Chelsea707 commited on
Commit
09bda78
·
verified ·
1 Parent(s): 2d08de5

Add Batch 1a6457c7-bb08-47a1-8891-1273ea9f0f5e data

Browse files
This view is limited to 50 files because it contains too many changes.   See raw diff
Files changed (50) hide show
  1. .gitattributes +64 -0
  2. 2023/Robust Unsupervised StyleGAN Image Restoration/22cf3860-c0a9-4cc3-ae76-e2aceeacd7e1_content_list.json +1710 -0
  3. 2023/Robust Unsupervised StyleGAN Image Restoration/22cf3860-c0a9-4cc3-ae76-e2aceeacd7e1_model.json +0 -0
  4. 2023/Robust Unsupervised StyleGAN Image Restoration/22cf3860-c0a9-4cc3-ae76-e2aceeacd7e1_origin.pdf +3 -0
  5. 2023/Robust Unsupervised StyleGAN Image Restoration/full.md +374 -0
  6. 2023/Robust Unsupervised StyleGAN Image Restoration/images.zip +3 -0
  7. 2023/Robust Unsupervised StyleGAN Image Restoration/layout.json +0 -0
  8. 2023/Robust and Scalable Gaussian Process Regression and Its Applications/def98ceb-97d5-4394-a2b2-443884eb66dd_content_list.json +2097 -0
  9. 2023/Robust and Scalable Gaussian Process Regression and Its Applications/def98ceb-97d5-4394-a2b2-443884eb66dd_model.json +0 -0
  10. 2023/Robust and Scalable Gaussian Process Regression and Its Applications/def98ceb-97d5-4394-a2b2-443884eb66dd_origin.pdf +3 -0
  11. 2023/Robust and Scalable Gaussian Process Regression and Its Applications/full.md +434 -0
  12. 2023/Robust and Scalable Gaussian Process Regression and Its Applications/images.zip +3 -0
  13. 2023/Robust and Scalable Gaussian Process Regression and Its Applications/layout.json +0 -0
  14. 2023/RobustNeRF_ Ignoring Distractors With Robust Losses/2f329418-58fe-4ad7-b462-6d0d0ce341f1_content_list.json +2145 -0
  15. 2023/RobustNeRF_ Ignoring Distractors With Robust Losses/2f329418-58fe-4ad7-b462-6d0d0ce341f1_model.json +0 -0
  16. 2023/RobustNeRF_ Ignoring Distractors With Robust Losses/2f329418-58fe-4ad7-b462-6d0d0ce341f1_origin.pdf +3 -0
  17. 2023/RobustNeRF_ Ignoring Distractors With Robust Losses/full.md +424 -0
  18. 2023/RobustNeRF_ Ignoring Distractors With Robust Losses/images.zip +3 -0
  19. 2023/RobustNeRF_ Ignoring Distractors With Robust Losses/layout.json +0 -0
  20. 2023/Role of Transients in Two-Bounce Non-Line-of-Sight Imaging/dd8b3d32-1f20-46bb-9a36-d731b9d46ed2_content_list.json +1589 -0
  21. 2023/Role of Transients in Two-Bounce Non-Line-of-Sight Imaging/dd8b3d32-1f20-46bb-9a36-d731b9d46ed2_model.json +2178 -0
  22. 2023/Role of Transients in Two-Bounce Non-Line-of-Sight Imaging/dd8b3d32-1f20-46bb-9a36-d731b9d46ed2_origin.pdf +3 -0
  23. 2023/Role of Transients in Two-Bounce Non-Line-of-Sight Imaging/full.md +333 -0
  24. 2023/Role of Transients in Two-Bounce Non-Line-of-Sight Imaging/images.zip +3 -0
  25. 2023/Role of Transients in Two-Bounce Non-Line-of-Sight Imaging/layout.json +0 -0
  26. 2023/Rotation-Invariant Transformer for Point Cloud Matching/424bbd8d-fe04-4142-91bc-ea8eb4c92ccb_content_list.json +1538 -0
  27. 2023/Rotation-Invariant Transformer for Point Cloud Matching/424bbd8d-fe04-4142-91bc-ea8eb4c92ccb_model.json +0 -0
  28. 2023/Rotation-Invariant Transformer for Point Cloud Matching/424bbd8d-fe04-4142-91bc-ea8eb4c92ccb_origin.pdf +3 -0
  29. 2023/Rotation-Invariant Transformer for Point Cloud Matching/full.md +311 -0
  30. 2023/Rotation-Invariant Transformer for Point Cloud Matching/images.zip +3 -0
  31. 2023/Rotation-Invariant Transformer for Point Cloud Matching/layout.json +0 -0
  32. 2023/Run, Don't Walk_ Chasing Higher FLOPS for Faster Neural Networks/6a2198d1-cecd-4fd5-b373-fd2aeb690abb_content_list.json +1585 -0
  33. 2023/Run, Don't Walk_ Chasing Higher FLOPS for Faster Neural Networks/6a2198d1-cecd-4fd5-b373-fd2aeb690abb_model.json +0 -0
  34. 2023/Run, Don't Walk_ Chasing Higher FLOPS for Faster Neural Networks/6a2198d1-cecd-4fd5-b373-fd2aeb690abb_origin.pdf +3 -0
  35. 2023/Run, Don't Walk_ Chasing Higher FLOPS for Faster Neural Networks/full.md +338 -0
  36. 2023/Run, Don't Walk_ Chasing Higher FLOPS for Faster Neural Networks/images.zip +3 -0
  37. 2023/Run, Don't Walk_ Chasing Higher FLOPS for Faster Neural Networks/layout.json +0 -0
  38. 2023/S3C_ Semi-Supervised VQA Natural Language Explanation via Self-Critical Learning/6ead34e8-75d6-4d08-a4ab-1da1402619c2_content_list.json +1506 -0
  39. 2023/S3C_ Semi-Supervised VQA Natural Language Explanation via Self-Critical Learning/6ead34e8-75d6-4d08-a4ab-1da1402619c2_model.json +0 -0
  40. 2023/S3C_ Semi-Supervised VQA Natural Language Explanation via Self-Critical Learning/6ead34e8-75d6-4d08-a4ab-1da1402619c2_origin.pdf +3 -0
  41. 2023/S3C_ Semi-Supervised VQA Natural Language Explanation via Self-Critical Learning/full.md +322 -0
  42. 2023/S3C_ Semi-Supervised VQA Natural Language Explanation via Self-Critical Learning/images.zip +3 -0
  43. 2023/S3C_ Semi-Supervised VQA Natural Language Explanation via Self-Critical Learning/layout.json +0 -0
  44. 2023/SAP-DETR_ Bridging the Gap Between Salient Points and Queries-Based Transformer Detector for Fast Model Convergency/909bbbe4-4e26-472f-ab37-1d60d8bab4b6_content_list.json +1378 -0
  45. 2023/SAP-DETR_ Bridging the Gap Between Salient Points and Queries-Based Transformer Detector for Fast Model Convergency/909bbbe4-4e26-472f-ab37-1d60d8bab4b6_model.json +1824 -0
  46. 2023/SAP-DETR_ Bridging the Gap Between Salient Points and Queries-Based Transformer Detector for Fast Model Convergency/909bbbe4-4e26-472f-ab37-1d60d8bab4b6_origin.pdf +3 -0
  47. 2023/SAP-DETR_ Bridging the Gap Between Salient Points and Queries-Based Transformer Detector for Fast Model Convergency/full.md +268 -0
  48. 2023/SAP-DETR_ Bridging the Gap Between Salient Points and Queries-Based Transformer Detector for Fast Model Convergency/images.zip +3 -0
  49. 2023/SAP-DETR_ Bridging the Gap Between Salient Points and Queries-Based Transformer Detector for Fast Model Convergency/layout.json +0 -0
  50. 2023/SCADE_ NeRFs from Space Carving With Ambiguity-Aware Depth Estimates/cb201944-e311-4584-9847-096959c3a596_content_list.json +1770 -0
.gitattributes CHANGED
@@ -7463,3 +7463,67 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
7463
  2023/Robust[[:space:]]Outlier[[:space:]]Rejection[[:space:]]for[[:space:]]3D[[:space:]]Registration[[:space:]]With[[:space:]]Variational[[:space:]]Bayes/65a0dc15-9eca-45e4-8931-aeda13fc78f9_origin.pdf filter=lfs diff=lfs merge=lfs -text
7464
  2023/Robust[[:space:]]Single[[:space:]]Image[[:space:]]Reflection[[:space:]]Removal[[:space:]]Against[[:space:]]Adversarial[[:space:]]Attacks/9967a3ae-116f-48d0-8a30-d1102ebae7fe_origin.pdf filter=lfs diff=lfs merge=lfs -text
7465
  2023/Robust[[:space:]]Test-Time[[:space:]]Adaptation[[:space:]]in[[:space:]]Dynamic[[:space:]]Scenarios/18d0ec2a-ba7b-4e7c-8b52-b26b805e62f6_origin.pdf filter=lfs diff=lfs merge=lfs -text
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
7463
  2023/Robust[[:space:]]Outlier[[:space:]]Rejection[[:space:]]for[[:space:]]3D[[:space:]]Registration[[:space:]]With[[:space:]]Variational[[:space:]]Bayes/65a0dc15-9eca-45e4-8931-aeda13fc78f9_origin.pdf filter=lfs diff=lfs merge=lfs -text
7464
  2023/Robust[[:space:]]Single[[:space:]]Image[[:space:]]Reflection[[:space:]]Removal[[:space:]]Against[[:space:]]Adversarial[[:space:]]Attacks/9967a3ae-116f-48d0-8a30-d1102ebae7fe_origin.pdf filter=lfs diff=lfs merge=lfs -text
7465
  2023/Robust[[:space:]]Test-Time[[:space:]]Adaptation[[:space:]]in[[:space:]]Dynamic[[:space:]]Scenarios/18d0ec2a-ba7b-4e7c-8b52-b26b805e62f6_origin.pdf filter=lfs diff=lfs merge=lfs -text
7466
+ 2023/Robust[[:space:]]Unsupervised[[:space:]]StyleGAN[[:space:]]Image[[:space:]]Restoration/22cf3860-c0a9-4cc3-ae76-e2aceeacd7e1_origin.pdf filter=lfs diff=lfs merge=lfs -text
7467
+ 2023/Robust[[:space:]]and[[:space:]]Scalable[[:space:]]Gaussian[[:space:]]Process[[:space:]]Regression[[:space:]]and[[:space:]]Its[[:space:]]Applications/def98ceb-97d5-4394-a2b2-443884eb66dd_origin.pdf filter=lfs diff=lfs merge=lfs -text
7468
+ 2023/RobustNeRF_[[:space:]]Ignoring[[:space:]]Distractors[[:space:]]With[[:space:]]Robust[[:space:]]Losses/2f329418-58fe-4ad7-b462-6d0d0ce341f1_origin.pdf filter=lfs diff=lfs merge=lfs -text
7469
+ 2023/Role[[:space:]]of[[:space:]]Transients[[:space:]]in[[:space:]]Two-Bounce[[:space:]]Non-Line-of-Sight[[:space:]]Imaging/dd8b3d32-1f20-46bb-9a36-d731b9d46ed2_origin.pdf filter=lfs diff=lfs merge=lfs -text
7470
+ 2023/Rotation-Invariant[[:space:]]Transformer[[:space:]]for[[:space:]]Point[[:space:]]Cloud[[:space:]]Matching/424bbd8d-fe04-4142-91bc-ea8eb4c92ccb_origin.pdf filter=lfs diff=lfs merge=lfs -text
7471
+ 2023/Run,[[:space:]]Don't[[:space:]]Walk_[[:space:]]Chasing[[:space:]]Higher[[:space:]]FLOPS[[:space:]]for[[:space:]]Faster[[:space:]]Neural[[:space:]]Networks/6a2198d1-cecd-4fd5-b373-fd2aeb690abb_origin.pdf filter=lfs diff=lfs merge=lfs -text
7472
+ 2023/S3C_[[:space:]]Semi-Supervised[[:space:]]VQA[[:space:]]Natural[[:space:]]Language[[:space:]]Explanation[[:space:]]via[[:space:]]Self-Critical[[:space:]]Learning/6ead34e8-75d6-4d08-a4ab-1da1402619c2_origin.pdf filter=lfs diff=lfs merge=lfs -text
7473
+ 2023/SAP-DETR_[[:space:]]Bridging[[:space:]]the[[:space:]]Gap[[:space:]]Between[[:space:]]Salient[[:space:]]Points[[:space:]]and[[:space:]]Queries-Based[[:space:]]Transformer[[:space:]]Detector[[:space:]]for[[:space:]]Fast[[:space:]]Model[[:space:]]Convergency/909bbbe4-4e26-472f-ab37-1d60d8bab4b6_origin.pdf filter=lfs diff=lfs merge=lfs -text
7474
+ 2023/SCADE_[[:space:]]NeRFs[[:space:]]from[[:space:]]Space[[:space:]]Carving[[:space:]]With[[:space:]]Ambiguity-Aware[[:space:]]Depth[[:space:]]Estimates/cb201944-e311-4584-9847-096959c3a596_origin.pdf filter=lfs diff=lfs merge=lfs -text
7475
+ 2023/SCConv_[[:space:]]Spatial[[:space:]]and[[:space:]]Channel[[:space:]]Reconstruction[[:space:]]Convolution[[:space:]]for[[:space:]]Feature[[:space:]]Redundancy/cde54b77-44c6-4dda-8761-9969150d4bec_origin.pdf filter=lfs diff=lfs merge=lfs -text
7476
+ 2023/SCOOP_[[:space:]]Self-Supervised[[:space:]]Correspondence[[:space:]]and[[:space:]]Optimization-Based[[:space:]]Scene[[:space:]]Flow/2d0fc651-4e0b-4b27-8a33-61605dae5a3b_origin.pdf filter=lfs diff=lfs merge=lfs -text
7477
+ 2023/SCOTCH[[:space:]]and[[:space:]]SODA_[[:space:]]A[[:space:]]Transformer[[:space:]]Video[[:space:]]Shadow[[:space:]]Detection[[:space:]]Framework/db4a242e-46af-47cc-9782-646b5ef8b2a3_origin.pdf filter=lfs diff=lfs merge=lfs -text
7478
+ 2023/SCPNet_[[:space:]]Semantic[[:space:]]Scene[[:space:]]Completion[[:space:]]on[[:space:]]Point[[:space:]]Cloud/a1134f69-6375-42b9-9a92-868bfcf809a3_origin.pdf filter=lfs diff=lfs merge=lfs -text
7479
+ 2023/SCoDA_[[:space:]]Domain[[:space:]]Adaptive[[:space:]]Shape[[:space:]]Completion[[:space:]]for[[:space:]]Real[[:space:]]Scans/a2de3a81-2619-4975-9f69-e49ad0b14643_origin.pdf filter=lfs diff=lfs merge=lfs -text
7480
+ 2023/SDC-UDA_[[:space:]]Volumetric[[:space:]]Unsupervised[[:space:]]Domain[[:space:]]Adaptation[[:space:]]Framework[[:space:]]for[[:space:]]Slice-Direction[[:space:]]Continuous[[:space:]]Cross-Modality[[:space:]]Medical[[:space:]]Image[[:space:]]Segmentation/20148cfd-bd1f-4726-9d6d-b13924bbcc1e_origin.pdf filter=lfs diff=lfs merge=lfs -text
7481
+ 2023/SDFusion_[[:space:]]Multimodal[[:space:]]3D[[:space:]]Shape[[:space:]]Completion,[[:space:]]Reconstruction,[[:space:]]and[[:space:]]Generation/f29ee3dd-8b63-4289-b6a4-d6e103a0e949_origin.pdf filter=lfs diff=lfs merge=lfs -text
7482
+ 2023/SE-ORNet_[[:space:]]Self-Ensembling[[:space:]]Orientation-Aware[[:space:]]Network[[:space:]]for[[:space:]]Unsupervised[[:space:]]Point[[:space:]]Cloud[[:space:]]Shape[[:space:]]Correspondence/94ea60f5-77d0-47dc-8f2c-3fe79d2f954a_origin.pdf filter=lfs diff=lfs merge=lfs -text
7483
+ 2023/SECAD-Net_[[:space:]]Self-Supervised[[:space:]]CAD[[:space:]]Reconstruction[[:space:]]by[[:space:]]Learning[[:space:]]Sketch-Extrude[[:space:]]Operations/94d1495e-0d51-4b2c-8690-ac04468c5645_origin.pdf filter=lfs diff=lfs merge=lfs -text
7484
+ 2023/SFD2_[[:space:]]Semantic-Guided[[:space:]]Feature[[:space:]]Detection[[:space:]]and[[:space:]]Description/6efa713f-1ae0-4651-bc92-c61912feb28d_origin.pdf filter=lfs diff=lfs merge=lfs -text
7485
+ 2023/SGLoc_[[:space:]]Scene[[:space:]]Geometry[[:space:]]Encoding[[:space:]]for[[:space:]]Outdoor[[:space:]]LiDAR[[:space:]]Localization/56c4bde5-7fa2-4496-a50a-885b12091caa_origin.pdf filter=lfs diff=lfs merge=lfs -text
7486
+ 2023/SHS-Net_[[:space:]]Learning[[:space:]]Signed[[:space:]]Hyper[[:space:]]Surfaces[[:space:]]for[[:space:]]Oriented[[:space:]]Normal[[:space:]]Estimation[[:space:]]of[[:space:]]Point[[:space:]]Clouds/b88f2048-c087-4b9c-9601-f88e3009b262_origin.pdf filter=lfs diff=lfs merge=lfs -text
7487
+ 2023/SIEDOB_[[:space:]]Semantic[[:space:]]Image[[:space:]]Editing[[:space:]]by[[:space:]]Disentangling[[:space:]]Object[[:space:]]and[[:space:]]Background/a369b37c-8e06-4d08-a268-de9d54816ab4_origin.pdf filter=lfs diff=lfs merge=lfs -text
7488
+ 2023/SIM_[[:space:]]Semantic-Aware[[:space:]]Instance[[:space:]]Mask[[:space:]]Generation[[:space:]]for[[:space:]]Box-Supervised[[:space:]]Instance[[:space:]]Segmentation/884f8e5b-a009-4236-8afe-619abfb36ed7_origin.pdf filter=lfs diff=lfs merge=lfs -text
7489
+ 2023/SINE_[[:space:]]SINgle[[:space:]]Image[[:space:]]Editing[[:space:]]With[[:space:]]Text-to-Image[[:space:]]Diffusion[[:space:]]Models/9e4aea54-e956-45c7-8e0d-e18b2b0dde98_origin.pdf filter=lfs diff=lfs merge=lfs -text
7490
+ 2023/SINE_[[:space:]]Semantic-Driven[[:space:]]Image-Based[[:space:]]NeRF[[:space:]]Editing[[:space:]]With[[:space:]]Prior-Guided[[:space:]]Editing[[:space:]]Field/98bd5d41-9994-4c5a-92c8-200bf021b945_origin.pdf filter=lfs diff=lfs merge=lfs -text
7491
+ 2023/SLACK_[[:space:]]Stable[[:space:]]Learning[[:space:]]of[[:space:]]Augmentations[[:space:]]With[[:space:]]Cold-Start[[:space:]]and[[:space:]]KL[[:space:]]Regularization/13983024-bdf8-451a-82ba-7c4e7f6f3424_origin.pdf filter=lfs diff=lfs merge=lfs -text
7492
+ 2023/SLOPER4D_[[:space:]]A[[:space:]]Scene-Aware[[:space:]]Dataset[[:space:]]for[[:space:]]Global[[:space:]]4D[[:space:]]Human[[:space:]]Pose[[:space:]]Estimation[[:space:]]in[[:space:]]Urban[[:space:]]Environments/6eb8c194-2d4e-4f05-826e-be6c8b0450e0_origin.pdf filter=lfs diff=lfs merge=lfs -text
7493
+ 2023/SMAE_[[:space:]]Few-Shot[[:space:]]Learning[[:space:]]for[[:space:]]HDR[[:space:]]Deghosting[[:space:]]With[[:space:]]Saturation-Aware[[:space:]]Masked[[:space:]]Autoencoders/1480ea63-f1bc-4488-8d0b-6bc8008633c3_origin.pdf filter=lfs diff=lfs merge=lfs -text
7494
+ 2023/SMOC-Net_[[:space:]]Leveraging[[:space:]]Camera[[:space:]]Pose[[:space:]]for[[:space:]]Self-Supervised[[:space:]]Monocular[[:space:]]Object[[:space:]]Pose[[:space:]]Estimation/7ed77cdd-0ec8-4cc4-8097-ec9f36276521_origin.pdf filter=lfs diff=lfs merge=lfs -text
7495
+ 2023/SMPConv_[[:space:]]Self-Moving[[:space:]]Point[[:space:]]Representations[[:space:]]for[[:space:]]Continuous[[:space:]]Convolution/a40d3fb1-72be-4c1d-a333-984a29892bca_origin.pdf filter=lfs diff=lfs merge=lfs -text
7496
+ 2023/SOOD_[[:space:]]Towards[[:space:]]Semi-Supervised[[:space:]]Oriented[[:space:]]Object[[:space:]]Detection/b6c29c1e-51a1-4935-bf5e-ec2e26810dec_origin.pdf filter=lfs diff=lfs merge=lfs -text
7497
+ 2023/SPARF_[[:space:]]Neural[[:space:]]Radiance[[:space:]]Fields[[:space:]]From[[:space:]]Sparse[[:space:]]and[[:space:]]Noisy[[:space:]]Poses/9768904d-8df6-48f0-a9ea-13cb333d26ef_origin.pdf filter=lfs diff=lfs merge=lfs -text
7498
+ 2023/SPIn-NeRF_[[:space:]]Multiview[[:space:]]Segmentation[[:space:]]and[[:space:]]Perceptual[[:space:]]Inpainting[[:space:]]With[[:space:]]Neural[[:space:]]Radiance[[:space:]]Fields/a5f839e7-988d-4934-9aef-4c494df6e55a_origin.pdf filter=lfs diff=lfs merge=lfs -text
7499
+ 2023/SQUID_[[:space:]]Deep[[:space:]]Feature[[:space:]]In-Painting[[:space:]]for[[:space:]]Unsupervised[[:space:]]Anomaly[[:space:]]Detection/c1d458f7-f8ad-480d-9ab9-60ca7bc1a759_origin.pdf filter=lfs diff=lfs merge=lfs -text
7500
+ 2023/STAR[[:space:]]Loss_[[:space:]]Reducing[[:space:]]Semantic[[:space:]]Ambiguity[[:space:]]in[[:space:]]Facial[[:space:]]Landmark[[:space:]]Detection/fb30f024-6bbe-44bb-baaf-0659043475bc_origin.pdf filter=lfs diff=lfs merge=lfs -text
7501
+ 2023/STDLens_[[:space:]]Model[[:space:]]Hijacking-Resilient[[:space:]]Federated[[:space:]]Learning[[:space:]]for[[:space:]]Object[[:space:]]Detection/970c86e7-a18e-4fa1-bf10-060bfc2535dd_origin.pdf filter=lfs diff=lfs merge=lfs -text
7502
+ 2023/STMT_[[:space:]]A[[:space:]]Spatial-Temporal[[:space:]]Mesh[[:space:]]Transformer[[:space:]]for[[:space:]]MoCap-Based[[:space:]]Action[[:space:]]Recognition/ecec9e25-9648-4975-9559-d389674e900b_origin.pdf filter=lfs diff=lfs merge=lfs -text
7503
+ 2023/STMixer_[[:space:]]A[[:space:]]One-Stage[[:space:]]Sparse[[:space:]]Action[[:space:]]Detector/7eef5975-ffbf-46c0-a31b-5b464eef7eef_origin.pdf filter=lfs diff=lfs merge=lfs -text
7504
+ 2023/SUDS_[[:space:]]Scalable[[:space:]]Urban[[:space:]]Dynamic[[:space:]]Scenes/e8fd5f67-d301-4dd3-8588-d4088862c6a6_origin.pdf filter=lfs diff=lfs merge=lfs -text
7505
+ 2023/SVFormer_[[:space:]]Semi-Supervised[[:space:]]Video[[:space:]]Transformer[[:space:]]for[[:space:]]Action[[:space:]]Recognition/2bd91319-3d84-428a-8335-5833e2e27f2c_origin.pdf filter=lfs diff=lfs merge=lfs -text
7506
+ 2023/SVGformer_[[:space:]]Representation[[:space:]]Learning[[:space:]]for[[:space:]]Continuous[[:space:]]Vector[[:space:]]Graphics[[:space:]]Using[[:space:]]Transformers/3621baa2-936b-4fa2-8381-1968bf4d3112_origin.pdf filter=lfs diff=lfs merge=lfs -text
7507
+ 2023/SViTT_[[:space:]]Temporal[[:space:]]Learning[[:space:]]of[[:space:]]Sparse[[:space:]]Video-Text[[:space:]]Transformers/a239fc53-51e4-48aa-8bdd-b397217aa71a_origin.pdf filter=lfs diff=lfs merge=lfs -text
7508
+ 2023/SadTalker_[[:space:]]Learning[[:space:]]Realistic[[:space:]]3D[[:space:]]Motion[[:space:]]Coefficients[[:space:]]for[[:space:]]Stylized[[:space:]]Audio-Driven[[:space:]]Single[[:space:]]Image[[:space:]]Talking[[:space:]]Face[[:space:]]Animation/aac86acd-650e-49a0-a046-054265e0fe1a_origin.pdf filter=lfs diff=lfs merge=lfs -text
7509
+ 2023/Safe[[:space:]]Latent[[:space:]]Diffusion_[[:space:]]Mitigating[[:space:]]Inappropriate[[:space:]]Degeneration[[:space:]]in[[:space:]]Diffusion[[:space:]]Models/27635077-d485-4fba-8dfd-7e03eb609fe0_origin.pdf filter=lfs diff=lfs merge=lfs -text
7510
+ 2023/Sample-Level[[:space:]]Multi-View[[:space:]]Graph[[:space:]]Clustering/2dbc917c-e6ba-403e-80e3-a2fe15c249a2_origin.pdf filter=lfs diff=lfs merge=lfs -text
7511
+ 2023/Samples[[:space:]]With[[:space:]]Low[[:space:]]Loss[[:space:]]Curvature[[:space:]]Improve[[:space:]]Data[[:space:]]Efficiency/0d4bc9ae-1b91-4517-8e70-147be86f88d2_origin.pdf filter=lfs diff=lfs merge=lfs -text
7512
+ 2023/Sampling[[:space:]]Is[[:space:]]Matter_[[:space:]]Point-Guided[[:space:]]3D[[:space:]]Human[[:space:]]Mesh[[:space:]]Reconstruction/ede83042-4d23-4ef9-a09b-ef8f9683611d_origin.pdf filter=lfs diff=lfs merge=lfs -text
7513
+ 2023/Scalable,[[:space:]]Detailed[[:space:]]and[[:space:]]Mask-Free[[:space:]]Universal[[:space:]]Photometric[[:space:]]Stereo/b4164742-7be0-4231-ae8a-d7fbbefac394_origin.pdf filter=lfs diff=lfs merge=lfs -text
7514
+ 2023/ScaleDet_[[:space:]]A[[:space:]]Scalable[[:space:]]Multi-Dataset[[:space:]]Object[[:space:]]Detector/483d2038-de2f-43cc-b2e7-cfb0359bee59_origin.pdf filter=lfs diff=lfs merge=lfs -text
7515
+ 2023/ScaleFL_[[:space:]]Resource-Adaptive[[:space:]]Federated[[:space:]]Learning[[:space:]]With[[:space:]]Heterogeneous[[:space:]]Clients/cb97df44-ce50-4c4d-9f39-8532b06bcaf1_origin.pdf filter=lfs diff=lfs merge=lfs -text
7516
+ 2023/ScaleKD_[[:space:]]Distilling[[:space:]]Scale-Aware[[:space:]]Knowledge[[:space:]]in[[:space:]]Small[[:space:]]Object[[:space:]]Detector/f4a95fb4-74c7-47ae-867c-7193aca0ac49_origin.pdf filter=lfs diff=lfs merge=lfs -text
7517
+ 2023/Scaling[[:space:]]Language-Image[[:space:]]Pre-Training[[:space:]]via[[:space:]]Masking/72402de8-32e2-4c16-9bec-c8e1554c3fec_origin.pdf filter=lfs diff=lfs merge=lfs -text
7518
+ 2023/Scaling[[:space:]]Up[[:space:]]GANs[[:space:]]for[[:space:]]Text-to-Image[[:space:]]Synthesis/cc6d67d2-1178-4cd7-a0a4-6ace5123194c_origin.pdf filter=lfs diff=lfs merge=lfs -text
7519
+ 2023/ScanDMM_[[:space:]]A[[:space:]]Deep[[:space:]]Markov[[:space:]]Model[[:space:]]of[[:space:]]Scanpath[[:space:]]Prediction[[:space:]]for[[:space:]]360deg[[:space:]]Images/0e9438ec-529d-4504-8952-ad1887816ae6_origin.pdf filter=lfs diff=lfs merge=lfs -text
7520
+ 2023/ScarceNet_[[:space:]]Animal[[:space:]]Pose[[:space:]]Estimation[[:space:]]With[[:space:]]Scarce[[:space:]]Annotations/b512070b-870b-466d-ae4d-037690406953_origin.pdf filter=lfs diff=lfs merge=lfs -text
7521
+ 2023/Scene-Aware[[:space:]]Egocentric[[:space:]]3D[[:space:]]Human[[:space:]]Pose[[:space:]]Estimation/3a95f507-702b-4e70-9e9b-d135557d5128_origin.pdf filter=lfs diff=lfs merge=lfs -text
7522
+ 2023/SceneComposer_[[:space:]]Any-Level[[:space:]]Semantic[[:space:]]Image[[:space:]]Synthesis/a420eecc-24ac-471a-873a-4cbedabb0ab3_origin.pdf filter=lfs diff=lfs merge=lfs -text
7523
+ 2023/SceneTrilogy_[[:space:]]On[[:space:]]Human[[:space:]]Scene-Sketch[[:space:]]and[[:space:]]Its[[:space:]]Complementarity[[:space:]]With[[:space:]]Photo[[:space:]]and[[:space:]]Text/61c06fa2-0cf3-4cfc-a6de-25fbe7140dea_origin.pdf filter=lfs diff=lfs merge=lfs -text
7524
+ 2023/Score[[:space:]]Jacobian[[:space:]]Chaining_[[:space:]]Lifting[[:space:]]Pretrained[[:space:]]2D[[:space:]]Diffusion[[:space:]]Models[[:space:]]for[[:space:]]3D[[:space:]]Generation/6ed9361a-e1bd-4c7d-893b-e63657bb23f8_origin.pdf filter=lfs diff=lfs merge=lfs -text
7525
+ 2023/SeSDF_[[:space:]]Self-Evolved[[:space:]]Signed[[:space:]]Distance[[:space:]]Field[[:space:]]for[[:space:]]Implicit[[:space:]]3D[[:space:]]Clothed[[:space:]]Human[[:space:]]Reconstruction/568571cd-b563-47cc-be1f-30a698863adb_origin.pdf filter=lfs diff=lfs merge=lfs -text
7526
+ 2023/SeaThru-NeRF_[[:space:]]Neural[[:space:]]Radiance[[:space:]]Fields[[:space:]]in[[:space:]]Scattering[[:space:]]Media/769a6e31-fc05-4ad4-bfda-b4fc158792ae_origin.pdf filter=lfs diff=lfs merge=lfs -text
7527
+ 2023/Search-Map-Search_[[:space:]]A[[:space:]]Frame[[:space:]]Selection[[:space:]]Paradigm[[:space:]]for[[:space:]]Action[[:space:]]Recognition/72f3e04e-8ac3-45ed-bc77-477759a59e7c_origin.pdf filter=lfs diff=lfs merge=lfs -text
7528
+ 2023/Seasoning[[:space:]]Model[[:space:]]Soups[[:space:]]for[[:space:]]Robustness[[:space:]]to[[:space:]]Adversarial[[:space:]]and[[:space:]]Natural[[:space:]]Distribution[[:space:]]Shifts/a6551453-1b77-4f6c-8df2-df18da134652_origin.pdf filter=lfs diff=lfs merge=lfs -text
7529
+ 2023/Seeing[[:space:]]Beyond[[:space:]]the[[:space:]]Brain_[[:space:]]Conditional[[:space:]]Diffusion[[:space:]]Model[[:space:]]With[[:space:]]Sparse[[:space:]]Masked[[:space:]]Modeling[[:space:]]for[[:space:]]Vision[[:space:]]Decoding/fe08ae3e-3e6a-4967-96af-a2a0ffcad985_origin.pdf filter=lfs diff=lfs merge=lfs -text
2023/Robust Unsupervised StyleGAN Image Restoration/22cf3860-c0a9-4cc3-ae76-e2aceeacd7e1_content_list.json ADDED
@@ -0,0 +1,1710 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ [
2
+ {
3
+ "type": "text",
4
+ "text": "Robust Unsupervised StyleGAN Image Restoration",
5
+ "text_level": 1,
6
+ "bbox": [
7
+ 240,
8
+ 130,
9
+ 756,
10
+ 152
11
+ ],
12
+ "page_idx": 0
13
+ },
14
+ {
15
+ "type": "text",
16
+ "text": "Yohan Poirier-Ginter\\*, Jean-François Lalonde\\* Inria, Université Côte d'Azur, Université Laval https://lvsn.github.io/RobustUnsupervised/",
17
+ "bbox": [
18
+ 303,
19
+ 180,
20
+ 692,
21
+ 233
22
+ ],
23
+ "page_idx": 0
24
+ },
25
+ {
26
+ "type": "image",
27
+ "img_path": "images/90898fdfe2be1005b020e1012c683e3e8e202c8c33fc2716ec151b598318270b.jpg",
28
+ "image_caption": [
29
+ "Figure 1. Our unsupervised StyleGAN image restoration method is robust and effective across a variety of tasks (left, top to bottom: upsampling, denoising, deartifacting, and inpainting) and a wide range of degradation levels. Since it avoids the need for task-specific hyperparameter tuning, it can directly handle combinations of such degradations, e.g., pairs (center) or even all four (right)."
30
+ ],
31
+ "image_footnote": [],
32
+ "bbox": [
33
+ 114,
34
+ 244,
35
+ 370,
36
+ 445
37
+ ],
38
+ "page_idx": 0
39
+ },
40
+ {
41
+ "type": "image",
42
+ "img_path": "images/8df8152cd13b675b505e8ce0a75faf491efddd25832127ae23f3b6fc55c58681.jpg",
43
+ "image_caption": [],
44
+ "image_footnote": [],
45
+ "bbox": [
46
+ 370,
47
+ 246,
48
+ 622,
49
+ 445
50
+ ],
51
+ "page_idx": 0
52
+ },
53
+ {
54
+ "type": "image",
55
+ "img_path": "images/828ad6f3fb4690e6efcdd78d359a743abc3bb1d0e379118fd8a574d7681cf80c.jpg",
56
+ "image_caption": [],
57
+ "image_footnote": [],
58
+ "bbox": [
59
+ 622,
60
+ 246,
61
+ 880,
62
+ 445
63
+ ],
64
+ "page_idx": 0
65
+ },
66
+ {
67
+ "type": "text",
68
+ "text": "Abstract",
69
+ "text_level": 1,
70
+ "bbox": [
71
+ 258,
72
+ 508,
73
+ 336,
74
+ 523
75
+ ],
76
+ "page_idx": 0
77
+ },
78
+ {
79
+ "type": "text",
80
+ "text": "GAN-based image restoration inverts the generative process to repair images corrupted by known degradations. Existing unsupervised methods must be carefully tuned for each task and degradation level. In this work, we make StyleGAN image restoration robust: a single set of hyperparameters works across a wide range of degradation levels. This makes it possible to handle combinations of several degradations, without the need to retune. Our proposed approach relies on a 3-phase progressive latent space extension and a conservative optimizer, which avoids the need for any additional regularization terms. Extensive experiments demonstrate robustness on inpainting, upsampling, denoising, and deartifacting at varying degradations levels, outperforming other StyleGAN-based inversion techniques. Our approach also favorably compares to diffusion-based restoration by yielding much more realistic inversion results. Code is available at the above URL.",
81
+ "bbox": [
82
+ 111,
83
+ 539,
84
+ 485,
85
+ 811
86
+ ],
87
+ "page_idx": 0
88
+ },
89
+ {
90
+ "type": "text",
91
+ "text": "1. Introduction",
92
+ "text_level": 1,
93
+ "bbox": [
94
+ 112,
95
+ 825,
96
+ 243,
97
+ 842
98
+ ],
99
+ "page_idx": 0
100
+ },
101
+ {
102
+ "type": "text",
103
+ "text": "Image restoration, the task of recovering a high quality image from a degraded input, is a long-standing problem in image processing and computer vision. Since different restoration tasks—such as denoising, upsampling,",
104
+ "bbox": [
105
+ 111,
106
+ 851,
107
+ 483,
108
+ 912
109
+ ],
110
+ "page_idx": 0
111
+ },
112
+ {
113
+ "type": "text",
114
+ "text": "deartifacting, etc.—can be quite distinct, many recent approaches [9, 10, 37, 54, 61, 62] propose to solve them in a supervised learning paradigm by leveraging curated datasets specifically designed for the task at hand. Unfortunately, designing task-specific approaches requires retraining large networks on each task separately.",
115
+ "bbox": [
116
+ 511,
117
+ 508,
118
+ 883,
119
+ 599
120
+ ],
121
+ "page_idx": 0
122
+ },
123
+ {
124
+ "type": "text",
125
+ "text": "In parallel, the advent of powerful generative models has enabled the emergence of unsupervised restoration methods [42], which do not require task-specific training. The idea is to invert the generative process to recover a clean image. Assuming a known (or approximate) degradation model, the optimization procedure therefore attempts to recover an image that both: 1) closely matches the target degraded image after undergoing a similar degradation model (fidelity); and 2) lies in the space of realistic images learned by the GAN (realism).",
126
+ "bbox": [
127
+ 511,
128
+ 604,
129
+ 883,
130
+ 756
131
+ ],
132
+ "page_idx": 0
133
+ },
134
+ {
135
+ "type": "text",
136
+ "text": "In the recent literature, StyleGAN [29-31] has been found to be particularly effective for unsupervised image restoration [13-15, 39, 40] because of the elegant design of its latent space. Indeed, these approaches leverage style inversion techniques [3, 4] to solve for a latent vector that, when given to the generator, creates an image close to the degraded target. Unfortunately, this only works when such a match actually exists in the model distribution, which is rarely the case in practice. Hence, effective methods extend the learned latent space to cre",
137
+ "bbox": [
138
+ 511,
139
+ 761,
140
+ 885,
141
+ 912
142
+ ],
143
+ "page_idx": 0
144
+ },
145
+ {
146
+ "type": "header",
147
+ "text": "CVF",
148
+ "bbox": [
149
+ 106,
150
+ 2,
151
+ 181,
152
+ 42
153
+ ],
154
+ "page_idx": 0
155
+ },
156
+ {
157
+ "type": "header",
158
+ "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.",
159
+ "bbox": [
160
+ 236,
161
+ 0,
162
+ 810,
163
+ 46
164
+ ],
165
+ "page_idx": 0
166
+ },
167
+ {
168
+ "type": "page_number",
169
+ "text": "22292",
170
+ "bbox": [
171
+ 478,
172
+ 935,
173
+ 519,
174
+ 946
175
+ ],
176
+ "page_idx": 0
177
+ },
178
+ {
179
+ "type": "text",
180
+ "text": "ate additional degrees of freedom to admit more images; this creates the need for additional regularization losses. Hyperparameters must therefore carefully be tuned for each specific task and degradation level.",
181
+ "bbox": [
182
+ 111,
183
+ 90,
184
+ 480,
185
+ 151
186
+ ],
187
+ "page_idx": 1
188
+ },
189
+ {
190
+ "type": "text",
191
+ "text": "In this work, we make unsupervised StyleGAN image restoration-by-inversion robust to the type and intensity of degradations. Our proposed method employs the same hyperparameters across all tasks and levels and does not rely on any regularization loss. Our approach leans on two key ideas. First, we rely on a 3-phase progressive latent space extension: we begin by optimizing over the learned (global) latent space, then expand it across individual layers of the generator, and finally expand it further across individual filters—where optimization at each phase is initialized with the result of the previous one. Second, we rely on a conservative, normalized gradient descent (NGD) [58] optimizer which is naturally constrained to stay close to its initial point compared to more sophisticated approaches such as Adam [33]. This combination of prudent optimization over a progressively richer latent space avoids additional regularization terms altogether and keeps the overall procedure simple and constant across all tasks. We evaluate our method on upsampling, inpainting, denoising and deartifacting on a wide range of degradation levels, achieving state-of-the-art results on most scenarios even when baselines are optimized on each independently. We also show that our approach outperforms existing techniques on compositions of these tasks without changing hyperparameters.",
192
+ "bbox": [
193
+ 114,
194
+ 154,
195
+ 486,
196
+ 530
197
+ ],
198
+ "page_idx": 1
199
+ },
200
+ {
201
+ "type": "list",
202
+ "sub_type": "text",
203
+ "list_items": [
204
+ "- We propose a robust 3-phase StyleGAN image restoration framework. Our optimization technique maintains: 1) strong realism when degradations level are high; and 2) high fidelity when they are low. Our method is fully unsupervised, requires no per-task training, and can handle different tasks at different levels without having to adjust hyperparameters.",
205
+ "- We demonstrate the effectiveness of the proposed method under diverse and composed degradations. We develop a benchmark of synthetic image restoration tasks—making their degradation levels easy to control—with care taken to avoid unrealistic assumptions. Our method outperforms existing unsupervised [13,40] and diffusion-based [32] approaches."
