|
| 625 |
+
"page_idx": 5
|
| 626 |
+
},
|
| 627 |
+
{
|
| 628 |
+
"type": "text",
|
| 629 |
+
"text": "4.1 Normal Flow Constraint",
|
| 630 |
+
"text_level": 1,
|
| 631 |
+
"bbox": [
|
| 632 |
+
215,
|
| 633 |
+
739,
|
| 634 |
+
465,
|
| 635 |
+
753
|
| 636 |
+
],
|
| 637 |
+
"page_idx": 5
|
| 638 |
+
},
|
| 639 |
+
{
|
| 640 |
+
"type": "text",
|
| 641 |
+
"text": "Normal flow can be computed easily from the event data, for example by fitting planes locally to the space-time \"cloud\" of events on image space [2, 32]. Specifically, assuming that the event data $e_k \\doteq \\{\\mathbf{x}_k, t_k\\}$ spans locally a surface in the spatio-temporal domain, the surface can be defined by a function $\\Sigma_e: \\mathbb{R}^2 \\to \\mathbb{R}; \\mathbf{x}_k \\mapsto t_k$ that maps the pixel coordinates to the latest event's",
|
| 642 |
+
"bbox": [
|
| 643 |
+
214,
|
| 644 |
+
763,
|
| 645 |
+
787,
|
| 646 |
+
842
|
| 647 |
+
],
|
| 648 |
+
"page_idx": 5
|
| 649 |
+
},
|
| 650 |
+
{
|
| 651 |
+
"type": "page_number",
|
| 652 |
+
"text": "6",
|
| 653 |
+
"bbox": [
|
| 654 |
+
217,
|
| 655 |
+
114,
|
| 656 |
+
228,
|
| 657 |
+
126
|
| 658 |
+
],
|
| 659 |
+
"page_idx": 5
|
| 660 |
+
},
|
| 661 |
+
{
|
| 662 |
+
"type": "header",
|
| 663 |
+
"text": "Z. Ren et al.",
|
| 664 |
+
"bbox": [
|
| 665 |
+
271,
|
| 666 |
+
114,
|
| 667 |
+
357,
|
| 668 |
+
127
|
| 669 |
+
],
|
| 670 |
+
"page_idx": 5
|
| 671 |
+
},
|
| 672 |
+
{
|
| 673 |
+
"type": "image",
|
| 674 |
+
"img_path": "images/4b6f04504e5b1e1da6053b87c1bffd9164d3a1fd887b5441ec3b879c7da47058.jpg",
|
| 675 |
+
"image_caption": [
|
| 676 |
+
"Fig. 2: Toy example of applying the proposed normal-flow constraint. (a): A 2D registration task using as input either optical flow or normal flow. The groundtruth displacement is defined by a global flow of $(1.732, -1)^{\\top}$ . (b)-(d): The loss landscape obtained using different geometric measurements in the registration task. The red dot denotes the resulting displacement, and the green cross the groundtruth displacement. The bottom-left regions in (b)-(d) display the corresponding registration results."
|
| 677 |
+
],
|
| 678 |
+
"image_footnote": [],
|
| 679 |
+
"bbox": [
|
| 680 |
+
215,
|
| 681 |
+
143,
|
| 682 |
+
341,
|
| 683 |
+
253
|
| 684 |
+
],
|
| 685 |
+
"page_idx": 6
|
| 686 |
+
},
|
| 687 |
+
{
|
| 688 |
+
"type": "image",
|
| 689 |
+
"img_path": "images/2d17edd895e022ec3c07138772e36bd4fa5c35063e614e26fea68de5c2b3b612.jpg",
|
| 690 |
+
"image_caption": [],
|
| 691 |
+
"image_footnote": [],
|
| 692 |
+
"bbox": [
|
| 693 |
+
341,
|
| 694 |
+
143,
|
| 695 |
+
477,
|
| 696 |
+
253
|
| 697 |
+
],
|
| 698 |
+
"page_idx": 6
|
| 699 |
+
},
|
| 700 |
+
{
|
| 701 |
+
"type": "image",
|
| 702 |
+
"img_path": "images/be2ae15f15e7913c2af0fcdd3c71127563bfd3f69bee8e501cce550b212cf651.jpg",
|
| 703 |
+
"image_caption": [],
|
| 704 |
+
"image_footnote": [],
|
| 705 |
+
"bbox": [
|
| 706 |
+
478,
|
| 707 |
+
143,
|
| 708 |
+
632,
|
| 709 |
+
253
|
| 710 |
+
],
|
| 711 |
+
"page_idx": 6
|
| 712 |
+
},
|
| 713 |
+
{
|
| 714 |
+
"type": "image",
|
| 715 |
+
"img_path": "images/52aecd3ac01157809c0e7563adeadf90b1b1db4698056f539adf29834c99358c.jpg",
|
| 716 |
+
"image_caption": [],
|
| 717 |
+
"image_footnote": [],
|
| 718 |
+
"bbox": [
|
| 719 |
+
633,
|
| 720 |
+
143,
|
| 721 |
+
785,
|
| 722 |
+
253
|
| 723 |
+
],
|
| 724 |
+
"page_idx": 6
|
| 725 |
+
},
|
| 726 |
+
{
|
| 727 |
+
"type": "text",
|
| 728 |
+
"text": "timestamp [2, 5]. The normal flow is parallel to the spatial derivative of $\\Sigma_{e}$ , namely $\\nabla \\Sigma_{e}(\\mathbf{x})$ . More specifically, it can be calculated by",
|
| 729 |
+
"bbox": [
|
| 730 |
+
214,
|
| 731 |
+
377,
|
| 732 |
+
782,
|
| 733 |
+
409
|
| 734 |
+
],
|
| 735 |
+
"page_idx": 6
|
| 736 |
+
},
|
| 737 |
+
{
|
| 738 |
+
"type": "equation",
|
| 739 |
+
"text": "\n$$\n\\mathbf {n} (\\mathbf {x}) = \\frac {\\nabla \\Sigma_ {e} (\\mathbf {x})}{\\left\\| \\nabla \\Sigma_ {e} (\\mathbf {x}) \\right\\| ^ {2}}. \\tag {6}\n$$\n",
|
| 740 |
+
"text_format": "latex",
|
| 741 |
+
"bbox": [
|
| 742 |
+
426,
|
| 743 |
+
420,
|
| 744 |
+
785,
|
| 745 |
+
455
|
| 746 |
+
],
|
| 747 |
+
"page_idx": 6
|
| 748 |
+
},
|
| 749 |
+
{
|
| 750 |
+
"type": "text",
|
| 751 |
+
"text": "Readers may refer to the supplementary material for a more detailed derivation. Surely there are other methods to extract event-based normal flow and the solvers we develop in this section are independent of them, i.e., our solvers can also be fed with the extracted normal flow data from such methods.",
|
| 752 |
+
"bbox": [
|
| 753 |
+
214,
|
| 754 |
+
465,
|
| 755 |
+
782,
|
| 756 |
+
525
|
| 757 |
+
],
|
| 758 |
+
"page_idx": 6
|
| 759 |
+
},
|
| 760 |
+
{
|
| 761 |
+
"type": "text",
|
| 762 |
+
"text": "Event-based normal flow cannot be used directly as a replacement for the full flow in (1) to solve for the motion parameters. However, we can leverage the relationship between both flows (3) to define an effective geometric error criterion that circumvents the partial observability of normal flow. If the flow $\\mathbf{u}$ depends on some unknown (\"state\") variables $\\pmb{\\theta}$ , we propose estimating them based on how well the normal flow constraint (3) is satisfied, e.g., by minimizing some function of the error terms",
|
| 763 |
+
"bbox": [
|
| 764 |
+
214,
|
| 765 |
+
526,
|
| 766 |
+
784,
|
| 767 |
+
630
|
| 768 |
+
],
|
| 769 |
+
"page_idx": 6
|
| 770 |
+
},
|
| 771 |
+
{
|
| 772 |
+
"type": "equation",
|
| 773 |
+
"text": "\n$$\n\\mathbf {n} (\\mathbf {x}) ^ {\\top} \\mathbf {u} (\\mathbf {x}; \\boldsymbol {\\theta}) - \\| \\mathbf {n} (\\mathbf {x}) \\| ^ {2} \\doteq \\epsilon_ {\\mathrm {n f}} \\in \\mathbb {R}. \\tag {7}\n$$\n",
|
| 774 |
+
"text_format": "latex",
|
| 775 |
+
"bbox": [
|
| 776 |
+
372,
|
| 777 |
+
643,
|
| 778 |
+
784,
|
| 779 |
+
662
|
| 780 |
+
],
|
| 781 |
+
"page_idx": 6
|
| 782 |
+
},
|
| 783 |
+
{
|
| 784 |
+
"type": "text",
|
| 785 |
+
"text": "To illustrate the validity of the proposed normal-flow constraint over the naive way of simply using normal flow to replace full optical flow, we use a 2D registration task as toy example. As shown in Fig. 2, the task consists of registering the solid-line triangle to the dashed-line one using as input either optical flow or normal flow. We observe that using the proposed normal-flow constraint as a geometric measurement leads to a registration result that is almost as accurate as the one given by the full optical flow. However, due to the partial observability of normal flow, simply using it as a replacement for full optical flow leads to a biased registration result, as shown in Fig. 2c. As we see in all following experiments, similar conclusions can be made in general problems of geometric model fitting.",
|
| 786 |
+
"bbox": [
|
| 787 |
+
214,
|
| 788 |
+
672,
|
| 789 |
+
787,
|
| 790 |
+
840
|
| 791 |
+
],
|
| 792 |
+
"page_idx": 6
|
| 793 |
+
},
|
| 794 |
+
{
|
| 795 |
+
"type": "header",
|
| 796 |
+
"text": "Motion and Structure from Event-based Normal Flow",