206
+ ],
207
+ "bbox": [
208
+ 112,
209
+ 532,
210
+ 483,
211
+ 743
212
+ ],
213
+ "page_idx": 1
214
+ },
215
+ {
216
+ "type": "text",
217
+ "text": "2. Related work",
218
+ "text_level": 1,
219
+ "bbox": [
220
+ 112,
221
+ 762,
222
+ 250,
223
+ 777
224
+ ],
225
+ "page_idx": 1
226
+ },
227
+ {
228
+ "type": "text",
229
+ "text": "This section covers previous work on StyleGAN inversion by optimization and its use for image restoration, and discusses related works on generative priors. Because the literature on general purpose image restoration is so extensive, we do not attempt to review it here and instead refer the reader to recent surveys [18, 19, 35, 57].",
230
+ "bbox": [
231
+ 111,
232
+ 789,
233
+ 480,
234
+ 880
235
+ ],
236
+ "page_idx": 1
237
+ },
238
+ {
239
+ "type": "text",
240
+ "text": "GANs and StyleGAN inversion Generative adversarial networks [20] (GANs) are a popular technique",
241
+ "bbox": [
242
+ 111,
243
+ 881,
244
+ 483,
245
+ 912
246
+ ],
247
+ "page_idx": 1
248
+ },
249
+ {
250
+ "type": "text",
251
+ "text": "for generative modeling. In particular, the StyleGAN family [28-31] learns a highly compressed latent representation of its domain and stands out for its popularity and generation quality. Great progress has been made in StyleGAN inversion: reversing the generation process to infer latent parameters that generate a given image [59].",
252
+ "bbox": [
253
+ 511,
254
+ 90,
255
+ 883,
256
+ 181
257
+ ],
258
+ "page_idx": 1
259
+ },
260
+ {
261
+ "type": "text",
262
+ "text": "Thanks to inversion, StyleGAN has found use in many image editing [23, 43, 51] and restoration [13-15, 39, 40] tasks. Purely optimization-based techniques were studied extensively [3, 4, 26, 44, 46, 66] in the context of image editing. These methods extend the learned latent space of a pretrained StyleGAN model (commonly named $\\mathcal{W}$ ) by adding additional parameters to optimize over it. The most common approach consists of using a different latent code for each layer [3] (dubbed $\\mathcal{W}^+$ ). Going beyond $\\mathcal{W}^+$ has also been explored in [46] who propose to fine-tune generator parameters, and [44] which uses different latent codes for each convolution filter. We build on these techniques by developing an inversion method designed specifically for robust image restoration.",
263
+ "bbox": [
264
+ 511,
265
+ 183,
266
+ 883,
267
+ 393
268
+ ],
269
+ "page_idx": 1
270
+ },
271
+ {
272
+ "type": "text",
273
+ "text": "Generative priors for unsupervised image restoration Since the initial work of Bora et al. [6], multiple papers have used generative priors for unsupervised image restoration [21, 25]. Pan et al. [42] obtained high-resolution results with BigGAN [7]. They relax their latent inversion method by introducing a coarse-to-fine, layer-by-layer generator fine-tuning. Our method differs in that it addresses robustness and compositionality.",
274
+ "bbox": [
275
+ 511,
276
+ 397,
277
+ 883,
278
+ 517
279
+ ],
280
+ "page_idx": 1
281
+ },
282
+ {
283
+ "type": "text",
284
+ "text": "StyleGAN image restoration Inversion methods designed for editing cannot be applied directly for restoration because loss functions are applied to degraded targets rather than clean images. Previous works address this with additional regularization terms and/or optimizer constraints. Of note, PULSE [40] proposed StyleGAN inversion for image upsampling, using a regularizer minimizing the latent extension to $\\mathcal{W}^+$ and a spherical optimization technique on a Gaussian approximation of $\\mathcal{W}$ . While PULSE succeeds in preserving high realism, it obtains low fidelity when the downsampling factor is small. ILO [15] uses similar regularization techniques, and further extends the latent space by optimizing intermediate feature maps in a progressive fashion (layer-by-layer) while constraining the solutions to sparse deviations from the range. More recently, SGILO [14] proposes to replace this sparsity constraint and instead trains a score-based model to learn the distribution of the outputs of an intermediate layer in the StyleGAN generator. BRGM [39] frames GAN inversion as Bayesian inference and estimates the maximum a-posteriori over the input latent vector that generated the reconstructed image given different regularization constraints. Its followup, L-BRGM [13], jointly optimizes the $\\mathcal{Z}$ (before the mapping network) and $\\mathcal{W}$ spaces, further improving quality. In contrast, our work achieves robustness by",
285
+ "bbox": [
286
+ 511,
287
+ 518,
288
+ 883,
289
+ 912
290
+ ],
291
+ "page_idx": 1
292
+ },
293
+ {
294
+ "type": "page_number",
295
+ "text": "22293",
296
+ "bbox": [
297
+ 478,
298
+ 935,
299
+ 517,
300
+ 946
301
+ ],
302
+ "page_idx": 1
303
+ },
304
+ {
305
+ "type": "image",
306
+ "img_path": "images/01269686a92a2f9db1a5fa56938d162ae30829b740759797a4d99eed45c43b57.jpg",
307
+ "image_caption": [
308
+ "Figure 2. Unsupervised image restoration with StyleGAN attempts to restore a degraded target image $\\pmb{y} = f(\\pmb{y}_{\\mathrm{clean}})$ by finding a generated image $\\pmb{x}_{\\mathrm{clean}}$ that matches this target once also degraded $\\pmb{x} = \\hat{f}(\\pmb{x}_{\\mathrm{clean}})$ (sec. 3); here, $\\hat{f}$ is a differentiable approximation to the degradation function $f$ and $\\pmb{y}_{\\mathrm{clean}}$ is the (unknown) ground truth image."
309
+ ],
310
+ "image_footnote": [],
311
+ "bbox": [
312
+ 114,
313
+ 88,
314
+ 482,
315
+ 239
316
+ ],
317
+ "page_idx": 2
318
+ },
319
+ {
320
+ "type": "text",
321
+ "text": "avoiding any such regularization losses.",
322
+ "bbox": [
323
+ 112,
324
+ 366,
325
+ 375,
326
+ 381
327
+ ],
328
+ "page_idx": 2
329
+ },
330
+ {
331
+ "type": "text",
332
+ "text": "Other similar works StyleGAN inversion was used to restore old photographs in [38]. Their method also restores a mixture of degradations, but is designed for real degradations (not robustness) and uses a task-specific encoder, while ours method is fully unsupervised. Multiple supervised methods [22, 34, 36, 56, 64, 65] focus on faces only, whereas our method does not use face-specific loss functions [16], is fully unsupervised, and works across a variety of image domains.",
333
+ "bbox": [
334
+ 111,
335
+ 383,
336
+ 483,
337
+ 518
338
+ ],
339
+ "page_idx": 2
340
+ },
341
+ {
342
+ "type": "text",
343
+ "text": "Diffusion models Denoising diffusion probabilistic models (DDPMs) [53], trained on extremely large datasets (e.g., billions of images for [45, 47, 49]), have recently been shown to outclass GANs when it comes to image generation quality [17]. New approaches have been proposed to tackle image restoration using DDPMs, e.g., Palette [48] obtains excellent results by training the same model for several different tasks. Denoising diffusion restoration models (DDRM) [32] show that pretrained DDPMs can be used for unsupervised restoration tasks, but are limited to linear inverse problems with additive Gaussian noise. In contrast, our method is more flexible since it only requires a differentiable approximation of the degradation function, which can be non-linear.",
344
+ "bbox": [
345
+ 111,
346
+ 521,
347
+ 483,
348
+ 733
349
+ ],
350
+ "page_idx": 2
351
+ },
352
+ {
353
+ "type": "text",
354
+ "text": "3. Image restoration with StyleGAN",
355
+ "text_level": 1,
356
+ "bbox": [
357
+ 112,
358
+ 753,
359
+ 418,
360
+ 771
361
+ ],
362
+ "page_idx": 2
363
+ },
364
+ {
365
+ "type": "text",
366
+ "text": "As illustrated in fig. 2, StyleGAN inversion attempts to recover an image $\\pmb{x}_{\\mathrm{clean}}$ that best matches an (unknown) ground truth image $\\pmb{y}_{\\mathrm{clean}}$ . To this end, we aim to search for a latent code $\\pmb{w} \\in \\mathcal{W}$ such that $\\pmb{x}_{\\mathrm{clean}} = G(\\pmb{w})$ best matches $\\pmb{y}_{\\mathrm{clean}}$ under some image distance function $\\ell$ . The resulting minimization problem,",
367
+ "bbox": [
368
+ 111,
369
+ 782,
370
+ 483,
371
+ 872
372
+ ],
373
+ "page_idx": 2
374
+ },
375
+ {
376
+ "type": "equation",
377
+ "text": "\n$$\n\\boldsymbol {w} = \\underset {\\tilde {\\boldsymbol {w}} \\in \\mathcal {W}} {\\arg \\min } \\ell \\left(G (\\tilde {\\boldsymbol {w}}), \\boldsymbol {y} _ {\\text {c l e a n}}\\right), \\tag {1}\n$$\n",
378
+ "text_format": "latex",
379
+ "bbox": [
380
+ 189,
381
+ 888,
382
+ 482,
383
+ 915
384
+ ],
385
+ "page_idx": 2
386
+ },
387
+ {
388
+ "type": "image",
389
+ "img_path": "images/e861fc957b4dc5c6debbbb7fa15903e15ebd59480c06371acceaff9b65850d54.jpg",
390
+ "image_caption": [
391
+ "Figure 3. Our 3-phase latent extension (sec. 4.1) approach. (a) Phase I uses a global latent code $\\pmb{w}$ , resulting in a prediction $\\pmb{x}$ . (b) Phase II performs layer-wise latent expansion, resulting in matrix $\\pmb{w}^{+}$ and in a prediction $\\pmb{x}^{+}$ . (c) Phase III performs filter-wise latent expansion, resulting in tensor $\\pmb{w}^{++}$ and the final prediction $\\pmb{x}^{++}$ ."
392
+ ],
393
+ "image_footnote": [],
394
+ "bbox": [
395
+ 517,
396
+ 74,
397
+ 880,
398
+ 286
399
+ ],
400
+ "page_idx": 2
401
+ },
402
+ {
403
+ "type": "text",
404
+ "text": "can be solved by gradient descent. In image restoration, the ground truth image $\\pmb{y}_{\\mathrm{clean}}$ is unknown: we are instead given a target image $\\pmb{y} = f(\\pmb{y}_{\\mathrm{clean}})$ , the result of a (non-injective, potentially non-differentiable) degradation function $f$ . Assuming that a differentiable approximation $\\hat{f}\\approx f$ can be constructed, restoration is performed by solving for",
405
+ "bbox": [
406
+ 511,
407
+ 402,
408
+ 883,
409
+ 508
410
+ ],
411
+ "page_idx": 2
412
+ },
413
+ {
414
+ "type": "equation",
415
+ "text": "\n$$\n\\boldsymbol {w} = \\underset {\\tilde {\\boldsymbol {w}} \\in \\mathcal {W}} {\\arg \\min } \\ell \\left(\\hat {f} (G (\\tilde {\\boldsymbol {w}})), \\boldsymbol {y}\\right). \\tag {2}\n$$\n",
416
+ "text_format": "latex",
417
+ "bbox": [
418
+ 594,
419
+ 513,
420
+ 883,
421
+ 541
422
+ ],
423
+ "page_idx": 2
424
+ },
425
+ {
426
+ "type": "text",
427
+ "text": "This can be generalized to compositions of $k$ different. \ndegradations,i.e., $\\{f_i\\}_{i = 1}^k$ ,by solving for",
428
+ "bbox": [
429
+ 511,
430
+ 547,
431
+ 883,
432
+ 578
433
+ ],
434
+ "page_idx": 2
435
+ },
436
+ {
437
+ "type": "equation",
438
+ "text": "\n$$\n\\boldsymbol {w} = \\underset {\\tilde {\\boldsymbol {w}} \\in \\mathcal {W}} {\\arg \\min } \\ell \\left(\\left[ \\hat {f} _ {k} \\circ \\dots \\circ \\hat {f} _ {2} \\circ \\hat {f} _ {1} \\right] (G (\\tilde {\\boldsymbol {w}})), \\boldsymbol {y}\\right), \\tag {3}\n$$\n",
439
+ "text_format": "latex",
440
+ "bbox": [
441
+ 532,
442
+ 584,
443
+ 883,
444
+ 612
445
+ ],
446
+ "page_idx": 2
447
+ },
448
+ {
449
+ "type": "text",
450
+ "text": "where $\\circ$ is the composition operator. Here, it is assumed that each subfunction $f_{i}$ has a differentiable approximation $\\hat{f}_i$ , and that the order of composition is known.",
451
+ "bbox": [
452
+ 511,
453
+ 618,
454
+ 883,
455
+ 662
456
+ ],
457
+ "page_idx": 2
458
+ },
459
+ {
460
+ "type": "text",
461
+ "text": "As described in sec. 2, this naive approach finds solutions that have high realism (look like real faces), but low fidelity (do not match the degraded target). Fidelity is most commonly improved by 1) performing latent extension [3], that is, solving for $\\boldsymbol{w}^{+} \\in \\mathcal{W}^{+}$ which has many more degrees of freedom; and 2) using a better performing optimizer like Adam [33]. These techniques both improve fidelity but also damage realism, motivating the use of regularization losses which must carefully be adjusted for different tasks [13, 39, 40]. In the next section, we show how the choice of latent extension and optimizer avoids the need for these regularizers.",
462
+ "bbox": [
463
+ 511,
464
+ 662,
465
+ 883,
466
+ 844
467
+ ],
468
+ "page_idx": 2
469
+ },
470
+ {
471
+ "type": "text",
472
+ "text": "4. Robust StyleGAN inversion",
473
+ "text_level": 1,
474
+ "bbox": [
475
+ 511,
476
+ 856,
477
+ 769,
478
+ 872
479
+ ],
480
+ "page_idx": 2
481
+ },
482
+ {
483
+ "type": "text",
484
+ "text": "We propose an inversion method designed specifically for robust image restoration, which revisits each aspect",
485
+ "bbox": [
486
+ 511,
487
+ 881,
488
+ 883,
489
+ 912
490
+ ],
491
+ "page_idx": 2
492
+ },
493
+ {
494
+ "type": "page_number",
495
+ "text": "22294",
496
+ "bbox": [
497
+ 478,
498
+ 936,
499
+ 517,
500
+ 948
501
+ ],
502
+ "page_idx": 2
503
+ },
504
+ {
505
+ "type": "text",
506
+ "text": "of the unsupervised StyleGAN optimization pipeline, namely latent extension, optimization, and loss functions.",
507
+ "bbox": [
508
+ 111,
509
+ 90,
510
+ 485,
511
+ 121
512
+ ],
513
+ "page_idx": 3
514
+ },
515
+ {
516
+ "type": "text",
517
+ "text": "4.1. Robust latent extension",
518
+ "text_level": 1,
519
+ "bbox": [
520
+ 112,
521
+ 130,
522
+ 330,
523
+ 145
524
+ ],
525
+ "page_idx": 3
526
+ },
527
+ {
528
+ "type": "text",
529
+ "text": "Inspired by the intuition that initialization is the best regularization [46], we propose a three-phase latent extension, where each phase is initialized by the outcome of the previous one, see fig. 3. Given a pretrained StyleGAN2 [31] model with $N_{L}$ layers in the generator, we denote style vectors $s_i^l \\in \\mathbb{R}^{512}$ at layer $l \\in [1,N_L]$ used to modulate convolution weights $\\pmb{\\theta}^{l} \\in \\mathbb{R}^{512 \\times 512}$ . Assuming $1 \\times 1$ filters to simplify notation<sup>1</sup>, each feature map pixel $p^l \\in \\mathbb{R}^{512}$ is processed by",
530
+ "bbox": [
531
+ 111,
532
+ 152,
533
+ 483,
534
+ 290
535
+ ],
536
+ "page_idx": 3
537
+ },
538
+ {
539
+ "type": "equation",
540
+ "text": "\n$$\np _ {i} ^ {l + 1} = \\left(\\boldsymbol {\\theta} _ {i} ^ {l} \\odot \\boldsymbol {s} _ {i} ^ {l}\\right) \\boldsymbol {p} ^ {l}, \\tag {4}\n$$\n",
541
+ "text_format": "latex",
542
+ "bbox": [
543
+ 225,
544
+ 299,
545
+ 483,
546
+ 319
547
+ ],
548
+ "page_idx": 3
549
+ },
550
+ {
551
+ "type": "text",
552
+ "text": "producing output feature $p_i^{l + 1}\\in \\mathbb{R}$ where $i\\in [0,512]$ .",
553
+ "bbox": [
554
+ 111,
555
+ 328,
556
+ 470,
557
+ 345
558
+ ],
559
+ "page_idx": 3
560
+ },
561
+ {
562
+ "type": "text",
563
+ "text": "Phase I performs global style modulation (fig. 3-(a)), and solves for a single latent vector $\\boldsymbol{w} \\in \\mathcal{W} = \\mathbb{R}^{512}$ shared across all layers as in eq. (3). Prior to optimization, $\\boldsymbol{w}$ is initialized to the mean of $\\mathcal{W}$ over the training set, i.e., $\\mathbb{E}_{\\tilde{\\boldsymbol{w}} \\in \\mathcal{W}}[\\tilde{\\boldsymbol{w}}]$ . Here, the style modulation vector $s_i^l$ in eq. (4) can be written as",
564
+ "bbox": [
565
+ 111,
566
+ 345,
567
+ 483,
568
+ 435
569
+ ],
570
+ "page_idx": 3
571
+ },
572
+ {
573
+ "type": "equation",
574
+ "text": "\n$$\n\\boldsymbol {s} _ {i} ^ {l} = \\mathcal {A} ^ {l} (\\boldsymbol {w}), \\tag {5}\n$$\n",
575
+ "text_format": "latex",
576
+ "bbox": [
577
+ 250,
578
+ 446,
579
+ 480,
580
+ 464
581
+ ],
582
+ "page_idx": 3
583
+ },
584
+ {
585
+ "type": "text",
586
+ "text": "where $\\mathcal{A}^l$ is the corresponding affine projection layer (which multiplies by a weight matrix and adds a bias).",
587
+ "bbox": [
588
+ 111,
589
+ 474,
590
+ 482,
591
+ 505
592
+ ],
593
+ "page_idx": 3
594
+ },
595
+ {
596
+ "type": "text",
597
+ "text": "Phase II performs layer-wise latent extension [3] (fig. 3-(b)), and solves for a latent matrix $\\boldsymbol{w}^{+} \\in \\mathcal{W}^{+} = \\mathbb{R}^{N_{L} \\times 512}$ . Each row of $\\boldsymbol{w}^{+}$ is initialized to $\\boldsymbol{w}$ . Style modulation for this phase becomes",
598
+ "bbox": [
599
+ 112,
600
+ 506,
601
+ 483,
602
+ 566
603
+ ],
604
+ "page_idx": 3
605
+ },
606
+ {
607
+ "type": "equation",
608
+ "text": "\n$$\n\\boldsymbol {s} _ {i} ^ {l} = \\mathcal {A} ^ {l} \\left(\\boldsymbol {w} _ {l} ^ {+}\\right). \\tag {6}\n$$\n",
609
+ "text_format": "latex",
610
+ "bbox": [
611
+ 245,
612
+ 577,
613
+ 480,
614
+ 595
615
+ ],
616
+ "page_idx": 3
617
+ },
618
+ {
619
+ "type": "text",
620
+ "text": "Phase III performs filter-wise latent extension (fig. 3-c), and solves for a latent tensor $\\boldsymbol{w}^{++} \\in \\mathcal{W}^{++} = \\mathbb{R}^{N_F \\times N_L \\times 512}$ , where a different latent code is used for each convolution filter and where $N_F$ is the number of such filters. Each submatrix of $\\boldsymbol{w}^{++}$ is initialized to $\\boldsymbol{w}^+$ . Style modulation for this phase becomes",
621
+ "bbox": [
622
+ 111,
623
+ 606,
624
+ 483,
625
+ 696
626
+ ],
627
+ "page_idx": 3
628
+ },
629
+ {
630
+ "type": "equation",
631
+ "text": "\n$$\n\\boldsymbol {s} _ {i} ^ {l} = \\mathcal {A} ^ {l} \\left(\\boldsymbol {w} _ {i, l} ^ {+ +}\\right). \\tag {7}\n$$\n",
632
+ "text_format": "latex",
633
+ "bbox": [
634
+ 240,
635
+ 707,
636
+ 480,
637
+ 726
638
+ ],
639
+ "page_idx": 3
640
+ },
641
+ {
642
+ "type": "text",
643
+ "text": "4.2. Robust optimization",
644
+ "text_level": 1,
645
+ "bbox": [
646
+ 112,
647
+ 747,
648
+ 305,
649
+ 763
650
+ ],
651
+ "page_idx": 3
652
+ },
653
+ {
654
+ "type": "text",
655
+ "text": "While most inversion approaches use the Adam [33] optimizer to solve eq. (3), we instead propose coupling our 3-phase latent extension with a weaker optimizer, normalized gradient decent (NGD) [58], a simple variant of SGD that normalizes the gradient before each step:",
656
+ "bbox": [
657
+ 111,
658
+ 768,
659
+ 483,
660
+ 845
661
+ ],
662
+ "page_idx": 3
663
+ },
664
+ {
665
+ "type": "equation",
666
+ "text": "\n$$\n\\bar {\\nabla} _ {\\boldsymbol {w}} \\ell (\\boldsymbol {w}) = \\frac {\\nabla_ {\\boldsymbol {w}} \\ell (\\boldsymbol {w})}{\\| \\nabla_ {\\boldsymbol {w}} \\ell (\\boldsymbol {w}) \\| _ {2}}, \\tag {8}\n$$\n",
667
+ "text_format": "latex",
668
+ "bbox": [
669
+ 205,
670
+ 856,
671
+ 480,
672
+ 888
673
+ ],
674
+ "page_idx": 3
675
+ },
676
+ {
677
+ "type": "code",
678
+ "sub_type": "algorithm",
679
+ "code_caption": [
680
+ "Algorithm 1: Robust StyleGAN inversion."
681
+ ],
682
+ "code_body": "Output: restored image $x^{++}$ # Phase I \n1 $\\pmb {w} = \\mathbb{E}_{\\tilde{\\pmb{w}}\\in \\mathcal{W}}[\\tilde{\\pmb{w}} ]$ \n2 for 1 to 150 do \n3 $\\begin{array}{r}\\pmb {x}\\gets G(\\pmb {w})\\\\ \\pmb {w}\\gets 0.08\\bar{\\nabla}_{\\pmb{w}}\\ell (\\hat{f} (\\pmb {x}),\\pmb {y}) \\end{array}$ \n4 Phase II \n5 $w^{+} =$ repeat $(\\pmb {w},N_L)$ \n6 for 1 to 150 do \n7 $\\begin{array}{r}\\pmb{x}^{+}\\gets G^{+}(\\pmb{w}^{+}) \\end{array}$ \n8 $\\begin{array}{r}\\pmb {w}^{+}\\gets 0.02\\bar{\\nabla}_{\\pmb{w}}\\ell (\\hat{f} (\\pmb{x}^{+}),\\pmb {y}) \\end{array}$ # Phase III \n9 $w^{+ + } =$ repeat $(\\pmb {w}^{+},N_{F})$ \n10 for 1 to 150 do \n11 $\\begin{array}{r}\\pmb{x}^{+ + }\\gets G^{+ + }(\\pmb{w}^{+ + })\\\\ \\pmb{w}^{+ + }\\gets 0.005\\bar{\\nabla}_{\\pmb{w}^{+ + }}\\ell (\\hat{f} (\\pmb{x}^{+ + }),\\pmb {y}) \\end{array}$ \n12 \n13 return $x^{+ + }$",
683
+ "bbox": [
684
+ 516,
685
+ 111,
686
+ 781,
687
+ 372
688
+ ],
689
+ "page_idx": 3
690
+ },
691
+ {
692
+ "type": "text",
693
+ "text": "Here, the gradient $\\bar{\\nabla}_{\\boldsymbol{w}}\\ell (\\boldsymbol {w})$ is explicitly set to 0 when $||\\nabla_{\\boldsymbol{w}}\\ell (\\boldsymbol {w})||_2 = 0$ . Following latent extension, we normalize each latent code separately (i.e. each row of $\\boldsymbol{w}^{+}$ and $\\boldsymbol{w}^{+ + }$ ). NGD conserves loss scale invariance, a key property of Adam that avoids learning rate adjustments following loss function changes.",
694
+ "bbox": [
695
+ 511,
696
+ 402,
697
+ 883,
698
+ 494
699
+ ],
700
+ "page_idx": 3
701
+ },
702
+ {
703
+ "type": "text",
704
+ "text": "4.3. Robust loss function",
705
+ "text_level": 1,
706
+ "bbox": [
707
+ 511,
708
+ 503,
709
+ 705,
710
+ 517
711
+ ],
712
+ "page_idx": 3
713
+ },
714
+ {
715
+ "type": "text",
716
+ "text": "The staple loss function in StyleGAN inversion is the LPIPS [63] perceptual loss combined with a L2 or L1 [4] pixel-wise loss. We found a multiresolution loss function to be much more robust, and use:",
717
+ "bbox": [
718
+ 511,
719
+ 526,
720
+ 883,
721
+ 585
722
+ ],
723
+ "page_idx": 3
724
+ },
725
+ {
726
+ "type": "equation",
727
+ "text": "\n$$\n\\ell_ {\\mathrm {M R}} (\\boldsymbol {x}, \\boldsymbol {y}) = \\sum_ {i = 1} ^ {k} \\ell_ {\\mathrm {L P I P S}} \\left(\\phi \\left(\\boldsymbol {x}, 2 ^ {i}\\right), \\phi \\left(\\boldsymbol {y}, 2 ^ {i}\\right)\\right), \\tag {9}\n$$\n",
728
+ "text_format": "latex",
729
+ "bbox": [
730
+ 537,
731
+ 594,
732
+ 882,
733
+ 635
734
+ ],
735
+ "page_idx": 3
736
+ },
737
+ {
738
+ "type": "text",
739
+ "text": "where $\\phi (\\cdot ,2^i)$ downsamples by a factor of $2^{i}$ using average pooling, and we set $k = 6$ for image resolution of $1024^{2}$ . All resolutions are weighted equally, giving our final loss function $\\ell = \\lambda_{\\mathrm{L1}}\\ell_{\\mathrm{L1}} + \\ell_{\\mathrm{MR}}$ with $\\lambda_{\\mathrm{L1}} = 0.1$ .",
740
+ "bbox": [
741
+ 511,
742
+ 645,
743
+ 883,
744
+ 707
745
+ ],
746
+ "page_idx": 3
747
+ },
748
+ {
749
+ "type": "text",
750
+ "text": "4.4. Overall algorithm",
751
+ "text_level": 1,
752
+ "bbox": [
753
+ 511,
754
+ 714,
755
+ 689,
756
+ 729
757
+ ],
758
+ "page_idx": 3
759
+ },
760
+ {
761
+ "type": "text",
762
+ "text": "Alg. 1 provides a detailed pseudo-code of our method. $G^{+}$ and $G^{++}$ denote the synthesis network $G$ after modification to accept $w^{+} \\in \\mathcal{W}^{+}$ and $w^{++} \\in \\mathcal{W}^{++}$ , respectively. Hyperparameters like learning rates and number of steps are explicitly provided since they are held constant across all tasks. Their values were found by cross-validation on a subset of the FFHQ training set.",
763
+ "bbox": [
764
+ 511,
765
+ 738,
766
+ 883,
767
+ 844
768
+ ],
769
+ "page_idx": 3
770
+ },
771
+ {
772
+ "type": "text",
773
+ "text": "5. Benchmarking restoration robustness",
774
+ "text_level": 1,
775
+ "bbox": [
776
+ 511,
777
+ 856,
778
+ 852,
779
+ 872
780
+ ],
781
+ "page_idx": 3
782
+ },
783
+ {
784
+ "type": "text",
785
+ "text": "This section first describes the proposed degradation models used in all experiments as well as their respec",
786
+ "bbox": [
787
+ 511,
788
+ 881,
789
+ 883,
790
+ 912
791
+ ],
792
+ "page_idx": 3
793
+ },
794
+ {
795
+ "type": "page_footnote",
796
+ "text": "In the $3 \\times 3$ case, operations are repeated spatially and summed.",
797
+ "bbox": [
798
+ 132,
799
+ 898,
800
+ 478,
801
+ 911
802
+ ],
803
+ "page_idx": 3
804
+ },
805
+ {
806
+ "type": "page_number",
807
+ "text": "22295",
808
+ "bbox": [
809
+ 478,
810
+ 935,
811
+ 517,
812
+ 946
813
+ ],
814
+ "page_idx": 3
815
+ },
816
+ {
817
+ "type": "text",
818
+ "text": "tive differentiable approximations. It then explains how compositions of tasks were formed.",
819
+ "bbox": [
820
+ 112,
821
+ 90,
822
+ 482,
823
+ 121
824
+ ],
825
+ "page_idx": 4
826
+ },
827
+ {
828
+ "type": "text",
829
+ "text": "5.1. Individual degradations",
830
+ "text_level": 1,
831
+ "bbox": [
832
+ 112,
833
+ 128,
834
+ 333,
835
+ 145
836
+ ],
837
+ "page_idx": 4
838
+ },
839
+ {
840
+ "type": "text",
841
+ "text": "Experiments are performed on four common sources of image degradation: upsampling, inpainting, denoising, and deartifacting. Synthetic models are used to facilitate comparison at different degradation levels. Each degradation is tested at five levels, which are referred to as extra-small (XS), small (S), medium (M), large (L), and extra-large (XL). Parameters for each degradation level will be provided below for each task in that same order. Fig. 4 shows examples of all four degradations at the XS and XL levels, see supp.",
842
+ "bbox": [
843
+ 111,
844
+ 152,
845
+ 483,
846
+ 303
847
+ ],
848
+ "page_idx": 4
849
+ },
850
+ {
851
+ "type": "text",
852
+ "text": "Upsampling targets $\\pmb{y}$ are produced by downsampling ground truth images $\\pmb{y}_{\\mathrm{clean}}$ by integer factors $k_{\\mathrm{down}} \\in \\{2,4,8,16,32\\}$ . Downsampling filters are uniformly sampled from the commonly used bilinear, bicubic, and Lanczos filters, which provides a coarse but broad range of aliasing profiles. During inversion, average pooling is used as the approximation $\\hat{f}$ in all cases.",
853
+ "bbox": [
854
+ 111,
855
+ 303,
856
+ 483,
857
+ 409
858
+ ],
859
+ "page_idx": 4
860
+ },
861
+ {
862
+ "type": "text",
863
+ "text": "Inpainting aims to predict missing regions in an image. To test a variety of masking conditions, random masks are generated by drawing $k_{\\mathrm{strokes}} \\in \\{1,5,9,13,17\\}$ random strokes of width $0.08r$ where $r$ is the image resolution, each stroke connecting two random points situated in the outer thirds of the image. In this task, $\\hat{f} = f$ i.e. we assume that the mask is known. To avoid the bias introduced by black pixels when evaluating the LPIPS loss, identical noise of distribution $\\mathcal{N}(0.5,1)$ is added to the masked region of both the prediction and the target before computing LPIPS.",
864
+ "bbox": [
865
+ 111,
866
+ 410,
867
+ 483,
868
+ 575
869
+ ],
870
+ "page_idx": 4
871
+ },
872
+ {
873
+ "type": "text",
874
+ "text": "Denoising targets are generated by using a mixture of Poisson and Bernoulli noise, simulating common sources of noise in cameras, namely shot noise and dead (or hot) pixels respectively. This is a challenging scenario because unlike Gaussian noise, Poisson noise is non-additive and signal-dependent, while Bernoulli noise is biased. Additionally, both are non-differentiable. We used the parameters $k_{\\mathrm{p}} \\in \\{96, 48, 24, 12, 6\\}$ and $k_{\\mathrm{b}} \\in \\{0.04, 0.08, 0.16, 0.32, 0.64\\}$ where $k_{\\mathrm{p}}$ gives the most likely value added to a pixel (independently for each channel) according to a Poisson distribution, and $k_{\\mathrm{b}}$ the probability that all channels of a pixel are replaced with black. We also account for clamping during image serialization, resulting in the overall noise model:",
875
+ "bbox": [
876
+ 111,
877
+ 575,
878
+ 483,
879
+ 787
880
+ ],
881
+ "page_idx": 4
882
+ },
883
+ {
884
+ "type": "equation",
885
+ "text": "\n$$\n\\frac {\\operatorname {c l a m p} \\left(p ^ {\\prime} \\cdot m\\right)}{2 5 5}, \\text {w h e r e} p ^ {\\prime} \\sim \\mathcal {P} \\left(k _ {\\mathrm {p}} p\\right), m \\sim \\mathcal {B} \\left(k _ {\\mathrm {b}}\\right), \\tag {10}\n$$\n",
886
+ "text_format": "latex",
887
+ "bbox": [
888
+ 133,
889
+ 794,
890
+ 482,
891
+ 835
892
+ ],
893
+ "page_idx": 4
894
+ },
895
+ {
896
+ "type": "text",
897
+ "text": "for a ground truth image pixel value $p \\in [0.0, 1.0]$ and where $\\mathrm{clamp}(\\cdot)$ saturates all values outside of [0, 255]. Note that $p'$ can only take discrete values. For the differentiable approximation $\\hat{f}$ , we replace the (non-differentiable, discrete) Poisson noise with a Gaussian ap",
898
+ "bbox": [
899
+ 111,
900
+ 835,
901
+ 483,
902
+ 912
903
+ ],
904
+ "page_idx": 4
905
+ },
906
+ {
907
+ "type": "text",
908
+ "text": "proximation $\\mathcal{N}(k_{\\mathrm{p}}p - 0.5,\\sqrt{k_{\\mathrm{p}}p})$ and treat the Bernoulli noise as an unknown mask. We also use a surrogate gradient for clamp:",
909
+ "bbox": [
910
+ 511,
911
+ 90,
912
+ 883,
913
+ 136
914
+ ],
915
+ "page_idx": 4
916
+ },
917
+ {
918
+ "type": "equation",
919
+ "text": "\n$$\n\\frac {d}{d p} \\operatorname {c l a m p} (p) \\approx \\frac {d}{d p} \\operatorname {s i g m o i d} (2 \\cdot (p - 0. 5)). \\tag {11}\n$$\n",
920
+ "text_format": "latex",
921
+ "bbox": [
922
+ 542,
923
+ 146,
924
+ 882,
925
+ 179
926
+ ],
927
+ "page_idx": 4
928
+ },
929
+ {
930
+ "type": "text",
931
+ "text": "Deartifacting is performed on JPEG images compressed with libjpeg [2, 12] at quality levels $k_{\\mathrm{jpeg}} \\in \\{18, 15, 12, 9, 6\\}$ . The lossy part of JPEG compression can be expressed as",
932
+ "bbox": [
933
+ 511,
934
+ 186,
935
+ 883,
936
+ 247
937
+ ],
938
+ "page_idx": 4
939
+ },
940
+ {
941
+ "type": "equation",
942
+ "text": "\n$$\n\\text {q u a n t i z e} \\circ \\mathrm {D C T} \\circ \\text {t o} _ {\\text {b l o c k s}} \\circ \\text {c h r o m a} _ {\\text {s u b s a m p l e}}. \\tag {12}\n$$\n",
943
+ "text_format": "latex",
944
+ "bbox": [
945
+ 521,
946
+ 258,
947
+ 882,
948
+ 289
949
+ ],
950
+ "page_idx": 4
951
+ },
952
+ {
953
+ "type": "text",
954
+ "text": "Here, 4:2:2 chroma subsampling is used. This final quantization step rounds each DCT component $c_{i,j}^{(k)}$ positioned in the $k^{\\text{th}}$ block at indices $i,j$ with",
955
+ "bbox": [
956
+ 511,
957
+ 290,
958
+ 883,
959
+ 339
960
+ ],
961
+ "page_idx": 4
962
+ },
963
+ {
964
+ "type": "equation",
965
+ "text": "\n$$\n\\operatorname {q u a n t i z e} \\left(c _ {i, j} ^ {(k)}\\right) = \\left\\lfloor c _ {i, j} ^ {(k)} / Q _ {i, j} \\right], \\tag {13}\n$$\n",
966
+ "text_format": "latex",
967
+ "bbox": [
968
+ 593,
969
+ 349,
970
+ 882,
971
+ 371
972
+ ],
973
+ "page_idx": 4
974
+ },
975
+ {
976
+ "type": "text",
977
+ "text": "where $Q_{i,j} \\in [1,255] \\subset \\mathbb{N}$ is the corresponding value from the quantization table. Rounding is the only non-differentiable part of this chain (the lossless part of JPEG compression does not need to be differentiated). We read the quantization tables from file headers and use Lomnitz's implementation [1] of [52], which proposes differentiable rounding:",
978
+ "bbox": [
979
+ 511,
980
+ 378,
981
+ 883,
982
+ 486
983
+ ],
984
+ "page_idx": 4
985
+ },
986
+ {
987
+ "type": "equation",
988
+ "text": "\n$$\n\\lfloor p \\rceil \\approx \\lfloor p \\rceil + (p - \\lfloor p \\rceil) ^ {3}. \\tag {14}\n$$\n",
989
+ "text_format": "latex",
990
+ "bbox": [
991
+ 612,
992
+ 496,
993
+ 882,
994
+ 513
995
+ ],
996
+ "page_idx": 4
997
+ },
998
+ {
999
+ "type": "text",
1000
+ "text": "We use this approximation to define a surrogate gradient (keeping the forward pass exact), and interpolate with a straight-through gradient estimator at $\\alpha = 0.8$ :",
1001
+ "bbox": [
1002
+ 511,
1003
+ 525,
1004
+ 883,
1005
+ 570
1006
+ ],
1007
+ "page_idx": 4
1008
+ },
1009
+ {
1010
+ "type": "equation",
1011
+ "text": "\n$$\n\\frac {d}{d p} \\lfloor p \\rceil \\approx \\frac {d}{d p} \\left[ (1 - \\alpha) p + \\alpha (p - \\lfloor p \\rceil) ^ {3} \\right]. \\tag {15}\n$$\n",
1012
+ "text_format": "latex",
1013
+ "bbox": [
1014
+ 547,
1015
+ 580,
1016
+ 880,
1017
+ 612
1018
+ ],
1019
+ "page_idx": 4
1020
+ },
1021
+ {
1022
+ "type": "text",
1023
+ "text": "5.2. Composed degradations",
1024
+ "text_level": 1,
1025
+ "bbox": [
1026
+ 513,
1027
+ 622,
1028
+ 735,
1029
+ 638
1030
+ ],
1031
+ "page_idx": 4
1032
+ },
1033
+ {
1034
+ "type": "text",
1035
+ "text": "Eq. (3) assumes knowledge of the order in which degradations are applied. In this work, the following restoration order is used:",
1036
+ "bbox": [
1037
+ 511,
1038
+ 645,
1039
+ 883,
1040
+ 689
1041
+ ],
1042
+ "page_idx": 4
1043
+ },
1044
+ {
1045
+ "type": "equation",
1046
+ "text": "\n$$\n\\text {i n p a i n t} \\circ \\text {d e a r i f a c t} \\circ \\text {d e n o i s e} \\circ \\text {u p s a m p l e}. \\tag {16}\n$$\n",
1047
+ "text_format": "latex",
1048
+ "bbox": [
1049
+ 535,
1050
+ 702,
1051
+ 880,
1052
+ 718
1053
+ ],
1054
+ "page_idx": 4
1055
+ },
1056
+ {
1057
+ "type": "text",
1058
+ "text": "Task compositions are created with (subsequences of) this ordering. For example, upsampling and deartifacting form a composition of length 2. All compositions were formed with tasks at degradation level medium (M).",
1059
+ "bbox": [
1060
+ 511,
1061
+ 729,
1062
+ 883,
1063
+ 790
1064
+ ],
1065
+ "page_idx": 4
1066
+ },
1067
+ {
1068
+ "type": "text",
1069
+ "text": "6. Comparison to StyleGAN-based methods",
1070
+ "text_level": 1,
1071
+ "bbox": [
1072
+ 511,
1073
+ 803,
1074
+ 880,
1075
+ 820
1076
+ ],
1077
+ "page_idx": 4
1078
+ },
1079
+ {
1080
+ "type": "text",
1081
+ "text": "6.1. Models, datasets, and metrics",
1082
+ "text_level": 1,
1083
+ "bbox": [
1084
+ 511,
1085
+ 828,
1086
+ 774,
1087
+ 843
1088
+ ],
1089
+ "page_idx": 4
1090
+ },
1091
+ {
1092
+ "type": "text",
1093
+ "text": "The PyTorch implementation [41] of StyleGAN2-ADA [28] is used in all experiments. In particular, we use the model pretrained on the FFHQ dataset [30] at $1024 \\times 1024$ resolution. Since the model is trained on",
1094
+ "bbox": [
1095
+ 511,
1096
+ 849,
1097
+ 883,
1098
+ 910
1099
+ ],
1100
+ "page_idx": 4
1101
+ },
1102
+ {
1103
+ "type": "page_number",
1104
+ "text": "22296",
1105
+ "bbox": [
1106
+ 478,
1107
+ 936,
1108
+ 517,
1109
+ 948
1110
+ ],
1111
+ "page_idx": 4
1112
+ },
1113
+ {
1114
+ "type": "image",
1115
+ "img_path": "images/8fc68f163d70564b3fc207f5a33bdf4e40fef58e95f288944860b7ff476552e3.jpg",
1116
+ "image_caption": [
1117
+ "Figure 4. Randomly selected qualitative restoration results on each independent task, for both extra-small and extra-large levels of degradation. Top to bottom: (degraded) target, ours, L-BRGM [13], and PULSE [40]. Left to right: upsampling, inpainting, denoising, and deartifacting. Our method shows better robustness to variations in levels of degradation. See supp. for more results."
1118
+ ],
1119
+ "image_footnote": [],
1120
+ "bbox": [
1121
+ 114,
1122
+ 88,
1123
+ 320,
1124
+ 391
1125
+ ],
1126
+ "page_idx": 5
1127
+ },
1128
+ {
1129
+ "type": "image",
1130
+ "img_path": "images/3a025d545e7280eb5becaf2be500242b31cd4fcfcd71cbf8e6b3cfe88ca87fbb.jpg",
1131
+ "image_caption": [],
1132
+ "image_footnote": [],
1133
+ "bbox": [
1134
+ 321,
1135
+ 88,
1136
+ 506,
1137
+ 390
1138
+ ],
1139
+ "page_idx": 5
1140
+ },
1141
+ {
1142
+ "type": "image",
1143
+ "img_path": "images/7eaa2780704093d6a469d855c3522662d736056400b15c0844f512de6e9ec19a.jpg",
1144
+ "image_caption": [],
1145
+ "image_footnote": [],
1146
+ "bbox": [
1147
+ 508,
1148
+ 88,
1149
+ 694,
1150
+ 390
1151
+ ],
1152
+ "page_idx": 5
1153
+ },
1154
+ {
1155
+ "type": "image",
1156
+ "img_path": "images/da7ae62f29d1cacfb9dd359a4aba43e4c8a300f78331604c4a13b2c13ccea84b.jpg",
1157
+ "image_caption": [],
1158
+ "image_footnote": [],
1159
+ "bbox": [
1160
+ 696,
1161
+ 89,
1162
+ 882,
1163
+ 390
1164
+ ],
1165
+ "page_idx": 5
1166
+ },
1167
+ {
1168
+ "type": "text",
1169
+ "text": "the entire FFHQ dataset, we gathered an additional 100 test images, dubbed “FFHQ extras” (FFHQ-X), using the same alignment script. Overlap with FFHQ was avoided by filtering by date (see supp.).",
1170
+ "bbox": [
1171
+ 109,
1172
+ 469,
1173
+ 482,
1174
+ 530
1175
+ ],
1176
+ "page_idx": 5
1177
+ },
1178
+ {
1179
+ "type": "text",
1180
+ "text": "Performance was measured via accuracy, realism, and fidelity metrics. Accuracy was measured with the LPIPS [63] between the predictions and the ground truth images. Realism was measured using the PatchFID [8] (pFID): FID [24] measured on image patches. More specifically, we extracted 1000 random crops of size $128 \\times 128$ per image, computing the FID of all crops from predictions to all crops from our FFHQ-X dataset. Finally, fidelity between degraded predictions and targets was measured using the LPIPS; this was done by degrading predictions with the true degradation to ignore the effect of differentiable approximations.",
1181
+ "bbox": [
1182
+ 109,
1183
+ 531,
1184
+ 482,
1185
+ 712
1186
+ ],
1187
+ "page_idx": 5
1188
+ },
1189
+ {
1190
+ "type": "text",
1191
+ "text": "6.2. Baselines",
1192
+ "text_level": 1,
1193
+ "bbox": [
1194
+ 112,
1195
+ 722,
1196
+ 220,
1197
+ 736
1198
+ ],
1199
+ "page_idx": 5
1200
+ },
1201
+ {
1202
+ "type": "text",
1203
+ "text": "Our method is compared on FFHQ-X against two unsupervised StyleGAN image restoration techniques, PULSE [40] and the current state of the art L-BRGM [13]. The same pretrained network is used for all approaches.",
1204
+ "bbox": [
1205
+ 109,
1206
+ 744,
1207
+ 482,
1208
+ 805
1209
+ ],
1210
+ "page_idx": 5
1211
+ },
1212
+ {
1213
+ "type": "text",
1214
+ "text": "PULSE [40] We adapted PULSE to use StyleGAN2-ADA and to tasks other than upsampling by increasing the number of steps from 100 to 250 and adjusting the learning rate to 0.8 (manual inspection on validation data) to increase its robustness to novel tasks.",
1215
+ "bbox": [
1216
+ 109,
1217
+ 805,
1218
+ 482,
1219
+ 880
1220
+ ],
1221
+ "page_idx": 5
1222
+ },
1223
+ {
1224
+ "type": "text",
1225
+ "text": "L-BRGM [13] We adjust its learning rate $\\lambda_{\\mathrm{lr}} = 0.05$ by manual inspection. L-BRGM uses best-of- $k$ random",
1226
+ "bbox": [
1227
+ 111,
1228
+ 881,
1229
+ 482,
1230
+ 912
1231
+ ],
1232
+ "page_idx": 5
1233
+ },
1234
+ {
1235
+ "type": "text",
1236
+ "text": "initialization; because initialization is difficult under high degradations, we doubled the number of samples from 100 to 200 and added a 0.8 truncation [30] to avoid outliers. L-BRGM also performs early stopping by inspecting the ground truth; for a fair comparison, we instead fixed the number of steps to 500 as in BRGM [39].",
1237
+ "bbox": [
1238
+ 511,
1239
+ 469,
1240
+ 883,
1241
+ 560
1242
+ ],
1243
+ "page_idx": 5
1244
+ },
1245
+ {
1246
+ "type": "text",
1247
+ "text": "Hyperparameter optimization for baselines We optimize the accuracy metric for the baselines hyperparameters for each (task, level) pair individually directly on the $FFHQ-X$ test set (see supp. for more details). In addition, baselines employ the ground truth Bernoulli mask (see sec. 5.1) in their differentiable approximation of eq. (10). Doing so was necessary for good performance.",
1248
+ "bbox": [
1249
+ 511,
1250
+ 561,
1251
+ 883,
1252
+ 666
1253
+ ],
1254
+ "page_idx": 5
1255
+ },
1256
+ {
1257
+ "type": "text",
1258
+ "text": "Our method We adjust hyperameters (learning rate and number of iterations, see alg. 1) once on validation data, then run our benchmark on the test set with no pertask or per-level hyperparameter adjustment: our method uses the same hyperparameters across all tasks and levels. Our method is not provided the Bernoulli mask. The total number of steps (450) is less than the 500 used for L-BRGM.",
1259
+ "bbox": [
1260
+ 511,
1261
+ 667,
1262
+ 883,
1263
+ 787
1264
+ ],
1265
+ "page_idx": 5
1266
+ },
1267
+ {
1268
+ "type": "text",
1269
+ "text": "6.3. Experimental results",
1270
+ "text_level": 1,
1271
+ "bbox": [
1272
+ 511,
1273
+ 797,
1274
+ 709,
1275
+ 813
1276
+ ],
1277
+ "page_idx": 5
1278
+ },
1279
+ {
1280
+ "type": "text",
1281
+ "text": "Robustness to degradation levels Our benchmark is run on our FFQH-X test set (sec. 6.1) on all five levels of degradation for each individual task, and quantitative results are presented in tab. 1. Recall from sec. 6.2 that our method, which employs a single set of hyperparameters across all experiments, is pitted against baselines",
1282
+ "bbox": [
1283
+ 511,
1284
+ 821,
1285
+ 883,
1286
+ 912
1287
+ ],
1288
+ "page_idx": 5
1289
+ },
1290
+ {
1291
+ "type": "page_number",
1292
+ "text": "22297",
1293
+ "bbox": [
1294
+ 478,
1295
+ 936,
1296
+ 517,
1297
+ 948
1298
+ ],
1299
+ "page_idx": 5
1300
+ },
1301
+ {
1302
+ "type": "table",
1303
+ "img_path": "images/2e22e53a6aeae4afc81548707d13e543292d81e3336ad937d363e127a4afff59.jpg",
1304
+ "table_caption": [],
1305
+ "table_footnote": [],
1306
+ "table_body": "<table><tr><td></td><td colspan=\"3\">Accur. (LPIPS) ↓</td><td colspan=\"3\">Fidelity (LPIPS) ↓</td><td colspan=\"3\">Realism (pFID) ↓</td></tr><tr><td></td><td colspan=\"3\">PULS L-BRG Ours</td><td colspan=\"3\">PULS L-BRG Ours</td><td colspan=\"3\">PULS L-BRG Ours</td></tr><tr><td colspan=\"10\">Upsampling (bilinear, bicubic or Lanczos)</td></tr><tr><td>XS</td><td>.493</td><td>.407</td><td>.414</td><td>.432</td><td>.295</td><td>.313</td><td>44.5</td><td>23.6</td><td>17.0</td></tr><tr><td>S</td><td>.492</td><td>.412</td><td>.449</td><td>.353</td><td>.140</td><td>.239</td><td>34.3</td><td>25.5</td><td>22.0</td></tr><tr><td>M</td><td>.495</td><td>.458</td><td>.472</td><td>.261</td><td>.124</td><td>.172</td><td>29.3</td><td>35.4</td><td>22.3</td></tr><tr><td>L</td><td>.501</td><td>.487</td><td>.490</td><td>.185</td><td>.129</td><td>.127</td><td>21.9</td><td>26.0</td><td>20.9</td></tr><tr><td>XL</td><td>.512</td><td>.506</td><td>.514</td><td>.083</td><td>.095</td><td>.090</td><td>24.9</td><td>21.3</td><td>21.3</td></tr><tr><td colspan=\"10\">Denoising (clamped Poisson and Bernoulli mixture)</td></tr><tr><td>XS</td><td>.501</td><td>.440</td><td>.425</td><td>.275</td><td>.152</td><td>.156</td><td>56.1</td><td>27.2</td><td>18.5</td></tr><tr><td>S</td><td>.499</td><td>.450</td><td>.434</td><td>.252</td><td>.138</td><td>.140</td><td>53.7</td><td>28.6</td><td>19.1</td></tr><tr><td>M</td><td>.500</td><td>.465</td><td>.446</td><td>.224</td><td>.155</td><td>.130</td><td>54.5</td><td>22.1</td><td>19.8</td></tr><tr><td>L</td><td>.501</td><td>.481</td><td>.457</td><td>.185</td><td>.138</td><td>.110</td><td>56.4</td><td>24.6</td><td>19.2</td></tr><tr><td>XL</td><td>.504</td><td>.511</td><td>.474</td><td>.134</td><td>.110</td><td>.084</td><td>49.4</td><td>25.1</td><td>17.9</td></tr><tr><td colspan=\"10\">Deartifacting (JPEG compression)</td></tr><tr><td>XS</td><td>.498</td><td>.442</td><td>.432</td><td>.404</td><td>.341</td><td>.349</td><td>52.3</td><td>26.3</td><td>14.8</td></tr><tr><td>S</td><td>.497</td><td>.448</td><td>.437</td><td>.398</td><td>.352</td><td>.350</td><td>49.6</td><td>22.4</td><td>15.4</td></tr><tr><td>M</td><td>.498</td><td>.461</td><td>.445</td><td>.413</td><td>.357</td><td>.357</td><td>33.2</td><td>24.1</td><td>15.4</td></tr><tr><td>L</td><td>.500</td><td>.475</td><td>.460</td><td>.395</td><td>.367</td><td>.374</td><td>46.9</td><td>25.2</td><td>16.0</td></tr><tr><td>XL</td><td>.508</td><td>.503</td><td>.490</td><td>.427</td><td>.418</td><td>.412</td><td>30.8</td><td>22.1</td><td>18.7</td></tr><tr><td colspan=\"10\">Inpainting (random strokes)</td></tr><tr><td>XS</td><td>.498</td><td>.409</td><td>.378</td><td>.464</td><td>.374</td><td>.348</td><td>46.9</td><td>24.4</td><td>12.9</td></tr><tr><td>S</td><td>.501</td><td>.425</td><td>.387</td><td>.356</td><td>.287</td><td>.264</td><td>42.3</td><td>27.2</td><td>14.2</td></tr><tr><td>M</td><td>.509</td><td>.438</td><td>.396</td><td>.283</td><td>.227</td><td>.206</td><td>38.5</td><td>30.1</td><td>14.5</td></tr><tr><td>L</td><td>.513</td><td>.452</td><td>.409</td><td>.231</td><td>.184</td><td>.163</td><td>32.6</td><td>33.1</td><td>15.3</td></tr><tr><td>XL</td><td>.524</td><td>.460</td><td>.422</td><td>.187</td><td>.157</td><td>.132</td><td>36.2</td><td>25.2</td><td>15.9</td></tr></table>",
1307
+ "bbox": [
1308
+ 122,
1309
+ 89,
1310
+ 475,
1311
+ 435
1312
+ ],
1313
+ "page_idx": 6
1314
+ },
1315
+ {
1316
+ "type": "text",
1317
+ "text": "with hyperparameters optimized for each task and level individually. Despite this obvious impediment, our proposed method outperforms baselines in the vast majority of scenarios. In particular, it reaches better accuracy in the majority of tasks against baselines optimized for accuracy, and realism is strictly better. Fig. 4 illustrates the improved visual quality obtained with our method.",
1318
+ "bbox": [
1319
+ 111,
1320
+ 556,
1321
+ 483,
1322
+ 661
1323
+ ],
1324
+ "page_idx": 6
1325
+ },
1326
+ {
1327
+ "type": "text",
1328
+ "text": "Robustness to compositions Robustness to compositions of degradations was evaluated by forming all possible subsequences of degradations (order in eq. (16)) at medium level. Quantitative results are aggregated in tab. 2, while fig. 5 presents qualitative examples. Our results show that both high fidelity and high realism are maintained with our method, while the baselines struggle under varied degradations.",
1329
+ "bbox": [
1330
+ 111,
1331
+ 662,
1332
+ 483,
1333
+ 782
1334
+ ],
1335
+ "page_idx": 6
1336
+ },
1337
+ {
1338
+ "type": "text",
1339
+ "text": "7. Comparison to diffusion models",
1340
+ "text_level": 1,
1341
+ "bbox": [
1342
+ 112,
1343
+ 795,
1344
+ 405,
1345
+ 811
1346
+ ],
1347
+ "page_idx": 6
1348
+ },
1349
+ {
1350
+ "type": "text",
1351
+ "text": "We compare our approach to the recent \"Denoising diffusion restoration model\" (DDRM) [32], which also does not require problem-specific supervised training and leverages a pre-trained (DDPM [53]) model. Importantly, DCRM is designed only for linear inverse problems with additive Gaussian noise, where the degradation model is",
1352
+ "bbox": [
1353
+ 111,
1354
+ 821,
1355
+ 483,
1356
+ 911
1357
+ ],
1358
+ "page_idx": 6
1359
+ },
1360
+ {
1361
+ "type": "table",
1362
+ "img_path": "images/af9504541b21cbcce36836c54f2a37f5e47fdc279b6813465a78485deb92a162.jpg",
1363
+ "table_caption": [
1364
+ "Table 1. Quantitative comparison on individual tasks against baselines (\"PULS\" [40] and \"L-BRG\" [13]). Baselines are optimized for accuracy on each row separately, while ours uses the same set of hyperparameters across all. Despite this clear handicap, our method is either on par or outperforms baselines. Results color-coded as best and second best."
1365
+ ],
1366
+ "table_footnote": [],
1367
+ "table_body": "<table><tr><td></td><td colspan=\"3\">Accur. (LPIPS) ↓</td><td colspan=\"3\">Fidelity (LPIPS) ↓</td><td colspan=\"3\">Realism (pFID) ↓</td></tr><tr><td></td><td colspan=\"3\">PULS L-BRG OUR5</td><td colspan=\"3\">PULS L-BRG OUR5</td><td colspan=\"3\">PULS L-BRG OUR5</td></tr><tr><td colspan=\"10\">2 degradations</td></tr><tr><td>NA</td><td>.517</td><td>.485</td><td>.459</td><td>.328</td><td>.301</td><td>.290</td><td>43.4</td><td>24.2</td><td>17.3</td></tr><tr><td>AP</td><td>.511</td><td>.478</td><td>.457</td><td>.270</td><td>.231</td><td>.204</td><td>29.7</td><td>17.6</td><td>17.0</td></tr><tr><td>UA</td><td>.510</td><td>.518</td><td>.508</td><td>.307</td><td>.348</td><td>.287</td><td>23.7</td><td>20.5</td><td>19.7</td></tr><tr><td>NP</td><td>.511</td><td>.480</td><td>.458</td><td>.125</td><td>.079</td><td>.062</td><td>47.0</td><td>20.9</td><td>19.2</td></tr><tr><td>UN</td><td>.501</td><td>.519</td><td>.511</td><td>.178</td><td>.149</td><td>.153</td><td>33.4</td><td>26.2</td><td>21.1</td></tr><tr><td>UP</td><td>.510</td><td>.478</td><td>.485</td><td>.140</td><td>.061</td><td>.089</td><td>23.9</td><td>35.3</td><td>20.7</td></tr><tr><td colspan=\"10\">3 degradations</td></tr><tr><td>UNP</td><td>.510</td><td>.526</td><td>.507</td><td>.086</td><td>.062</td><td>.051</td><td>28.5</td><td>22.1</td><td>20.1</td></tr><tr><td>UAP</td><td>.525</td><td>.523</td><td>.513</td><td>.205</td><td>.154</td><td>.119</td><td>23.0</td><td>18.4</td><td>20.5</td></tr><tr><td>UNA</td><td>.521</td><td>.535</td><td>.533</td><td>.265</td><td>.310</td><td>.290</td><td>26.2</td><td>20.7</td><td>22.8</td></tr><tr><td>NAP</td><td>.526</td><td>.502</td><td>.470</td><td>.210</td><td>.197</td><td>.160</td><td>38.7</td><td>18.4</td><td>18.5</td></tr><tr><td colspan=\"10\">4 degradations</td></tr><tr><td>UNAP</td><td>.533</td><td>.546</td><td>.525</td><td>.192</td><td>.177</td><td>.131</td><td>25.5</td><td>21.1</td><td>21.8</td></tr></table>",
1368
+ "bbox": [
1369
+ 516,
1370
+ 89,
1371
+ 885,
1372
+ 309
1373
+ ],
1374
+ "page_idx": 6
1375
+ },
1376
+ {
1377
+ "type": "table",
1378
+ "img_path": "images/0c236041033f81890c2e66ef6adb80379f25716b63f49370e26ddb861dd42b63.jpg",
1379
+ "table_caption": [
1380
+ "Table 2. Quantitative comparison on composed tasks (at medium levels) against baselines (\"PULS\" [40] and \"L-BRG\" [13]). Baselines are optimized for accuracy on each row separately, while ours uses the same set of hyperparameters across all. Despite this clear handicap, our method is either on par or outperforms baselines. Acronyms indicate restorations: Upsampling, deNoising, deArtifacting, and/or inPainting. Results color-coded as best and second best."
1381
+ ],
1382
+ "table_footnote": [],
1383
+ "table_body": "<table><tr><td></td><td colspan=\"2\">Accur. (LPIPS) ↓</td><td colspan=\"2\">Realism (pFID) ↓</td></tr><tr><td></td><td>DDRM</td><td>OURS</td><td>DDRM</td><td>OURS</td></tr><tr><td colspan=\"5\">Exact Upsampling (known average pooling)</td></tr><tr><td>0.412</td><td>0.443</td><td></td><td>44.75</td><td>18.92</td></tr><tr><td colspan=\"5\">Inexact Upsampling (unknown nearest-neighbor)</td></tr><tr><td>0.559</td><td>0.539</td><td></td><td>67.06</td><td>32.81</td></tr><tr><td colspan=\"5\">Linear Denoising (additive Gaussian noise)</td></tr><tr><td>0.336</td><td>0.452</td><td></td><td>43.28</td><td>19.25</td></tr><tr><td colspan=\"5\">Nonlinear Denoising (clamped additive Gaussian noise)</td></tr><tr><td>0.530</td><td>0.450</td><td></td><td>156.1</td><td>19.06</td></tr><tr><td colspan=\"5\">Inpainting (masking random strokes)</td></tr><tr><td>0.289</td><td>0.301</td><td></td><td>26.12</td><td>25.71</td></tr><tr><td colspan=\"5\">Uncropping (top left corner unmasked)</td></tr><tr><td>0.468</td><td>0.418</td><td></td><td>34.33</td><td>17.645</td></tr></table>",
1384
+ "bbox": [
1385
+ 557,
1386
+ 449,
1387
+ 841,
1388
+ 646
1389
+ ],
1390
+ "page_idx": 6
1391
+ },
1392
+ {
1393
+ "type": "text",
1394
+ "text": "Table 3. Quantitative comparison to DDRM [32] on accuracy and realism (see sec. 7). Results are color-coded as best .",
1395
+ "bbox": [
1396
+ 511,
1397
+ 656,
1398
+ 883,
1399
+ 686
1400
+ ],
1401
+ "page_idx": 6
1402
+ },
1403
+ {
1404
+ "type": "text",
1405
+ "text": "perfectly known. Therefore, for each of the upsampling, denoising and inpainting tasks, we experiment with both: 1) a linear inverse problem with known parameters; and 2) an adjusted version which is either non-linear, where the extract degradation model is unknown, or a more extreme version of the problem. Specifically, we test: average pooling (known, linear) vs nearest-neighbor (unknown) $32 \\times$ (XL) upsampling; additive Gaussian noise (linear) vs clamped (to [0, 255]) additive Gaussian noise (non-linear) with $\\sigma^2 = 0.8$ (XL) for both; and masking strokes (at XL level) vs uncropping the top-left corner (XXL). DDRM is available on CelebA (not FFHQ) at $256 \\times 256$ , so we use the corresponding StyleGAN2-ADA pretrained",
1406
+ "bbox": [
1407
+ 511,
1408
+ 715,
1409
+ 883,
1410
+ 912
1411
+ ],
1412
+ "page_idx": 6
1413
+ },
1414
+ {
1415
+ "type": "page_number",
1416
+ "text": "22298",
1417
+ "bbox": [
1418
+ 478,
1419
+ 936,
1420
+ 517,
1421
+ 948
1422
+ ],
1423
+ "page_idx": 6
1424
+ },
1425
+ {
1426
+ "type": "image",
1427
+ "img_path": "images/0dc25b0c6d19b5a73690935d182ba793285ca28427df5f89d63b4ab736204646.jpg",
1428
+ "image_caption": [
1429
+ "Figure 5. Qualitative comparisons on all possible compositions of the four tasks given medium levels of degradation. Samples are chosen randomly, and acronyms indicate the degradations: Upsampling, inPainting, deNoising and/or deArtifacting (sec. 6.3)."
1430
+ ],
1431
+ "image_footnote": [],
1432
+ "bbox": [
1433
+ 114,
1434
+ 88,
1435
+ 537,
1436
+ 311
1437
+ ],
1438
+ "page_idx": 7
1439
+ },
1440
+ {
1441
+ "type": "image",
1442
+ "img_path": "images/6e0f0b184a35cae0ad3bec11f10fc46215c8222bb5c2e5711e61f848e2ef27c7.jpg",
1443
+ "image_caption": [],
1444
+ "image_footnote": [],
1445
+ "bbox": [
1446
+ 537,
1447
+ 88,
1448
+ 883,
1449
+ 311
1450
+ ],
1451
+ "page_idx": 7
1452
+ },
1453
+ {
1454
+ "type": "image",
1455
+ "img_path": "images/9788772c3424bc9e816961daf2dbe2abb2a33c464b3ac9a2acead706f2fb127d.jpg",
1456
+ "image_caption": [
1457
+ "Figure 6. Results on LSUN horses and churches [60], AFHQ-v2 Dogs [11] and BreCaHAD [5] at resolution $256 \\times 256$ for small upsampling (U), small denoising (N), medium deartifacting (A), and medium inpainting (P); the left/right halves are the target/restored images. See supp. for details and more results."
1458
+ ],
1459
+ "image_footnote": [],
1460
+ "bbox": [
1461
+ 114,
1462
+ 375,
1463
+ 482,
1464
+ 657
1465
+ ],
1466
+ "page_idx": 7
1467
+ },
1468
+ {
1469
+ "type": "text",
1470
+ "text": "model. Due to the reduced resolution, pFID is measured on 250 patches of size $64 \\times 64$ .",
1471
+ "bbox": [
1472
+ 111,
1473
+ 772,
1474
+ 482,
1475
+ 801
1476
+ ],
1477
+ "page_idx": 7
1478
+ },
1479
+ {
1480
+ "type": "text",
1481
+ "text": "Interestingly, we find that DDRM reproduces the targets almost perfectly (fidelity is near zero, see supp. for details and qualitative results), to the point where it overfits to any deviation between its degradation model and the true degradation. Tab. 3 presents the accuracy and realism metrics for all the aforementioned tasks. While the accuracy favors DDRM for the tasks which perfectly",
1482
+ "bbox": [
1483
+ 111,
1484
+ 806,
1485
+ 483,
1486
+ 912
1487
+ ],
1488
+ "page_idx": 7
1489
+ },
1490
+ {
1491
+ "type": "text",
1492
+ "text": "match its assumptions, our approach outperforms it when this is not the case. In addition, our method provides significantly better realism across all scenarios. Overall, our approach yields high resolution results with more detail, can adapt to non-linear tasks, and it can also cope with larger degradations.",
1493
+ "bbox": [
1494
+ 511,
1495
+ 378,
1496
+ 883,
1497
+ 468
1498
+ ],
1499
+ "page_idx": 7
1500
+ },
1501
+ {
1502
+ "type": "text",
1503
+ "text": "8. Discussion",
1504
+ "text_level": 1,
1505
+ "bbox": [
1506
+ 513,
1507
+ 494,
1508
+ 625,
1509
+ 510
1510
+ ],
1511
+ "page_idx": 7
1512
+ },
1513
+ {
1514
+ "type": "text",
1515
+ "text": "**Limitations** Our method is not restricted to faces and works with any pretrained StyleGAN, as demonstrated in fig. 6. However, it is inherently limited to the domain learned by the GAN. Training StyleGAN on a large unstructured dataset like ImageNet is difficult but recent papers [27,50,55] have attained some success—applying our method on some of these large-scale models is a promising research avenue. Our method also faces the same ethical considerations as the GAN it relies on. Finally, as with other methods [13,32,40], we also require knowledge of an (approximate) degradation function.",
1516
+ "bbox": [
1517
+ 511,
1518
+ 525,
1519
+ 883,
1520
+ 691
1521
+ ],
1522
+ "page_idx": 7
1523
+ },
1524
+ {
1525
+ "type": "text",
1526
+ "text": "Conclusion This paper presents a method which makes StyleGAN-based image restoration robust to both variability in degradation levels and to compositions of different degradations. Our proposed method relies on a conservative optimization procedure over a progressively richer latent space and avoids regularization terms altogether. Using a single set of hyperparameters, we obtain competitive and even state-of-the-art results on several challenging scenarios when compared to baselines that are optimized for each task/level individually.",
1527
+ "bbox": [
1528
+ 511,
1529
+ 696,
1530
+ 883,
1531
+ 847
1532
+ ],
1533
+ "page_idx": 7
1534
+ },
1535
+ {
1536
+ "type": "text",
1537
+ "text": "Acknowledgements This research was supported by NSERC grant RGPIN-2020-04799, a MITACS Globalink internship to Y. Poirier-Ginter, and by the Digital Research Alliance Canada.",
1538
+ "bbox": [
1539
+ 511,
1540
+ 851,
1541
+ 883,
1542
+ 910
1543
+ ],
1544
+ "page_idx": 7
1545
+ },
1546
+ {
1547
+ "type": "page_number",
1548
+ "text": "22299",
1549
+ "bbox": [
1550
+ 478,
1551
+ 936,
1552
+ 517,
1553
+ 948
1554
+ ],
1555
+ "page_idx": 7
1556
+ },
1557
+ {
1558
+ "type": "text",
1559
+ "text": "References",
1560
+ "text_level": 1,
1561
+ "bbox": [
1562
+ 114,
1563
+ 89,
1564
+ 209,
1565
+ 104
1566
+ ],
1567
+ "page_idx": 8
1568
+ },
1569
+ {
1570
+ "type": "list",
1571
+ "sub_type": "ref_text",
1572
+ "list_items": [
1573
+ "[1] Diffjpeg. https://github.com/mlomnitz/DiffJPEG.",
1574
+ "[2] Libjpeg. https://libjpeg-turbo.org.",
1575
+ "[3] Rameen Abdal, Yipeng Qin, and Peter Wonka. Image2StyleGAN: How to embed images into the StyleGAN latent space? In Int. Conf. Comput. Vis., 2019.",
1576
+ "[4] Rameen Abdul, Yipeng Qin, and Peter Wonka. Image2stylegan++: How to edit the embedded images? In IEEE Conf. Comput. Vis. Pattern Recog., 2020.",
1577
+ "[5] Alper Aksac, Douglas J Demetrick, Tansel Ozyer, and Reda Alhajj. Brecahad: a dataset for breast cancer histopathological annotation and diagnosis. BMC research notes, 12(1):1-3, 2019.",
1578
+ "[6] Ashish Bora, Ajil Jalal, Eric Price, and Alexandros G Dimakis. Compressed sensing using generative models. In Int. Conf. Mach. Learn., 2017.",
1579
+ "[7] Andrew Brock, Jeff Donahue, and Karen Simonyan. Large scale GAN training for high fidelity natural image synthesis. In Int. Conf. Learn. Represent., 2019.",
1580
+ "[8] Lucy Chai, Michael Gharbi, Eli Shechtman, Phillip Isola, and Richard Zhang. Any-resolution training for high-resolution image synthesis. In Eur. Conf. Comput. Vis., 2022.",
1581
+ "[9] Hanting Chen, Yunhe Wang, Tianyu Guo, Chang Xu, Yiping Deng, Zhenhua Liu, Siwei Ma, Chunjing Xu, Chao Xu, and Wen Gao. Pre-trained image processing transformer. In IEEE Conf. Comput. Vis. Pattern Recog., 2021.",
1582
+ "[10] Li-Heng Chen, Christos G Bampis, Zhi Li, Andrey Norkin, and Alan C Bovik. Proxiqa: A proxy approach to perceptual optimization of learned image compression. IEEE Trans. Image Process., 30:360-373, 2020.",
1583
+ "[11] Yunjey Choi, Youngjung Uh, Jaejun Yoo, and Jung-Woo Ha. Stargan v2: Diverse image synthesis for multiple domains. In IEEE Conf. Comput. Vis. Pattern Recog., 2020.",
1584
+ "[12] Alex Clark. Pillow (pil fork) documentation, 2015.",
1585
+ "[13] Arthur Conmy, Subhadip Mukherjee, and Carola-Bibiane Schonlieb. Stylegan-induced data-driven regularization for inverse problems. In Int. Conf. Acous. Sp. Sig. Proc., 2022.",
1586
+ "[14] Giannis Daras, Yuval Dagan, Alexandros G Dimakis, and Constantinos Daskalakis. Score-guided intermediate layer optimization: Fast Langevin mixing for inverse problems. In Int. Conf. Mach. Learn., 2022.",
1587
+ "[15] Giannis Daras, Joseph Dean, Ajil Jalal, and Alexandros G Dimakis. Intermediate layer optimization for inverse problems using deep generative models. In Int. Conf. Mach. Learn., 2021.",
1588
+ "[16] Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In IEEE Conf. Comput. Vis. Pattern Recog., 2019.",
1589
+ "[17] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. In Adv. Neural Inform. Process. Syst., 2021.",
1590
+ "[18] Omar Elharrouss, Noor Almaadeed, Somaya Ali Al-Maadeed, and Younes Akbari. Image inpainting: A review. Neural Processing Letters, 51:2007-2028, 2019."