|
| 797 |
+
"bbox": [
|
| 798 |
+
372,
|
| 799 |
+
114,
|
| 800 |
+
732,
|
| 801 |
+
127
|
| 802 |
+
],
|
| 803 |
+
"page_idx": 6
|
| 804 |
+
},
|
| 805 |
+
{
|
| 806 |
+
"type": "page_number",
|
| 807 |
+
"text": "7",
|
| 808 |
+
"bbox": [
|
| 809 |
+
774,
|
| 810 |
+
116,
|
| 811 |
+
784,
|
| 812 |
+
126
|
| 813 |
+
],
|
| 814 |
+
"page_idx": 6
|
| 815 |
+
},
|
| 816 |
+
{
|
| 817 |
+
"type": "text",
|
| 818 |
+
"text": "4.2 Models of Motion and Structure",
|
| 819 |
+
"text_level": 1,
|
| 820 |
+
"bbox": [
|
| 821 |
+
215,
|
| 822 |
+
146,
|
| 823 |
+
529,
|
| 824 |
+
161
|
| 825 |
+
],
|
| 826 |
+
"page_idx": 7
|
| 827 |
+
},
|
| 828 |
+
{
|
| 829 |
+
"type": "text",
|
| 830 |
+
"text": "Based on the derivation in Sec. 4.1, we establish linear models to a family of motion and structure estimation problems using the event-based normal flow as input. In particular, we first discuss how to recover optical flow and depth by introducing the normal-flow constraint into the differential epipolar geometry. Then we show that the combination of the normal-flow constraint with the motion flow equation (1) results in solutions to the problem of instantaneous angular velocity estimation, and 6-DoF motion estimation given depth as prior, respectively. Finally, we show how to compute the differential homography.",
|
| 831 |
+
"bbox": [
|
| 832 |
+
212,
|
| 833 |
+
170,
|
| 834 |
+
787,
|
| 835 |
+
292
|
| 836 |
+
],
|
| 837 |
+
"page_idx": 7
|
| 838 |
+
},
|
| 839 |
+
{
|
| 840 |
+
"type": "text",
|
| 841 |
+
"text": "Optical Flow. Given as input the accurate normal flow, and camera velocities, the full optical flow $\\mathbf{u}(\\mathbf{x})$ at pixel location $\\mathbf{x}$ can be calculated by solving the following linear system obtained from the combination of the normal-flow constraint (3) and the differential epipolar geometry (5):",
|
| 842 |
+
"bbox": [
|
| 843 |
+
214,
|
| 844 |
+
313,
|
| 845 |
+
787,
|
| 846 |
+
376
|
| 847 |
+
],
|
| 848 |
+
"page_idx": 7
|
| 849 |
+
},
|
| 850 |
+
{
|
| 851 |
+
"type": "equation",
|
| 852 |
+
"text": "\n$$\n\\binom {\\mathbf {n} (\\mathbf {x}) ^ {\\top}} {- \\hat {\\mathbf {x}} ^ {\\top} [ \\boldsymbol {\\nu} ] _ {\\times}} \\mathbf {u} (\\mathbf {x}) = \\binom {\\| \\mathbf {n} (\\mathbf {x}) \\| ^ {2}} {\\hat {\\mathbf {x}} ^ {\\top} \\mathbf {s} \\hat {\\mathbf {x}}}. \\tag {8}\n$$\n",
|
| 853 |
+
"text_format": "latex",
|
| 854 |
+
"bbox": [
|
| 855 |
+
385,
|
| 856 |
+
385,
|
| 857 |
+
785,
|
| 858 |
+
421
|
| 859 |
+
],
|
| 860 |
+
"page_idx": 7
|
| 861 |
+
},
|
| 862 |
+
{
|
| 863 |
+
"type": "text",
|
| 864 |
+
"text": "Depth Estimation. Similarly, if the normal flow and camera velocities are known, the scene depth $Z(\\mathbf{x})$ can be determined by combining the normal-flow constraint (3) with the motion field eq. (1), yielding the closed-form solution:",
|
| 865 |
+
"bbox": [
|
| 866 |
+
214,
|
| 867 |
+
440,
|
| 868 |
+
787,
|
| 869 |
+
487
|
| 870 |
+
],
|
| 871 |
+
"page_idx": 7
|
| 872 |
+
},
|
| 873 |
+
{
|
| 874 |
+
"type": "equation",
|
| 875 |
+
"text": "\n$$\nZ (\\mathbf {x}) = \\frac {\\mathbf {n} (\\mathbf {x}) ^ {\\top} \\mathbf {A} (\\mathbf {x}) \\boldsymbol {\\nu}}{\\| \\mathbf {n} (\\mathbf {x}) \\| ^ {2} - \\mathbf {n} (\\mathbf {x}) ^ {\\top} \\mathbf {B} (\\mathbf {x}) \\boldsymbol {\\omega}}. \\tag {9}\n$$\n",
|
| 876 |
+
"text_format": "latex",
|
| 877 |
+
"bbox": [
|
| 878 |
+
383,
|
| 879 |
+
497,
|
| 880 |
+
785,
|
| 881 |
+
532
|
| 882 |
+
],
|
| 883 |
+
"page_idx": 7
|
| 884 |
+
},
|
| 885 |
+
{
|
| 886 |
+
"type": "text",
|
| 887 |
+
"text": "Angular Velocity Estimation. Assuming that an event camera undergoes a pure rotational motion, the optical flow $\\mathbf{u}$ in (1) can be simplified as",
|
| 888 |
+
"bbox": [
|
| 889 |
+
214,
|
| 890 |
+
549,
|
| 891 |
+
787,
|
| 892 |
+
580
|
| 893 |
+
],
|
| 894 |
+
"page_idx": 7
|
| 895 |
+
},
|
| 896 |
+
{
|
| 897 |
+
"type": "equation",
|
| 898 |
+
"text": "\n$$\n\\mathbf {u} (\\mathbf {x}) = \\mathbf {B} (\\mathbf {x}) \\boldsymbol {\\omega}. \\tag {10}\n$$\n",
|
| 899 |
+
"text_format": "latex",
|
| 900 |
+
"bbox": [
|
| 901 |
+
447,
|
| 902 |
+
592,
|
| 903 |
+
785,
|
| 904 |
+
608
|
| 905 |
+
],
|
| 906 |
+
"page_idx": 7
|
| 907 |
+
},
|
| 908 |
+
{
|
| 909 |
+
"type": "text",
|
| 910 |
+
"text": "By substituting (10) into (3), we have the linear equation in $\\omega$ :",
|
| 911 |
+
"bbox": [
|
| 912 |
+
215,
|
| 913 |
+
618,
|
| 914 |
+
674,
|
| 915 |
+
633
|
| 916 |
+
],
|
| 917 |
+
"page_idx": 7
|
| 918 |
+
},
|
| 919 |
+
{
|
| 920 |
+
"type": "equation",
|
| 921 |
+
"text": "\n$$\n\\mathbf {n} (\\mathbf {x}) ^ {\\top} \\mathbf {B} (\\mathbf {x}) \\omega = \\| \\mathbf {n} (\\mathbf {x}) \\| ^ {2}. \\tag {11}\n$$\n",
|
| 922 |
+
"text_format": "latex",
|
| 923 |
+
"bbox": [
|
| 924 |
+
413,
|
| 925 |
+
643,
|
| 926 |
+
785,
|
| 927 |
+
662
|
| 928 |
+
],
|
| 929 |
+
"page_idx": 7
|
| 930 |
+
},
|
| 931 |
+
{
|
| 932 |
+
"type": "text",
|
| 933 |
+
"text": "Since each normal flow provides only one constraint, the minimal problem requires three normal flow vectors as input.",
|
| 934 |
+
"bbox": [
|
| 935 |
+
214,
|
| 936 |
+
672,
|
| 937 |
+
785,
|
| 938 |
+
704
|
| 939 |
+
],
|
| 940 |
+
"page_idx": 7
|
| 941 |
+
},
|
| 942 |
+
{
|
| 943 |
+
"type": "text",
|
| 944 |
+
"text": "6-DoF Motion Estimation. Assume the event camera undergoes a general 6-DoF motion with known depth of the observed scene, we can solve the first-order kinematics, namely instantaneous linear and angular velocities, by solving the following linear system obtained from the normal-flow constraint (3) and the motion-flow equation (1):",
|
| 945 |
+
"bbox": [
|
| 946 |
+
214,
|
| 947 |
+
724,
|
| 948 |
+
787,
|
| 949 |
+
801
|
| 950 |
+
],
|
| 951 |
+
"page_idx": 7
|
| 952 |
+
},
|
| 953 |
+
{
|
| 954 |
+
"type": "equation",
|
| 955 |
+
"text": "\n$$\n\\mathbf {n} (\\mathbf {x}) ^ {\\top} \\mathbf {D} (\\mathbf {x}) \\left[ \\begin{array}{l} \\boldsymbol {\\nu} \\\\ \\boldsymbol {\\omega} \\end{array} \\right] = \\| \\mathbf {n} (\\mathbf {x}) \\| ^ {2}, \\tag {12}\n$$\n",
|
| 956 |
+
"text_format": "latex",
|
| 957 |
+
"bbox": [
|
| 958 |
+
401,
|
| 959 |
+
811,
|
| 960 |
+
785,
|
| 961 |
+
844
|
| 962 |
+
],
|
| 963 |
+
"page_idx": 7
|
| 964 |
+
},
|
| 965 |
+
{
|
| 966 |
+
"type": "page_number",
|
| 967 |
+
"text": "8",
|
| 968 |
+
"bbox": [
|
| 969 |
+
217,
|
| 970 |
+
114,
|
| 971 |
+
228,
|
| 972 |
+
126
|
| 973 |
+
],
|
| 974 |
+
"page_idx": 7
|
| 975 |
+
},
|
| 976 |
+
{
|
| 977 |
+
"type": "header",
|
| 978 |
+
"text": "Z. Ren et al.",
|
| 979 |
+
"bbox": [
|
| 980 |
+
271,
|
| 981 |
+
114,
|
| 982 |
+
357,
|
| 983 |
+
127
|
| 984 |
+
],
|
| 985 |
+
"page_idx": 7
|
| 986 |
+
},
|
| 987 |
+
{
|
| 988 |
+
"type": "text",
|
| 989 |
+
"text": "where $\\mathbf{D}(\\mathbf{x}) = \\left[\\frac{1}{Z(\\mathbf{x})}\\mathbf{A}(\\mathbf{x})\\mathbf{B}(\\mathbf{x})\\right]$ is the $2\\times 6$ feature sensitivity matrix [4, 7]. The minimal problem in this context requires six normal flow measurements to determine the 6-DoF motion parameters.",