1591
+ ],
1592
+ "bbox": [
1593
+ 116,
1594
+ 114,
1595
+ 482,
1596
+ 910
1597
+ ],
1598
+ "page_idx": 8
1599
+ },
1600
+ {
1601
+ "type": "list",
1602
+ "sub_type": "ref_text",
1603
+ "list_items": [
1604
+ "[19] Linwei Fan, Fan Zhang, Hui Fan, and Cai ming Zhang. Brief review of image denoising techniques. Visual Computing for Industry, Biomedicine and Art, 2, 2019.",
1605
+ "[20] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. In Adv. Neural Inform. Process. Syst., 2014.",
1606
+ "[21] Jinjin Gu, Yujun Shen, and Bolei Zhou. Image processing using multi-code gan prior. In IEEE Conf. Comput. Vis. Pattern Recog., 2020.",
1607
+ "[22] Yuchao Gu, Xintao Wang, Liangbin Xie, Chao Dong, Gen Li, Ying Shan, and Ming-Ming Cheng. Vqfr: Blind face restoration with vector-quantized dictionary and parallel decoder. In Eur. Conf. Comput. Vis., 2022.",
1608
+ "[23] Erik Härkönen, Aaron Hertzmann, Jaakko Lehtinen, and Sylvain Paris. Ganspace: Discovering interpretable gan controls. In Adv. Neural Inform. Process. Syst., 2020.",
1609
+ "[24] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. GANs trained by a two time-scale update rule converge to a local nash equilibrium. In Adv. Neural Inform. Process. Syst., 2017.",
1610
+ "[25] Shady Abu Hussein, Tom Tirer, and Raja Giryes. Image-adaptive gan based reconstruction. In Ass. Adv. Artif. Intel., 2020.",
1611
+ "[26] Kyoungkook Kang, Seongtae Kim, and Sunghyun Cho. Gan inversion for out-of-range images with geometric transformations. In Int. Conf. Comput. Vis., 2021.",
1612
+ "[27] Minguk Kang, Jun-Yan Zhu, Richard Zhang, Jaesik Park, Eli Shechtman, Sylvain Paris, and Taesung Park. Scaling up GANs for text-to-image synthesis. In IEEE Conf. Comput. Vis. Pattern Recog., 2023.",
1613
+ "[28] Tero Karras, Miika Aittala, Janne Hellsten, Samuli Laine, Jaakko Lehtinen, and Timo Aila. Training generative adversarial networks with limited data. In Adv. Neural Inform. Process. Syst., 2020.",
1614
+ "[29] Tero Karras, Miika Aittala, Samuli Laine, Erik Härkönen, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Alias-free generative adversarial networks. In Adv. Neural Inform. Process. Syst., 2021.",
1615
+ "[30] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In IEEE Conf. Comput. Vis. Pattern Recog., 2019.",
1616
+ "[31] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of StyleGAN. In IEEE Conf. Comput. Vis. Pattern Recog., 2020.",
1617
+ "[32] Bahjat Kawar, Michael Elad, Stefano Ermon, and Jiaming Song. Denoising diffusion restoration models. In Adv. Neural Inform. Process. Syst., 2022.",
1618
+ "[33] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014.",
1619
+ "[34] Aijin Li, Gengyan Li, Lei Sun, and Xintao Wang. Faceformer: Scale-aware blind face restoration with transformers. ArXiv, abs/2207.09790, 2022.",
1620
+ "[35] Juncheng Li, Zehua Pei, and Tieyong Zeng. From beginner to master: A survey for deep learning-based single-image super-resolution. arXiv preprint arXiv:2109.14335, 2021."
1621
+ ],
1622
+ "bbox": [
1623
+ 516,
1624
+ 92,
1625
+ 883,
1626
+ 910
1627
+ ],
1628
+ "page_idx": 8
1629
+ },
1630
+ {
1631
+ "type": "page_number",
1632
+ "text": "22300",
1633
+ "bbox": [
1634
+ 478,
1635
+ 936,
1636
+ 517,
1637
+ 948
1638
+ ],
1639
+ "page_idx": 8
1640
+ },
1641
+ {
1642
+ "type": "list",
1643
+ "sub_type": "ref_text",
1644
+ "list_items": [
1645
+ "[36] Xiaoming Li, Wenyu Li, Dongwei Ren, Hongzhi Zhang, Meng Wang, and Wangmeng Zuo. Enhanced blind face restoration with multi-exemplar images and adaptive spatial feature fusion. In IEEE Conf. Comput. Vis. Pattern Recog., 2020.",
1646
+ "[37] Jingyun Liang, Jiezhang Cao, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. Swinir: Image restoration using swim transformer. In Int. Conf. Comput. Vis., 2021.",
1647
+ "[38] Xuan Luo, Xuaner Zhang, Paul Yoo, Ricardo Martin-Brualla, Jason Lawrence, and Steven M Seitz. Time-travel rephotography. ACM Trans. Graph., 40(6):1-12, 2021.",
1648
+ "[39] Razvan V Marinescu, Daniel Moyer, and Polina Golland. Bayesian image reconstruction using deep generative models. arXiv preprint arXiv:2012.04567, 2020.",
1649
+ "[40] Sachit Menon, Alexandru Damian, Shijia Hu, Nikhil Ravi, and Cynthia Rudin. Pulse: Self-supervised photo upsampling via latent space exploration of generative models. In IEEE Conf. Comput. Vis. Pattern Recog., 2020.",
1650
+ "[41] NVlabs. Stylegan2 ada - python. https://github.com/NVlabs/stylegan2-ada-pytorch.",
1651
+ "[42] Xingang Pan, Xiaohang Zhan, Bo Dai, Dahua Lin, Chen Change Loy, and Ping Luo. Exploiting deep generative prior for versatile image restoration and manipulation. IEEE Trans. Pattern Anal. Mach. Intell., 2021.",
1652
+ "[43] Or Patashnik, Zongze Wu, Eli Shechtman, Daniel Cohen-Or, and Dani Lischinski. Styleclip: Text-driven manipulation of stylegan imagery. In Int. Conf. Comput. Vis., 2021.",
1653
+ "[44] Yohan Poirier-Ginter, Alexandre Lessard, Ryan Smith, and Jean-François Lalonde. Overparameterization improves StyleGAN inversion. In IEEE Conf. Comput. Vis. Pattern Recog., 2022.",
1654
+ "[45] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022.",
1655
+ "[46] Daniel Roich, Ron Mokady, Amit H Bermano, and Daniel Cohen-Or. Pivotal tuning for latent-based editing of real images. ACM Trans. Graph., 42(1), Feb. 2023.",
1656
+ "[47] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In IEEE Conf. Comput. Vis. Pattern Recog., 2022.",
1657
+ "[48] Chitwan Sahara, William Chan, Huiwen Chang, Chris Lee, Jonathan Ho, Tim Salimans, David Fleet, and Mohammad Norouzi. Palette: Image-to-image diffusion models. ACM Trans. Graph., pages 1-10, 2022.",
1658
+ "[49] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S Sara Mahdavi, Rapha Gontijo Lopes, et al. Photorealistic text-to-image diffusion models with deep language understanding. arXiv preprint arXiv:2205.11487, 2022.",
1659
+ "[50] Axel Sauer, Katja Schwarz, and Andreas Geiger. StyleGAN-XL: Scaling stylegan to large diverse datasets. ACM Trans. Graph., 1, 2022.",
1660
+ "[51] Yujun Shen, Ceyuan Yang, Xiaou Tang, and Bolei Zhou. Interfacegan: Interpreting the disentangled face represent"
1661
+ ],
1662
+ "bbox": [
1663
+ 114,
1664
+ 90,
1665
+ 485,
1666
+ 912
1667
+ ],
1668
+ "page_idx": 9
1669
+ },
1670
+ {
1671
+ "type": "list",
1672
+ "sub_type": "ref_text",
1673
+ "list_items": [
1674
+ "tation learned by gans. IEEE Trans. Pattern Anal. Mach. Intell., 2020.",
1675
+ "[52] Richard Shin and Dawn Song. Jpeg-resistant adversarial images. In NeurlPS Workshop Mach. Learn. Comp. Sec., 2017.",
1676
+ "[53] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In Int. Conf. Mach. Learn., 2015.",
1677
+ "[54] Zhengzhong Tu, Hossein Talebi, Han Zhang, Feng Yang, Peyman Milanfar, Alan Bovik, and Yinxiao Li. Maxim: Multi-axis MLP for image processing. In IEEE Conf. Comput. Vis. Pattern Recog., 2022.",
1678
+ "[55] Yuri Viazovetskyi, Vladimir Ivashkin, and Evgeny Kashin. Stylegan2 distillation for feed-forward image manipulation. In Eur. Conf. Comput. Vis., 2020.",
1679
+ "[56] Xintao Wang, Yu Li, Honglun Zhang, and Ying Shan. Towards real-world blind face restoration with generative facial prior. In IEEE Conf. Comput. Vis. Pattern Recog., 2021.",
1680
+ "[57] Zhihao Wang, Jian Chen, and Steven CH Hoi. Deep learning for image super-resolution: A survey. IEEE Trans. Pattern Anal. Mach. Intell., 43(10):3365-3387, 2020.",
1681
+ "[58] Jeremy Watt, Reza Borhani, and Aggelos K. Katsaggelos. Machine Learning Refined: Foundations, Algorithms, and Applications. Cambridge University Press, 2016.",
1682
+ "[59] Weihao Xia, Yulun Zhang, Yujiu Yang, Jing-Hao Xue, Bolei Zhou, and Ming-Hsuan Yang. GAN inversion: A survey. IEEE Trans. Pattern Anal. Mach. Intell., 2022.",
1683
+ "[60] Fisher Yu, Yinda Zhang, Shuran Song, Ari Seff, and Jianxiong Xiao. Lsun: Construction of a large-scale image dataset using deep learning with humans in the loop. arXiv preprint arXiv:1506.03365, 2015.",
1684
+ "[61] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, Ming-Hsuan Yang, and Ling Shao. Learning enriched features for real image restoration and enhancement. In Eur. Conf. Comput. Vis., 2020.",
1685
+ "[62] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, Ming-Hsuan Yang, and Ling Shao. Multi-stage progressive image restoration. In IEEE Conf. Comput. Vis. Pattern Recog., 2021.",
1686
+ "[63] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In IEEE Conf. Comput. Vis. Pattern Recog., 2018.",
1687
+ "[64] Shangchen Zhou, Kelvin CK Chan, Chongyi Li, and Chen Change Loy. Towards robust blind face restoration with codebook lookup transformer, 2022.",
1688
+ "[65] Feida Zhu, Junwei Zhu, Wenqing Chu, Xinyi Zhang, Xiaozhong Ji, Chengjie Wang, and Ying Tai. Blind face restoration via integrating face shape and generative priors. In IEEE Conf. Comput. Vis. Pattern Recog., 2022.",
1689
+ "[66] Peihao Zhu, Rameen Abdal, Yipeng Qin, John Femiani, and Peter Wonka. Improved stylegan embedding: Where are the good latents? arXiv preprint arXiv:2012.09036, 2020."
1690
+ ],
1691
+ "bbox": [
1692
+ 516,
1693
+ 92,
1694
+ 883,
1695
+ 904
1696
+ ],
1697
+ "page_idx": 9
1698
+ },
1699
+ {
1700
+ "type": "page_number",
1701
+ "text": "22301",
1702
+ "bbox": [
1703
+ 478,
1704
+ 936,
1705
+ 517,
1706
+ 948
1707
+ ],
1708
+ "page_idx": 9
1709
+ }
1710
+ ]
2023/Robust Unsupervised StyleGAN Image Restoration/22cf3860-c0a9-4cc3-ae76-e2aceeacd7e1_model.json ADDED
The diff for this file is too large to render. See raw diff
 
2023/Robust Unsupervised StyleGAN Image Restoration/22cf3860-c0a9-4cc3-ae76-e2aceeacd7e1_origin.pdf ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:25612d10cb945466a49610074bec24d15462a3dbf568f7095597757d89948233
3
+ size 2998742
2023/Robust Unsupervised StyleGAN Image Restoration/full.md ADDED
@@ -0,0 +1,374 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Robust Unsupervised StyleGAN Image Restoration
2
+
3
+ Yohan Poirier-Ginter\*, Jean-François Lalonde\* Inria, Université Côte d'Azur, Université Laval https://lvsn.github.io/RobustUnsupervised/
4
+
5
+ ![](images/90898fdfe2be1005b020e1012c683e3e8e202c8c33fc2716ec151b598318270b.jpg)
6
+ Figure 1. Our unsupervised StyleGAN image restoration method is robust and effective across a variety of tasks (left, top to bottom: upsampling, denoising, deartifacting, and inpainting) and a wide range of degradation levels. Since it avoids the need for task-specific hyperparameter tuning, it can directly handle combinations of such degradations, e.g., pairs (center) or even all four (right).
7
+
8
+ ![](images/8df8152cd13b675b505e8ce0a75faf491efddd25832127ae23f3b6fc55c58681.jpg)
9
+
10
+ ![](images/828ad6f3fb4690e6efcdd78d359a743abc3bb1d0e379118fd8a574d7681cf80c.jpg)
11
+
12
+ # Abstract
13
+
14
+ GAN-based image restoration inverts the generative process to repair images corrupted by known degradations. Existing unsupervised methods must be carefully tuned for each task and degradation level. In this work, we make StyleGAN image restoration robust: a single set of hyperparameters works across a wide range of degradation levels. This makes it possible to handle combinations of several degradations, without the need to retune. Our proposed approach relies on a 3-phase progressive latent space extension and a conservative optimizer, which avoids the need for any additional regularization terms. Extensive experiments demonstrate robustness on inpainting, upsampling, denoising, and deartifacting at varying degradations levels, outperforming other StyleGAN-based inversion techniques. Our approach also favorably compares to diffusion-based restoration by yielding much more realistic inversion results. Code is available at the above URL.
15
+
16
+ # 1. Introduction
17
+
18
+ Image restoration, the task of recovering a high quality image from a degraded input, is a long-standing problem in image processing and computer vision. Since different restoration tasks—such as denoising, upsampling,
19
+
20
+ deartifacting, etc.—can be quite distinct, many recent approaches [9, 10, 37, 54, 61, 62] propose to solve them in a supervised learning paradigm by leveraging curated datasets specifically designed for the task at hand. Unfortunately, designing task-specific approaches requires retraining large networks on each task separately.
21
+
22
+ In parallel, the advent of powerful generative models has enabled the emergence of unsupervised restoration methods [42], which do not require task-specific training. The idea is to invert the generative process to recover a clean image. Assuming a known (or approximate) degradation model, the optimization procedure therefore attempts to recover an image that both: 1) closely matches the target degraded image after undergoing a similar degradation model (fidelity); and 2) lies in the space of realistic images learned by the GAN (realism).
23
+
24
+ In the recent literature, StyleGAN [29-31] has been found to be particularly effective for unsupervised image restoration [13-15, 39, 40] because of the elegant design of its latent space. Indeed, these approaches leverage style inversion techniques [3, 4] to solve for a latent vector that, when given to the generator, creates an image close to the degraded target. Unfortunately, this only works when such a match actually exists in the model distribution, which is rarely the case in practice. Hence, effective methods extend the learned latent space to cre
25
+
26
+ ate additional degrees of freedom to admit more images; this creates the need for additional regularization losses. Hyperparameters must therefore carefully be tuned for each specific task and degradation level.
27
+
28
+ In this work, we make unsupervised StyleGAN image restoration-by-inversion robust to the type and intensity of degradations. Our proposed method employs the same hyperparameters across all tasks and levels and does not rely on any regularization loss. Our approach leans on two key ideas. First, we rely on a 3-phase progressive latent space extension: we begin by optimizing over the learned (global) latent space, then expand it across individual layers of the generator, and finally expand it further across individual filters—where optimization at each phase is initialized with the result of the previous one. Second, we rely on a conservative, normalized gradient descent (NGD) [58] optimizer which is naturally constrained to stay close to its initial point compared to more sophisticated approaches such as Adam [33]. This combination of prudent optimization over a progressively richer latent space avoids additional regularization terms altogether and keeps the overall procedure simple and constant across all tasks. We evaluate our method on upsampling, inpainting, denoising and deartifacting on a wide range of degradation levels, achieving state-of-the-art results on most scenarios even when baselines are optimized on each independently. We also show that our approach outperforms existing techniques on compositions of these tasks without changing hyperparameters.
29
+
30
+ - We propose a robust 3-phase StyleGAN image restoration framework. Our optimization technique maintains: 1) strong realism when degradations level are high; and 2) high fidelity when they are low. Our method is fully unsupervised, requires no per-task training, and can handle different tasks at different levels without having to adjust hyperparameters.
31
+ - We demonstrate the effectiveness of the proposed method under diverse and composed degradations. We develop a benchmark of synthetic image restoration tasks—making their degradation levels easy to control—with care taken to avoid unrealistic assumptions. Our method outperforms existing unsupervised [13,40] and diffusion-based [32] approaches.
32
+
33
+ # 2. Related work
34
+
35
+ This section covers previous work on StyleGAN inversion by optimization and its use for image restoration, and discusses related works on generative priors. Because the literature on general purpose image restoration is so extensive, we do not attempt to review it here and instead refer the reader to recent surveys [18, 19, 35, 57].
36
+
37
+ GANs and StyleGAN inversion Generative adversarial networks [20] (GANs) are a popular technique
38
+
39
+ for generative modeling. In particular, the StyleGAN family [28-31] learns a highly compressed latent representation of its domain and stands out for its popularity and generation quality. Great progress has been made in StyleGAN inversion: reversing the generation process to infer latent parameters that generate a given image [59].
40
+
41
+ Thanks to inversion, StyleGAN has found use in many image editing [23, 43, 51] and restoration [13-15, 39, 40] tasks. Purely optimization-based techniques were studied extensively [3, 4, 26, 44, 46, 66] in the context of image editing. These methods extend the learned latent space of a pretrained StyleGAN model (commonly named $\mathcal{W}$ ) by adding additional parameters to optimize over it. The most common approach consists of using a different latent code for each layer [3] (dubbed $\mathcal{W}^+$ ). Going beyond $\mathcal{W}^+$ has also been explored in [46] who propose to fine-tune generator parameters, and [44] which uses different latent codes for each convolution filter. We build on these techniques by developing an inversion method designed specifically for robust image restoration.
42
+
43
+ Generative priors for unsupervised image restoration Since the initial work of Bora et al. [6], multiple papers have used generative priors for unsupervised image restoration [21, 25]. Pan et al. [42] obtained high-resolution results with BigGAN [7]. They relax their latent inversion method by introducing a coarse-to-fine, layer-by-layer generator fine-tuning. Our method differs in that it addresses robustness and compositionality.
44
+
45
+ StyleGAN image restoration Inversion methods designed for editing cannot be applied directly for restoration because loss functions are applied to degraded targets rather than clean images. Previous works address this with additional regularization terms and/or optimizer constraints. Of note, PULSE [40] proposed StyleGAN inversion for image upsampling, using a regularizer minimizing the latent extension to $\mathcal{W}^+$ and a spherical optimization technique on a Gaussian approximation of $\mathcal{W}$ . While PULSE succeeds in preserving high realism, it obtains low fidelity when the downsampling factor is small. ILO [15] uses similar regularization techniques, and further extends the latent space by optimizing intermediate feature maps in a progressive fashion (layer-by-layer) while constraining the solutions to sparse deviations from the range. More recently, SGILO [14] proposes to replace this sparsity constraint and instead trains a score-based model to learn the distribution of the outputs of an intermediate layer in the StyleGAN generator. BRGM [39] frames GAN inversion as Bayesian inference and estimates the maximum a-posteriori over the input latent vector that generated the reconstructed image given different regularization constraints. Its followup, L-BRGM [13], jointly optimizes the $\mathcal{Z}$ (before the mapping network) and $\mathcal{W}$ spaces, further improving quality. In contrast, our work achieves robustness by
46
+
47
+ ![](images/01269686a92a2f9db1a5fa56938d162ae30829b740759797a4d99eed45c43b57.jpg)
48
+ Figure 2. Unsupervised image restoration with StyleGAN attempts to restore a degraded target image $\pmb{y} = f(\pmb{y}_{\mathrm{clean}})$ by finding a generated image $\pmb{x}_{\mathrm{clean}}$ that matches this target once also degraded $\pmb{x} = \hat{f}(\pmb{x}_{\mathrm{clean}})$ (sec. 3); here, $\hat{f}$ is a differentiable approximation to the degradation function $f$ and $\pmb{y}_{\mathrm{clean}}$ is the (unknown) ground truth image.
49
+
50
+ avoiding any such regularization losses.
51
+
52
+ Other similar works StyleGAN inversion was used to restore old photographs in [38]. Their method also restores a mixture of degradations, but is designed for real degradations (not robustness) and uses a task-specific encoder, while ours method is fully unsupervised. Multiple supervised methods [22, 34, 36, 56, 64, 65] focus on faces only, whereas our method does not use face-specific loss functions [16], is fully unsupervised, and works across a variety of image domains.
53
+
54
+ Diffusion models Denoising diffusion probabilistic models (DDPMs) [53], trained on extremely large datasets (e.g., billions of images for [45, 47, 49]), have recently been shown to outclass GANs when it comes to image generation quality [17]. New approaches have been proposed to tackle image restoration using DDPMs, e.g., Palette [48] obtains excellent results by training the same model for several different tasks. Denoising diffusion restoration models (DDRM) [32] show that pretrained DDPMs can be used for unsupervised restoration tasks, but are limited to linear inverse problems with additive Gaussian noise. In contrast, our method is more flexible since it only requires a differentiable approximation of the degradation function, which can be non-linear.
55
+
56
+ # 3. Image restoration with StyleGAN
57
+
58
+ As illustrated in fig. 2, StyleGAN inversion attempts to recover an image $\pmb{x}_{\mathrm{clean}}$ that best matches an (unknown) ground truth image $\pmb{y}_{\mathrm{clean}}$ . To this end, we aim to search for a latent code $\pmb{w} \in \mathcal{W}$ such that $\pmb{x}_{\mathrm{clean}} = G(\pmb{w})$ best matches $\pmb{y}_{\mathrm{clean}}$ under some image distance function $\ell$ . The resulting minimization problem,
59
+
60
+ $$
61
+ \boldsymbol {w} = \underset {\tilde {\boldsymbol {w}} \in \mathcal {W}} {\arg \min } \ell \left(G (\tilde {\boldsymbol {w}}), \boldsymbol {y} _ {\text {c l e a n}}\right), \tag {1}
62
+ $$
63
+
64
+ ![](images/e861fc957b4dc5c6debbbb7fa15903e15ebd59480c06371acceaff9b65850d54.jpg)
65
+ Figure 3. Our 3-phase latent extension (sec. 4.1) approach. (a) Phase I uses a global latent code $\pmb{w}$ , resulting in a prediction $\pmb{x}$ . (b) Phase II performs layer-wise latent expansion, resulting in matrix $\pmb{w}^{+}$ and in a prediction $\pmb{x}^{+}$ . (c) Phase III performs filter-wise latent expansion, resulting in tensor $\pmb{w}^{++}$ and the final prediction $\pmb{x}^{++}$ .
66
+
67
+ can be solved by gradient descent. In image restoration, the ground truth image $\pmb{y}_{\mathrm{clean}}$ is unknown: we are instead given a target image $\pmb{y} = f(\pmb{y}_{\mathrm{clean}})$ , the result of a (non-injective, potentially non-differentiable) degradation function $f$ . Assuming that a differentiable approximation $\hat{f}\approx f$ can be constructed, restoration is performed by solving for
68
+
69
+ $$
70
+ \boldsymbol {w} = \underset {\tilde {\boldsymbol {w}} \in \mathcal {W}} {\arg \min } \ell \left(\hat {f} (G (\tilde {\boldsymbol {w}})), \boldsymbol {y}\right). \tag {2}
71
+ $$
72
+
73
+ This can be generalized to compositions of $k$ different.
74
+ degradations,i.e., $\{f_i\}_{i = 1}^k$ ,by solving for
75
+
76
+ $$
77
+ \boldsymbol {w} = \underset {\tilde {\boldsymbol {w}} \in \mathcal {W}} {\arg \min } \ell \left(\left[ \hat {f} _ {k} \circ \dots \circ \hat {f} _ {2} \circ \hat {f} _ {1} \right] (G (\tilde {\boldsymbol {w}})), \boldsymbol {y}\right), \tag {3}
78
+ $$
79
+
80
+ where $\circ$ is the composition operator. Here, it is assumed that each subfunction $f_{i}$ has a differentiable approximation $\hat{f}_i$ , and that the order of composition is known.
81
+
82
+ As described in sec. 2, this naive approach finds solutions that have high realism (look like real faces), but low fidelity (do not match the degraded target). Fidelity is most commonly improved by 1) performing latent extension [3], that is, solving for $\boldsymbol{w}^{+} \in \mathcal{W}^{+}$ which has many more degrees of freedom; and 2) using a better performing optimizer like Adam [33]. These techniques both improve fidelity but also damage realism, motivating the use of regularization losses which must carefully be adjusted for different tasks [13, 39, 40]. In the next section, we show how the choice of latent extension and optimizer avoids the need for these regularizers.
83
+
84
+ # 4. Robust StyleGAN inversion
85
+
86
+ We propose an inversion method designed specifically for robust image restoration, which revisits each aspect
87
+
88
+ of the unsupervised StyleGAN optimization pipeline, namely latent extension, optimization, and loss functions.
89
+
90
+ # 4.1. Robust latent extension
91
+
92
+ Inspired by the intuition that initialization is the best regularization [46], we propose a three-phase latent extension, where each phase is initialized by the outcome of the previous one, see fig. 3. Given a pretrained StyleGAN2 [31] model with $N_{L}$ layers in the generator, we denote style vectors $s_i^l \in \mathbb{R}^{512}$ at layer $l \in [1,N_L]$ used to modulate convolution weights $\pmb{\theta}^{l} \in \mathbb{R}^{512 \times 512}$ . Assuming $1 \times 1$ filters to simplify notation<sup>1</sup>, each feature map pixel $p^l \in \mathbb{R}^{512}$ is processed by
93
+
94
+ $$
95
+ p _ {i} ^ {l + 1} = \left(\boldsymbol {\theta} _ {i} ^ {l} \odot \boldsymbol {s} _ {i} ^ {l}\right) \boldsymbol {p} ^ {l}, \tag {4}
96
+ $$
97
+
98
+ producing output feature $p_i^{l + 1}\in \mathbb{R}$ where $i\in [0,512]$ .
99
+
100
+ Phase I performs global style modulation (fig. 3-(a)), and solves for a single latent vector $\boldsymbol{w} \in \mathcal{W} = \mathbb{R}^{512}$ shared across all layers as in eq. (3). Prior to optimization, $\boldsymbol{w}$ is initialized to the mean of $\mathcal{W}$ over the training set, i.e., $\mathbb{E}_{\tilde{\boldsymbol{w}} \in \mathcal{W}}[\tilde{\boldsymbol{w}}]$ . Here, the style modulation vector $s_i^l$ in eq. (4) can be written as
101
+
102
+ $$
103
+ \boldsymbol {s} _ {i} ^ {l} = \mathcal {A} ^ {l} (\boldsymbol {w}), \tag {5}
104
+ $$
105
+
106
+ where $\mathcal{A}^l$ is the corresponding affine projection layer (which multiplies by a weight matrix and adds a bias).
107
+
108
+ Phase II performs layer-wise latent extension [3] (fig. 3-(b)), and solves for a latent matrix $\boldsymbol{w}^{+} \in \mathcal{W}^{+} = \mathbb{R}^{N_{L} \times 512}$ . Each row of $\boldsymbol{w}^{+}$ is initialized to $\boldsymbol{w}$ . Style modulation for this phase becomes
109
+
110
+ $$
111
+ \boldsymbol {s} _ {i} ^ {l} = \mathcal {A} ^ {l} \left(\boldsymbol {w} _ {l} ^ {+}\right). \tag {6}
112
+ $$
113
+
114
+ Phase III performs filter-wise latent extension (fig. 3-c), and solves for a latent tensor $\boldsymbol{w}^{++} \in \mathcal{W}^{++} = \mathbb{R}^{N_F \times N_L \times 512}$ , where a different latent code is used for each convolution filter and where $N_F$ is the number of such filters. Each submatrix of $\boldsymbol{w}^{++}$ is initialized to $\boldsymbol{w}^+$ . Style modulation for this phase becomes
115
+
116
+ $$
117
+ \boldsymbol {s} _ {i} ^ {l} = \mathcal {A} ^ {l} \left(\boldsymbol {w} _ {i, l} ^ {+ +}\right). \tag {7}
118
+ $$
119
+
120
+ # 4.2. Robust optimization
121
+
122
+ While most inversion approaches use the Adam [33] optimizer to solve eq. (3), we instead propose coupling our 3-phase latent extension with a weaker optimizer, normalized gradient decent (NGD) [58], a simple variant of SGD that normalizes the gradient before each step:
123
+
124
+ $$
125
+ \bar {\nabla} _ {\boldsymbol {w}} \ell (\boldsymbol {w}) = \frac {\nabla_ {\boldsymbol {w}} \ell (\boldsymbol {w})}{\| \nabla_ {\boldsymbol {w}} \ell (\boldsymbol {w}) \| _ {2}}, \tag {8}
126
+ $$
127
+
128
+ Algorithm 1: Robust StyleGAN inversion.
129
+ Output: restored image $x^{++}$ # Phase I
130
+ 1 $\pmb {w} = \mathbb{E}_{\tilde{\pmb{w}}\in \mathcal{W}}[\tilde{\pmb{w}} ]$
131
+ 2 for 1 to 150 do
132
+ 3 $\begin{array}{r}\pmb {x}\gets G(\pmb {w})\\ \pmb {w}\gets 0.08\bar{\nabla}_{\pmb{w}}\ell (\hat{f} (\pmb {x}),\pmb {y}) \end{array}$
133
+ 4 Phase II
134
+ 5 $w^{+} =$ repeat $(\pmb {w},N_L)$
135
+ 6 for 1 to 150 do
136
+ 7 $\begin{array}{r}\pmb{x}^{+}\gets G^{+}(\pmb{w}^{+}) \end{array}$
137
+ 8 $\begin{array}{r}\pmb {w}^{+}\gets 0.02\bar{\nabla}_{\pmb{w}}\ell (\hat{f} (\pmb{x}^{+}),\pmb {y}) \end{array}$ # Phase III
138
+ 9 $w^{+ + } =$ repeat $(\pmb {w}^{+},N_{F})$
139
+ 10 for 1 to 150 do
140
+ 11 $\begin{array}{r}\pmb{x}^{+ + }\gets G^{+ + }(\pmb{w}^{+ + })\\ \pmb{w}^{+ + }\gets 0.005\bar{\nabla}_{\pmb{w}^{+ + }}\ell (\hat{f} (\pmb{x}^{+ + }),\pmb {y}) \end{array}$
141
+ 12
142
+ 13 return $x^{+ + }$
143
+
144
+ Here, the gradient $\bar{\nabla}_{\boldsymbol{w}}\ell (\boldsymbol {w})$ is explicitly set to 0 when $||\nabla_{\boldsymbol{w}}\ell (\boldsymbol {w})||_2 = 0$ . Following latent extension, we normalize each latent code separately (i.e. each row of $\boldsymbol{w}^{+}$ and $\boldsymbol{w}^{+ + }$ ). NGD conserves loss scale invariance, a key property of Adam that avoids learning rate adjustments following loss function changes.
145
+
146
+ # 4.3. Robust loss function
147
+
148
+ The staple loss function in StyleGAN inversion is the LPIPS [63] perceptual loss combined with a L2 or L1 [4] pixel-wise loss. We found a multiresolution loss function to be much more robust, and use:
149
+
150
+ $$
151
+ \ell_ {\mathrm {M R}} (\boldsymbol {x}, \boldsymbol {y}) = \sum_ {i = 1} ^ {k} \ell_ {\mathrm {L P I P S}} \left(\phi \left(\boldsymbol {x}, 2 ^ {i}\right), \phi \left(\boldsymbol {y}, 2 ^ {i}\right)\right), \tag {9}
152
+ $$
153
+
154
+ where $\phi (\cdot ,2^i)$ downsamples by a factor of $2^{i}$ using average pooling, and we set $k = 6$ for image resolution of $1024^{2}$ . All resolutions are weighted equally, giving our final loss function $\ell = \lambda_{\mathrm{L1}}\ell_{\mathrm{L1}} + \ell_{\mathrm{MR}}$ with $\lambda_{\mathrm{L1}} = 0.1$ .
155
+
156
+ # 4.4. Overall algorithm
157
+
158
+ Alg. 1 provides a detailed pseudo-code of our method. $G^{+}$ and $G^{++}$ denote the synthesis network $G$ after modification to accept $w^{+} \in \mathcal{W}^{+}$ and $w^{++} \in \mathcal{W}^{++}$ , respectively. Hyperparameters like learning rates and number of steps are explicitly provided since they are held constant across all tasks. Their values were found by cross-validation on a subset of the FFHQ training set.
159
+
160
+ # 5. Benchmarking restoration robustness
161
+
162
+ This section first describes the proposed degradation models used in all experiments as well as their respec
163
+
164
+ tive differentiable approximations. It then explains how compositions of tasks were formed.
165
+
166
+ # 5.1. Individual degradations
167
+
168
+ Experiments are performed on four common sources of image degradation: upsampling, inpainting, denoising, and deartifacting. Synthetic models are used to facilitate comparison at different degradation levels. Each degradation is tested at five levels, which are referred to as extra-small (XS), small (S), medium (M), large (L), and extra-large (XL). Parameters for each degradation level will be provided below for each task in that same order. Fig. 4 shows examples of all four degradations at the XS and XL levels, see supp.
169
+
170
+ Upsampling targets $\pmb{y}$ are produced by downsampling ground truth images $\pmb{y}_{\mathrm{clean}}$ by integer factors $k_{\mathrm{down}} \in \{2,4,8,16,32\}$ . Downsampling filters are uniformly sampled from the commonly used bilinear, bicubic, and Lanczos filters, which provides a coarse but broad range of aliasing profiles. During inversion, average pooling is used as the approximation $\hat{f}$ in all cases.
171
+
172
+ Inpainting aims to predict missing regions in an image. To test a variety of masking conditions, random masks are generated by drawing $k_{\mathrm{strokes}} \in \{1,5,9,13,17\}$ random strokes of width $0.08r$ where $r$ is the image resolution, each stroke connecting two random points situated in the outer thirds of the image. In this task, $\hat{f} = f$ i.e. we assume that the mask is known. To avoid the bias introduced by black pixels when evaluating the LPIPS loss, identical noise of distribution $\mathcal{N}(0.5,1)$ is added to the masked region of both the prediction and the target before computing LPIPS.
173
+
174
+ Denoising targets are generated by using a mixture of Poisson and Bernoulli noise, simulating common sources of noise in cameras, namely shot noise and dead (or hot) pixels respectively. This is a challenging scenario because unlike Gaussian noise, Poisson noise is non-additive and signal-dependent, while Bernoulli noise is biased. Additionally, both are non-differentiable. We used the parameters $k_{\mathrm{p}} \in \{96, 48, 24, 12, 6\}$ and $k_{\mathrm{b}} \in \{0.04, 0.08, 0.16, 0.32, 0.64\}$ where $k_{\mathrm{p}}$ gives the most likely value added to a pixel (independently for each channel) according to a Poisson distribution, and $k_{\mathrm{b}}$ the probability that all channels of a pixel are replaced with black. We also account for clamping during image serialization, resulting in the overall noise model:
175
+
176
+ $$
177
+ \frac {\operatorname {c l a m p} \left(p ^ {\prime} \cdot m\right)}{2 5 5}, \text {w h e r e} p ^ {\prime} \sim \mathcal {P} \left(k _ {\mathrm {p}} p\right), m \sim \mathcal {B} \left(k _ {\mathrm {b}}\right), \tag {10}
178
+ $$
179
+
180
+ for a ground truth image pixel value $p \in [0.0, 1.0]$ and where $\mathrm{clamp}(\cdot)$ saturates all values outside of [0, 255]. Note that $p'$ can only take discrete values. For the differentiable approximation $\hat{f}$ , we replace the (non-differentiable, discrete) Poisson noise with a Gaussian ap
181
+
182
+ proximation $\mathcal{N}(k_{\mathrm{p}}p - 0.5,\sqrt{k_{\mathrm{p}}p})$ and treat the Bernoulli noise as an unknown mask. We also use a surrogate gradient for clamp:
183
+
184
+ $$
185
+ \frac {d}{d p} \operatorname {c l a m p} (p) \approx \frac {d}{d p} \operatorname {s i g m o i d} (2 \cdot (p - 0. 5)). \tag {11}
186
+ $$
187
+
188
+ Deartifacting is performed on JPEG images compressed with libjpeg [2, 12] at quality levels $k_{\mathrm{jpeg}} \in \{18, 15, 12, 9, 6\}$ . The lossy part of JPEG compression can be expressed as
189
+
190
+ $$
191
+ \text {q u a n t i z e} \circ \mathrm {D C T} \circ \text {t o} _ {\text {b l o c k s}} \circ \text {c h r o m a} _ {\text {s u b s a m p l e}}. \tag {12}
192
+ $$
193
+
194
+ Here, 4:2:2 chroma subsampling is used. This final quantization step rounds each DCT component $c_{i,j}^{(k)}$ positioned in the $k^{\text{th}}$ block at indices $i,j$ with
195
+
196
+ $$
197
+ \operatorname {q u a n t i z e} \left(c _ {i, j} ^ {(k)}\right) = \left\lfloor c _ {i, j} ^ {(k)} / Q _ {i, j} \right], \tag {13}
198
+ $$
199
+
200
+ where $Q_{i,j} \in [1,255] \subset \mathbb{N}$ is the corresponding value from the quantization table. Rounding is the only non-differentiable part of this chain (the lossless part of JPEG compression does not need to be differentiated). We read the quantization tables from file headers and use Lomnitz's implementation [1] of [52], which proposes differentiable rounding:
201
+
202
+ $$
203
+ \lfloor p \rceil \approx \lfloor p \rceil + (p - \lfloor p \rceil) ^ {3}. \tag {14}
204
+ $$
205
+
206
+ We use this approximation to define a surrogate gradient (keeping the forward pass exact), and interpolate with a straight-through gradient estimator at $\alpha = 0.8$ :
207
+
208
+ $$
209
+ \frac {d}{d p} \lfloor p \rceil \approx \frac {d}{d p} \left[ (1 - \alpha) p + \alpha (p - \lfloor p \rceil) ^ {3} \right]. \tag {15}
210
+ $$
211
+
212
+ # 5.2. Composed degradations
213
+
214
+ Eq. (3) assumes knowledge of the order in which degradations are applied. In this work, the following restoration order is used:
215
+
216
+ $$
217
+ \text {i n p a i n t} \circ \text {d e a r i f a c t} \circ \text {d e n o i s e} \circ \text {u p s a m p l e}. \tag {16}
218
+ $$
219
+
220
+ Task compositions are created with (subsequences of) this ordering. For example, upsampling and deartifacting form a composition of length 2. All compositions were formed with tasks at degradation level medium (M).