|
| 990 |
+
"bbox": [
|
| 991 |
+
212,
|
| 992 |
+
143,
|
| 993 |
+
782,
|
| 994 |
+
196
|
| 995 |
+
],
|
| 996 |
+
"page_idx": 8
|
| 997 |
+
},
|
| 998 |
+
{
|
| 999 |
+
"type": "text",
|
| 1000 |
+
"text": "Motion and Structure Estimation in Planar Scenes. For a moving camera that observes a planar scene, the motion and structure parameters are compatibly encoded in a homography matrix. In our case, we attempt to recover the differential homography $\\mathbf{h}_d$ (the vector form of $\\mathbf{H}_d$ in (4)) because the instantaneous linear and angular velocities are considered. By expanding (4) and rearranging equations, we have",
|
| 1001 |
+
"bbox": [
|
| 1002 |
+
212,
|
| 1003 |
+
215,
|
| 1004 |
+
782,
|
| 1005 |
+
306
|
| 1006 |
+
],
|
| 1007 |
+
"page_idx": 8
|
| 1008 |
+
},
|
| 1009 |
+
{
|
| 1010 |
+
"type": "equation",
|
| 1011 |
+
"text": "\n$$\n\\mathbf {u} (\\mathbf {x}) = \\mathbf {C} (\\mathbf {x}) \\mathbf {h} _ {d}, \\tag {13}\n$$\n",
|
| 1012 |
+
"text_format": "latex",
|
| 1013 |
+
"bbox": [
|
| 1014 |
+
446,
|
| 1015 |
+
316,
|
| 1016 |
+
784,
|
| 1017 |
+
330
|
| 1018 |
+
],
|
| 1019 |
+
"page_idx": 8
|
| 1020 |
+
},
|
| 1021 |
+
{
|
| 1022 |
+
"type": "text",
|
| 1023 |
+
"text": "where $\\mathbf{C}(\\mathbf{x})$ is a $2 \\times 9$ matrix that solely depends on the pixel location $\\mathbf{x}$ . Subsequently, substituting (3) into (13), gives:",
|
| 1024 |
+
"bbox": [
|
| 1025 |
+
212,
|
| 1026 |
+
338,
|
| 1027 |
+
782,
|
| 1028 |
+
369
|
| 1029 |
+
],
|
| 1030 |
+
"page_idx": 8
|
| 1031 |
+
},
|
| 1032 |
+
{
|
| 1033 |
+
"type": "equation",
|
| 1034 |
+
"text": "\n$$\n\\mathbf {n} (\\mathbf {x}) ^ {\\top} \\mathbf {C} (\\mathbf {x}) \\mathbf {h} _ {d} = \\| \\mathbf {n} (\\mathbf {x}) \\| ^ {2}. \\tag {14}\n$$\n",
|
| 1035 |
+
"text_format": "latex",
|
| 1036 |
+
"bbox": [
|
| 1037 |
+
410,
|
| 1038 |
+
375,
|
| 1039 |
+
784,
|
| 1040 |
+
393
|
| 1041 |
+
],
|
| 1042 |
+
"page_idx": 8
|
| 1043 |
+
},
|
| 1044 |
+
{
|
| 1045 |
+
"type": "text",
|
| 1046 |
+
"text": "As discussed in [44], retrieving the motion and structure parameters from the decomposition of $\\mathbf{h}_d$ is not trivial and beyond the scope of this paper. Therefore, we only focus on the numerical solution of $\\mathbf{h}_d$ .",
|
| 1047 |
+
"bbox": [
|
| 1048 |
+
212,
|
| 1049 |
+
400,
|
| 1050 |
+
782,
|
| 1051 |
+
445
|
| 1052 |
+
],
|
| 1053 |
+
"page_idx": 8
|
| 1054 |
+
},
|
| 1055 |
+
{
|
| 1056 |
+
"type": "text",
|
| 1057 |
+
"text": "4.3 Our Solvers",
|
| 1058 |
+
"text_level": 1,
|
| 1059 |
+
"bbox": [
|
| 1060 |
+
214,
|
| 1061 |
+
465,
|
| 1062 |
+
359,
|
| 1063 |
+
478
|
| 1064 |
+
],
|
| 1065 |
+
"page_idx": 8
|
| 1066 |
+
},
|
| 1067 |
+
{
|
| 1068 |
+
"type": "text",
|
| 1069 |
+
"text": "According to the motion-and-structure models derived in Sec. 4.2, we have established certain constraints between the model parameters and the normal flow observations. This allows us to simply develop linear solvers for each specific problem. Besides, we further propose a continuous-time nonlinear solver that can effectively handle the presence of sudden variations in motion and does not rely on the constant-motion assumption. To convey our solvers concisely, we generalize the motion-and-structure models into a unified form, as",
|
| 1070 |
+
"bbox": [
|
| 1071 |
+
212,
|
| 1072 |
+
486,
|
| 1073 |
+
784,
|
| 1074 |
+
592
|
| 1075 |
+
],
|
| 1076 |
+
"page_idx": 8
|
| 1077 |
+
},
|
| 1078 |
+
{
|
| 1079 |
+
"type": "equation",
|
| 1080 |
+
"text": "\n$$\n\\mathbf {n} (\\mathbf {x}) ^ {\\top} \\mathbf {O} (\\mathbf {x}) \\boldsymbol {\\theta} = \\| \\mathbf {n} (\\mathbf {x}) \\| ^ {2}, \\tag {15}\n$$\n",
|
| 1081 |
+
"text_format": "latex",
|
| 1082 |
+
"bbox": [
|
| 1083 |
+
415,
|
| 1084 |
+
599,
|
| 1085 |
+
784,
|
| 1086 |
+
616
|
| 1087 |
+
],
|
| 1088 |
+
"page_idx": 8
|
| 1089 |
+
},
|
| 1090 |
+
{
|
| 1091 |
+
"type": "text",
|
| 1092 |
+
"text": "where $\\mathbf{O}(\\mathbf{x})$ is the system matrix and $\\theta$ are the motion/structure parameters.",
|
| 1093 |
+
"bbox": [
|
| 1094 |
+
212,
|
| 1095 |
+
625,
|
| 1096 |
+
779,
|
| 1097 |
+
640
|
| 1098 |
+
],
|
| 1099 |
+
"page_idx": 8
|
| 1100 |
+
},
|
| 1101 |
+
{
|
| 1102 |
+
"type": "text",
|
| 1103 |
+
"text": "Linear Solver. Given as input a set (e.g., K) of sparse normal-flow estimates obtained at the same time, our linear solver works out the following (overdetermined) linear system of equations:",
|
| 1104 |
+
"bbox": [
|
| 1105 |
+
212,
|
| 1106 |
+
657,
|
| 1107 |
+
782,
|
| 1108 |
+
704
|
| 1109 |
+
],
|
| 1110 |
+
"page_idx": 8
|
| 1111 |
+
},
|
| 1112 |
+
{
|
| 1113 |
+
"type": "equation",
|
| 1114 |
+
"text": "\n$$\n\\left( \\begin{array}{c} \\mathbf {n} \\left(\\mathbf {x} _ {1}\\right) ^ {\\top} \\mathbf {O} \\left(\\mathbf {x} _ {1}\\right) \\\\ \\vdots \\\\ \\mathbf {n} \\left(\\mathbf {x} _ {\\mathrm {K}}\\right) ^ {\\top} \\mathbf {O} \\left(\\mathbf {x} _ {\\mathrm {K}}\\right) \\end{array} \\right) \\boldsymbol {\\theta} = \\left( \\begin{array}{c} \\| \\mathbf {n} \\left(\\mathbf {x} _ {1}\\right) \\| ^ {2} \\\\ \\vdots \\\\ \\| \\mathbf {n} \\left(\\mathbf {x} _ {\\mathrm {K}}\\right) \\| ^ {2} \\end{array} \\right). \\tag {16}\n$$\n",
|
| 1115 |
+
"text_format": "latex",
|
| 1116 |
+
"bbox": [
|
| 1117 |
+
370,
|
| 1118 |
+
710,
|
| 1119 |
+
784,
|
| 1120 |
+
768
|
| 1121 |
+
],
|
| 1122 |
+
"page_idx": 8
|
| 1123 |
+
},
|
| 1124 |
+
{
|
| 1125 |
+
"type": "text",
|
| 1126 |
+
"text": "For robust estimation, RANSAC [12] is adopted to deal with noise and outliers.",
|
| 1127 |
+
"bbox": [
|
| 1128 |
+
212,
|
| 1129 |
+
773,
|
| 1130 |
+
782,
|
| 1131 |
+
790
|
| 1132 |
+
],
|
| 1133 |
+
"page_idx": 8
|
| 1134 |
+
},
|
| 1135 |
+
{
|
| 1136 |
+
"type": "header",
|
| 1137 |
+
"text": "Motion and Structure from Event-based Normal Flow",
|
| 1138 |
+
"bbox": [
|
| 1139 |
+
370,
|
| 1140 |
+
114,
|
| 1141 |
+
732,
|
| 1142 |
+
127
|
| 1143 |
+
],
|
| 1144 |
+
"page_idx": 8
|
| 1145 |
+
},
|
| 1146 |
+
{
|
| 1147 |
+
"type": "page_number",
|
| 1148 |
+
"text": "9",
|
| 1149 |
+
"bbox": [
|
| 1150 |
+
774,
|
| 1151 |
+
114,
|
| 1152 |
+
784,
|
| 1153 |
+
126
|
| 1154 |
+
],
|
| 1155 |
+
"page_idx": 8
|
| 1156 |
+
},
|
| 1157 |
+
{
|
| 1158 |
+
"type": "page_footnote",
|
| 1159 |
+
"text": "$^{5}$ $\\mathbf{O}(\\mathbf{x})$ degenerates to 1 in the problem of optical flow estimation.",
|
| 1160 |
+
"bbox": [
|
| 1161 |
+
217,
|
| 1162 |
+
796,
|
| 1163 |
+
658,
|
| 1164 |
+
811
|
| 1165 |
+
],
|
| 1166 |
+
"page_idx": 8
|
| 1167 |
+
},
|
| 1168 |
+
{
|
| 1169 |
+
"type": "page_footnote",
|
| 1170 |
+
"text": "An approximation can be made based on the assumption of constant camera motion. In this way, the normal flow estimates can be collected within a short time interval.",
|
| 1171 |
+
"bbox": [
|
| 1172 |
+
220,
|
| 1173 |
+
811,
|
| 1174 |
+
782,
|
| 1175 |
+
839
|
| 1176 |
+
],
|
| 1177 |
+
"page_idx": 8
|
| 1178 |
+
},
|
| 1179 |
+
{
|
| 1180 |
+
"type": "text",
|
| 1181 |
+