221
+
222
+ # 6. Comparison to StyleGAN-based methods
223
+
224
+ # 6.1. Models, datasets, and metrics
225
+
226
+ The PyTorch implementation [41] of StyleGAN2-ADA [28] is used in all experiments. In particular, we use the model pretrained on the FFHQ dataset [30] at $1024 \times 1024$ resolution. Since the model is trained on
227
+
228
+ ![](images/8fc68f163d70564b3fc207f5a33bdf4e40fef58e95f288944860b7ff476552e3.jpg)
229
+ Figure 4. Randomly selected qualitative restoration results on each independent task, for both extra-small and extra-large levels of degradation. Top to bottom: (degraded) target, ours, L-BRGM [13], and PULSE [40]. Left to right: upsampling, inpainting, denoising, and deartifacting. Our method shows better robustness to variations in levels of degradation. See supp. for more results.
230
+
231
+ ![](images/3a025d545e7280eb5becaf2be500242b31cd4fcfcd71cbf8e6b3cfe88ca87fbb.jpg)
232
+
233
+ ![](images/7eaa2780704093d6a469d855c3522662d736056400b15c0844f512de6e9ec19a.jpg)
234
+
235
+ ![](images/da7ae62f29d1cacfb9dd359a4aba43e4c8a300f78331604c4a13b2c13ccea84b.jpg)
236
+
237
+ the entire FFHQ dataset, we gathered an additional 100 test images, dubbed “FFHQ extras” (FFHQ-X), using the same alignment script. Overlap with FFHQ was avoided by filtering by date (see supp.).
238
+
239
+ Performance was measured via accuracy, realism, and fidelity metrics. Accuracy was measured with the LPIPS [63] between the predictions and the ground truth images. Realism was measured using the PatchFID [8] (pFID): FID [24] measured on image patches. More specifically, we extracted 1000 random crops of size $128 \times 128$ per image, computing the FID of all crops from predictions to all crops from our FFHQ-X dataset. Finally, fidelity between degraded predictions and targets was measured using the LPIPS; this was done by degrading predictions with the true degradation to ignore the effect of differentiable approximations.
240
+
241
+ # 6.2. Baselines
242
+
243
+ Our method is compared on FFHQ-X against two unsupervised StyleGAN image restoration techniques, PULSE [40] and the current state of the art L-BRGM [13]. The same pretrained network is used for all approaches.
244
+
245
+ PULSE [40] We adapted PULSE to use StyleGAN2-ADA and to tasks other than upsampling by increasing the number of steps from 100 to 250 and adjusting the learning rate to 0.8 (manual inspection on validation data) to increase its robustness to novel tasks.
246
+
247
+ L-BRGM [13] We adjust its learning rate $\lambda_{\mathrm{lr}} = 0.05$ by manual inspection. L-BRGM uses best-of- $k$ random
248
+
249
+ initialization; because initialization is difficult under high degradations, we doubled the number of samples from 100 to 200 and added a 0.8 truncation [30] to avoid outliers. L-BRGM also performs early stopping by inspecting the ground truth; for a fair comparison, we instead fixed the number of steps to 500 as in BRGM [39].
250
+
251
+ Hyperparameter optimization for baselines We optimize the accuracy metric for the baselines hyperparameters for each (task, level) pair individually directly on the $FFHQ-X$ test set (see supp. for more details). In addition, baselines employ the ground truth Bernoulli mask (see sec. 5.1) in their differentiable approximation of eq. (10). Doing so was necessary for good performance.
252
+
253
+ Our method We adjust hyperameters (learning rate and number of iterations, see alg. 1) once on validation data, then run our benchmark on the test set with no pertask or per-level hyperparameter adjustment: our method uses the same hyperparameters across all tasks and levels. Our method is not provided the Bernoulli mask. The total number of steps (450) is less than the 500 used for L-BRGM.
254
+
255
+ # 6.3. Experimental results
256
+
257
+ Robustness to degradation levels Our benchmark is run on our FFQH-X test set (sec. 6.1) on all five levels of degradation for each individual task, and quantitative results are presented in tab. 1. Recall from sec. 6.2 that our method, which employs a single set of hyperparameters across all experiments, is pitted against baselines
258
+
259
+ <table><tr><td></td><td colspan="3">Accur. (LPIPS) ↓</td><td colspan="3">Fidelity (LPIPS) ↓</td><td colspan="3">Realism (pFID) ↓</td></tr><tr><td></td><td colspan="3">PULS L-BRG Ours</td><td colspan="3">PULS L-BRG Ours</td><td colspan="3">PULS L-BRG Ours</td></tr><tr><td colspan="10">Upsampling (bilinear, bicubic or Lanczos)</td></tr><tr><td>XS</td><td>.493</td><td>.407</td><td>.414</td><td>.432</td><td>.295</td><td>.313</td><td>44.5</td><td>23.6</td><td>17.0</td></tr><tr><td>S</td><td>.492</td><td>.412</td><td>.449</td><td>.353</td><td>.140</td><td>.239</td><td>34.3</td><td>25.5</td><td>22.0</td></tr><tr><td>M</td><td>.495</td><td>.458</td><td>.472</td><td>.261</td><td>.124</td><td>.172</td><td>29.3</td><td>35.4</td><td>22.3</td></tr><tr><td>L</td><td>.501</td><td>.487</td><td>.490</td><td>.185</td><td>.129</td><td>.127</td><td>21.9</td><td>26.0</td><td>20.9</td></tr><tr><td>XL</td><td>.512</td><td>.506</td><td>.514</td><td>.083</td><td>.095</td><td>.090</td><td>24.9</td><td>21.3</td><td>21.3</td></tr><tr><td colspan="10">Denoising (clamped Poisson and Bernoulli mixture)</td></tr><tr><td>XS</td><td>.501</td><td>.440</td><td>.425</td><td>.275</td><td>.152</td><td>.156</td><td>56.1</td><td>27.2</td><td>18.5</td></tr><tr><td>S</td><td>.499</td><td>.450</td><td>.434</td><td>.252</td><td>.138</td><td>.140</td><td>53.7</td><td>28.6</td><td>19.1</td></tr><tr><td>M</td><td>.500</td><td>.465</td><td>.446</td><td>.224</td><td>.155</td><td>.130</td><td>54.5</td><td>22.1</td><td>19.8</td></tr><tr><td>L</td><td>.501</td><td>.481</td><td>.457</td><td>.185</td><td>.138</td><td>.110</td><td>56.4</td><td>24.6</td><td>19.2</td></tr><tr><td>XL</td><td>.504</td><td>.511</td><td>.474</td><td>.134</td><td>.110</td><td>.084</td><td>49.4</td><td>25.1</td><td>17.9</td></tr><tr><td colspan="10">Deartifacting (JPEG compression)</td></tr><tr><td>XS</td><td>.498</td><td>.442</td><td>.432</td><td>.404</td><td>.341</td><td>.349</td><td>52.3</td><td>26.3</td><td>14.8</td></tr><tr><td>S</td><td>.497</td><td>.448</td><td>.437</td><td>.398</td><td>.352</td><td>.350</td><td>49.6</td><td>22.4</td><td>15.4</td></tr><tr><td>M</td><td>.498</td><td>.461</td><td>.445</td><td>.413</td><td>.357</td><td>.357</td><td>33.2</td><td>24.1</td><td>15.4</td></tr><tr><td>L</td><td>.500</td><td>.475</td><td>.460</td><td>.395</td><td>.367</td><td>.374</td><td>46.9</td><td>25.2</td><td>16.0</td></tr><tr><td>XL</td><td>.508</td><td>.503</td><td>.490</td><td>.427</td><td>.418</td><td>.412</td><td>30.8</td><td>22.1</td><td>18.7</td></tr><tr><td colspan="10">Inpainting (random strokes)</td></tr><tr><td>XS</td><td>.498</td><td>.409</td><td>.378</td><td>.464</td><td>.374</td><td>.348</td><td>46.9</td><td>24.4</td><td>12.9</td></tr><tr><td>S</td><td>.501</td><td>.425</td><td>.387</td><td>.356</td><td>.287</td><td>.264</td><td>42.3</td><td>27.2</td><td>14.2</td></tr><tr><td>M</td><td>.509</td><td>.438</td><td>.396</td><td>.283</td><td>.227</td><td>.206</td><td>38.5</td><td>30.1</td><td>14.5</td></tr><tr><td>L</td><td>.513</td><td>.452</td><td>.409</td><td>.231</td><td>.184</td><td>.163</td><td>32.6</td><td>33.1</td><td>15.3</td></tr><tr><td>XL</td><td>.524</td><td>.460</td><td>.422</td><td>.187</td><td>.157</td><td>.132</td><td>36.2</td><td>25.2</td><td>15.9</td></tr></table>
260
+
261
+ with hyperparameters optimized for each task and level individually. Despite this obvious impediment, our proposed method outperforms baselines in the vast majority of scenarios. In particular, it reaches better accuracy in the majority of tasks against baselines optimized for accuracy, and realism is strictly better. Fig. 4 illustrates the improved visual quality obtained with our method.
262
+
263
+ Robustness to compositions Robustness to compositions of degradations was evaluated by forming all possible subsequences of degradations (order in eq. (16)) at medium level. Quantitative results are aggregated in tab. 2, while fig. 5 presents qualitative examples. Our results show that both high fidelity and high realism are maintained with our method, while the baselines struggle under varied degradations.
264
+
265
+ # 7. Comparison to diffusion models
266
+
267
+ We compare our approach to the recent "Denoising diffusion restoration model" (DDRM) [32], which also does not require problem-specific supervised training and leverages a pre-trained (DDPM [53]) model. Importantly, DCRM is designed only for linear inverse problems with additive Gaussian noise, where the degradation model is
268
+
269
+ Table 1. Quantitative comparison on individual tasks against baselines ("PULS" [40] and "L-BRG" [13]). Baselines are optimized for accuracy on each row separately, while ours uses the same set of hyperparameters across all. Despite this clear handicap, our method is either on par or outperforms baselines. Results color-coded as best and second best.
270
+
271
+ <table><tr><td></td><td colspan="3">Accur. (LPIPS) ↓</td><td colspan="3">Fidelity (LPIPS) ↓</td><td colspan="3">Realism (pFID) ↓</td></tr><tr><td></td><td colspan="3">PULS L-BRG OUR5</td><td colspan="3">PULS L-BRG OUR5</td><td colspan="3">PULS L-BRG OUR5</td></tr><tr><td colspan="10">2 degradations</td></tr><tr><td>NA</td><td>.517</td><td>.485</td><td>.459</td><td>.328</td><td>.301</td><td>.290</td><td>43.4</td><td>24.2</td><td>17.3</td></tr><tr><td>AP</td><td>.511</td><td>.478</td><td>.457</td><td>.270</td><td>.231</td><td>.204</td><td>29.7</td><td>17.6</td><td>17.0</td></tr><tr><td>UA</td><td>.510</td><td>.518</td><td>.508</td><td>.307</td><td>.348</td><td>.287</td><td>23.7</td><td>20.5</td><td>19.7</td></tr><tr><td>NP</td><td>.511</td><td>.480</td><td>.458</td><td>.125</td><td>.079</td><td>.062</td><td>47.0</td><td>20.9</td><td>19.2</td></tr><tr><td>UN</td><td>.501</td><td>.519</td><td>.511</td><td>.178</td><td>.149</td><td>.153</td><td>33.4</td><td>26.2</td><td>21.1</td></tr><tr><td>UP</td><td>.510</td><td>.478</td><td>.485</td><td>.140</td><td>.061</td><td>.089</td><td>23.9</td><td>35.3</td><td>20.7</td></tr><tr><td colspan="10">3 degradations</td></tr><tr><td>UNP</td><td>.510</td><td>.526</td><td>.507</td><td>.086</td><td>.062</td><td>.051</td><td>28.5</td><td>22.1</td><td>20.1</td></tr><tr><td>UAP</td><td>.525</td><td>.523</td><td>.513</td><td>.205</td><td>.154</td><td>.119</td><td>23.0</td><td>18.4</td><td>20.5</td></tr><tr><td>UNA</td><td>.521</td><td>.535</td><td>.533</td><td>.265</td><td>.310</td><td>.290</td><td>26.2</td><td>20.7</td><td>22.8</td></tr><tr><td>NAP</td><td>.526</td><td>.502</td><td>.470</td><td>.210</td><td>.197</td><td>.160</td><td>38.7</td><td>18.4</td><td>18.5</td></tr><tr><td colspan="10">4 degradations</td></tr><tr><td>UNAP</td><td>.533</td><td>.546</td><td>.525</td><td>.192</td><td>.177</td><td>.131</td><td>25.5</td><td>21.1</td><td>21.8</td></tr></table>
272
+
273
+ Table 2. Quantitative comparison on composed tasks (at medium levels) against baselines ("PULS" [40] and "L-BRG" [13]). Baselines are optimized for accuracy on each row separately, while ours uses the same set of hyperparameters across all. Despite this clear handicap, our method is either on par or outperforms baselines. Acronyms indicate restorations: Upsampling, deNoising, deArtifacting, and/or inPainting. Results color-coded as best and second best.
274
+
275
+ <table><tr><td></td><td colspan="2">Accur. (LPIPS) ↓</td><td colspan="2">Realism (pFID) ↓</td></tr><tr><td></td><td>DDRM</td><td>OURS</td><td>DDRM</td><td>OURS</td></tr><tr><td colspan="5">Exact Upsampling (known average pooling)</td></tr><tr><td>0.412</td><td>0.443</td><td></td><td>44.75</td><td>18.92</td></tr><tr><td colspan="5">Inexact Upsampling (unknown nearest-neighbor)</td></tr><tr><td>0.559</td><td>0.539</td><td></td><td>67.06</td><td>32.81</td></tr><tr><td colspan="5">Linear Denoising (additive Gaussian noise)</td></tr><tr><td>0.336</td><td>0.452</td><td></td><td>43.28</td><td>19.25</td></tr><tr><td colspan="5">Nonlinear Denoising (clamped additive Gaussian noise)</td></tr><tr><td>0.530</td><td>0.450</td><td></td><td>156.1</td><td>19.06</td></tr><tr><td colspan="5">Inpainting (masking random strokes)</td></tr><tr><td>0.289</td><td>0.301</td><td></td><td>26.12</td><td>25.71</td></tr><tr><td colspan="5">Uncropping (top left corner unmasked)</td></tr><tr><td>0.468</td><td>0.418</td><td></td><td>34.33</td><td>17.645</td></tr></table>
276
+
277
+ Table 3. Quantitative comparison to DDRM [32] on accuracy and realism (see sec. 7). Results are color-coded as best .
278
+
279
+ perfectly known. Therefore, for each of the upsampling, denoising and inpainting tasks, we experiment with both: 1) a linear inverse problem with known parameters; and 2) an adjusted version which is either non-linear, where the extract degradation model is unknown, or a more extreme version of the problem. Specifically, we test: average pooling (known, linear) vs nearest-neighbor (unknown) $32 \times$ (XL) upsampling; additive Gaussian noise (linear) vs clamped (to [0, 255]) additive Gaussian noise (non-linear) with $\sigma^2 = 0.8$ (XL) for both; and masking strokes (at XL level) vs uncropping the top-left corner (XXL). DDRM is available on CelebA (not FFHQ) at $256 \times 256$ , so we use the corresponding StyleGAN2-ADA pretrained
280
+
281
+ ![](images/0dc25b0c6d19b5a73690935d182ba793285ca28427df5f89d63b4ab736204646.jpg)
282
+ Figure 5. Qualitative comparisons on all possible compositions of the four tasks given medium levels of degradation. Samples are chosen randomly, and acronyms indicate the degradations: Upsampling, inPainting, deNoising and/or deArtifacting (sec. 6.3).
283
+
284
+ ![](images/6e0f0b184a35cae0ad3bec11f10fc46215c8222bb5c2e5711e61f848e2ef27c7.jpg)
285
+
286
+ ![](images/9788772c3424bc9e816961daf2dbe2abb2a33c464b3ac9a2acead706f2fb127d.jpg)
287
+ Figure 6. Results on LSUN horses and churches [60], AFHQ-v2 Dogs [11] and BreCaHAD [5] at resolution $256 \times 256$ for small upsampling (U), small denoising (N), medium deartifacting (A), and medium inpainting (P); the left/right halves are the target/restored images. See supp. for details and more results.
288
+
289
+ model. Due to the reduced resolution, pFID is measured on 250 patches of size $64 \times 64$ .
290
+
291
+ Interestingly, we find that DDRM reproduces the targets almost perfectly (fidelity is near zero, see supp. for details and qualitative results), to the point where it overfits to any deviation between its degradation model and the true degradation. Tab. 3 presents the accuracy and realism metrics for all the aforementioned tasks. While the accuracy favors DDRM for the tasks which perfectly
292
+
293
+ match its assumptions, our approach outperforms it when this is not the case. In addition, our method provides significantly better realism across all scenarios. Overall, our approach yields high resolution results with more detail, can adapt to non-linear tasks, and it can also cope with larger degradations.
294
+
295
+ # 8. Discussion
296
+
297
+ **Limitations** Our method is not restricted to faces and works with any pretrained StyleGAN, as demonstrated in fig. 6. However, it is inherently limited to the domain learned by the GAN. Training StyleGAN on a large unstructured dataset like ImageNet is difficult but recent papers [27,50,55] have attained some success—applying our method on some of these large-scale models is a promising research avenue. Our method also faces the same ethical considerations as the GAN it relies on. Finally, as with other methods [13,32,40], we also require knowledge of an (approximate) degradation function.
298
+
299
+ Conclusion This paper presents a method which makes StyleGAN-based image restoration robust to both variability in degradation levels and to compositions of different degradations. Our proposed method relies on a conservative optimization procedure over a progressively richer latent space and avoids regularization terms altogether. Using a single set of hyperparameters, we obtain competitive and even state-of-the-art results on several challenging scenarios when compared to baselines that are optimized for each task/level individually.
300
+
301
+ Acknowledgements This research was supported by NSERC grant RGPIN-2020-04799, a MITACS Globalink internship to Y. Poirier-Ginter, and by the Digital Research Alliance Canada.
302
+
303
+ # References
304
+
305
+ [1] Diffjpeg. https://github.com/mlomnitz/DiffJPEG.
306
+ [2] Libjpeg. https://libjpeg-turbo.org.
307
+ [3] Rameen Abdal, Yipeng Qin, and Peter Wonka. Image2StyleGAN: How to embed images into the StyleGAN latent space? In Int. Conf. Comput. Vis., 2019.
308
+ [4] Rameen Abdul, Yipeng Qin, and Peter Wonka. Image2stylegan++: How to edit the embedded images? In IEEE Conf. Comput. Vis. Pattern Recog., 2020.
309
+ [5] Alper Aksac, Douglas J Demetrick, Tansel Ozyer, and Reda Alhajj. Brecahad: a dataset for breast cancer histopathological annotation and diagnosis. BMC research notes, 12(1):1-3, 2019.
310
+ [6] Ashish Bora, Ajil Jalal, Eric Price, and Alexandros G Dimakis. Compressed sensing using generative models. In Int. Conf. Mach. Learn., 2017.
311
+ [7] Andrew Brock, Jeff Donahue, and Karen Simonyan. Large scale GAN training for high fidelity natural image synthesis. In Int. Conf. Learn. Represent., 2019.
312
+ [8] Lucy Chai, Michael Gharbi, Eli Shechtman, Phillip Isola, and Richard Zhang. Any-resolution training for high-resolution image synthesis. In Eur. Conf. Comput. Vis., 2022.
313
+ [9] Hanting Chen, Yunhe Wang, Tianyu Guo, Chang Xu, Yiping Deng, Zhenhua Liu, Siwei Ma, Chunjing Xu, Chao Xu, and Wen Gao. Pre-trained image processing transformer. In IEEE Conf. Comput. Vis. Pattern Recog., 2021.
314
+ [10] Li-Heng Chen, Christos G Bampis, Zhi Li, Andrey Norkin, and Alan C Bovik. Proxiqa: A proxy approach to perceptual optimization of learned image compression. IEEE Trans. Image Process., 30:360-373, 2020.
315
+ [11] Yunjey Choi, Youngjung Uh, Jaejun Yoo, and Jung-Woo Ha. Stargan v2: Diverse image synthesis for multiple domains. In IEEE Conf. Comput. Vis. Pattern Recog., 2020.
316
+ [12] Alex Clark. Pillow (pil fork) documentation, 2015.
317
+ [13] Arthur Conmy, Subhadip Mukherjee, and Carola-Bibiane Schonlieb. Stylegan-induced data-driven regularization for inverse problems. In Int. Conf. Acous. Sp. Sig. Proc., 2022.
318
+ [14] Giannis Daras, Yuval Dagan, Alexandros G Dimakis, and Constantinos Daskalakis. Score-guided intermediate layer optimization: Fast Langevin mixing for inverse problems. In Int. Conf. Mach. Learn., 2022.
319
+ [15] Giannis Daras, Joseph Dean, Ajil Jalal, and Alexandros G Dimakis. Intermediate layer optimization for inverse problems using deep generative models. In Int. Conf. Mach. Learn., 2021.
320
+ [16] Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In IEEE Conf. Comput. Vis. Pattern Recog., 2019.
321
+ [17] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. In Adv. Neural Inform. Process. Syst., 2021.
322
+ [18] Omar Elharrouss, Noor Almaadeed, Somaya Ali Al-Maadeed, and Younes Akbari. Image inpainting: A review. Neural Processing Letters, 51:2007-2028, 2019.
323
+
324
+ [19] Linwei Fan, Fan Zhang, Hui Fan, and Cai ming Zhang. Brief review of image denoising techniques. Visual Computing for Industry, Biomedicine and Art, 2, 2019.
325
+ [20] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. In Adv. Neural Inform. Process. Syst., 2014.
326
+ [21] Jinjin Gu, Yujun Shen, and Bolei Zhou. Image processing using multi-code gan prior. In IEEE Conf. Comput. Vis. Pattern Recog., 2020.
327
+ [22] Yuchao Gu, Xintao Wang, Liangbin Xie, Chao Dong, Gen Li, Ying Shan, and Ming-Ming Cheng. Vqfr: Blind face restoration with vector-quantized dictionary and parallel decoder. In Eur. Conf. Comput. Vis., 2022.
328
+ [23] Erik Härkönen, Aaron Hertzmann, Jaakko Lehtinen, and Sylvain Paris. Ganspace: Discovering interpretable gan controls. In Adv. Neural Inform. Process. Syst., 2020.
329
+ [24] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. GANs trained by a two time-scale update rule converge to a local nash equilibrium. In Adv. Neural Inform. Process. Syst., 2017.
330
+ [25] Shady Abu Hussein, Tom Tirer, and Raja Giryes. Image-adaptive gan based reconstruction. In Ass. Adv. Artif. Intel., 2020.
331
+ [26] Kyoungkook Kang, Seongtae Kim, and Sunghyun Cho. Gan inversion for out-of-range images with geometric transformations. In Int. Conf. Comput. Vis., 2021.
332
+ [27] Minguk Kang, Jun-Yan Zhu, Richard Zhang, Jaesik Park, Eli Shechtman, Sylvain Paris, and Taesung Park. Scaling up GANs for text-to-image synthesis. In IEEE Conf. Comput. Vis. Pattern Recog., 2023.
333
+ [28] Tero Karras, Miika Aittala, Janne Hellsten, Samuli Laine, Jaakko Lehtinen, and Timo Aila. Training generative adversarial networks with limited data. In Adv. Neural Inform. Process. Syst., 2020.
334
+ [29] Tero Karras, Miika Aittala, Samuli Laine, Erik Härkönen, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Alias-free generative adversarial networks. In Adv. Neural Inform. Process. Syst., 2021.
335
+ [30] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In IEEE Conf. Comput. Vis. Pattern Recog., 2019.
336
+ [31] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of StyleGAN. In IEEE Conf. Comput. Vis. Pattern Recog., 2020.
337
+ [32] Bahjat Kawar, Michael Elad, Stefano Ermon, and Jiaming Song. Denoising diffusion restoration models. In Adv. Neural Inform. Process. Syst., 2022.
338
+ [33] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014.
339
+ [34] Aijin Li, Gengyan Li, Lei Sun, and Xintao Wang. Faceformer: Scale-aware blind face restoration with transformers. ArXiv, abs/2207.09790, 2022.
340
+ [35] Juncheng Li, Zehua Pei, and Tieyong Zeng. From beginner to master: A survey for deep learning-based single-image super-resolution. arXiv preprint arXiv:2109.14335, 2021.
341
+
342
+ [36] Xiaoming Li, Wenyu Li, Dongwei Ren, Hongzhi Zhang, Meng Wang, and Wangmeng Zuo. Enhanced blind face restoration with multi-exemplar images and adaptive spatial feature fusion. In IEEE Conf. Comput. Vis. Pattern Recog., 2020.
343
+ [37] Jingyun Liang, Jiezhang Cao, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. Swinir: Image restoration using swim transformer. In Int. Conf. Comput. Vis., 2021.
344
+ [38] Xuan Luo, Xuaner Zhang, Paul Yoo, Ricardo Martin-Brualla, Jason Lawrence, and Steven M Seitz. Time-travel rephotography. ACM Trans. Graph., 40(6):1-12, 2021.
345
+ [39] Razvan V Marinescu, Daniel Moyer, and Polina Golland. Bayesian image reconstruction using deep generative models. arXiv preprint arXiv:2012.04567, 2020.
346
+ [40] Sachit Menon, Alexandru Damian, Shijia Hu, Nikhil Ravi, and Cynthia Rudin. Pulse: Self-supervised photo upsampling via latent space exploration of generative models. In IEEE Conf. Comput. Vis. Pattern Recog., 2020.
347
+ [41] NVlabs. Stylegan2 ada - python. https://github.com/NVlabs/stylegan2-ada-pytorch.
348
+ [42] Xingang Pan, Xiaohang Zhan, Bo Dai, Dahua Lin, Chen Change Loy, and Ping Luo. Exploiting deep generative prior for versatile image restoration and manipulation. IEEE Trans. Pattern Anal. Mach. Intell., 2021.
349
+ [43] Or Patashnik, Zongze Wu, Eli Shechtman, Daniel Cohen-Or, and Dani Lischinski. Styleclip: Text-driven manipulation of stylegan imagery. In Int. Conf. Comput. Vis., 2021.
350
+ [44] Yohan Poirier-Ginter, Alexandre Lessard, Ryan Smith, and Jean-François Lalonde. Overparameterization improves StyleGAN inversion. In IEEE Conf. Comput. Vis. Pattern Recog., 2022.
351
+ [45] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022.
352
+ [46] Daniel Roich, Ron Mokady, Amit H Bermano, and Daniel Cohen-Or. Pivotal tuning for latent-based editing of real images. ACM Trans. Graph., 42(1), Feb. 2023.
353
+ [47] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In IEEE Conf. Comput. Vis. Pattern Recog., 2022.
354
+ [48] Chitwan Sahara, William Chan, Huiwen Chang, Chris Lee, Jonathan Ho, Tim Salimans, David Fleet, and Mohammad Norouzi. Palette: Image-to-image diffusion models. ACM Trans. Graph., pages 1-10, 2022.
355
+ [49] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S Sara Mahdavi, Rapha Gontijo Lopes, et al. Photorealistic text-to-image diffusion models with deep language understanding. arXiv preprint arXiv:2205.11487, 2022.
356
+ [50] Axel Sauer, Katja Schwarz, and Andreas Geiger. StyleGAN-XL: Scaling stylegan to large diverse datasets. ACM Trans. Graph., 1, 2022.
357
+ [51] Yujun Shen, Ceyuan Yang, Xiaou Tang, and Bolei Zhou. Interfacegan: Interpreting the disentangled face represent
358
+
359
+ tation learned by gans. IEEE Trans. Pattern Anal. Mach. Intell., 2020.
360
+ [52] Richard Shin and Dawn Song. Jpeg-resistant adversarial images. In NeurlPS Workshop Mach. Learn. Comp. Sec., 2017.
361
+ [53] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In Int. Conf. Mach. Learn., 2015.
362
+ [54] Zhengzhong Tu, Hossein Talebi, Han Zhang, Feng Yang, Peyman Milanfar, Alan Bovik, and Yinxiao Li. Maxim: Multi-axis MLP for image processing. In IEEE Conf. Comput. Vis. Pattern Recog., 2022.
363
+ [55] Yuri Viazovetskyi, Vladimir Ivashkin, and Evgeny Kashin. Stylegan2 distillation for feed-forward image manipulation. In Eur. Conf. Comput. Vis., 2020.
364
+ [56] Xintao Wang, Yu Li, Honglun Zhang, and Ying Shan. Towards real-world blind face restoration with generative facial prior. In IEEE Conf. Comput. Vis. Pattern Recog., 2021.
365
+ [57] Zhihao Wang, Jian Chen, and Steven CH Hoi. Deep learning for image super-resolution: A survey. IEEE Trans. Pattern Anal. Mach. Intell., 43(10):3365-3387, 2020.
366
+ [58] Jeremy Watt, Reza Borhani, and Aggelos K. Katsaggelos. Machine Learning Refined: Foundations, Algorithms, and Applications. Cambridge University Press, 2016.
367
+ [59] Weihao Xia, Yulun Zhang, Yujiu Yang, Jing-Hao Xue, Bolei Zhou, and Ming-Hsuan Yang. GAN inversion: A survey. IEEE Trans. Pattern Anal. Mach. Intell., 2022.
368
+ [60] Fisher Yu, Yinda Zhang, Shuran Song, Ari Seff, and Jianxiong Xiao. Lsun: Construction of a large-scale image dataset using deep learning with humans in the loop. arXiv preprint arXiv:1506.03365, 2015.
369
+ [61] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, Ming-Hsuan Yang, and Ling Shao. Learning enriched features for real image restoration and enhancement. In Eur. Conf. Comput. Vis., 2020.
370
+ [62] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, Ming-Hsuan Yang, and Ling Shao. Multi-stage progressive image restoration. In IEEE Conf. Comput. Vis. Pattern Recog., 2021.
371
+ [63] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In IEEE Conf. Comput. Vis. Pattern Recog., 2018.
372
+ [64] Shangchen Zhou, Kelvin CK Chan, Chongyi Li, and Chen Change Loy. Towards robust blind face restoration with codebook lookup transformer, 2022.
373
+ [65] Feida Zhu, Junwei Zhu, Wenqing Chu, Xinyi Zhang, Xiaozhong Ji, Chengjie Wang, and Ying Tai. Blind face restoration via integrating face shape and generative priors. In IEEE Conf. Comput. Vis. Pattern Recog., 2022.
374
+ [66] Peihao Zhu, Rameen Abdal, Yipeng Qin, John Femiani, and Peter Wonka. Improved stylegan embedding: Where are the good latents? arXiv preprint arXiv:2012.09036, 2020.
2023/Robust Unsupervised StyleGAN Image Restoration/images.zip ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:80eeae4a0b738f8bbed26a3bbdfb982c1597246814b25e31af0d498d9c48f617
3
+ size 785334
2023/Robust Unsupervised StyleGAN Image Restoration/layout.json ADDED
The diff for this file is too large to render. See raw diff
 
2023/Robust and Scalable Gaussian Process Regression and Its Applications/def98ceb-97d5-4394-a2b2-443884eb66dd_content_list.json ADDED
@@ -0,0 +1,2097 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ [
2
+ {
3
+ "type": "text",
4
+ "text": "Robust and Scalable Gaussian Process Regression and Its Applications",
5
+ "text_level": 1,
6
+ "bbox": [
7
+ 129,
8
+ 130,
9
+ 841,
10
+ 152
11
+ ],
12
+ "page_idx": 0
13
+ },
14
+ {
15
+ "type": "text",
16
+ "text": "Yifan Lu $^{1}$ , Jiayi Ma $^{1*}$ , Leyuan Fang $^{2}$ , Xin Tian $^{1}$ , and Junjun Jiang $^{3}$",
17
+ "bbox": [
18
+ 199,
19
+ 179,
20
+ 767,
21
+ 199
22
+ ],
23
+ "page_idx": 0
24
+ },
25
+ {
26
+ "type": "text",
27
+ "text": "$^{1}$ Wuhan University, China $^{2}$ Hunan University, China $^{3}$ Harbin Institute of Technology, China",
28
+ "bbox": [
29
+ 93,
30
+ 199,
31
+ 875,
32
+ 215
33
+ ],
34
+ "page_idx": 0
35
+ },
36
+ {
37
+ "type": "text",
38
+ "text": "{lyf048, xin.tian}@whu.edu.cn, {jyma2010, fangleyuan}@gmail.com, jiangjunjun@hit.edu.cn",
39
+ "bbox": [
40
+ 98,
41
+ 218,
42
+ 866,
43
+ 234
44
+ ],
45
+ "page_idx": 0
46
+ },
47
+ {
48
+ "type": "text",
49
+ "text": "Abstract",
50
+ "text_level": 1,
51
+ "bbox": [
52
+ 233,
53
+ 268,
54
+ 313,
55
+ 284
56
+ ],
57
+ "page_idx": 0
58
+ },
59
+ {
60
+ "type": "text",
61
+ "text": "This paper introduces a robust and scalable Gaussian process regression (GPR) model via variational learning. This enables the application of Gaussian processes to a wide range of real data, which are often large-scale and contaminated by outliers. Towards this end, we employ a mixture likelihood model where outliers are assumed to be sampled from a uniform distribution. We next derive a variational formulation that jointly infers the mode of data, i.e., inlier or outlier, as well as hyperparameters by maximizing a lower bound of the true log marginal likelihood. Compared to previous robust GPR, our formulation approximates the exact posterior distribution. The inducing variable approximation and stochastic variational inference are further introduced to our variational framework, extending our model to large-scale data. We apply our model to two challenging real-world applications, namely feature matching and dense gene expression imputation. Extensive experiments demonstrate the superiority of our model in terms of robustness and speed. Notably, when matching 4k feature points, its inference is completed in milliseconds with almost no false matches. The code is at github.com/YifanLu2000/Robust-Scalable-GPR.",
62
+ "bbox": [
63
+ 75,
64
+ 299,
65
+ 473,
66
+ 606
67
+ ],
68
+ "page_idx": 0
69
+ },
70
+ {
71
+ "type": "text",
72
+ "text": "1. Introduction",
73
+ "text_level": 1,
74
+ "bbox": [
75
+ 76,
76
+ 636,
77
+ 209,
78
+ 652
79
+ ],
80
+ "page_idx": 0
81
+ },
82
+ {
83
+ "type": "text",
84
+ "text": "Gaussian processes (GPs) [31] are probably the primary non-parametric method for inference on latent functions. They have a wide range of applications from biology [3] to computer vision [41]. A commonly used observation model for Gaussian process regression (GPR) is the Normal distribution, which brings great convenience to the inference. Unfortunately, a well-known limitation of the Gaussian observation model is its sensitivity to outliers in data. As illustrated in Fig. 1 (b), a few outliers can drastically destroy the entire posterior regression result. This hinders the real-world applications of GPR for many domains, where outliers are often inevitable. This paper intends to conquer the GPR with outlier contaminated data.",
85
+ "bbox": [
86
+ 75,
87
+ 662,
88
+ 468,
89
+ 858
90
+ ],
91
+ "page_idx": 0
92
+ },
93
+ {
94
+ "type": "text",
95
+ "text": "The idea of robust regression is not new. Outlier detec",
96
+ "bbox": [
97
+ 96,
98
+ 859,
99
+ 468,
100
+ 875
101
+ ],
102
+ "page_idx": 0
103
+ },
104
+ {
105
+ "type": "image",
106
+ "img_path": "images/422eeaa15b9fa07532c6070ace79536295c679cfce8350a007917cc3996bfd1c.jpg",
107
+ "image_caption": [],
108
+ "image_footnote": [],
109
+ "bbox": [
110
+ 506,
111
+ 268,
112
+ 633,
113
+ 354
114
+ ],
115
+ "page_idx": 0
116
+ },
117
+ {
118
+ "type": "image",
119
+ "img_path": "images/ba4e85d82fd04d290f99c63a1570bd1cf7083263db456c972362c9aedd7729cc.jpg",
120
+ "image_caption": [],
121
+ "image_footnote": [],
122
+ "bbox": [
123
+ 633,
124
+ 268,
125
+ 756,
126
+ 354
127
+ ],
128
+ "page_idx": 0
129
+ },
130
+ {
131
+ "type": "image",
132
+ "img_path": "images/13370d06014ebd7bdf4c720c3cb41b4eaa3dfb96ed42358fecb539782426169d.jpg",
133
+ "image_caption": [],
134
+ "image_footnote": [],
135
+ "bbox": [
136
+ 758,
137
+ 268,
138
+ 883,
139
+ 354
140
+ ],
141
+ "page_idx": 0
142
+ },
143
+ {
144
+ "type": "image",
145
+ "img_path": "images/bab8f13f8194b03b38e708713d7d3fc42c3a0268df8001011ea10a883e557ea2.jpg",
146
+ "image_caption": [
147
+ "Figure 1. Regression with our model. (a) Perform exact GPR from 100 inliers. (b) When there are only 6 outliers in the data, the exact GPR leads to completely wrong results. (c) By comparison, our model is able to recover the exact posterior even facing 100 outliers. (d) The feature matching result using our model. (e) The dense spatial gene expression imputation result using our model."