"text": "Continuous-Time Nonlinear Solver. To handle the case of a sudden variation in speed that may violate the constant-motion assumption, we further propose a continuous-time nonlinear solver. In particular, we employ cubic B-splines to represent the continuous-time model parameters as",
|
| 1182 |
+
"bbox": [
|
| 1183 |
+
212,
|
| 1184 |
+
146,
|
| 1185 |
+
782,
|
| 1186 |
+
205
|
| 1187 |
+
],
|
| 1188 |
+
"page_idx": 9
|
| 1189 |
+
},
|
| 1190 |
+
{
|
| 1191 |
+
"type": "equation",
|
| 1192 |
+
"text": "\n$$\n\\boldsymbol {\\theta} (u (t)) = \\sum_ {i = 0} ^ {3} B _ {i, 3} (u (t)) \\mathbf {c} _ {i}, \\tag {17}\n$$\n",
|
| 1193 |
+
"text_format": "latex",
|
| 1194 |
+
"bbox": [
|
| 1195 |
+
406,
|
| 1196 |
+
219,
|
| 1197 |
+
785,
|
| 1198 |
+
258
|
| 1199 |
+
],
|
| 1200 |
+
"page_idx": 9
|
| 1201 |
+
},
|
| 1202 |
+
{
|
| 1203 |
+
"type": "text",
|
| 1204 |
+
"text": "where $B_{i,k}$ denotes the basis function, $i$ the index of control points, $k (= 3)$ the order of the spline and $\\mathbf{c}_i \\in \\mathbb{R}^N$ the corresponding control point defined in the $N$ -dimensional space of model parameters. Examples of cubic B-splines applied to event camera motion estimation are found in [18, 28]. The function $u(\\cdot)$ is a normalization operator, which transfers time $t$ to the spline's parameter domain by means of basis translation [8]. In this way, the model parameters at any time instant $t$ can be obtained via an interpolation using the known control points.",
|
| 1205 |
+
"bbox": [
|
| 1206 |
+
212,
|
| 1207 |
+
268,
|
| 1208 |
+
784,
|
| 1209 |
+
375
|
| 1210 |
+
],
|
| 1211 |
+
"page_idx": 9
|
| 1212 |
+
},
|
| 1213 |
+
{
|
| 1214 |
+
"type": "text",
|
| 1215 |
+
"text": "The formulation of the continuous-time model fitting problem goes as follows. The unknown variable consists of all the involved control points, $\\mathcal{X} \\doteq [\\mathbf{c}_0, \\mathbf{c}_1, \\dots, \\mathbf{c}_n]$ , where $n$ denotes the total number of control points. By introducing the continuous-time representation of model parameters to (15), we finally create the following objective function:",
|
| 1216 |
+
"bbox": [
|
| 1217 |
+
212,
|
| 1218 |
+
376,
|
| 1219 |
+
784,
|
| 1220 |
+
450
|
| 1221 |
+
],
|
| 1222 |
+
"page_idx": 9
|
| 1223 |
+
},
|
| 1224 |
+
{
|
| 1225 |
+
"type": "equation",
|
| 1226 |
+
"text": "\n$$\n\\boldsymbol {\\mathcal {X}} ^ {*} = \\arg \\min _ {\\boldsymbol {\\mathcal {X}}} \\sum_ {e _ {k} \\in \\mathcal {E}} \\left\\| \\mathbf {n} \\left(\\mathbf {x} _ {k}\\right) ^ {\\top} \\mathbf {O} \\left(\\mathbf {x} _ {k}\\right) \\boldsymbol {\\theta} \\left(u \\left(t _ {k}\\right)\\right) - \\| \\mathbf {n} \\left(\\mathbf {x} _ {k}\\right) \\| ^ {2} \\right\\| ^ {2}, \\tag {18}\n$$\n",
|
| 1227 |
+
"text_format": "latex",
|
| 1228 |
+
"bbox": [
|
| 1229 |
+
297,
|
| 1230 |
+
467,
|
| 1231 |
+
785,
|
| 1232 |
+
500
|
| 1233 |
+
],
|
| 1234 |
+
"page_idx": 9
|
| 1235 |
+
},
|
| 1236 |
+
{
|
| 1237 |
+
"type": "text",
|
| 1238 |
+
"text": "where $\\mathcal{E}$ denotes the event set in which the norm flow observation is available for every involved event $e_k$ . Note that this continuous-time formulation aligns better with the asynchronous nature of event-based normal flow, enabling a state estimation over a larger temporal window that may lead to a better signal-to-noise ratio. The linear solver can be used to initialize the continuous-time nonlinear solver. An illustrative explanation of our solvers is given in Fig. 1.",
|
| 1239 |
+
"bbox": [
|
| 1240 |
+
212,
|
| 1241 |
+
508,
|
| 1242 |
+
782,
|
| 1243 |
+
599
|
| 1244 |
+
],
|
| 1245 |
+
"page_idx": 9
|
| 1246 |
+
},
|
| 1247 |
+
{
|
| 1248 |
+
"type": "text",
|
| 1249 |
+
"text": "5 Experiments",
|
| 1250 |
+
"text_level": 1,
|
| 1251 |
+
"bbox": [
|
| 1252 |
+
215,
|
| 1253 |
+
623,
|
| 1254 |
+
375,
|
| 1255 |
+
641
|
| 1256 |
+
],
|
| 1257 |
+
"page_idx": 9
|
| 1258 |
+
},
|
| 1259 |
+
{
|
| 1260 |
+
"type": "text",
|
| 1261 |
+
"text": "To justify the effectiveness of the proposed algorithms, experiments are conducted on both simulation and real data. First, we introduce the datasets used and the applied metrics for evaluation (Sec. 5.1). Then we report the results of our linear solver and continuous-time nonlinear solver (Sec. 5.2 and Sec. 5.3). Finally, a discussion on the computational complexity and the limitation of our method is provided in Sec. 5.4.",
|
| 1262 |
+
"bbox": [
|
| 1263 |
+
212,
|
| 1264 |
+
655,
|
| 1265 |
+
784,
|
| 1266 |
+
746
|
| 1267 |
+
],
|
| 1268 |
+
"page_idx": 9
|
| 1269 |
+
},
|
| 1270 |
+
{
|
| 1271 |
+
"type": "text",
|
| 1272 |
+
"text": "5.1 Experimental Setup",
|
| 1273 |
+
"text_level": 1,
|
| 1274 |
+
"bbox": [
|
| 1275 |
+
215,
|
| 1276 |
+
768,
|
| 1277 |
+
426,
|
| 1278 |
+
784
|
| 1279 |
+
],
|
| 1280 |
+
"page_idx": 9
|
| 1281 |
+
},
|
| 1282 |
+
{
|
| 1283 |
+
"type": "text",
|
| 1284 |
+
"text": "Our evaluation utilizes four datasets, including the indoor event-camera dataset (ECD) captured with a DAVIS240C camera and motion-capture system for groundtruth, and a high-resolution dataset with two rotational motion sequences",
|
| 1285 |
+
"bbox": [
|
| 1286 |
+
212,
|
| 1287 |
+
794,
|
| 1288 |
+
782,
|
| 1289 |
+
840
|
| 1290 |
+
],
|
| 1291 |
+
"page_idx": 9
|
| 1292 |
+
},
|
| 1293 |
+
{
|
| 1294 |
+
"type": "page_number",
|
| 1295 |
+
"text": "10",
|
| 1296 |
+
"bbox": [
|
| 1297 |
+
217,
|
| 1298 |
+
114,
|
| 1299 |
+
235,
|
| 1300 |
+
126
|
| 1301 |
+
],
|
| 1302 |
+
"page_idx": 9
|
| 1303 |
+
},
|
| 1304 |
+
{
|
| 1305 |
+
"type": "header",
|
| 1306 |
+
"text": "Z. Ren et al.",
|
| 1307 |
+
"bbox": [
|
| 1308 |
+
271,
|
| 1309 |
+
114,
|
| 1310 |
+
357,
|
| 1311 |
+
127
|
| 1312 |
+
],
|
| 1313 |
+
"page_idx": 9
|
| 1314 |
+
},
|
| 1315 |
+
{
|
| 1316 |
+
"type": "image",
|
| 1317 |
+
"img_path": "images/6d5184372b1a0e4636446f974281d9e865262d9d32e175502a23e5068bfac67b.jpg",
|
| 1318 |
+
"image_caption": [
|
| 1319 |
+
"(a) Estimated optical flow over a map of event accumulation."
|
| 1320 |
+
],
|
| 1321 |
+
"image_footnote": [],
|
| 1322 |
+
"bbox": [
|
| 1323 |
+
217,
|
| 1324 |
+
145,
|
| 1325 |
+
415,
|
| 1326 |
+
248
|
| 1327 |
+
],
|
| 1328 |
+
"page_idx": 10
|
| 1329 |
+
},
|
| 1330 |
+
{
|
| 1331 |
+
"type": "image",
|
| 1332 |
+
"img_path": "images/b484b8710ef572824360f0a0292a6cea7e4fcfc3d925bc5dcd9f1dbedecd369b.jpg",
|
| 1333 |
+
"image_caption": [
|
| 1334 |
+
"(b) Reconstruction result displayed over groundtruth structures (a v-shape roof) and viewed from different perspectives."
|
| 1335 |
+
],
|
| 1336 |
+
"image_footnote": [],
|
| 1337 |
+
"bbox": [
|
| 1338 |
+
423,
|
| 1339 |
+
143,
|
| 1340 |
+
602,
|
| 1341 |
+
246
|
| 1342 |
+
],
|
| 1343 |
+
"page_idx": 10
|
| 1344 |
+
},
|
| 1345 |
+
{
|
| 1346 |
+
"type": "image",
|
| 1347 |
+
"img_path": "images/dde771b4898659557cdb422227caec1f8c8fce581a2f70c55816fd436d16d6d3.jpg",
|
| 1348 |
+
"image_caption": [
|
| 1349 |
+
"Fig. 3: Qualitative result of optical flow and depth on sequence three_wall Translation of our synthetic data. Note that our method returns sparse results."