148
+ ],
149
+ "image_footnote": [],
150
+ "bbox": [
151
+ 508,
152
+ 357,
153
+ 689,
154
+ 426
155
+ ],
156
+ "page_idx": 0
157
+ },
158
+ {
159
+ "type": "image",
160
+ "img_path": "images/67dc0e22054a7d3fbba3e5c8064ce53ee21035d388c5175cb432719fd82e7b0e.jpg",
161
+ "image_caption": [],
162
+ "image_footnote": [],
163
+ "bbox": [
164
+ 696,
165
+ 358,
166
+ 883,
167
+ 431
168
+ ],
169
+ "page_idx": 0
170
+ },
171
+ {
172
+ "type": "text",
173
+ "text": "tion has been extensively and systematically described in [6,9,10,29]. In the context of GPR, many efforts tried to replace the Gaussian likelihood with other distributions showing heavy-tail behaviors, including Student- $t$ [16,21,28,30], Laplace [22, 30], Gaussian mixture [8, 22, 27], and data-dependent noise model [17]. The challenge with these non-Gaussian likelihoods lies in the inference, which is analytically intractable. To this end, many approximation schemes have been applied, despite having high computational complexity, e.g., Markov Chain Monte Carlo (MCMC) sampling and Expectation Propagation (EP) [22].",
174
+ "bbox": [
175
+ 496,
176
+ 551,
177
+ 892,
178
+ 718
179
+ ],
180
+ "page_idx": 0
181
+ },
182
+ {
183
+ "type": "text",
184
+ "text": "In this paper, we propose a more effective mixture likelihood model, where uniform distribution accounts for the outliers and Gaussian for inliers. In our formulation, the outliers are independent of the GP and do not affect the computation of the posterior GP, thereby allowing to tolerate more outliers. We next introduce a variational method that jointly determines the modes of data (i.e., inlier or outlier) as well as hyperparameters by maximizing a lower bound to the marginal likelihood. We highlight that the difference between our variational formulation and pervious methods is that the modes of data now become variational parameters and are obtained by minimizing the Kullback-",
185
+ "bbox": [
186
+ 496,
187
+ 719,
188
+ 893,
189
+ 901
190
+ ],
191
+ "page_idx": 0
192
+ },
193
+ {
194
+ "type": "header",
195
+ "text": "CVF",
196
+ "bbox": [
197
+ 106,
198
+ 2,
199
+ 181,
200
+ 42
201
+ ],
202
+ "page_idx": 0
203
+ },
204
+ {
205
+ "type": "header",
206
+ "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.",
207
+ "bbox": [
208
+ 236,
209
+ 0,
210
+ 810,
211
+ 46
212
+ ],
213
+ "page_idx": 0
214
+ },
215
+ {
216
+ "type": "page_footnote",
217
+ "text": "*Corresponding Author",
218
+ "bbox": [
219
+ 94,
220
+ 886,
221
+ 223,
222
+ 898
223
+ ],
224
+ "page_idx": 0
225
+ },
226
+ {
227
+ "type": "page_number",
228
+ "text": "21950",
229
+ "bbox": [
230
+ 478,
231
+ 944,
232
+ 519,
233
+ 957
234
+ ],
235
+ "page_idx": 0
236
+ },
237
+ {
238
+ "type": "text",
239
+ "text": "Leibler (KL) divergence between the variational and the true posterior distribution. Thus, the proposed formulation is less likely to overfit and is able to approximate the exact posterior GP only from inliers, as in Fig. 1 (c).",
240
+ "bbox": [
241
+ 75,
242
+ 90,
243
+ 468,
244
+ 151
245
+ ],
246
+ "page_idx": 1
247
+ },
248
+ {
249
+ "type": "text",
250
+ "text": "Inspired by [37], the sparse inducing variable approximation is integrated into our variational framework, which retains the exact GP prior but performs posterior approximation, and reduces the time complexity from $\\mathcal{O}(n^3)$ to $\\mathcal{O}(nm^2)$ . By treating the inducing variables as global variables [18], our variational model enjoys the acceleration by Stochastic Variational Inference (SVI) [20]. It performs stochastic optimization from natural gradient and further decreases the time complexity to $\\mathcal{O}(km^2)$ . This provides a guarantee for our model to scale to large-scale data.",
251
+ "bbox": [
252
+ 75,
253
+ 152,
254
+ 467,
255
+ 301
256
+ ],
257
+ "page_idx": 1
258
+ },
259
+ {
260
+ "type": "text",
261
+ "text": "We apply our robust GPR model to two real-world applications, say feature matching and dense spatial gene expression imputation, as illustrated in the Figs. 1 (d) and (e). Extensive experiments demonstrate the superiority of our method on both numerical data and real applications.",
262
+ "bbox": [
263
+ 75,
264
+ 301,
265
+ 467,
266
+ 378
267
+ ],
268
+ "page_idx": 1
269
+ },
270
+ {
271
+ "type": "text",
272
+ "text": "To summarize, our contributions include the following. (i) We present a robust Gaussian process regression model, which uses variational learning to approximate the true exact posterior. (ii) We leverage inducing variables and SVI to adapt our model to large-scale data. (iii) Two applications of our model are described. Extensive experimental validation demonstrates the superiority of our model.",
273
+ "bbox": [
274
+ 75,
275
+ 378,
276
+ 467,
277
+ 483
278
+ ],
279
+ "page_idx": 1
280
+ },
281
+ {
282
+ "type": "text",
283
+ "text": "2. Related Works",
284
+ "text_level": 1,
285
+ "bbox": [
286
+ 76,
287
+ 497,
288
+ 225,
289
+ 512
290
+ ],
291
+ "page_idx": 1
292
+ },
293
+ {
294
+ "type": "text",
295
+ "text": "Robust GPR typically employs non-Gaussian likelihood with heavy tails. Representatively, [22] investigated the Student's- $t$ noise model. They further developed variational inference and MCMC for GPR with Student's- $t$ . In [39], Laplace approximation was used to approximate the log marginal likelihood of Student's- $t$ . The Laplace noise model is also exploited in [22]. [30] introduced Expectation Maximization (EM) algorithm for Student's- $t$ and Laplace. Recently, [8] used mixtures of two Gaussians assuming a lower variance for regular noise and a higher one for outliers, with EM to learn the hyperparameters. In this work, we also use a non-Gaussian likelihood, i.e., mixture model. The difference is that here the outliers are considered to follow a uniform distribution, which is a more reasonable assumption since the sampling distribution of outliers should be independent of the true latent function. Besides, we derive a variational formulation in the GPR setting, which approximates the posterior distribution.",
296
+ "bbox": [
297
+ 75,
298
+ 522,
299
+ 467,
300
+ 792
301
+ ],
302
+ "page_idx": 1
303
+ },
304
+ {
305
+ "type": "text",
306
+ "text": "Scalable GPR improves the scalability of GPR while maintaining prediction quality for big data. The seminal work [40] adopted Nyström approximation to approximate the kernel matrix using $m$ sparse points. Later, the idea was further promoted by [7, 15, 23, 33], which approximated the prior and performed exact inference. Unlike the prior approximations, the other line is the posterior approximations,",
307
+ "bbox": [
308
+ 75,
309
+ 794,
310
+ 467,
311
+ 900
312
+ ],
313
+ "page_idx": 1
314
+ },
315
+ {
316
+ "type": "text",
317
+ "text": "which retain the exact prior but perform approximate inference. The most well-known work is the elegant variational free energy [37]. It directly approximates the posterior by using variational inference. With the advances in variational inference [20], SVI is introduced to GP [18], which uses natural gradients and results in remarkable computational efficiency. However, as the exact GPR, these scalable GPRs are sensitive to outliers. In contrast, our model can tolerate massive outliers while retaining highly scalability.",
318
+ "bbox": [
319
+ 496,
320
+ 90,
321
+ 890,
322
+ 227
323
+ ],
324
+ "page_idx": 1
325
+ },
326
+ {
327
+ "type": "text",
328
+ "text": "3. Gaussian Process Regression Revisited",
329
+ "text_level": 1,
330
+ "bbox": [
331
+ 498,
332
+ 239,
333
+ 846,
334
+ 255
335
+ ],
336
+ "page_idx": 1
337
+ },
338
+ {
339
+ "type": "text",
340
+ "text": "Gaussian process is a collection of random variables, for which any finite subset has a joint Gaussian distribution [31]. It is completely specified by a mean function $m(\\mathbf{x})$ and a covariance function $k(\\mathbf{x},\\mathbf{x}^{\\prime})$ where the covariance function typically depends on a set of hyperparameters $\\varphi$ . The GP is usually used as a prior over latent function $f(\\mathbf{x})\\sim \\mathcal{GP}(m,k)$ . We can combine the GP prior with observed data to give a posterior over desired latent function.",
341
+ "bbox": [
342
+ 496,
343
+ 263,
344
+ 890,
345
+ 383
346
+ ],
347
+ "page_idx": 1
348
+ },
349
+ {
350
+ "type": "text",
351
+ "text": "Consider a regression problem, where the training data is $\\mathcal{D} = \\{(\\mathbf{x}_i,y_i)\\}_{i = 1}^n$ . Each $y_{i}$ is a noisy realization of the latent function at location $\\mathbf{x}_i$ such that $y_{i} = f_{i} + \\epsilon_{i}$ , where $\\epsilon_{i}\\sim \\mathcal{N}(\\epsilon_{i}|0,\\sigma^{2})$ is typically assumed to be an i.i.d. Gaussian noise and $f_{i} = f(\\mathbf{x}_{i})$ . Denoting all training inputs as $\\mathbf{X}$ , all training outputs $\\mathbf{y}$ , and all latent function values at $\\mathbf{X}$ as $\\mathbf{f}$ . Placing a GP prior on the latent function $f$ , we obtain",
352
+ "bbox": [
353
+ 496,
354
+ 383,
355
+ 890,
356
+ 491
357
+ ],
358
+ "page_idx": 1
359
+ },
360
+ {
361
+ "type": "equation",
362
+ "text": "\n$$\np (\\mathbf {f} | \\mathbf {X}, \\varphi) \\sim \\mathcal {N} (\\mathbf {f} | m (\\mathbf {X}), \\mathbf {K} _ {n n}), \\tag {1}\n$$\n",
363
+ "text_format": "latex",
364
+ "bbox": [
365
+ 584,
366
+ 500,
367
+ 890,
368
+ 515
369
+ ],
370
+ "page_idx": 1
371
+ },
372
+ {
373
+ "type": "text",
374
+ "text": "where $\\mathbf{K}_{nn}$ is the covariance function evaluated between all the training points. For simplicity of exposition, the mean function is set to zero $m(\\mathbf{x}) = 0$ . It does not reduce the generalizability of the GP as long as the problem is correctly transformed, e.g., by subtracting the mean of $\\mathbf{y}$ . One can induce a posterior distribution on $\\mathbf{f}$ according to Bayes' rule",
375
+ "bbox": [
376
+ 496,
377
+ 522,
378
+ 890,
379
+ 613
380
+ ],
381
+ "page_idx": 1
382
+ },
383
+ {
384
+ "type": "equation",
385
+ "text": "\n$$\np (\\mathbf {f} | \\mathcal {D}, \\sigma^ {2}, \\varphi) \\sim \\mathcal {N} (\\mathbf {f} | \\mathbf {K} _ {n n} (\\mathbf {K} _ {n n} ^ {\\epsilon}) ^ {- 1} \\mathbf {y}, \\boldsymbol {\\Sigma}), \\tag {2}\n$$\n",
386
+ "text_format": "latex",
387
+ "bbox": [
388
+ 550,
389
+ 627,
390
+ 890,
391
+ 643
392
+ ],
393
+ "page_idx": 1
394
+ },
395
+ {
396
+ "type": "text",
397
+ "text": "which is also a multivariate normal distribution, where $\\mathbf{K}_{nn}^{\\epsilon} = \\mathbf{K}_{nn} + \\sigma^{2}\\mathbf{I}$ and $\\boldsymbol {\\Sigma} = (\\mathbf{K}_{nn}^{-1} + \\sigma^{-2}\\mathbf{I})^{-1}$ . The posterior distribution of $\\mathbf{f}$ can help to compute the posterior predictive distribution of $\\mathbf{f}_*$ at any test location $\\mathbf{X}_{*}$ :",
398
+ "bbox": [
399
+ 496,
400
+ 648,
401
+ 890,
402
+ 709
403
+ ],
404
+ "page_idx": 1
405
+ },
406
+ {
407
+ "type": "equation",
408
+ "text": "\n$$\n\\begin{array}{l} p \\left(\\mathbf {f} _ {*} \\mid \\mathcal {D}, \\mathbf {X} _ {*}\\right) = \\int p \\left(\\mathbf {f} _ {*} \\mid \\mathbf {f}, \\mathbf {X} _ {*}\\right) p (\\mathbf {f} \\mid \\mathcal {D}) d \\mathbf {f} \\tag {3} \\\\ = \\mathcal {N} \\left(\\mathbf {f} _ {*} | \\mathbf {K} _ {* n} \\left(\\mathbf {K} _ {n n} ^ {\\epsilon}\\right) ^ {- 1} \\mathbf {y}, \\mathbf {K} _ {* *} - \\mathbf {K} _ {* n} \\left(\\mathbf {K} _ {n n} ^ {\\epsilon}\\right) ^ {- 1} \\mathbf {K} _ {n *}\\right), \\\\ \\end{array}\n$$\n",
409
+ "text_format": "latex",
410
+ "bbox": [
411
+ 506,
412
+ 715,
413
+ 890,
414
+ 766
415
+ ],
416
+ "page_idx": 1
417
+ },
418
+ {
419
+ "type": "text",
420
+ "text": "where, $\\sigma^2$ and $\\varphi$ are omitted for brevity. The above posterior depends on the values of $\\sigma^2$ and $\\varphi$ , which can be inferred by maximizing the log marginal likelihood:",
421
+ "bbox": [
422
+ 496,
423
+ 773,
424
+ 890,
425
+ 820
426
+ ],
427
+ "page_idx": 1
428
+ },
429
+ {
430
+ "type": "equation",
431
+ "text": "\n$$\n\\log p (\\mathbf {y} \\mid \\sigma^ {2}, \\varphi) = \\log \\mathcal {N} (\\mathbf {y} \\mid \\mathbf {0}, \\mathbf {K} _ {n n} ^ {\\epsilon}), \\tag {4}\n$$\n",
432
+ "text_format": "latex",
433
+ "bbox": [
434
+ 573,
435
+ 827,
436
+ 890,
437
+ 844
438
+ ],
439
+ "page_idx": 1
440
+ },
441
+ {
442
+ "type": "text",
443
+ "text": "which automatically achieves the bias-variance trade-off.",
444
+ "bbox": [
445
+ 500,
446
+ 852,
447
+ 874,
448
+ 866
449
+ ],
450
+ "page_idx": 1
451
+ },
452
+ {
453
+ "type": "page_footnote",
454
+ "text": "<sup>1</sup>Note that the derivation in this paper extends to multiple independent output dimensions is straightforward.",
455
+ "bbox": [
456
+ 500,
457
+ 875,
458
+ 890,
459
+ 898
460
+ ],
461
+ "page_idx": 1
462
+ },
463
+ {
464
+ "type": "page_number",
465
+ "text": "21951",
466
+ "bbox": [
467
+ 478,
468
+ 944,
469
+ 517,
470
+ 955
471
+ ],
472
+ "page_idx": 1
473
+ },
474
+ {
475
+ "type": "text",
476
+ "text": "4. Robust Gaussian Process Regression Model",
477
+ "text_level": 1,
478
+ "bbox": [
479
+ 76,
480
+ 89,
481
+ 465,
482
+ 107
483
+ ],
484
+ "page_idx": 2
485
+ },
486
+ {
487
+ "type": "text",
488
+ "text": "In real-world scenarios, a huge discrepancy between the model and data-generating process often occurs, leading to outliers. However, a critical flaw of Gaussian observation model is its non-robustness. This poses a great challenge to GPR in practice. To this end, this section introduces a robust GPR, which identifies the mode of data and simultaneously approximates the posterior of the desired model.",
489
+ "bbox": [
490
+ 75,
491
+ 114,
492
+ 468,
493
+ 220
494
+ ],
495
+ "page_idx": 2
496
+ },
497
+ {
498
+ "type": "text",
499
+ "text": "4.1. Problem Formulation",
500
+ "text_level": 1,
501
+ "bbox": [
502
+ 76,
503
+ 229,
504
+ 279,
505
+ 244
506
+ ],
507
+ "page_idx": 2
508
+ },
509
+ {
510
+ "type": "text",
511
+ "text": "We explicitly consider the generation process of outliers. Since outliers are independent of the latent function, a reasonable assumption is that they obey a uniform distribution $\\frac{1}{a}$ , where the output space is a bounded region that covers the output $\\mathbf{y}$ and $a$ is a constant denoting the volume of the region. The inliers follow a regular Gaussian noise $\\mathcal{N}(y_i|f_i,\\sigma^2)$ . We then associate each data with a latent variable $z_{i} \\in \\{0,1\\}$ , where $z_{i} = 0$ indicates that the observation is generated by outlier distribution and $z_{i} = 1$ inlier. The joint distribution of $(y_{i},z_{i})$ is then a mixture of two models:",
512
+ "bbox": [
513
+ 75,
514
+ 253,
515
+ 470,
516
+ 417
517
+ ],
518
+ "page_idx": 2
519
+ },
520
+ {
521
+ "type": "equation",
522
+ "text": "\n$$\np \\left(y _ {i}, z _ {i} \\mid f _ {i}, \\boldsymbol {\\theta}\\right) = \\left\\{\\left(1 - \\gamma\\right) \\frac {1}{a} \\right\\} ^ {1 - z _ {i}} \\left\\{\\gamma \\mathcal {N} \\left(y _ {i} \\mid f _ {i}, \\sigma^ {2}\\right) \\right\\} ^ {z _ {i}}, \\tag {5}\n$$\n",
523
+ "text_format": "latex",
524
+ "bbox": [
525
+ 86,
526
+ 426,
527
+ 468,
528
+ 455
529
+ ],
530
+ "page_idx": 2
531
+ },
532
+ {
533
+ "type": "text",
534
+ "text": "where $\\theta = (\\sigma^2, \\gamma, \\varphi)$ denotes the latent variables set and $\\gamma = p(z_i = 1)$ is the mixing coefficient.",
535
+ "bbox": [
536
+ 76,
537
+ 465,
538
+ 468,
539
+ 497
540
+ ],
541
+ "page_idx": 2
542
+ },
543
+ {
544
+ "type": "text",
545
+ "text": "Prior distributions describe beliefs about the model variables before the inference. As a robust GPR problem, we place a GP prior over the latent function, i.e., Eq. (1). The variable $\\gamma$ controls the probability of occurrence of two models. We suppose it follows a Beta distribution",
546
+ "bbox": [
547
+ 75,
548
+ 497,
549
+ 468,
550
+ 571
551
+ ],
552
+ "page_idx": 2
553
+ },
554
+ {
555
+ "type": "equation",
556
+ "text": "\n$$\np (\\gamma) \\sim \\operatorname {B e t a} (\\gamma | B _ {a}, B _ {b}), \\tag {6}\n$$\n",
557
+ "text_format": "latex",
558
+ "bbox": [
559
+ 189,
560
+ 585,
561
+ 468,
562
+ 601
563
+ ],
564
+ "page_idx": 2
565
+ },
566
+ {
567
+ "type": "text",
568
+ "text": "where $B_{a}$ and $B_{b}$ are hyperparameters. As we will see, $p(\\gamma)$ is conjugate to the posterior and hence easier to handle. For other variables $(\\sigma^2,\\varphi)$ , we introduce noninformative prior.",
569
+ "bbox": [
570
+ 75,
571
+ 612,
572
+ 468,
573
+ 657
574
+ ],
575
+ "page_idx": 2
576
+ },
577
+ {
578
+ "type": "text",
579
+ "text": "Full joint distribution is obtained by incorporating the above prior distributions as follows:",
580
+ "bbox": [
581
+ 75,
582
+ 657,
583
+ 468,
584
+ 686
585
+ ],
586
+ "page_idx": 2
587
+ },
588
+ {
589
+ "type": "equation",
590
+ "text": "\n$$\np (\\mathcal {D}, \\mathbf {f}, \\boldsymbol {\\theta}) \\propto p (\\mathbf {f} | \\mathbf {X}) p (\\gamma) \\prod_ {i = 1} ^ {n} p \\left(y _ {i}, z _ {i} \\mid f _ {i}, \\sigma^ {2}, \\gamma\\right), \\tag {7}\n$$\n",
591
+ "text_format": "latex",
592
+ "bbox": [
593
+ 114,
594
+ 698,
595
+ 468,
596
+ 738
597
+ ],
598
+ "page_idx": 2
599
+ },
600
+ {
601
+ "type": "text",
602
+ "text": "where $\\pmb{\\theta} = (\\sigma^2, \\gamma, \\mathbf{Z}, \\varphi)$ and $\\mathbf{Z} \\in \\{0,1\\}^n$ is the indicator variable vector. Our goal is to obtain the posterior $p(\\mathbf{f}, \\pmb{\\theta}|\\mathcal{D})$ , which can be induced from $p(\\mathcal{D}, \\mathbf{f}, \\pmb{\\theta})$ . Nevertheless, as there are two possibilities for each observation, i.e., inlier and outlier, it will generate $2^n$ combinations to be evaluated. Thus, the exact estimation is analytically intractable and approximation is needed. Next, we use the variational inference to approximate the exact posterior. Once the variational distribution is obtained, the posterior predictive distribution can be determined.",
603
+ "bbox": [
604
+ 75,
605
+ 748,
606
+ 468,
607
+ 900
608
+ ],
609
+ "page_idx": 2
610
+ },
611
+ {
612
+ "type": "text",
613
+ "text": "4.2. Variational Inference",
614
+ "text_level": 1,
615
+ "bbox": [
616
+ 500,
617
+ 90,
618
+ 700,
619
+ 104
620
+ ],
621
+ "page_idx": 2
622
+ },
623
+ {
624
+ "type": "text",
625
+ "text": "We wish to directly approximate the posterior $p(\\mathbf{f},\\boldsymbol {\\theta}|\\mathcal{D})$ using a variational posterior $q(\\mathbf{f},\\boldsymbol {\\theta})$ such that the KL divergence between $p(\\mathbf{f},\\boldsymbol {\\theta}|\\mathcal{D})$ and $q(\\mathbf{f},\\boldsymbol {\\theta})$ is minimized. The minimization is equivalently expressed as maximization of lower bound for the true log marginal likelihood:",
626
+ "bbox": [
627
+ 498,
628
+ 113,
629
+ 893,
630
+ 189
631
+ ],
632
+ "page_idx": 2
633
+ },
634
+ {
635
+ "type": "equation",
636
+ "text": "\n$$\n\\mathcal {L} (\\mathbf {f}, \\boldsymbol {\\theta}) = \\int q (\\mathbf {f}, \\boldsymbol {\\theta}) \\log \\left(\\frac {p (\\mathbf {f} , \\boldsymbol {\\theta} , \\mathcal {D})}{q (\\mathbf {f} , \\boldsymbol {\\theta})}\\right) d \\mathbf {f} d \\boldsymbol {\\theta}. \\tag {8}\n$$\n",
637
+ "text_format": "latex",
638
+ "bbox": [
639
+ 544,
640
+ 196,
641
+ 890,
642
+ 231
643
+ ],
644
+ "page_idx": 2
645
+ },
646
+ {
647
+ "type": "text",
648
+ "text": "The maximization is, however, intractable since the optimal solution is $\\hat{q} (\\mathbf{f},\\pmb {\\theta}) = p(\\mathbf{f},\\pmb {\\theta}|\\mathcal{D})$ . To this end, we consider instead a restricted distribution of $q(\\mathbf{f},\\pmb {\\theta})$",
649
+ "bbox": [
650
+ 498,
651
+ 237,
652
+ 890,
653
+ 282
654
+ ],
655
+ "page_idx": 2
656
+ },
657
+ {
658
+ "type": "text",
659
+ "text": "Mean-field factorization supposes $q(\\mathbf{f},\\pmb {\\theta})$ can be partitioned into disjoint groups. Considering the ease of calculation, we factorize $q(\\mathbf{f},\\pmb {\\theta}) = q_{1}(\\mathbf{f})q_{2}(\\gamma)q_{3}(\\mathbf{Z})q_{4}(\\sigma^{2},\\pmb {\\varphi})$ . Besides, we assume $q_{4}$ is Dirac delta function, i.e., it has a point mass at $(\\sigma^2,\\varphi)$ . With this factorization, the maximization of Eq. (8) is obtained by optimizing each of the factors. For $q_{i}$ , the minimization becomes",
660
+ "bbox": [
661
+ 498,
662
+ 282,
663
+ 890,
664
+ 388
665
+ ],
666
+ "page_idx": 2
667
+ },
668
+ {
669
+ "type": "equation",
670
+ "text": "\n$$\n\\hat {q} _ {i} = \\underset {q _ {i}} {\\arg \\min } \\int q _ {i} \\log \\left(\\frac {\\exp \\mathbb {E} _ {j \\neq i} [ \\log p (\\mathbf {f} , \\boldsymbol {\\theta} , \\mathcal {D}) ]}{q _ {i}}\\right) d \\boldsymbol {\\theta} _ {i}, \\tag {9}\n$$\n",
671
+ "text_format": "latex",
672
+ "bbox": [
673
+ 508,
674
+ 395,
675
+ 890,
676
+ 441
677
+ ],
678
+ "page_idx": 2
679
+ },
680
+ {
681
+ "type": "text",
682
+ "text": "where $\\mathbb{E}_{j\\neq i}[\\cdot ]$ denotes the expectation over $q_{j}$ for $j\\neq i$ Note that Eq. (9) is the negative KL divergence between $q_{i}$ and $\\mathbb{E}_{j\\neq i}\\left[\\log p(\\mathbf{f},\\boldsymbol {\\theta},\\mathcal{D})\\right]$ . Thus, the optimal $q_{i}$ is given by",
683
+ "bbox": [
684
+ 498,
685
+ 441,
686
+ 890,
687
+ 489
688
+ ],
689
+ "page_idx": 2
690
+ },
691
+ {
692
+ "type": "equation",
693
+ "text": "\n$$\n\\log \\hat {q} _ {i} = \\mathbb {E} _ {j \\neq i} [ \\log p (\\mathbf {f}, \\boldsymbol {\\theta}, \\mathcal {D}) ] + \\text {c o n s t .} \\tag {10}\n$$\n",
694
+ "text_format": "latex",
695
+ "bbox": [
696
+ 566,
697
+ 496,
698
+ 890,
699
+ 513
700
+ ],
701
+ "page_idx": 2
702
+ },
703
+ {
704
+ "type": "text",
705
+ "text": "4.3. Sparse Inducing Variable Approximation",
706
+ "text_level": 1,
707
+ "bbox": [
708
+ 500,
709
+ 521,
710
+ 852,
711
+ 537
712
+ ],
713
+ "page_idx": 2
714
+ },
715
+ {
716
+ "type": "text",
717
+ "text": "The variational posterior approximation presented above can easily integrate sparse inducing variables [37], which allows the reduction of time complexity of GPR from $\\mathcal{O}(n^3)$ to $\\mathcal{O}(nm^2)$ . Introducing $m$ pseudo-inputs $\\mathbf{X}_m$ with inducing variables $\\mathbf{f}_m$ , our goal is to use $\\mathbf{f}_m$ to approximate $q(\\mathbf{f})$ , where $\\mathbf{f}_m$ akin to $\\mathbf{f}$ follows the same GP prior that $p(\\mathbf{f}_m) \\sim \\mathcal{N}(\\mathbf{0}, \\mathbf{K}_{mm})$ . We suppose $\\mathbf{f}_m$ is a sufficient statistic for $\\mathbf{f}$ , i.e., for any value $u$ , $p(u|\\mathbf{f}, \\mathbf{f}_m) = p(u|\\mathbf{f}_m)$ holds [37]. In this setting, the augmented variational posterior $q(\\mathbf{f}, \\mathbf{f}_m) = p(\\mathbf{f}|\\mathbf{f}_m)\\phi(\\mathbf{f}_m)$ and augmented latent variables set $\\boldsymbol{\\theta} = (\\gamma, \\sigma^2, \\mathbf{Z}, \\varphi, \\mathbf{X}_m)$ with factorization $q(\\mathbf{f}, \\mathbf{f}_m, \\boldsymbol{\\theta}) = q_1(\\mathbf{f}, \\mathbf{f}_m)q_2(\\gamma)q_3(\\mathbf{Z})q_4(\\sigma^2, \\varphi, \\mathbf{X}_m)$ . The bound (8) becomes",
718
+ "bbox": [
719
+ 498,
720
+ 544,
721
+ 890,
722
+ 739
723
+ ],
724
+ "page_idx": 2
725
+ },
726
+ {
727
+ "type": "equation",
728
+ "text": "\n$$\n\\mathcal {L} = \\int q (\\mathbf {f}, \\mathbf {f} _ {m}, \\boldsymbol {\\theta}) \\log \\left(\\frac {p \\left(\\mathbf {f} _ {m}\\right) p (\\gamma) p (\\mathbf {y} , \\mathbf {Z} \\mid \\mathbf {f})}{\\phi \\left(\\mathbf {f} _ {m}\\right) q _ {2 - 4}}\\right) d \\mathbf {f} d \\mathbf {f} _ {m} d \\boldsymbol {\\theta}, \\tag {11}\n$$\n",
729
+ "text_format": "latex",
730
+ "bbox": [
731
+ 504,
732
+ 747,
733
+ 890,
734
+ 792
735
+ ],
736
+ "page_idx": 2
737
+ },
738
+ {
739
+ "type": "text",
740
+ "text": "where the term $p(\\mathbf{f}|\\mathbf{f}_m)$ inside the log is eliminated. In what follows, we maximize the bound analytically by solving the optimal choice of the variational distribution $q(\\mathbf{f},\\mathbf{f}_m,\\boldsymbol {\\theta})$",
741
+ "bbox": [
742
+ 498,
743
+ 792,
744
+ 890,
745
+ 839
746
+ ],
747
+ "page_idx": 2
748
+ },
749
+ {
750
+ "type": "text",
751
+ "text": "4.4. Maximize Lower Bound",
752
+ "text_level": 1,
753
+ "bbox": [
754
+ 500,
755
+ 847,
756
+ 720,
757
+ 861
758
+ ],
759
+ "page_idx": 2
760
+ },
761
+ {
762
+ "type": "text",
763
+ "text": "To optimize the bound (11) in mean-field factorization, a common approach is to use the coordinate ascent technique,",
764
+ "bbox": [
765
+ 500,
766
+ 869,
767
+ 890,
768
+ 901
769
+ ],
770
+ "page_idx": 2
771
+ },
772
+ {
773
+ "type": "page_number",
774
+ "text": "21952",
775
+ "bbox": [
776
+ 478,
777
+ 944,
778
+ 519,
779
+ 955
780
+ ],
781
+ "page_idx": 2
782
+ },
783
+ {
784
+ "type": "text",
785
+ "text": "i.e., optimizing one factor $q_{i}$ while keeping the others $q_{j\\neq i}$ fixed, and cycling through these factors in turn. Convergence is guaranteed because the bound is convex and monotonically increasing with respect to each factor [5]. Next, we give closed-form expressions for factor updates without detailed derivations, which can be found in suppl. material.",
786
+ "bbox": [
787
+ 75,
788
+ 90,
789
+ 467,
790
+ 181
791
+ ],
792
+ "page_idx": 3
793
+ },
794
+ {
795
+ "type": "text",
796
+ "text": "Updating $q_{1}(\\mathbf{f},\\mathbf{f}_{m})$ means finding the optimal posterior GP given $q_{2}$ $q_{3}$ , and $q_{4}$ . We denote $p_i = \\mathbb{E}[z_i]$ and $\\mathbf{P} =$ diag $(p_1,p_2,\\dots ,p_n)$ . Integrating over f, (11) becomes",
797
+ "bbox": [
798
+ 75,
799
+ 181,
800
+ 467,
801
+ 228
802
+ ],
803
+ "page_idx": 3
804
+ },
805
+ {
806
+ "type": "equation",
807
+ "text": "\n$$\n\\mathcal {L} = \\int \\phi (\\mathbf {f} _ {m}) \\log \\left(\\frac {p \\left(\\mathbf {f} _ {m}\\right) Q \\left(\\mathbf {f} _ {m} , \\mathbf {y}\\right)}{\\phi \\left(\\mathbf {f} _ {m}\\right)}\\right) d \\mathbf {f} _ {m} + \\text {c o n s t .}, \\tag {12}\n$$\n",
808
+ "text_format": "latex",
809
+ "bbox": [
810
+ 86,
811
+ 238,
812
+ 467,
813
+ 272
814
+ ],
815
+ "page_idx": 3
816
+ },
817
+ {
818
+ "type": "text",
819
+ "text": "where",
820
+ "bbox": [
821
+ 76,
822
+ 282,
823
+ 122,
824
+ 294
825
+ ],
826
+ "page_idx": 3
827
+ },
828
+ {
829
+ "type": "equation",
830
+ "text": "\n$$\n\\log Q \\left(\\mathbf {f} _ {m}, \\mathbf {y}\\right) = \\log \\mathcal {N} \\left(\\mathbf {y} \\mid \\boldsymbol {\\mu} _ {\\mathbf {f} \\mid \\mathbf {f} _ {m}}, \\sigma^ {2} \\mathbf {P} ^ {- 1}\\right) - \\frac {1}{2 \\sigma^ {2}} \\operatorname {t r} \\left(\\mathbf {P K} _ {\\mathbf {f} \\mid \\mathbf {f} _ {m}}\\right), \\tag {13}\n$$\n",
831
+ "text_format": "latex",
832
+ "bbox": [
833
+ 78,
834
+ 294,
835
+ 467,
836
+ 334
837
+ ],
838
+ "page_idx": 3
839
+ },
840
+ {
841
+ "type": "text",
842
+ "text": "$\\mu_{\\mathbf{f}|\\mathbf{f}_m} = \\mathbf{K}_{nm}\\mathbf{K}_{mm}^{-1}\\mathbf{f}_m,$ $\\mathbf{K}_{\\mathbf{f}|\\mathbf{f}_m} = \\mathbf{K}_{nn} - \\mathbf{K}_{nm}\\mathbf{K}_{mm}^{-1}\\mathbf{K}_{mn}$ are the expectation and covariance of $p(\\mathbf{f}|\\mathbf{f}_m)$ , respectively. With some linear algebraic manipulation, the optimal $\\hat{\\phi} (\\mathbf{f}_m)$ is given by a multivariate Gaussian distribution",
843
+ "bbox": [
844
+ 76,
845
+ 334,
846
+ 467,
847
+ 396
848
+ ],
849
+ "page_idx": 3
850
+ },
851
+ {
852
+ "type": "equation",
853
+ "text": "\n$$\n\\begin{array}{l} \\hat {\\phi} (\\mathbf {f} _ {m}) \\propto p (\\mathbf {f} _ {m}) Q (\\mathbf {f} _ {m}, \\mathbf {y}) \\tag {14} \\\\ = \\mathcal {N} (\\mathbf {f} _ {m} | \\boldsymbol {\\mu} _ {m}, \\mathbf {A} _ {m}), \\\\ \\end{array}\n$$\n",
854
+ "text_format": "latex",
855
+ "bbox": [
856
+ 184,
857
+ 397,
858
+ 467,
859
+ 434
860
+ ],
861
+ "page_idx": 3
862
+ },
863
+ {
864
+ "type": "text",
865
+ "text": "where $\\pmb{\\mu}_{m} = \\sigma^{-2}\\mathbf{K}_{mm}\\pmb{\\Sigma}\\mathbf{K}_{mn}\\mathbf{P}\\mathbf{y},\\mathbf{A}_{m} = \\mathbf{K}_{mm}\\pmb{\\Sigma}\\mathbf{K}_{mm},$ and $\\pmb {\\Sigma} = (\\mathbf{K}_{mm} + \\sigma^{-2}\\mathbf{K}_{mn}\\mathbf{P}\\mathbf{K}_{nm})^{-1}$ Note that we can always recover $q(\\mathbf{f})$ from $q(\\mathbf{f},\\mathbf{f}_m) = p(\\mathbf{f}|\\mathbf{f}_m)\\phi (\\mathbf{f}_m)$ by marginalizing out $\\mathbf{f}_m$ , which gives optimal $\\hat{q} (\\mathbf{f}) =$ $\\mathcal{N}(\\mathbf{f}|\\boldsymbol {\\mu}_{\\mathbf{f}},\\mathbf{A})$ , where $\\pmb {\\mu_{\\mathrm{f}}} = \\mathbf{K}_{nm}\\mathbf{K}_{mm}^{-1}\\pmb {\\mu}_m$ and $\\mathbf{A} =$ $\\mathbf{K}_{\\mathbf{f}|\\mathbf{f}_m} + \\mathbf{K}_{nm}\\pmb{\\Sigma}\\mathbf{K}_{mn}$",
866
+ "bbox": [
867
+ 75,
868
+ 435,
869
+ 467,
870
+ 527
871
+ ],
872
+ "page_idx": 3
873
+ },
874
+ {
875
+ "type": "text",
876
+ "text": "Remark 1 With the optimal $\\hat{\\phi}(\\mathbf{f}_m)$ , we obtain $\\langle \\mathcal{N}_i \\rangle = \\exp(\\mathbb{E}[\\log \\mathcal{N}_i])$ as follows",
877
+ "bbox": [
878
+ 75,
879
+ 540,
880
+ 467,
881
+ 571
882
+ ],
883
+ "page_idx": 3
884
+ },
885
+ {
886
+ "type": "equation",
887
+ "text": "\n$$\n\\langle \\mathcal {N} _ {i} \\rangle = \\mathcal {N} \\left(y _ {i} \\mid \\boldsymbol {\\mu} _ {\\mathbf {f} i}, \\sigma^ {2}\\right) \\exp \\left(- \\frac {1}{2 \\sigma^ {2}} \\mathbf {A} _ {i i}\\right), \\tag {15}\n$$\n",
888
+ "text_format": "latex",
889
+ "bbox": [
890
+ 142,
891
+ 571,
892
+ 467,
893
+ 599
894
+ ],
895
+ "page_idx": 3
896
+ },
897
+ {
898
+ "type": "text",
899
+ "text": "where $\\mathcal{N}_i$ denotes $\\mathcal{N}(y_i|f_i,\\sigma^2)$ , $\\pmb{\\mu}_{\\mathbf{fi}}$ is the $i$ -th element of $\\pmb{\\mu}_{\\mathbf{f}}$ , and $\\mathbf{A}_{ii}$ is the $i$ -th diagonal element of $\\mathbf{A}$ .",
900
+ "bbox": [
901
+ 75,
902
+ 599,
903
+ 468,
904
+ 631
905
+ ],
906