|
| 1350 |
+
],
|
| 1351 |
+
"image_footnote": [],
|
| 1352 |
+
"bbox": [
|
| 1353 |
+
604,
|
| 1354 |
+
143,
|
| 1355 |
+
784,
|
| 1356 |
+
247
|
| 1357 |
+
],
|
| 1358 |
+
"page_idx": 10
|
| 1359 |
+
},
|
| 1360 |
+
{
|
| 1361 |
+
"type": "text",
|
| 1362 |
+
"text": "using an iniVation DAVIS346 camera and Xsens IMU. The VECTor dataset, captured with a Prophesee Gen 3 camera and an accompanying RGB-D depth camera, is used for 6-DoF motion tracking. Additionally, a synthetic dataset generated in Blender is employed for motion-and-structure pattern analysis. Evaluation metrics for kinematics estimation include average error (AE) and root mean squared error (RMSE), while differential homography estimation is assessed using the Frobenius norm. Best results are in bold, and second-best are underlined. The supplementary material provides detailed descriptions of datasets and metrics, as well as implementation details of our methods.",
|
| 1363 |
+
"bbox": [
|
| 1364 |
+
212,
|
| 1365 |
+
335,
|
| 1366 |
+
787,
|
| 1367 |
+
472
|
| 1368 |
+
],
|
| 1369 |
+
"page_idx": 10
|
| 1370 |
+
},
|
| 1371 |
+
{
|
| 1372 |
+
"type": "text",
|
| 1373 |
+
"text": "5.2 Evaluation of the Linear Solver",
|
| 1374 |
+
"text_level": 1,
|
| 1375 |
+
"bbox": [
|
| 1376 |
+
214,
|
| 1377 |
+
491,
|
| 1378 |
+
519,
|
| 1379 |
+
505
|
| 1380 |
+
],
|
| 1381 |
+
"page_idx": 10
|
| 1382 |
+
},
|
| 1383 |
+
{
|
| 1384 |
+
"type": "text",
|
| 1385 |
+
"text": "In the following evaluation, we qualitatively and quantitatively assess the performance of our linear method (Ours) in each specific task, and compare against different state-of-the-art (SOTA) methods (CMax [14] and Pro-STR [19]) and their customized version initialized using our linear solver.",
|
| 1386 |
+
"bbox": [
|
| 1387 |
+
212,
|
| 1388 |
+
512,
|
| 1389 |
+
785,
|
| 1390 |
+
571
|
| 1391 |
+
],
|
| 1392 |
+
"page_idx": 10
|
| 1393 |
+
},
|
| 1394 |
+
{
|
| 1395 |
+
"type": "text",
|
| 1396 |
+
"text": "Optical Flow and Depth Estimation. Due to a lack of benchmark for extensive evaluation, we qualitatively evaluate our method on the tasks of optical flow and depth estimation. As illustrated in Fig. 3a, our linear solver demonstrates the capability to accurately reconstruct full flow using normal flow and known motion parameters. An example of the depth estimation result is given in Fig. 3b, showing that the recovered structure aligns well with the groundtruth.",
|
| 1397 |
+
"bbox": [
|
| 1398 |
+
212,
|
| 1399 |
+
590,
|
| 1400 |
+
787,
|
| 1401 |
+
681
|
| 1402 |
+
],
|
| 1403 |
+
"page_idx": 10
|
| 1404 |
+
},
|
| 1405 |
+
{
|
| 1406 |
+
"type": "text",
|
| 1407 |
+
"text": "Angular Velocity Estimation. As shown in Tab. 1, we see that our linear solver outperforms CMax and Pro-STR. Besides, we find the combination of our linear solver and CMax leads to the best result, indicating that the proposed linear solver is a complementary method to existing nonlinear methods.",
|
| 1408 |
+
"bbox": [
|
| 1409 |
+
212,
|
| 1410 |
+
700,
|
| 1411 |
+
787,
|
| 1412 |
+
761
|
| 1413 |
+
],
|
| 1414 |
+
"page_idx": 10
|
| 1415 |
+
},
|
| 1416 |
+
{
|
| 1417 |
+
"type": "text",
|
| 1418 |
+
"text": "Differential Homography Estimation. The result is consistent with that in the above task. Table 2 reveals that our linear method outperforms CMax and Pro-STR in terms of numerical accuracy. The combination of our linear solver and state-of-the-art nonlinear solver leads to the best result.",
|
| 1419 |
+
"bbox": [
|
| 1420 |
+
212,
|
| 1421 |
+
779,
|
| 1422 |
+
787,
|
| 1423 |
+
839
|
| 1424 |
+
],
|
| 1425 |
+
"page_idx": 10
|
| 1426 |
+
},
|
| 1427 |
+
{
|
| 1428 |
+
"type": "header",
|
| 1429 |
+
"text": "Motion and Structure from Event-based Normal Flow",
|
| 1430 |
+
"bbox": [
|
| 1431 |
+
372,
|
| 1432 |
+
114,
|
| 1433 |
+
730,
|
| 1434 |
+
127
|
| 1435 |
+
],
|
| 1436 |
+
"page_idx": 10
|
| 1437 |
+
},
|
| 1438 |
+
{
|
| 1439 |
+
"type": "page_number",
|
| 1440 |
+
"text": "11",
|
| 1441 |
+
"bbox": [
|
| 1442 |
+
767,
|
| 1443 |
+
114,
|
| 1444 |
+
782,
|
| 1445 |
+
126
|
| 1446 |
+
],
|
| 1447 |
+
"page_idx": 10
|
| 1448 |
+
},
|
| 1449 |
+
{
|
| 1450 |
+
"type": "table",
|
| 1451 |
+
"img_path": "images/8b7702183272a27efcd7157645f94b84df46160219da6cde800b334df3814271.jpg",
|
| 1452 |
+
"table_caption": [
|
| 1453 |
+
"Table 1: Evaluation of our linear solver on the task of angular velocity estimation."
|
| 1454 |
+
],
|
| 1455 |
+
"table_footnote": [],
|
| 1456 |
+
"table_body": "<table><tr><td rowspan=\"2\">Method</td><td colspan=\"2\">shapesRotation [29]</td><td colspan=\"2\">groundRotation (our data)</td><td colspan=\"2\">patternsRotation (our data)</td></tr><tr><td>eω(deg/s)</td><td>RMSEω(deg/s)</td><td>eω(deg/s)</td><td>RMSEω(deg/s)</td><td>eω(deg/s)</td><td>RMSEω(deg/s)</td></tr><tr><td>CMax [14]</td><td>15.60</td><td>26.38</td><td>9.48</td><td>15.58</td><td>5.95</td><td>9.07</td></tr><tr><td>Pro-STR [19]</td><td>90.61</td><td>145.81</td><td>31.07</td><td>44.74</td><td>10.47</td><td>11.88</td></tr><tr><td>Ours</td><td>15.36</td><td>23.98</td><td>3.30</td><td>8.01</td><td>4.70</td><td>6.08</td></tr><tr><td>Ours+CMax [14]</td><td>10.12</td><td>14.29</td><td>2.85</td><td>3.74</td><td>0.35</td><td>0.73</td></tr><tr><td>Ours+Pro-STR [19]</td><td>19.23</td><td>40.44</td><td>9.13</td><td>17.07</td><td>4.31</td><td>5.23</td></tr><tr><td rowspan=\"2\">Method</td><td colspan=\"2\">dynamicRotation [29]</td><td colspan=\"2\">boxesRotation (our data)</td><td colspan=\"2\">cubesRotation (our data)</td></tr><tr><td>eω(deg/s)</td><td>RMSEω(deg/s)</td><td>eω(deg/s)</td><td>RMSEω(deg/s)</td><td>eω(deg/s)</td><td>RMSEω(deg/s)</td></tr><tr><td>CMax [14]</td><td>16.14</td><td>26.48</td><td>8.18</td><td>13.56</td><td>6.77</td><td>10.62</td></tr><tr><td>Pro-STR [19]</td><td>54.64</td><td>85.63</td><td>22.72</td><td>32.05</td><td>11.07</td><td>12.79</td></tr><tr><td>Ours</td><td>9.41</td><td>18.06</td><td>2.60</td><td>4.78</td><td>2.31</td><td>3.02</td></tr><tr><td>Ours+CMax [14]</td><td>7.71</td><td>10.85</td><td>2.64</td><td>3.80</td><td>0.49</td><td>0.77</td></tr><tr><td>Ours+Pro-STR [19]</td><td>18.87</td><td>32.65</td><td>3.49</td><td>5.87</td><td>2.04</td><td>2.76</td></tr></table>",
|
| 1457 |
+
"bbox": [
|
| 1458 |
+
228,
|
| 1459 |
+
193,
|
| 1460 |
+
782,
|
| 1461 |
+
380
|
| 1462 |
+
],
|
| 1463 |
+
"page_idx": 11
|
| 1464 |
+
},
|
| 1465 |
+
{
|
| 1466 |
+
"type": "table",
|
| 1467 |
+
"img_path": "images/298324b7af0d69de2ce3e06b90bfbacbba88d7ee4dfebf6d38a6a0c4e312b05b.jpg",
|
| 1468 |
+
"table_caption": [
|
| 1469 |
+
"Table 2: Evaluation of our linear solver for differential homography estimation."
|
| 1470 |
+
],
|
| 1471 |
+
"table_footnote": [],
|
| 1472 |
+
"table_body": "<table><tr><td rowspan=\"2\">Method</td><td>board Slow [15]</td><td>patterns rotation</td><td>patterns 6 dof</td></tr><tr><td>F-Norm</td><td>F-Norm</td><td>F-Norm</td></tr><tr><td>CMax [14]</td><td>1.16</td><td>1.33</td><td>1.27</td></tr><tr><td>Pro-STR [19]</td><td>2.07</td><td>1.22</td><td>1.37</td></tr><tr><td>Ours</td><td>0.24</td><td>0.20</td><td>0.43</td></tr><tr><td>Ours+CMax [14]</td><td>0.23</td><td>0.15</td><td>0.41</td></tr><tr><td>Ours+Pro-STR [19]</td><td>0.19</td><td>0.17</td><td>0.42</td></tr></table>",
|
| 1473 |
+
"bbox": [
|
| 1474 |
+
295,
|
| 1475 |
+
459,
|
| 1476 |
+
710,
|
| 1477 |
+
545
|
| 1478 |
+
],
|
| 1479 |
+
"page_idx": 11
|
| 1480 |
+
},
|
| 1481 |
+
{
|
| 1482 |
+
"type": "table",
|
| 1483 |
+
"img_path": "images/c1f86fcb6b160ce345ee4f36711bd686c2c77260d9a5055dbc7d1967b9d4ff75.jpg",
|
| 1484 |
+
"table_caption": [
|
| 1485 |
+
"Table 3: Evaluation of our linear solver on the task of 6-DoF motion tracking."