+ "page_idx": 3
907
+ },
908
+ {
909
+ "type": "text",
910
+ "text": "Updating $q_{2}(\\gamma)$ indicates updating the inlier ratio or mixture. Suppose $q_{1}, q_{3}$ , and $q_{4}$ are given and note $\\hat{n} = \\operatorname{tr}(\\mathbf{P})$ . According to Eq. (10), we conclude that $\\hat{q}_{2}(\\gamma)$ again follows a Beta distribution",
911
+ "bbox": [
912
+ 75,
913
+ 643,
914
+ 467,
915
+ 703
916
+ ],
917
+ "page_idx": 3
918
+ },
919
+ {
920
+ "type": "equation",
921
+ "text": "\n$$\n\\hat {q} _ {2} (\\gamma) = \\operatorname {B e t a} (\\gamma | B _ {a} + \\hat {n}, B _ {b} + n - \\hat {n}). \\tag {16}\n$$\n",
922
+ "text_format": "latex",
923
+ "bbox": [
924
+ 142,
925
+ 717,
926
+ 467,
927
+ 732
928
+ ],
929
+ "page_idx": 3
930
+ },
931
+ {
932
+ "type": "text",
933
+ "text": "Remark 2 With the optimal $\\hat{q}_2(\\gamma)$ , according to the properties of Beta distribution, we obtain $\\langle \\gamma \\rangle = \\exp (\\mathbb{E}[\\log \\gamma ])$ and $\\langle 1 - \\gamma \\rangle = \\exp (\\mathbb{E}[\\log (1 - \\gamma)])$ as follows",
934
+ "bbox": [
935
+ 75,
936
+ 744,
937
+ 467,
938
+ 789
939
+ ],
940
+ "page_idx": 3
941
+ },
942
+ {
943
+ "type": "equation",
944
+ "text": "\n$$\n\\langle \\gamma \\rangle = \\exp (\\psi (B _ {a} + \\hat {n}) - \\psi (B _ {a} + B _ {b} + n)), \\tag {17}\n$$\n",
945
+ "text_format": "latex",
946
+ "bbox": [
947
+ 117,
948
+ 791,
949
+ 467,
950
+ 813
951
+ ],
952
+ "page_idx": 3
953
+ },
954
+ {
955
+ "type": "equation",
956
+ "text": "\n$$\n\\langle 1 - \\gamma \\rangle = \\exp (\\psi (B _ {b} + n - \\hat {n}) - \\psi (B _ {a} + B _ {b} + n)), \\tag {17}\n$$\n",
957
+ "text_format": "latex",
958
+ "bbox": [
959
+ 89,
960
+ 806,
961
+ 467,
962
+ 825
963
+ ],
964
+ "page_idx": 3
965
+ },
966
+ {
967
+ "type": "text",
968
+ "text": "where $\\psi (\\cdot)$ is the digamma function.",
969
+ "bbox": [
970
+ 76,
971
+ 827,
972
+ 316,
973
+ 842
974
+ ],
975
+ "page_idx": 3
976
+ },
977
+ {
978
+ "type": "text",
979
+ "text": "Updating $q_{3}(\\mathbf{Z})$ represents determining the probability of each data being an inlier given $q_{1}, q_{2}$ , and $q_{4}$ . It encodes the mode of each data point. From Eq. (10), it leads to a",
980
+ "bbox": [
981
+ 75,
982
+ 854,
983
+ 467,
984
+ 900
985
+ ],
986
+ "page_idx": 3
987
+ },
988
+ {
989
+ "type": "text",
990
+ "text": "further factorization $\\hat{q}_3(\\mathbf{Z}) = \\prod_{i=1}^n \\hat{q}_3^{[i]}(z_i)$ and $\\hat{q}_3^{[i]}(z_i)$ is shown to follow a Bernoulli distribution:",
991
+ "bbox": [
992
+ 498,
993
+ 88,
994
+ 890,
995
+ 119
996
+ ],
997
+ "page_idx": 3
998
+ },
999
+ {
1000
+ "type": "equation",
1001
+ "text": "\n$$\n\\hat {q} _ {3} ^ {[ i ]} \\left(z _ {i}\\right) = \\left(1 - p _ {i}\\right) ^ {1 - z _ {i}} p _ {i} ^ {z _ {i}}, \\tag {18}\n$$\n",
1002
+ "text_format": "latex",
1003
+ "bbox": [
1004
+ 604,
1005
+ 119,
1006
+ 890,
1007
+ 138
1008
+ ],
1009
+ "page_idx": 3
1010
+ },
1011
+ {
1012
+ "type": "equation",
1013
+ "text": "\n$$\n\\text {w h e r e} \\quad p _ {i} = \\frac {\\langle \\gamma \\rangle \\left\\langle \\mathcal {N} _ {i} \\right\\rangle}{\\left. \\langle 1 - \\gamma \\rangle / a + \\langle \\gamma \\rangle \\left\\langle \\mathcal {N} _ {i} \\right\\rangle \\right.}. \\tag {19}\n$$\n",
1014
+ "text_format": "latex",
1015
+ "bbox": [
1016
+ 500,
1017
+ 151,
1018
+ 890,
1019
+ 185
1020
+ ],
1021
+ "page_idx": 3
1022
+ },
1023
+ {
1024
+ "type": "text",
1025
+ "text": "Remark 3 $p_i$ shows the posterior probability of the $i$ -th data being an inlier. It depends on both the deviation between $y_i$ and the posterior mean $\\mu_{\\mathrm{fi}}$ , and the term containing the posterior variance $\\mathbf{A}_{ii}$ . The importance of this variance term is that it encodes the uncertainty of the latent function at $\\mathbf{x}_i$ , which means if we are uncertain about that data, then we cannot identify it as an inlier, even though $y_i$ and $\\mu_{\\mathrm{fi}}$ are close. This is helpful to avoid overfitting.",
1026
+ "bbox": [
1027
+ 496,
1028
+ 193,
1029
+ 890,
1030
+ 314
1031
+ ],
1032
+ "page_idx": 3
1033
+ },
1034
+ {
1035
+ "type": "text",
1036
+ "text": "Updating $q_{4}(\\sigma^{2},\\varphi ,\\mathbf{X}_{m})$ involves the optimization of hyperparameters. Specifically, $\\sigma^2$ controls the noise level, $\\varphi$ represents the shape of the kernel function, and $\\mathbf{X}_m$ is the location of the pseudo-inputs. As assumed previously, $q_{4}(\\sigma^{2},\\varphi ,\\mathbf{X}_{m})$ obeys the Dirac delta distribution, which means we directly maximize the lower bound (11) with respect to $(\\sigma^2,\\varphi ,\\mathbf{X}_m)$ , rather than using Eq. (10).",
1037
+ "bbox": [
1038
+ 496,
1039
+ 324,
1040
+ 890,
1041
+ 430
1042
+ ],
1043
+ "page_idx": 3
1044
+ },
1045
+ {
1046
+ "type": "text",
1047
+ "text": "Given $q_{1}$ , $q_{2}$ , and $q_{3}$ , taking derivative of (11) over $\\sigma^2$ and setting to zero, we obtain a closed-form expression",
1048
+ "bbox": [
1049
+ 498,
1050
+ 430,
1051
+ 890,
1052
+ 460
1053
+ ],
1054
+ "page_idx": 3
1055
+ },
1056
+ {
1057
+ "type": "equation",
1058
+ "text": "\n$$\n\\hat {\\sigma} ^ {2} = \\frac {1}{\\hat {n}} \\left(\\mathbf {y} - \\boldsymbol {\\mu} _ {\\mathbf {f}}\\right) ^ {\\top} \\mathbf {P} \\left(\\mathbf {y} - \\boldsymbol {\\mu} _ {\\mathbf {f}}\\right) + \\frac {1}{\\hat {n}} \\operatorname {t r} (\\mathbf {P A}), \\tag {20}\n$$\n",
1059
+ "text_format": "latex",
1060
+ "bbox": [
1061
+ 535,
1062
+ 460,
1063
+ 890,
1064
+ 489
1065
+ ],
1066
+ "page_idx": 3
1067
+ },
1068
+ {
1069
+ "type": "text",
1070
+ "text": "where $\\operatorname{tr}(\\cdot)$ is the trace. The lower bound (11) is, however, difficult to optimize for the remaining hyperparameters $(\\varphi, \\mathbf{X}_m)$ since the integral is intractable. To this end, we use the reverse Jensen's inequality and obtain",
1071
+ "bbox": [
1072
+ 496,
1073
+ 489,
1074
+ 890,
1075
+ 547
1076
+ ],
1077
+ "page_idx": 3
1078
+ },
1079
+ {
1080
+ "type": "equation",
1081
+ "text": "\n$$\n\\mathcal {L} _ {2} := \\log \\mathcal {N} (\\mathbf {y} | 0, \\mathbf {K} _ {\\mathbf {y}}) - \\frac {1}{2 \\sigma^ {2}} \\operatorname {t r} \\left(\\mathbf {P K} _ {\\mathbf {f} | \\mathbf {f} _ {m}}\\right) + \\text {c o n s t .} \\leq \\mathcal {L}, \\tag {21}\n$$\n",
1082
+ "text_format": "latex",
1083
+ "bbox": [
1084
+ 504,
1085
+ 558,
1086
+ 890,
1087
+ 585
1088
+ ],
1089
+ "page_idx": 3
1090
+ },
1091
+ {
1092
+ "type": "text",
1093
+ "text": "where $\\mathbf{K_y} = \\sigma^2\\mathbf{P}^{-1} + \\mathbf{K}_{nm}\\mathbf{K}_{mm}^{-1}\\mathbf{K}_{mn}$ . To set the optimized hyperparameters $(\\varphi, \\mathbf{X}_m)$ , we seek partial derivatives of $\\mathcal{L}_2$ with respect to them. Thereafter, the standard gradient descent algorithm provides the value that maximizes $\\mathcal{L}_2$ .",
1094
+ "bbox": [
1095
+ 496,
1096
+ 595,
1097
+ 890,
1098
+ 657
1099
+ ],
1100
+ "page_idx": 3
1101
+ },
1102
+ {
1103
+ "type": "text",
1104
+ "text": "4.5. Stochastic Variational Inference",
1105
+ "text_level": 1,
1106
+ "bbox": [
1107
+ 500,
1108
+ 665,
1109
+ 781,
1110
+ 680
1111
+ ],
1112
+ "page_idx": 3
1113
+ },
1114
+ {
1115
+ "type": "text",
1116
+ "text": "Modern applications often involve handling massive data. However, the GPR equipped with variational inference does not easily scale to big data, even with inducing variables. This is because we need to go through the entire training data at each coordinate update. As the training data size grows, the computational overhead becomes larger.",
1117
+ "bbox": [
1118
+ 496,
1119
+ 688,
1120
+ 890,
1121
+ 779
1122
+ ],
1123
+ "page_idx": 3
1124
+ },
1125
+ {
1126
+ "type": "text",
1127
+ "text": "An alternative to coordinate ascent is to use SVI [20], which is similar to stochastic gradient descent (SGD), that performs stochastic optimization from the noisy but cheap-to-compute estimates of the gradient. It iterates between subsampling a subset of data and adjusting the hidden structure solely based on the subset. Thus, SVI is much more efficient than conventional variational inference. To apply SVI to the probabilistic model, a set of global variables is",
1128
+ "bbox": [
1129
+ 496,
1130
+ 780,
1131
+ 890,
1132
+ 900
1133
+ ],
1134
+ "page_idx": 3
1135
+ },
1136
+ {
1137
+ "type": "page_number",
1138
+ "text": "21953",
1139
+ "bbox": [
1140
+ 478,
1141
+ 944,
1142
+ 517,
1143
+ 955
1144
+ ],
1145
+ "page_idx": 3
1146
+ },
1147
+ {
1148
+ "type": "text",
1149
+ "text": "needed [18]. In our robust GPR setting, the inducing variables $\\mathbf{f}_m$ and mixing coefficient variable $\\gamma$ fill this role.",
1150
+ "bbox": [
1151
+ 75,
1152
+ 90,
1153
+ 468,
1154
+ 119
1155
+ ],
1156
+ "page_idx": 4
1157
+ },
1158
+ {
1159
+ "type": "text",
1160
+ "text": "Natural Gradient. In gradient-based optimization, the natural gradient accounts for the geometric structure of probability parameters and is used in SVI to replace the Euclidean gradient. In exponential families, it is given by premultiplying the usual gradient by the inverse Fisher information matrix $G(\\eta)^{-1}$ , where $\\eta$ is the canonical parameter of the exponential families. Since both $\\phi(\\mathbf{f}_m)$ and $q(\\gamma)$ are exponential family distributions, they enjoy simple natural gradients [19] of the bound (11) such that $g(\\eta) = \\mathbb{E}[\\eta] - \\eta$ , where $g(\\cdot)$ denotes the natural gradient. For Gaussian distribution $\\phi(\\mathbf{f}_m)$ , its canonical parameter is $[\\mathbf{A}_m^{-1}\\boldsymbol{\\mu}_m, -\\frac{1}{2}\\mathbf{A}_m^{-1}]$ . For Beta distribution $q(\\gamma)$ , its canonical parameter is $[B_a + \\hat{n}, B_b + n - \\hat{n}]$ . Thereafter, we obtain the update rule using natural gradient with a step size $\\epsilon_t$",
1161
+ "bbox": [
1162
+ 75,
1163
+ 121,
1164
+ 472,
1165
+ 333
1166
+ ],
1167
+ "page_idx": 4
1168
+ },
1169
+ {
1170
+ "type": "equation",
1171
+ "text": "\n$$\n\\boldsymbol {\\eta} ^ {t} = \\boldsymbol {\\eta} ^ {t - 1} + \\varepsilon_ {t} g (\\boldsymbol {\\eta} ^ {t - 1}) = (1 - \\varepsilon_ {t}) \\boldsymbol {\\eta} ^ {t - 1} + \\varepsilon_ {t} \\mathbb {E} [ \\boldsymbol {\\eta} ]. \\tag {22}\n$$\n",
1172
+ "text_format": "latex",
1173
+ "bbox": [
1174
+ 88,
1175
+ 342,
1176
+ 468,
1177
+ 362
1178
+ ],
1179
+ "page_idx": 4
1180
+ },
1181
+ {
1182
+ "type": "text",
1183
+ "text": "Note that when the step size $\\varepsilon_{t} = 1$ , we recover the original variational update in Eq. (14) and Eq. (16). Now, the noisy natural gradient can be easily computed by sampling either individual or mini-batch of the data.",
1184
+ "bbox": [
1185
+ 75,
1186
+ 372,
1187
+ 468,
1188
+ 431
1189
+ ],
1190
+ "page_idx": 4
1191
+ },
1192
+ {
1193
+ "type": "text",
1194
+ "text": "4.6. Making Predictions",
1195
+ "text_level": 1,
1196
+ "bbox": [
1197
+ 76,
1198
+ 441,
1199
+ 264,
1200
+ 458
1201
+ ],
1202
+ "page_idx": 4
1203
+ },
1204
+ {
1205
+ "type": "text",
1206
+ "text": "So far we have described how to infer the approximate posterior of the latent variables. To make predictions, we use the predictive distribution similar to Eq. (3):",
1207
+ "bbox": [
1208
+ 75,
1209
+ 465,
1210
+ 468,
1211
+ 511
1212
+ ],
1213
+ "page_idx": 4
1214
+ },
1215
+ {
1216
+ "type": "equation",
1217
+ "text": "\n$$\n\\begin{array}{l} p \\left(\\mathbf {f} _ {*} \\mid \\mathcal {D}, \\mathbf {X} _ {*}\\right) = \\int p \\left(\\mathbf {f} _ {*} \\mid \\mathbf {f} _ {m}\\right) p \\left(\\mathbf {f} \\mid \\mathbf {f} _ {m}\\right) \\phi \\left(\\mathbf {f} _ {m}\\right) d \\mathbf {f} _ {m} d \\mathbf {f} \\tag {23} \\\\ = \\mathcal {N} \\left(\\mathbf {f} _ {*} | \\mathbf {K} _ {* m} \\mathbf {K} _ {m m} ^ {- 1} \\boldsymbol {\\mu} _ {m}, \\mathbf {K} _ {\\mathbf {f} _ {*} | \\mathbf {f} _ {m}} + \\mathbf {K} _ {* m} \\boldsymbol {\\Sigma} \\mathbf {K} _ {m *}\\right), \\\\ \\end{array}\n$$\n",
1218
+ "text_format": "latex",
1219
+ "bbox": [
1220
+ 98,
1221
+ 511,
1222
+ 468,
1223
+ 561
1224
+ ],
1225
+ "page_idx": 4
1226
+ },
1227
+ {
1228
+ "type": "text",
1229
+ "text": "where $\\mathbf{K}_{\\mathbf{f}_{*}|\\mathbf{f}_{m}} = \\mathbf{K}_{**} - \\mathbf{K}_{*m}\\mathbf{K}_{mm}^{-1}\\mathbf{K}_{m*}$ .",
1230
+ "bbox": [
1231
+ 76,
1232
+ 571,
1233
+ 352,
1234
+ 589
1235
+ ],
1236
+ "page_idx": 4
1237
+ },
1238
+ {
1239
+ "type": "text",
1240
+ "text": "5. Applications",
1241
+ "text_level": 1,
1242
+ "bbox": [
1243
+ 76,
1244
+ 601,
1245
+ 207,
1246
+ 618
1247
+ ],
1248
+ "page_idx": 4
1249
+ },
1250
+ {
1251
+ "type": "text",
1252
+ "text": "In this section, we describe how to use the proposed GPR model to deal with real-world tasks, including feature matching and dense gene expression imputation.",
1253
+ "bbox": [
1254
+ 75,
1255
+ 626,
1256
+ 468,
1257
+ 672
1258
+ ],
1259
+ "page_idx": 4
1260
+ },
1261
+ {
1262
+ "type": "text",
1263
+ "text": "5.1. Feature Matching",
1264
+ "text_level": 1,
1265
+ "bbox": [
1266
+ 76,
1267
+ 681,
1268
+ 251,
1269
+ 698
1270
+ ],
1271
+ "page_idx": 4
1272
+ },
1273
+ {
1274
+ "type": "text",
1275
+ "text": "Feature matching aims at establishing reliable feature point correspondences. To solve the matching problem in a regression manner, we first construct a putative correspondence set $\\mathcal{P} = \\{(\\mathbf{x}_i,\\mathbf{x}_i')\\}_{i = 1}^n$ by nearest neighbor (NN) matching, where $\\mathbf{x}_i$ and $\\mathbf{x}_i'$ represent the coordinates of feature points. The putative set is then converted to motion vector set, i.e., $\\mathcal{D} = \\{(\\mathbf{x}_i,\\mathbf{y}_i)\\}_{i = 1}^n$ , where $\\mathbf{y}_i = \\mathbf{x}_i' - \\mathbf{x}_i$ . The motion of an object projected on the image plane is known to be slow-and-smooth [38], which is well described by GP. Thus, the motion vectors of inliers can be seen as sampling on the underlying model $\\mathbf{f}$ , while the outliers are spuriously distributed. Our goal is to determine the correctness of each data. The task, however, is non-trivial as there are often",
1276
+ "bbox": [
1277
+ 75,
1278
+ 704,
1279
+ 470,
1280
+ 900
1281
+ ],
1282
+ "page_idx": 4
1283
+ },
1284
+ {
1285
+ "type": "text",
1286
+ "text": "massive outliers (even up to $90\\%$ ). Moreover, the feature points in each image can reach several thousands, which requires high computational efficiency. As we will see, our model is robust and fast enough to handle such cases.",
1287
+ "bbox": [
1288
+ 496,
1289
+ 90,
1290
+ 890,
1291
+ 151
1292
+ ],
1293
+ "page_idx": 4
1294
+ },
1295
+ {
1296
+ "type": "text",
1297
+ "text": "Inference Setting. To apply our model to feature matching, some special considerations need to be set. The feature coordinates are first normalized. For the covariance kernel, we choose the squared exponential (SE) kernel $k(\\mathbf{x},\\mathbf{x}^{\\prime}) = \\lambda^{-1}\\exp (-\\beta \\| \\mathbf{x} - \\mathbf{x}^{\\prime}\\|^{2})$ . We find that fixed hyperparameters work well for most situations, which are empirically set to $\\lambda = 4$ and $\\beta = 0.1$ . The positions of the inducing variables are randomly selected from the training data and kept fixed during inference. The ratio information is also considered as a prior. The mini-batch size is set to $\\max \\left\\{\\frac{n}{8},200\\right\\}$ . After inference convergences, the $i$ -th data is decided to be inlier if $p_i > \\tau$ , where $\\tau$ is a threshold and is set to $\\tau = 0.75$ .",
1298
+ "bbox": [
1299
+ 496,
1300
+ 152,
1301
+ 892,
1302
+ 332
1303
+ ],
1304
+ "page_idx": 4
1305
+ },
1306
+ {
1307
+ "type": "text",
1308
+ "text": "Predictions. One of the attractive properties of our model for this task is that when the training is completed, the resulting posterior GP is able to predict the motion vector of each feature point. Combined with the descriptors' information, we can yield much more correct correspondences rather than being limited to putative set. More correct correspondences are crucial for downstream tasks, as they can significantly improve the accuracy.",
1309
+ "bbox": [
1310
+ 496,
1311
+ 333,
1312
+ 890,
1313
+ 455
1314
+ ],
1315
+ "page_idx": 4
1316
+ },
1317
+ {
1318
+ "type": "text",
1319
+ "text": "5.2. Dense Gene Expression Imputation",
1320
+ "text_level": 1,
1321
+ "bbox": [
1322
+ 498,
1323
+ 460,
1324
+ 807,
1325
+ 478
1326
+ ],
1327
+ "page_idx": 4
1328
+ },
1329
+ {
1330
+ "type": "text",
1331
+ "text": "Recently, technological advances have made it possible to measure spatially resolved transcriptome-wide mRNA expressions with spatial information in tissue samples. GPR is thus utilized to infer gene expression in discrete samples, generating dense gene expression in 2D, 3D, and even spatiotemporal [3]. Unfortunately, due to equipment malfunctions and process disturbances, the readout data may contain bad spots. In this case, a vanilla GPR may lead to completely wrong results and thus mislead the subsequent analysis. By comparison, the robust GPR developed in this paper copes well with outliers and is able to give meaningful imputation results. For this application, the implementation is simple and requires little extra consideration.",
1332
+ "bbox": [
1333
+ 496,
1334
+ 484,
1335
+ 890,
1336
+ 680
1337
+ ],
1338
+ "page_idx": 4
1339
+ },
1340
+ {
1341
+ "type": "text",
1342
+ "text": "6. Experiments",
1343
+ "text_level": 1,
1344
+ "bbox": [
1345
+ 500,
1346
+ 693,
1347
+ 633,
1348
+ 709
1349
+ ],
1350
+ "page_idx": 4
1351
+ },
1352
+ {
1353
+ "type": "text",
1354
+ "text": "In this section, we first conduct numerical experiments to evaluate the performance in a fully controlled environment. After that, we apply our model to tackle feature matching problem and compare it with many state-of-the-art methods tailored for feature matching. In the end, the evaluation of gene expression imputation is performed. All experiments are conducted on a desktop with Intel i7-10700 CPU, 16 GB memory with MATLAB except deep learning methods.",
1355
+ "bbox": [
1356
+ 496,
1357
+ 718,
1358
+ 890,
1359
+ 839
1360
+ ],
1361
+ "page_idx": 4
1362
+ },
1363
+ {
1364
+ "type": "text",
1365
+ "text": "6.1. Numerical Experiments",
1366
+ "text_level": 1,
1367
+ "bbox": [
1368
+ 500,
1369
+ 847,
1370
+ 718,
1371
+ 863
1372
+ ],
1373
+ "page_idx": 4
1374
+ },
1375
+ {
1376
+ "type": "text",
1377
+ "text": "We study two datasets, i.e., Neal dataset [28] and Friedman dataset [14]. The Neal dataset is one-dimensional data,",
1378
+ "bbox": [
1379
+ 500,
1380
+ 869,
1381
+ 890,
1382
+ 900
1383
+ ],
1384
+ "page_idx": 4
1385
+ },
1386
+ {
1387
+ "type": "page_number",
1388
+ "text": "21954",
1389
+ "bbox": [
1390
+ 478,
1391
+ 944,
1392
+ 519,
1393
+ 955
1394
+ ],
1395
+ "page_idx": 4
1396
+ },
1397
+ {
1398
+ "type": "image",
1399
+ "img_path": "images/bf6975106d089e2f181b198558389f9a26f504bb91fd7013d7378b6516549c3a.jpg",
1400
+ "image_caption": [
1401
+ "Figure 2. Visual examples on Neal data. Our model approximates the exact posterior GP using only inliers, even with massive outliers. The legend is shown in the bottom."
1402
+ ],
1403
+ "image_footnote": [],
1404
+ "bbox": [
1405
+ 83,
1406
+ 88,
1407
+ 444,
1408
+ 292
1409
+ ],
1410
+ "page_idx": 5
1411
+ },
1412
+ {
1413
+ "type": "text",
1414
+ "text": "where the input $x$ is drawn from a uniform distribution within $[-2.5, 2.5]$ . The target value is calculated as:",
1415
+ "bbox": [
1416
+ 75,
1417
+ 349,
1418
+ 468,
1419
+ 378
1420
+ ],
1421
+ "page_idx": 5
1422
+ },
1423
+ {
1424
+ "type": "equation",
1425
+ "text": "\n$$\nf (x) = 0. 3 + 0. 4 x + 0. 5 \\sin (2. 7 x) + 1. 1 / \\left(1 + x ^ {2}\\right).\n$$\n",
1426
+ "text_format": "latex",
1427
+ "bbox": [
1428
+ 99,
1429
+ 381,
1430
+ 444,
1431
+ 397
1432
+ ],
1433
+ "page_idx": 5
1434
+ },
1435
+ {
1436
+ "type": "text",
1437
+ "text": "The Gaussian noise is added to the target value with zero mean and variance of 0.1. Next, outliers are injected into the training data, where the value is drawn from uniform distribution within $[-5, 5]$ . The Friedman is 10-dimensional data and the function value depends on the first five dimensions",
1438
+ "bbox": [
1439
+ 75,
1440
+ 398,
1441
+ 468,
1442
+ 472
1443
+ ],
1444
+ "page_idx": 5
1445
+ },
1446
+ {
1447
+ "type": "equation",
1448
+ "text": "\n$$\nf (\\mathbf {x}) = 1 0 \\sin \\left(\\pi x _ {1} x _ {2}\\right) + 2 0 \\left(x _ {3} - 0. 5\\right) ^ {2} + 1 0 x _ {4} + 5 x _ {5},\n$$\n",
1449
+ "text_format": "latex",
1450
+ "bbox": [
1451
+ 89,
1452
+ 473,
1453
+ 455,
1454
+ 489
1455
+ ],
1456
+ "page_idx": 5
1457
+ },
1458
+ {
1459
+ "type": "text",
1460
+ "text": "while the remaining dimensions complicate the task by adding a feature selection process. The input $\\mathbf{x}$ is sampled uniformly within $[0,1]^{10}$ . Gaussian noise $\\mathcal{N}(0,0.01)$ is added and random outliers are distributed among [5, 25].",
1461
+ "bbox": [
1462
+ 75,
1463
+ 491,
1464
+ 468,
1465
+ 550
1466
+ ],
1467
+ "page_idx": 5
1468
+ },
1469
+ {
1470
+ "type": "text",
1471
+ "text": "We fix the inlier number to 100 and change the outlier ratio. GPR with Student's- $t$ (St) [32], Laplace (Lap) [32], and Gaussian mixture distribution (MEM) [8], as well as variational free energy (VEF) [37], are chosen as competitors. Some intuitive examples on Neal data are shown in Fig. 2, where outlier ratio increases from 0 to $80\\%$ . We see that the prediction by our model almost exactly reproduces the full GP prediction using only inliers. In contrast, GPR-St and -Lap give poor approximations. GPR-MEM is able to give a better answer, yet still not as good as ours, especially at extremely high outlier ratios.",
1472
+ "bbox": [
1473
+ 75,
1474
+ 551,
1475
+ 468,
1476
+ 715
1477
+ ],
1478
+ "page_idx": 5
1479
+ },
1480
+ {
1481
+ "type": "text",
1482
+ "text": "To statistically verify performance, we select mean absolute error (MAE, m), root mean square error (RMSE, r), and negative log of predictive probability (NLP, n) as metrics. 1000 testing points are generated in equal intervals. The statistical results on Neal and Friedman in Table 1 suggest that our method outperforms other alternatives in terms of robustness and accuracy. More results including KL and statistical variance are provided in suppl. material.",
1483
+ "bbox": [
1484
+ 75,
1485
+ 717,
1486
+ 468,
1487
+ 838
1488
+ ],
1489
+ "page_idx": 5
1490
+ },
1491
+ {
1492
+ "type": "text",
1493
+ "text": "6.2. Feature Matching",
1494
+ "text_level": 1,
1495
+ "bbox": [
1496
+ 76,
1497
+ 847,
1498
+ 250,
1499
+ 863
1500
+ ],
1501
+ "page_idx": 5
1502
+ },
1503
+ {
1504
+ "type": "text",
1505
+ "text": "We collect three large and diverse feature matching datasets with three different vision tasks, say YFCC100M",
1506
+ "bbox": [
1507
+ 76,
1508
+ 869,
1509
+ 468,
1510
+ 898
1511
+ ],
1512
+ "page_idx": 5
1513
+ },
1514
+ {
1515
+ "type": "table",
1516
+ "img_path": "images/7ac37f3d520986f63f35272987d6142a96b39cf6cf680f8bd40e8f213296d7bf.jpg",
1517
+ "table_caption": [
1518
+ "Table 1. Quantitative results on Neal and Friedman data with different outlier ratios. Bold indicates the best."
1519
+ ],
1520
+ "table_footnote": [],
1521
+ "table_body": "<table><tr><td colspan=\"10\">Neal [28]</td></tr><tr><td rowspan=\"2\">Method</td><td colspan=\"3\">10%</td><td colspan=\"3\">50%</td><td colspan=\"3\">80%</td></tr><tr><td>m</td><td>r</td><td>n</td><td>m</td><td>r</td><td>n</td><td>m</td><td>r</td><td>n</td></tr><tr><td>GPR-VFE [37]</td><td>0.18</td><td>0.21</td><td>0.73</td><td>0.58</td><td>0.77</td><td>1.75</td><td>0.76</td><td>1.08</td><td>1.96</td></tr><tr><td>GPR-St [32]</td><td>0.03</td><td>0.03</td><td>-0.41</td><td>0.09</td><td>0.19</td><td>0.39</td><td>0.77</td><td>1.15</td><td>1.66</td></tr><tr><td>GPR-Lap [32]</td><td>0.03</td><td>0.07</td><td>-1.17</td><td>0.42</td><td>0.56</td><td>1.62</td><td>0.35</td><td>0.72</td><td>2.03</td></tr><tr><td>GPR-MEM [8]</td><td>0.04</td><td>0.08</td><td>-1.36</td><td>0.06</td><td>0.12</td><td>-0.66</td><td>0.77</td><td>1.52</td><td>2.04</td></tr><tr><td>Ours</td><td>0.01</td><td>0.02</td><td>-1.38</td><td>0.03</td><td>0.04</td><td>-1.25</td><td>0.03</td><td>0.03</td><td>-0.89</td></tr><tr><td colspan=\"10\">Friedman [14]</td></tr><tr><td rowspan=\"2\">Method</td><td colspan=\"3\">20%</td><td colspan=\"3\">30%</td><td colspan=\"3\">50%</td></tr><tr><td>m</td><td>r</td><td>n</td><td>m</td><td>r</td><td>n</td><td>m</td><td>r</td><td>n</td></tr><tr><td>GPR-VFE [37]</td><td>1.12</td><td>1.24</td><td>2.28</td><td>1.47</td><td>1.62</td><td>3.08</td><td>2.88</td><td>3.16</td><td>6.93</td></tr><tr><td>GPR-St [32]</td><td>0.04</td><td>0.04</td><td>-0.08</td><td>0.91</td><td>1.10</td><td>3.59</td><td>1.47</td><td>2.18</td><td>7.16</td></tr><tr><td>GPR-Lap [32]</td><td>0.23</td><td>0.29</td><td>0.37</td><td>0.46</td><td>0.61</td><td>0.94</td><td>2.03</td><td>2.23</td><td>3.29</td></tr><tr><td>GPR-MEM [8]</td><td>0.11</td><td>0.21</td><td>-0.99</td><td>0.13</td><td>0.24</td><td>-0.66</td><td>2.13</td><td>2.32</td><td>2.98</td></tr><tr><td>Ours</td><td>0.04</td><td>0.06</td><td>-1.46</td><td>0.06</td><td>0.08</td><td>-0.35</td><td>0.10</td><td>0.14</td><td>0.55</td></tr></table>",
1522
+ "bbox": [
1523
+ 503,
1524
+ 119,
1525
+ 890,
1526
+ 282
1527
+ ],
1528
+ "page_idx": 5
1529
+ },
1530
+ {
1531
+ "type": "image",
1532
+ "img_path": "images/d63bae63fad5a96af15bc558e42713b63d1df2c55e121e76a9bfee1978ae7202.jpg",
1533
+ "image_caption": [
1534
+ "Figure 3. Qualitative illustration of relative pose estimation. We show the true positive in blue and false positive in red. Best viewed in color with $200\\%$ zoom in."
1535
+ ],
1536
+ "image_footnote": [],
1537
+ "bbox": [
1538
+ 503,
1539
+ 295,
1540
+ 883,
1541
+ 479
1542
+ ],
1543
+ "page_idx": 5
1544
+ },
1545
+ {
1546
+ "type": "text",
1547
+ "text": "dataset [36] (4000 pairs) for relative pose estimation, CPC dataset [42] (1000 pairs) for fundamental matrix estimation, and HPatches dataset [1] (580 pairs) for homography estimation. Six state-of-the-arts from handcrafted to deep learning are chosen for comparison, namely, MAGSAC++ [2], LPM [26], MCDM [25], CRC [12], PointCN [43], and OANet [44], with baseline ratio test. Briefly, MAGSAC++ is resampling-based method, LPM is a heuristic method that preserves local structure of feature points, CRC interpolates a smooth vector field using Fourier bases, PointCN uses deep neural networks with context normalization to encode global contextual information, OANet further captures local contextual information by clustering unordered correspondences. Note that we assess currently other robust GPR methods are not capable to handle the feature matching problem, thus none of them are chosen for comparison.",
1548
+ "bbox": [
1549
+ 496,
1550
+ 540,
1551
+ 890,
1552
+ 782
1553
+ ],
1554
+ "page_idx": 5
1555
+ },
1556
+ {
1557
+ "type": "text",
1558
+ "text": "6.2.1 Relative Pose Estimation",
1559
+ "text_level": 1,
1560
+ "bbox": [
1561
+ 500,
1562
+ 801,
1563
+ 725,
1564
+ 814
1565
+ ],
1566
+ "page_idx": 5
1567
+ },
1568
+ {
1569
+ "type": "text",
1570
+ "text": "Following [44], Yahoo's YFCC100M dataset [36] generates 4000 testing image pairs. SIFT [24] with NN matching is adopted to establish putative correspondences. The maximum number of keypoints for each image is limited to 4000. RANSAC [13] is used to estimate the relative camera pose.",
1571
+ "bbox": [
1572
+ 496,
1573
+ 824,
1574
+ 890,
1575
+ 900
1576
+ ],
1577
+ "page_idx": 5
1578
+ },
1579
+ {
1580
+ "type": "page_number",
1581
+ "text": "21955",
1582
+ "bbox": [
1583
+ 478,
1584
+ 944,
1585
+ 517,
1586
+ 955
1587
+ ],
1588
+ "page_idx": 5
1589
+ },
1590
+ {
1591
+ "type": "table",
1592
+ "img_path": "images/a2a4c42330d858a74bd3de50c323d75f71b527294e5f94b926e4572308fe6f64.jpg",
1593
+ "table_caption": [
1594
+ "Table 2. Quantitative comparison for relative pose estimation. Bold indicates the best and underline ranks the second."
1595
+ ],
1596
+ "table_footnote": [],
1597
+ "table_body": "<table><tr><td colspan=\"7\">YFCC100M [36]</td></tr><tr><td rowspan=\"2\">Method</td><td colspan=\"3\">AUC</td><td rowspan=\"2\">P</td><td rowspan=\"2\">R</td><td rowspan=\"2\">F</td></tr><tr><td>@5°</td><td>@10°</td><td>@20°</td></tr><tr><td>Ratio test</td><td>28.8</td><td>39.2</td><td>52.1</td><td>24.3</td><td>55.3</td><td>32.8</td></tr><tr><td>MSC++ [2]</td><td>24.9</td><td>33.7</td><td>45.7</td><td>24.3</td><td>55.3</td><td>32.8</td></tr><tr><td>PointCN [43]</td><td>45.1</td><td>54.6</td><td>65.0</td><td>47.2</td><td>73.4</td><td>54.9</td></tr><tr><td>OANet [44]</td><td>53.6</td><td>64.1</td><td>75.8</td><td>51.2</td><td>87.0</td><td>61.7</td></tr><tr><td>LPM [26]</td><td>38.6</td><td>48.8</td><td>59.7</td><td>51.4</td><td>43.6</td><td>45.9</td></tr><tr><td>MCDM [25]</td><td>47.8</td><td>59.4</td><td>70.7</td><td>63.0</td><td>46.5</td><td>52.7</td></tr><tr><td>CRC [12]</td><td>38.1</td><td>48.2</td><td>58.6</td><td>41.2</td><td>57.8</td><td>44.5</td></tr><tr><td>Ours</td><td>53.8</td><td>63.7</td><td>73.8</td><td>85.3</td><td>54.0</td><td>63.7</td></tr></table>",
1598
+ "bbox": [
1599
+ 114,
1600
+ 119,
1601
+ 428,
1602
+ 258
1603
+ ],
1604
+ "page_idx": 6
1605
+ },
1606
+ {
1607
+ "type": "image",
1608
+ "img_path": "images/dc04c4bb8db41b27359a91b479060ac3afea0f3337e0a7442160a15dd2c9426b.jpg",
1609
+ "image_caption": [
1610
+ "Figure 4. Visual examples of homography and fundamental matrix estimation. We show the true positive in blue and false positive in red. Best viewed in color with $200\\%$ zoom in."