|
| 1486 |
+
],
|
| 1487 |
+
"table_footnote": [],
|
| 1488 |
+
"table_body": "<table><tr><td rowspan=\"2\">Method</td><td colspan=\"4\">corner Slow [15]</td><td colspan=\"4\">patterns _6 dof (our data)</td></tr><tr><td>eω(deg/s)</td><td>RMSEω(deg/s)</td><td>ev(m/s)</td><td>RMSEv(m/s)</td><td>eω(deg/s)</td><td>RMSEω(deg/s)</td><td>ev(m/s)</td><td>RMSEv(m/s)</td></tr><tr><td>CMax [14]</td><td>12.39</td><td>14.39</td><td>0.34</td><td>0.41</td><td>12.94</td><td>15.68</td><td>1.02</td><td>1.23</td></tr><tr><td>Pro-STR [19]</td><td>5.19</td><td>7.84</td><td>0.13</td><td>0.22</td><td>11.21</td><td>13.49</td><td>1.33</td><td>1.49</td></tr><tr><td>Ours</td><td>1.31</td><td>1.79</td><td>0.01</td><td>0.02</td><td>0.56</td><td>0.89</td><td>0.12</td><td>0.16</td></tr><tr><td>Ours+CMax [14]</td><td>2.19</td><td>2.82</td><td>0.11</td><td>0.17</td><td>0.55</td><td>0.87</td><td>0.18</td><td>0.30</td></tr><tr><td>Ours+Pro-STR [19]</td><td>1.06</td><td>1.34</td><td>0.04</td><td>0.05</td><td>1.07</td><td>1.56</td><td>0.12</td><td>0.15</td></tr><tr><td rowspan=\"2\">Method</td><td colspan=\"4\">mountain_normal [15]</td><td colspan=\"4\">cubes _6 dof (our data)</td></tr><tr><td>eω(deg/s)</td><td>RMSEω(deg/s)</td><td>ev(m/s)</td><td>RMSEv(m/s)</td><td>eω(deg/s)</td><td>RMSEω(deg/s)</td><td>ev(m/s)</td><td>RMSEv(m/s)</td></tr><tr><td>CMax [14]</td><td>11.01</td><td>14.17</td><td>0.44</td><td>0.52</td><td>10.74</td><td>14.33</td><td>0.78</td><td>1.05</td></tr><tr><td>Pro-STR [19]</td><td>5.23</td><td>10.53</td><td>0.14</td><td>0.25</td><td>9.75</td><td>11.40</td><td>1.29</td><td>1.44</td></tr><tr><td>Ours</td><td>1.28</td><td>1.83</td><td>0.07</td><td>0.09</td><td>0.79</td><td>1.02</td><td>0.11</td><td>0.14</td></tr><tr><td>Ours+CMax [14]</td><td>3.84</td><td>6.30</td><td>0.21</td><td>0.30</td><td>1.14</td><td>1.72</td><td>0.12</td><td>0.14</td></tr><tr><td>Ours+Pro-STR [19]</td><td>1.18</td><td>1.58</td><td>0.06</td><td>0.09</td><td>1.05</td><td>1.33</td><td>0.11</td><td>0.14</td></tr></table>",
|
| 1489 |
+
"bbox": [
|
| 1490 |
+
228,
|
| 1491 |
+
625,
|
| 1492 |
+
782,
|
| 1493 |
+
816
|
| 1494 |
+
],
|
| 1495 |
+
"page_idx": 11
|
| 1496 |
+
},
|
| 1497 |
+
{
|
| 1498 |
+
"type": "page_number",
|
| 1499 |
+
"text": "12",
|
| 1500 |
+
"bbox": [
|
| 1501 |
+
217,
|
| 1502 |
+
114,
|
| 1503 |
+
235,
|
| 1504 |
+
126
|
| 1505 |
+
],
|
| 1506 |
+
"page_idx": 11
|
| 1507 |
+
},
|
| 1508 |
+
{
|
| 1509 |
+
"type": "header",
|
| 1510 |
+
"text": "Z. Ren et al.",
|
| 1511 |
+
"bbox": [
|
| 1512 |
+
271,
|
| 1513 |
+
114,
|
| 1514 |
+
357,
|
| 1515 |
+
126
|
| 1516 |
+
],
|
| 1517 |
+
"page_idx": 11
|
| 1518 |
+
},
|
| 1519 |
+
{
|
| 1520 |
+
"type": "image",
|
| 1521 |
+
"img_path": "images/ed1bf464dcf21f8e30c6fc27c6c4a7026b849de293306d384b183408bec5592c.jpg",
|
| 1522 |
+
"image_caption": [],
|
| 1523 |
+
"image_footnote": [],
|
| 1524 |
+
"bbox": [
|
| 1525 |
+
238,
|
| 1526 |
+
146,
|
| 1527 |
+
763,
|
| 1528 |
+
287
|
| 1529 |
+
],
|
| 1530 |
+
"page_idx": 12
|
| 1531 |
+
},
|
| 1532 |
+
{
|
| 1533 |
+
"type": "image",
|
| 1534 |
+
"img_path": "images/9dbb303a4ba45eb7fccc2c30a6525065060729a1d1244f7888c7ce80fb08dfcd.jpg",
|
| 1535 |
+
"image_caption": [
|
| 1536 |
+
"(a) Results of the continuous-time angular velocity estimator."
|
| 1537 |
+
],
|
| 1538 |
+
"image_footnote": [],
|
| 1539 |
+
"bbox": [
|
| 1540 |
+
243,
|
| 1541 |
+
301,
|
| 1542 |
+
759,
|
| 1543 |
+
335
|
| 1544 |
+
],
|
| 1545 |
+
"page_idx": 12
|
| 1546 |
+
},
|
| 1547 |
+
{
|
| 1548 |
+
"type": "image",
|
| 1549 |
+
"img_path": "images/80db2d36fcde1c23580460abe1e6b7dc7d0034b32ea1f45813e038fc04168c64.jpg",
|
| 1550 |
+
"image_caption": [
|
| 1551 |
+
"(b) Comparison against two state-of-the-art nonlinear solutions using data segments that involve sudden variations of angular velocity.",
|
| 1552 |
+
"Fig. 4: Angular velocity estimation using the continuous-time nonlinear solver (Ours)."
|
| 1553 |
+
],
|
| 1554 |
+
"image_footnote": [],
|
| 1555 |
+
"bbox": [
|
| 1556 |
+
243,
|
| 1557 |
+
337,
|
| 1558 |
+
761,
|
| 1559 |
+
435
|
| 1560 |
+
],
|
| 1561 |
+
"page_idx": 12
|
| 1562 |
+
},
|
| 1563 |
+
{
|
| 1564 |
+
"type": "text",
|
| 1565 |
+
"text": "6-DoF Motion Tracking. Given known depth, we recover the linear and angular camera velocities on the task of 6-DoF motion estimation. As shown in Tab. 3, our linear solver reports more accurate results than both CMax and Pro-STR. In addition, we surprisingly find that the linear solver can sometimes outperform the combination of itself with the two nonlinear solvers. This is, according to our investigation, attributed to the violation of constant-motion assumption in some sequences, which is held for the two nonlinear solvers.",
|
| 1566 |
+
"bbox": [
|
| 1567 |
+
212,
|
| 1568 |
+
518,
|
| 1569 |
+
787,
|
| 1570 |
+
625
|
| 1571 |
+
],
|
| 1572 |
+
"page_idx": 12
|
| 1573 |
+
},
|
| 1574 |
+
{
|
| 1575 |
+
"type": "text",
|
| 1576 |
+
"text": "5.3 Evaluation of the Continuous-Time Solver",
|
| 1577 |
+
"text_level": 1,
|
| 1578 |
+
"bbox": [
|
| 1579 |
+
215,
|
| 1580 |
+
647,
|
| 1581 |
+
609,
|
| 1582 |
+
662
|
| 1583 |
+
],
|
| 1584 |
+
"page_idx": 12
|
| 1585 |
+
},
|
| 1586 |
+
{
|
| 1587 |
+
"type": "text",
|
| 1588 |
+
"text": "To justify the claimed advantage of the proposed continuous-time nonlinear solver, we use the exemplary task of angular velocity estimation. The sequence used (shapesRotation [29]) captures a purely rotating event camera, and some parts of the data involves aggressive motion. As illustrated in Fig. 4a, our nonlinear solver gives accurate estimation results with a high stability. Moreover, we select three segments that witness a sudden variation of angular speed within a short time interval and compare our continuous-time nonlinear solver against CMax and Pro-STR. Note that all of them are initialized by our linear method for a fair comparison. As Fig. 4b shows, our results are more consistent with the groundtruth angular velocity than the other two methods, which are less accurate due to the dependence on the constant-motion assumption.",
|
| 1589 |
+
"bbox": [
|
| 1590 |
+
212,
|
| 1591 |
+
672,
|
| 1592 |
+
787,
|
| 1593 |
+
840
|
| 1594 |
+
],
|
| 1595 |
+
"page_idx": 12
|
| 1596 |
+
},
|
| 1597 |
+
{
|
| 1598 |
+
"type": "header",
|
| 1599 |
+
"text": "Motion and Structure from Event-based Normal Flow",
|
| 1600 |
+
"bbox": [
|
| 1601 |
+
372,
|
| 1602 |
+
114,
|
| 1603 |
+
730,
|
| 1604 |
+
127
|
| 1605 |
+
],
|
| 1606 |
+
"page_idx": 12
|
| 1607 |
+
},
|
| 1608 |
+
{
|
| 1609 |
+
"type": "page_number",
|
| 1610 |
+
"text": "13",
|
| 1611 |
+
"bbox": [
|
| 1612 |
+
767,
|
| 1613 |
+
114,
|
| 1614 |
+
784,
|
| 1615 |
+
126
|
| 1616 |
+
],
|
| 1617 |
+
"page_idx": 12
|
| 1618 |
+
},
|
| 1619 |
+
{
|
| 1620 |
+
"type": "table",
|
| 1621 |
+
"img_path": "images/9656dabb1e96a819276ad7d926f7181243e5a58113f2f83a3e7e730899958425.jpg",
|
| 1622 |
+
"table_caption": [
|
| 1623 |
+
"Table 4: Runtime of 6-DoF tracking on sequence mountain_normal (640×480 px)."
|
| 1624 |
+
],
|
| 1625 |
+
"table_footnote": [],
|
| 1626 |
+
"table_body": "<table><tr><td>Component</td><td>Runtime</td></tr><tr><td>Normal flow extraction</td><td>20 ms (for extracting 2000 normal flow vectors)</td></tr><tr><td>Linear solver</td><td>25.97 ms</td></tr><tr><td>CT solver</td><td>28.29 ms</td></tr></table>",
|
| 1627 |
+
"bbox": [
|
| 1628 |
+
282,
|
| 1629 |
+
171,
|
| 1630 |
+
723,
|
| 1631 |
+
224
|
| 1632 |
+
],
|
| 1633 |
+
"page_idx": 13
|
| 1634 |
+
},
|
| 1635 |
+
{
|
| 1636 |
+
"type": "text",
|
| 1637 |
+
"text": "5.4 Discussion",
|
| 1638 |
+
"text_level": 1,
|
| 1639 |
+
"bbox": [
|
| 1640 |
+
215,
|
| 1641 |
+
250,
|
| 1642 |
+
349,
|
| 1643 |
+
263
|
| 1644 |
+
],
|
| 1645 |
+
"page_idx": 13
|
| 1646 |
+
},
|
| 1647 |
+
{
|
| 1648 |
+
"type": "text",
|
| 1649 |
+
"text": "Computational Complexity. The computational complexity of the linear solver depends on the specific task of model fitting. For optical flow and depth estimation, it is linearly proportional to the size of the input event flow, and thus, the complexity is $\\mathcal{O}(N)$ . For the other tasks, the time cost predominantly hinges on the total number of RANSAC iterations and the computations performed per iteration. In each iteration, we typically solve a linear overdetermined system via QR decomposition at a complexity of $\\mathcal{O}(cN^2)$ , where $c$ is the cardinality of a minimal problem (e.g., $c = 3$ for the problem of angular velocity estimation).",
|
| 1650 |
+
"bbox": [
|
| 1651 |
+
212,
|
| 1652 |
+
271,
|
| 1653 |
+
787,
|
| 1654 |
+
393
|
| 1655 |
+
],
|
| 1656 |
+
"page_idx": 13
|
| 1657 |
+
},
|
| 1658 |
+
{
|
| 1659 |
+
"type": "text",
|
| 1660 |
+
"text": "Runtime. All experiments are conducted on a standard computer with Intel i7-13700F CPU, 16G memory and implemented with MATLAB (version R2022a). Using the 6-DoF tracking task as an example, we measure the runtime of the proposed linear solver and continuous-time (CT) solver, as reported in Tab. 4.",
|
| 1661 |
+
"bbox": [
|
| 1662 |
+
212,
|
| 1663 |
+
410,
|
| 1664 |
+
785,
|
| 1665 |
+
472
|
| 1666 |
+
],
|
| 1667 |
+
"page_idx": 13
|
| 1668 |
+
},
|
| 1669 |
+
{
|
| 1670 |
+
"type": "text",
|
| 1671 |
+
"text": "Limitations. To make it clear, we emphasize that our method, like other compared methods, does not resolve the motion and structure parameters simultaneously. It should also be noted that the success of our method relies heavily on the quality of the input normal flow. Particularly, our normal flow extractor (including some other open source solutions for computing normal flow from event data) may fail in the presence of densely repetitive textures, leading to a failure of the modeling fitting scheme. To assess the sensitivity of our method to the quality of event-based normal flow, we propose an analysis on numerical stability using simulation data in the supplementary material.",
|
| 1672 |
+
"bbox": [
|
| 1673 |
+
212,
|
| 1674 |
+
489,
|
| 1675 |
+
787,
|
| 1676 |
+
626
|
| 1677 |
+
],
|
| 1678 |
+
"page_idx": 13
|
| 1679 |
+
},
|
| 1680 |
+
{
|
| 1681 |
+
"type": "text",
|
| 1682 |
+
"text": "6 Conclusions",
|
| 1683 |
+
"text_level": 1,
|
| 1684 |
+
"bbox": [
|
| 1685 |
+
215,
|
| 1686 |
+
645,
|
| 1687 |
+
370,
|
| 1688 |
+
662
|
| 1689 |
+
],
|
| 1690 |
+
"page_idx": 13
|
| 1691 |
+
},
|
| 1692 |
+
{
|
| 1693 |
+
"type": "text",
|
| 1694 |
+
"text": "This work presents a solution to the general problem of geometric model fitting with event data. Considering the differential working principle of event cameras, our method establishes specific constraints between instantaneous motion-and-structure parameters and event-based normal flow. Experiments show that the proposed linear solver leads to closed-form and deterministic solutions, which can be used to initialize existing state-of-the-art nonlinear methods. The proposed continuous-time nonlinear solver, furthermore, demonstrates the capability to handle sudden speed variations, obviating the dependence on the constant-motion assumption. Finally, we hope this work can inspire future research on event-based geometric model fitting, a paramount topic to many downstream techniques, such as event-based 3-D vision, SLAM, and scene segmentation.",
|
| 1695 |
+
"bbox": [
|
| 1696 |
+
212,
|
| 1697 |
+
672,
|
| 1698 |
+
787,
|
| 1699 |
+
840
|
| 1700 |
+
],
|
| 1701 |
+
"page_idx": 13
|
| 1702 |
+
},
|
| 1703 |
+
{
|
| 1704 |
+
"type": "page_number",
|
| 1705 |
+
"text": "14",
|
| 1706 |
+
"bbox": [
|
| 1707 |
+
217,
|
| 1708 |
+
114,
|
| 1709 |
+
235,
|
| 1710 |
+
126
|
| 1711 |
+
],
|
| 1712 |
+
"page_idx": 13
|
| 1713 |
+
},
|
| 1714 |
+
{
|
| 1715 |
+
"type": "header",
|
| 1716 |
+
"text": "Z. Ren et al.",
|
| 1717 |
+
"bbox": [
|
| 1718 |
+
271,
|
| 1719 |
+
114,
|
| 1720 |
+
357,
|
| 1721 |
+
126
|
| 1722 |
+
],
|
| 1723 |
+
"page_idx": 13
|
| 1724 |
+
},
|
| 1725 |
+
{
|
| 1726 |
+
"type": "text",
|
| 1727 |
+
"text": "Acknowledgments",
|
| 1728 |
+
"text_level": 1,
|
| 1729 |
+
"bbox": [
|
| 1730 |
+
215,
|
| 1731 |
+
143,
|
| 1732 |
+
392,
|
| 1733 |
+
162
|
| 1734 |
+
],
|
| 1735 |
+
"page_idx": 14
|
| 1736 |
+
},
|
| 1737 |
+
{
|
| 1738 |
+
"type": "text",
|
| 1739 |
+
"text": "We thank Dr. Yi Yu for proof reading. This work was supported by the National Key Research and Development Project of China under Grant 2023YFB4706600. Funded by the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation) under Germany's Excellence Strategy - EXC 2002/1 \"Science of Intelligence\" - project number 390523135.",
|
| 1740 |
+
"bbox": [
|
| 1741 |
+
212,
|
| 1742 |
+
180,
|
| 1743 |
+
787,
|
| 1744 |
+
256
|
| 1745 |
+
],
|
| 1746 |
+
"page_idx": 14
|
| 1747 |
+
},
|
| 1748 |
+
{
|
| 1749 |
+
"type": "text",
|
| 1750 |
+
"text": "References",
|
| 1751 |
+
"text_level": 1,
|
| 1752 |
+
"bbox": [
|
| 1753 |
+
215,
|
| 1754 |
+
282,
|
| 1755 |
+
321,
|
| 1756 |
+
299
|
| 1757 |
+
],
|
| 1758 |
+
"page_idx": 14
|
| 1759 |
+
},
|
| 1760 |
+
{
|
| 1761 |
+
"type": "list",
|
| 1762 |
+
"sub_type": "ref_text",
|
| 1763 |
+
"list_items": [
|
| 1764 |
+
"1. Barranco, F., Fermüller, C., Aloimonos, Y., Ros, E.: Joint direct estimation of 3D geometry and 3D motion using spatio temporal gradients. Pattern Recog. 113, 107759 (2021). https://doi.org/10.1016/j.patcog.2020.107759",
|
| 1765 |
+
"2. Benosman, R., Clercq, C., Lagorce, X., Ieng, S.H., Bartolozzi, C.: Event-based visual flow. IEEE Trans. Neural Netw. Learn. Syst. 25(2), 407-417 (2014). https://doi.org/10.1109/TNNLS.2013.2273537",
|
| 1766 |
+
"3. Brodsky, T., Fermüller, C., Aloimonos, Y.: Structure from motion: Beyond the epipolar constraint. Int. J. Comput. Vis. 37, 231-258 (2000)",
|
| 1767 |
+
"4. Bryner, S., Gallego, G., Rebecq, H., Scaramuzza, D.: Event-based, direct camera tracking from a photometric 3D map using nonlinear optimization. In: IEEE Int. Conf. Robot. Autom. (ICRA) (2019). https://doi.org/10.1109/ICRA.2019.8794255",
|
| 1768 |
+
"5. Clady, X., Clercq, C., Ieng, S.H., Houseini, F., Randazzo, M., Natale, L., Bartolozzi, C., Benosman, R.: Asynchronous visual event-based time-to-contact. Front. Neurosci. 8(9) (2014). https://doi.org/10.3389/fnins.2014.00009",
|
| 1769 |
+
"6. Conradt, J., Cook, M., Berner, R., Lichtsteiner, P., Douglas, R.J., Delbruck, T.: A pencil balancing robot using a pair of AER dynamic vision sensors. In: IEEE Int. Symp. Circuits Syst. (ISCAS). pp. 781-784 (2009). https://doi.org/10.1109/ISCAS.2009.5117867",
|
| 1770 |
+
"7. Corke, P.: Robotics, Vision and Control: Fundamental Algorithms in MATLAB. Springer Tracts in Advanced Robotics, Springer (2017). https://doi.org/10.1007/978-3-319-54413-7",
|
| 1771 |
+
"8. De Boor, C.: On calculating with b-splines. Journal of Approximation theory 6(1), 50-62 (1972)",
|
| 1772 |
+
"9. Delbruck, T., Lang, M.: Robotic goalie with 3ms reaction time at $4\\%$ CPU load using event-based dynamic vision sensor. Front. Neurosci. 7, 223 (2013). https://doi.org/10.3389/fnins.2013.00223",
|
| 1773 |
+
"0. Dinaux, R., Wessendorp, N., Dupeyroux, J., De Croon, G.C.: Faith: Fast iterative half-plane focus of expansion estimation using optic flow. IEEE Robotics and Automation Letters 6(4), 7627-7634 (2021)",
|
| 1774 |
+
"1. Fermüller, C.: Passive navigation as a pattern recognition problem. Int. J. Comput. Vis. 14(2), 147-158 (1995)",
|
| 1775 |
+
"12. Fischler, M.A., Bolles, R.C.: Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. Commun. ACM 24(6), 381-395 (1981). https://doi.org/10.1145/358669.358692",
|
| 1776 |
+
"13. Gallego, G., Gehrig, M., Scaramuzza, D.: Focus is all you need: Loss functions for event-based vision. In: IEEE Conf. Comput. Vis. Pattern Recog. (CVPR). pp. 12272-12281 (2019). https://doi.org/10.1109/CVPR.2019.01256"
|
| 1777 |
+
],
|
| 1778 |
+
"bbox": [
|
| 1779 |
+
218,
|
| 1780 |
+
316,
|
| 1781 |
+
785,
|
| 1782 |
+
839
|
| 1783 |
+
],
|
| 1784 |
+
"page_idx": 14
|
| 1785 |
+
},
|
| 1786 |
+
{
|
| 1787 |
+
"type": "header",
|
| 1788 |
+
"text": "Motion and Structure from Event-based Normal Flow",
|
| 1789 |
+
"bbox": [
|
| 1790 |
+
370,
|
| 1791 |
+
114,
|
| 1792 |
+
730,
|
| 1793 |
+
127
|
| 1794 |
+
],
|
| 1795 |
+
"page_idx": 14
|
| 1796 |
+
},
|
| 1797 |
+
{
|
| 1798 |
+
"type": "page_number",
|
| 1799 |
+
"text": "15",
|
| 1800 |
+
"bbox": [
|
| 1801 |
+
767,
|
| 1802 |
+
114,
|
| 1803 |
+
785,
|
| 1804 |
+
126
|
| 1805 |
+
],
|
| 1806 |
+
"page_idx": 14
|
| 1807 |
+
},
|
| 1808 |
+
{
|
| 1809 |
+
"type": "list",
|
| 1810 |
+
"sub_type": "ref_text",
|
| 1811 |
+
"list_items": [
|
| 1812 |
+
"14. Gallego, G., Rebecq, H., Scaramuzza, D.: A unifying contrast maximization framework for event cameras, with applications to motion, depth, and optical flow estimation. In: IEEE Conf. Comput. Vis. Pattern Recog. (CVPR). pp. 3867-3876 (2018). https://doi.org/10.1109/CVPR.2018.00407",
|
| 1813 |
+
"15. Gao, L., Liang, Y., Yang, J., Wu, S., Wang, C., Chen, J., Kneip, L.: Vector: A versatile event-centric benchmark for multi-sensor slam. IEEE Robotics and Automation Letters 7(3), 8217-8224 (2022). https://doi.org/10.1109/LRA.2022.3186770",
|
| 1814 |
+
"16. Gao, L., Su, H., Gehrig, D., Cannici, M., Scaramuzza, D., Kneip, L.: A 5-point minimal solver for event camera relative motion estimation. In: Int. Conf. Comput. Vis. (ICCV). pp. 8015-8025 (2023). https://doi.org/10.1109/ICCV51070.2023.00739",
|
| 1815 |
+
"17. Gu, C., Learned-Miller, E., Sheldon, D., Gallego, G., Bideau, P.: The spatiotemporal Poisson point process: A simple model for the alignment of event camera data. In: Int. Conf. Comput. Vis. (ICCV). pp. 13495-13504 (2021). https://doi.org/10.1109/ICCV48922.2021.01324",
|
| 1816 |
+
"18. Guo, S., Gallego, G.: CMax-SLAM: Event-based rotational-motion bundle adjustment and SLAM system using contrast maximization. IEEE Trans. Robot. pp. 1-20 (2024). https://doi.org/10.1109/TRO.2024.3378443",
|
| 1817 |
+
"19. Huang, X., Zhang, Y., Xiong, Z.: Progressive spatio-temporal alignment for efficient event-based motion estimation. In: IEEE Conf. Comput. Vis. Pattern Recog. (CVPR). pp. 1537-1546 (2023)",
|
| 1818 |
+
"20. Ji, H., Fermuller, C.: A 3D shape constraint on video. IEEE Trans. Pattern Anal. Mach. Intell. 28(6), 1018-1023 (2006)",
|
| 1819 |
+
"21. Kim, H., Leutenegger, S., Davison, A.J.: Real-time 3D reconstruction and 6-DoF tracking with an event camera. In: Eur. Conf. Comput. Vis. (ECCV). pp. 349-364 (2016). https://doi.org/\\protect{10.1007/978-3-319-46466-4_21}",
|
| 1820 |
+
"22. Liu, D., Parra, A., Chin, T.J.: Spatiotemporal registration for event-based visual odometry. In: IEEE Conf. Comput. Vis. Pattern Recog. (CVPR). pp. 4935-4944 (2021). https://doi.org/10.1109/CVPR46437.2021.00490",
|
| 1821 |
+
"23. Longuet-Higgins, H.C., Prazdny, K.: The interpretation of a moving retinal image. Proceedings of the Royal Society of London. Series B. Biological Sciences 208(1173), 385-397 (1980). https://doi.org/10.1098/rspb.1980.0057",
|
| 1822 |
+
"24. Lu, X., Zhou, Y., Niu, J., Zhong, S., Shen, S.: Event-based visual inertial velometer (2024)",
|
| 1823 |
+
"25. Ma, Y., Soatto, S., Košecká, J., Sastry, S.: An Invitation to 3-D Vision: From Images to Geometric Models. Springer (2004)",
|
| 1824 |
+
"26. Meyer, F., Bouthemy, P.: Estimation of time-to-collision maps from first order motion models and normal flows. In: 1992 11th IAPR International Conference on Pattern Recognition. vol. 1, pp. 78-82. IEEE Computer Society (1992)",
|
| 1825 |
+
"27. Meyer, F.G.: Time-to-collision from first-order models of the motion field. IEEE transactions on robotics and automation 10(6), 792-798 (1994)",
|
| 1826 |
+
"28. Mueggler, E., Gallego, G., Rebecq, H., Scaramuzza, D.: Continuous-time visual-inertial odometry for event cameras. IEEE Trans. Robot. 34(6), 1425-1440 (Dec 2018). https://doi.org/10.1109/tro.2018.2858287",
|
| 1827 |
+
"29. Mueggler, E., Rebecq, H., Gallego, G., Delbruck, T., Scaramuzza, D.: The event-camera dataset and simulator: Event-based data for pose estimation, visual odometry, and SLAM. Int. J. Robot. Research 36(2), 142-149 (2017). https://doi.org/10.1177/0278364917691115",
|
| 1828 |
+
"30. Nunes, U.M., Demiris, Y.: Entropy minimisation framework for event-based vision model estimation. In: Eur. Conf. Comput. Vis. (ECCV). pp. 161-176 (2020). https://doi.org/10.1007/978-3-030-58558-7_10"
|
| 1829 |
+
],
|
| 1830 |
+
"bbox": [
|
| 1831 |
+
217,
|
| 1832 |
+
146,
|
| 1833 |
+
785,
|
| 1834 |
+
840
|
| 1835 |
+
],
|
| 1836 |
+
"page_idx": 15
|
| 1837 |
+
},
|
| 1838 |
+
{
|
| 1839 |
+
"type": "page_number",
|
| 1840 |
+
"text": "16",
|
| 1841 |
+
"bbox": [
|
| 1842 |
+
217,
|
| 1843 |
+
114,
|
| 1844 |
+
235,
|
| 1845 |
+
126
|
| 1846 |
+
],
|
| 1847 |
+
"page_idx": 15
|
| 1848 |
+
},
|
| 1849 |
+
{
|
| 1850 |
+
"type": "header",
|
| 1851 |
+
"text": "Z. Ren et al.",
|
| 1852 |
+
"bbox": [
|
| 1853 |
+
271,
|
| 1854 |
+
114,
|
| 1855 |
+
357,
|
| 1856 |
+
126
|
| 1857 |
+
],
|
| 1858 |
+
"page_idx": 15
|
| 1859 |
+
},
|
| 1860 |
+
{
|
| 1861 |
+
"type": "list",
|
| 1862 |
+
"sub_type": "ref_text",
|
| 1863 |
+
"list_items": [
|
| 1864 |
+
"31. Parameshwara, C.M., Hari, G., Fermüller, C., Sanket, N.J., Aloimonos, Y.: Diff-posednet: direct differentiable camera pose estimation. In: IEEE Conf. Comput. Vis. Pattern Recog. (CVPR). pp. 6845-6854 (2022)",
|
| 1865 |
+
"32. Pijnacker Hordijk, B.J., Scheper, K.Y., De Croon, G.C.: Vertical landing for micro air vehicles using event-based optical flow. J. Field Robot. 35(1), 69-90 (2018)",
|
| 1866 |
+
"33. Rebecq, H., Gallego, G., Scaramuzza, D.: EMVS: Event-based multi-view stereo. In: British Mach. Vis. Conf. (BMVC) (2016). https://doi.org/10.5244/C.30.63",
|
| 1867 |
+
"34. Rebecq, H., Horstschaefer, T., Scaramuzza, D.: Real-time visual-inertial odometry for event cameras using keyframe-based nonlinear optimization. In: British Mach. Vis. Conf. (BMVC) (2017)",
|
| 1868 |
+
"35. Rebecq, H., Horstschäfer, T., Gallego, G., Scaramuzza, D.: EVO: A geometric approach to event-based 6-DOF parallel tracking and mapping in real-time. IEEE Robot. Autom. Lett. 2(2), 593-600 (2017). https://doi.org/10.1109/LRA.2016.2645143",
|
| 1869 |
+
"36. Shiba, S., Aoki, Y., Gallego, G.: Event collapse in contrast maximization frameworks. Sensors 22(14), 1-20 (2022). https://doi.org/10.3390/s22145190",
|
| 1870 |
+
"37. Shiba, S., Aoki, Y., Gallego, G.: A fast geometric regularizer to mitigate event collapse in the contrast maximization framework. Adv. Intell. Syst. p. 2200251 (2022). https://doi.org/10.1002/aisy.202200251",
|
| 1871 |
+
"38. Shiba, S., Aoki, Y., Gallego, G.: Secrets of event-based optical flow. In: Eur. Conf. Comput. Vis. (ECCV). pp. 628-645 (2022). https://doi.org/10.1007/978-3-031-19797-0 36",
|
| 1872 |
+
"39. Stoffregen, T., Gallego, G., Drummond, T., Kleeman, L., Scaramuzza, D.: Event-based motion segmentation by motion compensation. In: Int. Conf. Comput. Vis. (ICCV). pp. 7243-7252 (2019). https://doi.org/10.1109/ICCV.2019.00734",
|
| 1873 |
+
"40. Trucco, E., Verri, A.: Introductory Techniques for 3-D Computer Vision. Prentice Hall PTR, Upper Saddle River, NJ, USA (1998)",
|
| 1874 |
+
"41. Zhou, Y., Gallego, G., Rebecq, H., Kneip, L., Li, H., Scaramuzza, D.: Semidense 3D reconstruction with a stereo event camera. In: Eur. Conf. Comput. Vis. (ECCV). pp. 242-258 (2018). https://doi.org/10.1007/978-3-030-01246-5_15",
|
| 1875 |
+
"42. Zhou, Y., Gallego, G., Shen, S.: Event-based stereo visual odometry. IEEE Trans. Robot. 37(5), 1433-1450 (2021). https://doi.org/10.1109/TRO.2021.3062252",
|
| 1876 |
+
"43. Zhu, A.Z., Chen, Y., Daniilidis, K.: Realtime time synchronized event-based stereo. In: Eur. Conf. Comput. Vis. (ECCV). pp. 438-452 (2018). https://doi.org/10.1007/978-3-030-01231-1_27",
|
| 1877 |
+
"44. Zhuang, B., Tran, Q.H.: Image stitching and rectification for hand-held cameras. In: Eur. Conf. Comput. Vis. (ECCV) (2020). https://doi.org/10.1007/978-3-030-58571-6_15"
|
| 1878 |
+
],
|
| 1879 |
+
"bbox": [
|
| 1880 |
+
212,
|
| 1881 |
+
146,
|
| 1882 |
+
787,
|
| 1883 |
+
674
|
| 1884 |
+
],
|
| 1885 |
+
"page_idx": 16
|
| 1886 |
+
},
|
| 1887 |
+
{
|
| 1888 |
+
"type": "header",
|
| 1889 |
+
"text": "Motion and Structure from Event-based Normal Flow",
|
| 1890 |
+
"bbox": [
|
| 1891 |
+
372,
|
| 1892 |
+
114,
|
| 1893 |
+
730,
|
| 1894 |
+
127
|
| 1895 |
+
],
|
| 1896 |
+
"page_idx": 16
|
| 1897 |
+
},
|
| 1898 |
+
{
|
| 1899 |
+
"type": "page_number",
|
| 1900 |
+
"text": "17",
|
| 1901 |
+
"bbox": [
|
| 1902 |
+
767,
|
| 1903 |
+
116,
|
| 1904 |
+
784,
|
| 1905 |
+
126
|
| 1906 |
+
],
|
| 1907 |
+
"page_idx": 16
|
| 1908 |
+
}
|
| 1909 |
+
]
|