1611
+ ],
1612
+ "image_footnote": [],
1613
+ "bbox": [
1614
+ 84,
1615
+ 268,
1616
+ 455,
1617
+ 433
1618
+ ],
1619
+ "page_idx": 6
1620
+ },
1621
+ {
1622
+ "type": "text",
1623
+ "text": "We measure the maximum of rotation and translation errors in degree and report the approximate Area Under the Curve (AUC) with thresholds 5, 10, and 20 degrees. Precision, recall, and F1-score before RANSAC with respect to the ground-truth inliers are also presented. See suppl. material for a detailed description of the dataset and the metrics.",
1624
+ "bbox": [
1625
+ 75,
1626
+ 494,
1627
+ 467,
1628
+ 585
1629
+ ],
1630
+ "page_idx": 6
1631
+ },
1632
+ {
1633
+ "type": "text",
1634
+ "text": "Some visualized examples are demonstrated in Fig. 3. Statistics results are comprehensively reported in Table 2. Since deep learning methods such as PointCN and OANet are trained in such dataset with geometric loss, they can capture geometric information and therefore have higher AUC compared to other handcrafted methods. By comparison, our method achieves comparable results to these deep learning methods with much higher precision and F1-score.",
1635
+ "bbox": [
1636
+ 75,
1637
+ 585,
1638
+ 467,
1639
+ 705
1640
+ ],
1641
+ "page_idx": 6
1642
+ },
1643
+ {
1644
+ "type": "text",
1645
+ "text": "6.2.2 Homography & Fundamental Matrix Estimation",
1646
+ "text_level": 1,
1647
+ "bbox": [
1648
+ 76,
1649
+ 724,
1650
+ 467,
1651
+ 739
1652
+ ],
1653
+ "page_idx": 6
1654
+ },
1655
+ {
1656
+ "type": "text",
1657
+ "text": "Homography and fundamental matrix estimation are critical parts in computer vision. The maximum keypoints for HPatches and CPC are limited to 4000. For evaluation metrics, we choose the precision, recall, and F1-score as well as the accuracy of the estimation. For homography estimation, we adopt homography error defined in [11], and for fundamental matrix, we follow [4] and use normalized symmetric geometry distance (NSGD) as the metrics. An estimate is classified as accurate or not by a threshold, which is set to 4 pixels for homography error and 0.05 for the NSGD [4].",
1658
+ "bbox": [
1659
+ 75,
1660
+ 750,
1661
+ 467,
1662
+ 900
1663
+ ],
1664
+ "page_idx": 6
1665
+ },
1666
+ {
1667
+ "type": "table",
1668
+ "img_path": "images/1dbe48d164c483c47c9234ccffd747ca50d995d9c198642eda7b7eb48381e2af.jpg",
1669
+ "table_caption": [
1670
+ "Table 3. Statistical results for homography and fundamental matrix estimation. Bold ranks the first, underline the second."
1671
+ ],
1672
+ "table_footnote": [],
1673
+ "table_body": "<table><tr><td rowspan=\"2\">Method</td><td colspan=\"4\">HPatches [1]</td><td colspan=\"4\">CPC [42]</td></tr><tr><td>Acc.</td><td>P</td><td>R</td><td>F</td><td>Acc.</td><td>P</td><td>R</td><td>F</td></tr><tr><td>Ratio test</td><td>71.7</td><td>71.6</td><td>51.6</td><td>63.0</td><td>17.0</td><td>12.7</td><td>63.9</td><td>20.5</td></tr><tr><td>MSC++ [2]</td><td>74.4</td><td>87.3</td><td>92.4</td><td>86.0</td><td>6.9</td><td>19.7</td><td>29.6</td><td>19.4</td></tr><tr><td>PointCN [43]</td><td>73.1</td><td>76.9</td><td>85.7</td><td>78.6</td><td>29.3</td><td>19.8</td><td>70.2</td><td>28.8</td></tr><tr><td>OANet [44]</td><td>70.0</td><td>73.5</td><td>77.2</td><td>72.1</td><td>47.0</td><td>34.6</td><td>81.0</td><td>45.4</td></tr><tr><td>LPM [26]</td><td>60.6</td><td>72.5</td><td>50.5</td><td>52.3</td><td>23.8</td><td>21.9</td><td>48.9</td><td>26.1</td></tr><tr><td>MCDM [25]</td><td>77.7</td><td>62.9</td><td>91.0</td><td>73.0</td><td>34.8</td><td>23.0</td><td>61.0</td><td>32.5</td></tr><tr><td>CRC [12]</td><td>71.7</td><td>75.5</td><td>92.5</td><td>80.4</td><td>28.7</td><td>24.8</td><td>47.2</td><td>30.4</td></tr><tr><td>Ours</td><td>78.1</td><td>89.9</td><td>94.3</td><td>91.5</td><td>49.6</td><td>61.5</td><td>59.3</td><td>57.2</td></tr><tr><td>Ours*</td><td>82.4</td><td>73.9</td><td>109.6</td><td>-</td><td>51.9</td><td>39.2</td><td>97.3</td><td>-</td></tr></table>",
1674
+ "bbox": [
1675
+ 501,
1676
+ 116,
1677
+ 890,
1678
+ 255
1679
+ ],
1680
+ "page_idx": 6
1681
+ },
1682
+ {
1683
+ "type": "table",
1684
+ "img_path": "images/b869c03b445fc7fbf651123e7c30fef54ed2c422347dc29b8b6bcf8eff87aa53.jpg",
1685
+ "table_caption": [
1686
+ "Table 4. Ablation studies on feature matching. Bold is the best."
1687
+ ],
1688
+ "table_footnote": [],
1689
+ "table_body": "<table><tr><td colspan=\"2\">Method</td><td>Prec.</td><td>Recall</td><td>F1</td><td>Time</td></tr><tr><td colspan=\"2\">CRC [12]</td><td>50.7</td><td>67.8</td><td>54.9</td><td>68.1</td></tr><tr><td colspan=\"2\">OANet (on GPU) [44]</td><td>47.9</td><td>75.2</td><td>56.0</td><td>41.2</td></tr><tr><td rowspan=\"5\">RS-GPR</td><td>No SVI</td><td>76.3</td><td>66.1</td><td>70.0</td><td>47.1</td></tr><tr><td>No inducing variables</td><td>72.6</td><td>69.0</td><td>69.5</td><td>1.6e5</td></tr><tr><td>Opt. Hyperparameters</td><td>85.0</td><td>61.2</td><td>67.7</td><td>1556</td></tr><tr><td>EM optimization</td><td>67.6</td><td>69.8</td><td>65.5</td><td>43.7</td></tr><tr><td>Default</td><td>88.0</td><td>68.2</td><td>75.6</td><td>9.8</td></tr></table>",
1690
+ "bbox": [
1691
+ 519,
1692
+ 280,
1693
+ 870,
1694
+ 378
1695
+ ],
1696
+ "page_idx": 6
1697
+ },
1698
+ {
1699
+ "type": "text",
1700
+ "text": "Several visual examples are provided in Fig. 4. Note that our predictive model is also presented and indicated by the superscript ${}^{*}\\ast$ . It establishes significantly more correct correspondences compared to other methods. Table 3 reports the statistical results. As it indicates, our method outperforms all other methods in the two tasks. When using the predictive model, it can even produce more correct correspondences than the putative set (e.g., the recall in HPatches is higher than 100), thus further improving accuracy.",
1701
+ "bbox": [
1702
+ 496,
1703
+ 396,
1704
+ 890,
1705
+ 532
1706
+ ],
1707
+ "page_idx": 6
1708
+ },
1709
+ {
1710
+ "type": "text",
1711
+ "text": "6.2.3 Ablation Studies of Feature Matching",
1712
+ "text_level": 1,
1713
+ "bbox": [
1714
+ 500,
1715
+ 554,
1716
+ 815,
1717
+ 569
1718
+ ],
1719
+ "page_idx": 6
1720
+ },
1721
+ {
1722
+ "type": "text",
1723
+ "text": "To measure performance, especially runtime, in different settings, we conduct ablation studies on 100 image pairs from YFCC100M. It mainly includes four choices, i.e., not using SVI, not using inducing variables, optimizing the kernel hyperparameters, and using EM optimization instead of variational inference. We also include the CRC and OANet to compare the runtime. The results are reported in Table 4, where we see that our method performs the best with the default setting described in Sect. 5.1. In particular, the running time is much faster compared to other settings and methods.",
1724
+ "bbox": [
1725
+ 496,
1726
+ 578,
1727
+ 890,
1728
+ 729
1729
+ ],
1730
+ "page_idx": 6
1731
+ },
1732
+ {
1733
+ "type": "text",
1734
+ "text": "6.3. Dense Gene Expression Imputation",
1735
+ "text_level": 1,
1736
+ "bbox": [
1737
+ 500,
1738
+ 739,
1739
+ 805,
1740
+ 756
1741
+ ],
1742
+ "page_idx": 6
1743
+ },
1744
+ {
1745
+ "type": "text",
1746
+ "text": "We collect two spatial transcriptomics datasets from mouse olfactory bulb and human breast cancer biopsies [34], respectively. Briefly, the spatial transcriptomics gene expression data is measured from thin tissue sections placed on an array with ploy probes and spatially resolved DNA barcodes. These create a grid of circular spots where the mRNA abundance of each spot is measured using probes with the barcodes encoding the spatial locations. The mouse olfactory bulb and human breast cancer biopsies datasets",
1747
+ "bbox": [
1748
+ 496,
1749
+ 763,
1750
+ 890,
1751
+ 900
1752
+ ],
1753
+ "page_idx": 6
1754
+ },
1755
+ {
1756
+ "type": "page_number",
1757
+ "text": "21956",
1758
+ "bbox": [
1759
+ 478,
1760
+ 945,
1761
+ 517,
1762
+ 955
1763
+ ],
1764
+ "page_idx": 6
1765
+ },
1766
+ {
1767
+ "type": "image",
1768
+ "img_path": "images/f18ddbf02269356ae12c989aebb22acf2f900a111dfd7aa45aba705fdda444c2.jpg",
1769
+ "image_caption": [
1770
+ "Figure 5. Visualization of dense spatial gene expression imputation. The spatial transcriptomics data is plotted on the top row. The bottom row shows the imputation results of different methods. We stack the spatial transcriptomics (shown as dots) on the imputed results to show the quality of the imputation."
1771
+ ],
1772
+ "image_footnote": [],
1773
+ "bbox": [
1774
+ 83,
1775
+ 90,
1776
+ 271,
1777
+ 199
1778
+ ],
1779
+ "page_idx": 7
1780
+ },
1781
+ {
1782
+ "type": "image",
1783
+ "img_path": "images/55745084f69e7e663fc4106324058dc6db9f1fa989348d7b22a8e582325efa9a.jpg",
1784
+ "image_caption": [],
1785
+ "image_footnote": [],
1786
+ "bbox": [
1787
+ 274,
1788
+ 90,
1789
+ 462,
1790
+ 199
1791
+ ],
1792
+ "page_idx": 7
1793
+ },
1794
+ {
1795
+ "type": "table",
1796
+ "img_path": "images/7e2a87e4e76fbbe587afdb4002f46d7baea7049f056a8a401616c3b08ea3db79.jpg",
1797
+ "table_caption": [
1798
+ "Table 5. Statistical results of dense gene expression imputation. Bold indicates the best."
1799
+ ],
1800
+ "table_footnote": [],
1801
+ "table_body": "<table><tr><td colspan=\"7\">Mouse olfactory bulb</td></tr><tr><td rowspan=\"2\">Method</td><td colspan=\"3\">Pattern 1</td><td colspan=\"3\">Pattern 2</td></tr><tr><td>m</td><td>r</td><td>n</td><td>m</td><td>r</td><td>n</td></tr><tr><td>GPR-Exact</td><td>0.46</td><td>0.59</td><td>1.25</td><td>0.33</td><td>0.40</td><td>1.10</td></tr><tr><td>GPR-St [32]</td><td>0.07</td><td>0.14</td><td>-0.57</td><td>0.02</td><td>0.04</td><td>-1.33</td></tr><tr><td>GPR-Lap [32]</td><td>0.20</td><td>0.27</td><td>0.33</td><td>0.48</td><td>0.54</td><td>0.94</td></tr><tr><td>Ours</td><td>0.03</td><td>0.05</td><td>-1.99</td><td>0.02</td><td>0.04</td><td>-2.21</td></tr><tr><td colspan=\"7\">Breast cancer tissue</td></tr><tr><td rowspan=\"2\">Method</td><td colspan=\"3\">Pattern 1</td><td colspan=\"3\">Pattern 2</td></tr><tr><td>m</td><td>r</td><td>n</td><td>m</td><td>r</td><td>n</td></tr><tr><td>GPR-Exact</td><td>0.24</td><td>0.35</td><td>1.44</td><td>0.23</td><td>0.33</td><td>1.15</td></tr><tr><td>GPR-St [32]</td><td>0.19</td><td>0.32</td><td>0.37</td><td>0.18</td><td>0.30</td><td>0.22</td></tr><tr><td>GPR-Lap [32]</td><td>0.25</td><td>0.36</td><td>0.56</td><td>0.25</td><td>0.37</td><td>0.53</td></tr><tr><td>Ours</td><td>0.12</td><td>0.16</td><td>-0.54</td><td>0.11</td><td>0.17</td><td>-0.39</td></tr></table>",
1802
+ "bbox": [
1803
+ 117,
1804
+ 311,
1805
+ 429,
1806
+ 484
1807
+ ],
1808
+ "page_idx": 7
1809
+ },
1810
+ {
1811
+ "type": "text",
1812
+ "text": "contain 14859 and 12856 measured genes, respectively. To identify spatially variable genes, we adopt SpatialDE [35] and select two patterns for each dataset. We next add 5 random outliers to each identified pattern to simulate the outlier scenario. In the end, different GPR methods are applied to impute dense spatial gene expression.",
1813
+ "bbox": [
1814
+ 75,
1815
+ 500,
1816
+ 468,
1817
+ 589
1818
+ ],
1819
+ "page_idx": 7
1820
+ },
1821
+ {
1822
+ "type": "text",
1823
+ "text": "Two visual examples are presented in Fig. 5, where in each example, the second row shows the imputed results. As can be seen, our method accurately infers the correct gene expression regardless of the outliers. To give quantitative results, we randomly split each dataset into $80\\%$ training data and $20\\%$ test data. As mentioned above, we use the MAE, RMSE, and NLP to measure the performance, and report the statistical results in Table 5. We see that our method always has the lowest error.",
1824
+ "bbox": [
1825
+ 75,
1826
+ 590,
1827
+ 468,
1828
+ 726
1829
+ ],
1830
+ "page_idx": 7
1831
+ },
1832
+ {
1833
+ "type": "text",
1834
+ "text": "7. Analysis",
1835
+ "text_level": 1,
1836
+ "bbox": [
1837
+ 76,
1838
+ 739,
1839
+ 171,
1840
+ 757
1841
+ ],
1842
+ "page_idx": 7
1843
+ },
1844
+ {
1845
+ "type": "text",
1846
+ "text": "Optimization Process. We visualize the intermediate results of the optimization process on Neal data with $50\\%$ outliers and present it in the top row of Fig. 6. We see that the field of $p$ in Eq. (19) gradually converges to the inliers. After convergence, the posterior mean and variance functions are very close to the ground truth, and the inducing variables are restricted to the interval of the inliers.",
1847
+ "bbox": [
1848
+ 75,
1849
+ 763,
1850
+ 468,
1851
+ 868
1852
+ ],
1853
+ "page_idx": 7
1854
+ },
1855
+ {
1856
+ "type": "text",
1857
+ "text": "The Role of Term $\\mathbf{A}_{ii}$ . We remove the term containing $\\mathbf{A}_{ii}$ in Eq. (15) and show the results in the bottom of Fig. 6.",
1858
+ "bbox": [
1859
+ 76,
1860
+ 869,
1861
+ 468,
1862
+ 900
1863
+ ],
1864
+ "page_idx": 7
1865
+ },
1866
+ {
1867
+ "type": "image",
1868
+ "img_path": "images/63607d253c4cdef0168a9b0993d4d1c20ff65bda2af6dc79d4cd66053cea2219.jpg",
1869
+ "image_caption": [
1870
+ "Figure 6. Intermediate results of the optimization process. The top row shows the full model while the bottom removes term $\\mathbf{A}_{ii}$ . We visualize the field of $p$ , which indicates the inlier probability at any location. The deeper the blue color, the higher the probability."
1871
+ ],
1872
+ "image_footnote": [],
1873
+ "bbox": [
1874
+ 506,
1875
+ 90,
1876
+ 887,
1877
+ 204
1878
+ ],
1879
+ "page_idx": 7
1880
+ },
1881
+ {
1882
+ "type": "image",
1883
+ "img_path": "images/7675747a1a2dd41bde5fa61c995b2e69b4a3152c8469d6feffc58aea63ae31f3.jpg",
1884
+ "image_caption": [
1885
+ "Figure 7. Our model deal with different non-uniform outliers."
1886
+ ],
1887
+ "image_footnote": [],
1888
+ "bbox": [
1889
+ 519,
1890
+ 271,
1891
+ 867,
1892
+ 332
1893
+ ],
1894
+ "page_idx": 7
1895
+ },
1896
+ {
1897
+ "type": "image",
1898
+ "img_path": "images/62e6027f2054a4ce502523dbd7253a07b2680fd0a94db254e8cfea899846273d.jpg",
1899
+ "image_caption": [
1900
+ "Figure 8. The inference time of each GPR method. Our model with SVI can easily handle big data."
1901
+ ],
1902
+ "image_footnote": [],
1903
+ "bbox": [
1904
+ 545,
1905
+ 359,
1906
+ 841,
1907
+ 438
1908
+ ],
1909
+ "page_idx": 7
1910
+ },
1911
+ {
1912
+ "type": "text",
1913
+ "text": "Without variance term, the regression tends to overfit.",
1914
+ "bbox": [
1915
+ 500,
1916
+ 484,
1917
+ 852,
1918
+ 498
1919
+ ],
1920
+ "page_idx": 7
1921
+ },
1922
+ {
1923
+ "type": "text",
1924
+ "text": "Non-uniform Outliers. Different types of non-uniform outliers are tested in Fig. 7. More detailed results are given in suppl. material. Our model exhibits strong generalization and robustness. The reason is that uniform distribution is a very weak assumption with high flexibility. It accommodates many outlier distributions.",
1925
+ "bbox": [
1926
+ 496,
1927
+ 500,
1928
+ 890,
1929
+ 589
1930
+ ],
1931
+ "page_idx": 7
1932
+ },
1933
+ {
1934
+ "type": "text",
1935
+ "text": "**Scalability.** We test the runtime towards different problem sizes on Neal data with $50\\%$ outliers. Other methods as well as ours w. and w/o. SVI are included. The number of inducing variables is set to 15 for our method in both two settings. As plotted in Fig. 8, our method can easily scale to large-scale data, especially when equipped with SVI.",
1936
+ "bbox": [
1937
+ 496,
1938
+ 590,
1939
+ 890,
1940
+ 680
1941
+ ],
1942
+ "page_idx": 7
1943
+ },
1944
+ {
1945
+ "type": "text",
1946
+ "text": "8. Conclusion",
1947
+ "text_level": 1,
1948
+ "bbox": [
1949
+ 500,
1950
+ 694,
1951
+ 617,
1952
+ 709
1953
+ ],
1954
+ "page_idx": 7
1955
+ },
1956
+ {
1957
+ "type": "text",
1958
+ "text": "This paper presents a robust and scalable GPR using variational learning. It uses an outlier robust mixture likelihood model, where the uniform distribution accounts for the outliers. A variational formulation is introduced to learn the mode of data and hyperparameters by minimizing the KL divergence between the true posterior distribution and the approximate one. Its most attractive property is that it can rigorously approximate the exact posterior. Inducing variable approximation and SVI further extend our model to big data. We apply our model to feature matching and dense gene expression imputation. Extensive results show the significant improvement over existing robust GPRs.",
1959
+ "bbox": [
1960
+ 496,
1961
+ 719,
1962
+ 890,
1963
+ 900
1964
+ ],
1965
+ "page_idx": 7
1966
+ },
1967
+ {
1968
+ "type": "page_number",
1969
+ "text": "21957",
1970
+ "bbox": [
1971
+ 478,
1972
+ 944,
1973
+ 517,
1974
+ 955
1975
+ ],
1976
+ "page_idx": 7
1977
+ },
1978
+ {
1979
+ "type": "text",
1980
+ "text": "References",
1981
+ "text_level": 1,
1982
+ "bbox": [
1983
+ 78,
1984
+ 89,
1985
+ 174,
1986
+ 106
1987
+ ],
1988
+ "page_idx": 8
1989
+ },
1990
+ {
1991
+ "type": "list",
1992
+ "sub_type": "ref_text",
1993
+ "list_items": [
1994
+ "[1] Vassileios Balntas, Karel Lenc, Andrea Vedaldi, and Krystian Mikolajczyk. Hpatches: A benchmark and evaluation of handcrafted and learned local descriptors. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5173-5182, 2017. 6, 7",
1995
+ "[2] Daniel Barath, Jana Noskova, Maksym Ivashechkin, and Jiri Matas. Magsac++, a fast, reliable and accurate robust estimator. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1304-1312, 2020. 6, 7",
1996
+ "[3] Sophie Bergmann, Christopher A Penfold, Erin Slatery, Dylan Siriwardena, Charis Drummer, Stephen Clark, Stanley E Strawbridge, Keiko Kishimoto, Alice Vickers, Mukul Tewary, et al. Spatial profiling of early primate gastrulation in utero. Nature, 609(7925):136-143, 2022. 1, 5",
1997
+ "[4] Jia-Wang Bian, Yu-Huan Wu, Ji Zhao, Yun Liu, Le Zhang, Ming-Ming Cheng, and Ian Reid. An evaluation of feature matchers for fundamental matrix estimation. In Proceedings of the British Machine Vision Conference, 2019. 7",
1998
+ "[5] Christopher M Bishop and Nasser M Nasrabadi. Pattern Recognition and Machine Learning. Springer, 2006. 4",
1999
+ "[6] Michael J Black and Anand Rangarajan. On the unification of line processes, outlier rejection, and robust statistics with applications in early vision. International Journal of Computer Vision, 19(1):57-91, 1996. 1",
2000
+ "[7] Lehel Csato and Manfred Opper. Sparse on-line gaussian processes. Neural Computation, 14(3):641-668, 2002. 2",
2001
+ "[8] Atefeh Daemi, Hariprasad Kodamana, and Biao Huang. Gaussian process modelling with gaussian mixture likelihood. Journal of Process Control, 81:209-220, 2019. 1, 2, 6",
2002
+ "[9] A Philip Dawid. Posterior expectations for large observations. Biometrika, 60(3):664-667, 1973. 1",
2003
+ "[10] Bruno De Finetti. The bayesian approach to the rejection of outliers. In Proceedings of the fourth Berkeley Symposium on Probability and Statistics, pages 199–210, 1961. 1",
2004
+ "[11] Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superpoint: Self-supervised interest point detection and description. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops, pages 224-236, 2018. 7",
2005
+ "[12] Aoxiang Fan, Xingyu Jiang, Yong Ma, Xiaoguang Mei, and Jiayi Ma. Smoothness-driven consensus based on compact representation for robust feature matching. IEEE Transactions on Neural Networks and Learning Systems, 2021. 6, 7",
2006
+ "[13] Martin A Fischler and Robert C Bolles. Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. Communications of the ACM, 24(6):381-395, 1981. 6",
2007
+ "[14] Jerome H Friedman. Multivariate adaptive regression splines. The Annals of Statistics, 19(1):1-67, 1991. 5, 6",
2008
+ "[15] Yarin Gal and Richard Turner. Improving the gaussian process sparse spectrum approximation by representing uncertainty in frequency inputs. In Proceedings of the Interna"
2009
+ ],
2010
+ "bbox": [
2011
+ 78,
2012
+ 114,
2013
+ 470,
2014
+ 900
2015
+ ],
2016
+ "page_idx": 8
2017
+ },
2018
+ {
2019
+ "type": "list",
2020
+ "sub_type": "ref_text",
2021
+ "list_items": [
2022
+ "tional Conference on Machine Learning, pages 655-664, 2015. 2",
2023
+ "[16] John Geweke. Bayesian treatment of the independent student-t linear model. Journal of Applied Econometrics, 8(S1):S19-S40, 1993. 1",
2024
+ "[17] Paul Goldberg, Christopher Williams, and Christopher Bishop. Regression with input-dependent noise: A gaussian process treatment. Advances in Neural Information Processing Systems, 10, 1997. 1",
2025
+ "[18] James Hensman, Nicolo Fusi, and Neil D Lawrence. Gaussian processes for big data. arXiv preprint arXiv:1309.6835, 2013. 2, 5",
2026
+ "[19] James Hensman, Magnus Rattray, and Neil Lawrence. Fast variational inference in the conjugate exponential family. Advances in Neural Information Processing Systems, 25, 2012. 5",
2027
+ "[20] Matthew D Hoffman, David M Blei, Chong Wang, and John Paisley. Stochastic variational inference. Journal of Machine Learning Research, 14:1303-1347, 2013. 2, 4",
2028
+ "[21] Pasi Jylänki, Jarno Vanhatalo, and Aki Vehtari. Robust gaussian process regression with a student-t likelihood. Journal of Machine Learning Research, 12(11):3227-3257, 2011. 1",
2029
+ "[22] Malte Kuss. Gaussian process models for robust regression, classification, and reinforcement learning. PhD thesis, ech-nische Universität Darmstadt Darmstadt, Germany, 2006. 1, 2",
2030
+ "[23] Miguel Lázaro-Gredilla, Joaquin Quinonero-Candela, Carl Edward Rasmussen, and Aníbal R Figueiras-Vidal. Sparse spectrum gaussian process regression. The Journal of Machine Learning Research, 11:1865-1881, 2010. 2",
2031
+ "[24] David G Lowe. Distinctive image features from scale-invariant keypoints. International Journal of Computer Vision, 60(2):91-110, 2004. 6",
2032
+ "[25] Jiayi Ma, Aoxiang Fan, Xingyu Jiang, and Guobao Xiao. Feature matching via motion-consistency driven probabilistic graphical model. International Journal of Computer Vision, 130(9):2249-2264, 2022. 6, 7",
2033
+ "[26] Jiayi Ma, Ji Zhao, Junjun Jiang, Huabing Zhou, and Xiaojie Guo. Locality preserving matching. International Journal of Computer Vision, 127(5):512-531, 2019. 6, 7",
2034
+ "[27] Andrew Naish-Guzman and Sean Holden. Robust regression with twinned gaussian processes. Advances in Neural Information Processing Systems, 20, 2007. 1",
2035
+ "[28] Radford M Neal. Monte carlo implementation of gaussian process models for bayesian regression and classification. arXiv preprint physics/9701026, 1997. 1, 5, 6",
2036
+ "[29] Anthony O'Hagan. On outlier rejection phenomena in bayes inference. Journal of the Royal Statistical Society: Series B (Methodological), 41(3):358-367, 1979. 1",
2037
+ "[30] Rishik Ranjan, Biao Huang, and Alireza Fatehi. Robust gaussian process modeling using em algorithm. Journal of Process Control, 42:125-136, 2016. 1, 2",
2038
+ "[31] Carl Edward Rasmussen. Gaussian processes in machine learning. In Summer School on Machine Learning, pages 63-71. Springer, 2003. 1, 2",
2039
+ "[32] Carl Edward Rasmussen and Hannes Nickisch. Gaussian processes for machine learning (gpml) toolbox. The Jour-"
2040
+ ],
2041
+ "bbox": [
2042
+ 501,
2043
+ 92,
2044
+ 890,
2045
+ 900
2046
+ ],
2047
+ "page_idx": 8
2048
+ },
2049
+ {
2050
+ "type": "page_number",
2051
+ "text": "21958",
2052
+ "bbox": [
2053
+ 478,
2054
+ 944,
2055
+ 519,
2056
+ 955
2057
+ ],
2058
+ "page_idx": 8
2059
+ },
2060
+ {
2061
+ "type": "list",
2062
+ "sub_type": "ref_text",
2063
+ "list_items": [
2064
+ "nal of Machine Learning Research, 11:3011-3015, 2010. 6, 8",
2065
+ "[33] Carl Edward Rasmussen and Joaquin Quinonero-Candela. Healing the relevance vector machine through augmentation. In Proceedings of the International Conference on Machine learning, pages 689-696, 2005. 2",
2066
+ "[34] Patrik L Stahl, Fredrik Salmen, Sanja Vickovic, Anna Lundmark, José Fernández Navarro, Jens Magnusson, Stefania Giacomello, Michaela Asp, Jakub O Westholm, Mikael Huss, et al. Visualization and analysis of gene expression in tissue sections by spatial transcriptomics. Science, 353(6294):78-82, 2016. 7",
2067
+ "[35] Valentine Svensson, Sarah A Teichmann, and Oliver Stegle. Spatialde: identification of spatially variable genes. Nature Methods, 15(5):343-346, 2018. 8",
2068
+ "[36] Bart Thomee, David A Shamma, Gerald Friedland, Benjamin Elizalde, Karl Ni, Douglas Poland, Damian Borth, and Li-Jia Li. Yfcc100m: The new data in multimedia research. Communications of the ACM, 59(2):64-73, 2016. 6, 7",
2069
+ "[37] Michalis Titsias. Variational learning of inducing variables in sparse gaussian processes. In Proceedings of the Artificial Intelligence and Statistics, pages 567-574, 2009. 2, 3, 6",
2070
+ "[38] Shimon Ullman. The interpretation of structure from motion. Proceedings of the Royal Society of London. Series B. Biological Sciences, 203(1153):405-426, 1979. 5",
2071
+ "[39] Jarno Vanhatalo, Pasi Jylänk, and Aki Vehtari. Gaussian process regression with student-t likelihood. Advances in Neural Information Processing Systems, 22, 2009. 2",
2072
+ "[40] Christopher Williams and Matthias Seeger. Using the nystrom method to speed up kernel machines. Advances in Neural Information Processing Systems, 13, 2000. 2",
2073
+ "[41] Oliver Williams and Andrew Fitzgibbon. Gaussian process implicit surfaces. In *Gaussian Processes in Practice*, 2006. 1",
2074
+ "[42] Kyle Wilson and Noah Snavely. Robust global translations with 1dsfm. In Proceedings of the European Conference on Computer Vision, pages 61-75, 2014. 6, 7",
2075
+ "[43] Kwang Moo Yi, Eduard Trulls, Yuki Ono, Vincent Lepetit, Mathieu Salzmann, and Pascal Fua. Learning to find good correspondences. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2666-2674, 2018. 6, 7",
2076
+ "[44] Jiahui Zhang, Dawei Sun, Zixin Luo, Anbang Yao, Lei Zhou, Tianwei Shen, Yurong Chen, Long Quan, and Honggen Liao. Learning two-view correspondences and geometry using order-aware network. In Proceedings of the IEEE International Conference on Computer Vision, pages 5845-5854, 2019. 6, 7"
2077
+ ],
2078
+ "bbox": [
2079
+ 78,
2080
+ 90,
2081
+ 468,
2082
+ 770
2083
+ ],
2084
+ "page_idx": 9
2085
+ },
2086
+ {
2087
+ "type": "page_number",
2088
+ "text": "21959",
2089
+ "bbox": [
2090
+ 478,
2091
+ 944,
2092
+ 517,
2093
+ 955
2094
+ ],
2095
+ "page_idx": 9
2096
+ }
2097
+ ]
2023/Robust and Scalable Gaussian Process Regression and Its Applications/def98ceb-97d5-4394-a2b2-443884eb66dd_model.json ADDED
The diff for this file is too large to render. See raw diff
 
2023/Robust and Scalable Gaussian Process Regression and Its Applications/def98ceb-97d5-4394-a2b2-443884eb66dd_origin.pdf ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:69c32f79187fc896b416e5db011877f7eee69dcfc6a71ce828b8ae0501ef47aa
3
+ size 5418878
2023/Robust and Scalable Gaussian Process Regression and Its Applications/full.md ADDED
@@ -0,0 +1,434 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Robust and Scalable Gaussian Process Regression and Its Applications
2
+
3
+ Yifan Lu $^{1}$ , Jiayi Ma $^{1*}$ , Leyuan Fang $^{2}$ , Xin Tian $^{1}$ , and Junjun Jiang $^{3}$
4
+
5
+ $^{1}$ Wuhan University, China $^{2}$ Hunan University, China $^{3}$ Harbin Institute of Technology, China
6
+
7
+ {lyf048, xin.tian}@whu.edu.cn, {jyma2010, fangleyuan}@gmail.com, jiangjunjun@hit.edu.cn
8
+
9
+ # Abstract
10
+
11
+ This paper introduces a robust and scalable Gaussian process regression (GPR) model via variational learning. This enables the application of Gaussian processes to a wide range of real data, which are often large-scale and contaminated by outliers. Towards this end, we employ a mixture likelihood model where outliers are assumed to be sampled from a uniform distribution. We next derive a variational formulation that jointly infers the mode of data, i.e., inlier or outlier, as well as hyperparameters by maximizing a lower bound of the true log marginal likelihood. Compared to previous robust GPR, our formulation approximates the exact posterior distribution. The inducing variable approximation and stochastic variational inference are further introduced to our variational framework, extending our model to large-scale data. We apply our model to two challenging real-world applications, namely feature matching and dense gene expression imputation. Extensive experiments demonstrate the superiority of our model in terms of robustness and speed. Notably, when matching 4k feature points, its inference is completed in milliseconds with almost no false matches. The code is at github.com/YifanLu2000/Robust-Scalable-GPR.
12
+
13
+ # 1. Introduction
14
+
15
+ Gaussian processes (GPs) [31] are probably the primary non-parametric method for inference on latent functions. They have a wide range of applications from biology [3] to computer vision [41]. A commonly used observation model for Gaussian process regression (GPR) is the Normal distribution, which brings great convenience to the inference. Unfortunately, a well-known limitation of the Gaussian observation model is its sensitivity to outliers in data. As illustrated in Fig. 1 (b), a few outliers can drastically destroy the entire posterior regression result. This hinders the real-world applications of GPR for many domains, where outliers are often inevitable. This paper intends to conquer the GPR with outlier contaminated data.
16
+
17
+ The idea of robust regression is not new. Outlier detec
18
+
19
+ ![](images/422eeaa15b9fa07532c6070ace79536295c679cfce8350a007917cc3996bfd1c.jpg)
20
+
21
+ ![](images/ba4e85d82fd04d290f99c63a1570bd1cf7083263db456c972362c9aedd7729cc.jpg)
22
+
23
+ ![](images/13370d06014ebd7bdf4c720c3cb41b4eaa3dfb96ed42358fecb539782426169d.jpg)
24
+
25
+ ![](images/bab8f13f8194b03b38e708713d7d3fc42c3a0268df8001011ea10a883e557ea2.jpg)
26
+ Figure 1. Regression with our model. (a) Perform exact GPR from 100 inliers. (b) When there are only 6 outliers in the data, the exact GPR leads to completely wrong results. (c) By comparison, our model is able to recover the exact posterior even facing 100 outliers. (d) The feature matching result using our model. (e) The dense spatial gene expression imputation result using our model.
27
+
28
+ ![](images/67dc0e22054a7d3fbba3e5c8064ce53ee21035d388c5175cb432719fd82e7b0e.jpg)
29
+
30
+ tion has been extensively and systematically described in [6,9,10,29]. In the context of GPR, many efforts tried to replace the Gaussian likelihood with other distributions showing heavy-tail behaviors, including Student- $t$ [16,21,28,30], Laplace [22, 30], Gaussian mixture [8, 22, 27], and data-dependent noise model [17]. The challenge with these non-Gaussian likelihoods lies in the inference, which is analytically intractable. To this end, many approximation schemes have been applied, despite having high computational complexity, e.g., Markov Chain Monte Carlo (MCMC) sampling and Expectation Propagation (EP) [22].
31
+
32
+ In this paper, we propose a more effective mixture likelihood model, where uniform distribution accounts for the outliers and Gaussian for inliers. In our formulation, the outliers are independent of the GP and do not affect the computation of the posterior GP, thereby allowing to tolerate more outliers. We next introduce a variational method that jointly determines the modes of data (i.e., inlier or outlier) as well as hyperparameters by maximizing a lower bound to the marginal likelihood. We highlight that the difference between our variational formulation and pervious methods is that the modes of data now become variational parameters and are obtained by minimizing the Kullback-
33
+
34
+ Leibler (KL) divergence between the variational and the true posterior distribution. Thus, the proposed formulation is less likely to overfit and is able to approximate the exact posterior GP only from inliers, as in Fig. 1 (c).
35
+
36
+ Inspired by [37], the sparse inducing variable approximation is integrated into our variational framework, which retains the exact GP prior but performs posterior approximation, and reduces the time complexity from $\mathcal{O}(n^3)$ to $\mathcal{O}(nm^2)$ . By treating the inducing variables as global variables [18], our variational model enjoys the acceleration by Stochastic Variational Inference (SVI) [20]. It performs stochastic optimization from natural gradient and further decreases the time complexity to $\mathcal{O}(km^2)$ . This provides a guarantee for our model to scale to large-scale data.
37
+
38
+ We apply our robust GPR model to two real-world applications, say feature matching and dense spatial gene expression imputation, as illustrated in the Figs. 1 (d) and (e). Extensive experiments demonstrate the superiority of our method on both numerical data and real applications.
39
+
40
+ To summarize, our contributions include the following. (i) We present a robust Gaussian process regression model, which uses variational learning to approximate the true exact posterior. (ii) We leverage inducing variables and SVI to adapt our model to large-scale data. (iii) Two applications of our model are described. Extensive experimental validation demonstrates the superiority of our model.
41
+
42
+ # 2. Related Works
43
+
44
+ Robust GPR typically employs non-Gaussian likelihood with heavy tails. Representatively, [22] investigated the Student's- $t$ noise model. They further developed variational inference and MCMC for GPR with Student's- $t$ . In [39], Laplace approximation was used to approximate the log marginal likelihood of Student's- $t$ . The Laplace noise model is also exploited in [22]. [30] introduced Expectation Maximization (EM) algorithm for Student's- $t$ and Laplace. Recently, [8] used mixtures of two Gaussians assuming a lower variance for regular noise and a higher one for outliers, with EM to learn the hyperparameters. In this work, we also use a non-Gaussian likelihood, i.e., mixture model. The difference is that here the outliers are considered to follow a uniform distribution, which is a more reasonable assumption since the sampling distribution of outliers should be independent of the true latent function. Besides, we derive a variational formulation in the GPR setting, which approximates the posterior distribution.
45
+
46
+ Scalable GPR improves the scalability of GPR while maintaining prediction quality for big data. The seminal work [40] adopted Nyström approximation to approximate the kernel matrix using $m$ sparse points. Later, the idea was further promoted by [7, 15, 23, 33], which approximated the prior and performed exact inference. Unlike the prior approximations, the other line is the posterior approximations,
47
+
48
+ which retain the exact prior but perform approximate inference. The most well-known work is the elegant variational free energy [37]. It directly approximates the posterior by using variational inference. With the advances in variational inference [20], SVI is introduced to GP [18], which uses natural gradients and results in remarkable computational efficiency. However, as the exact GPR, these scalable GPRs are sensitive to outliers. In contrast, our model can tolerate massive outliers while retaining highly scalability.
49
+
50
+ # 3. Gaussian Process Regression Revisited
51
+
52
+ Gaussian process is a collection of random variables, for which any finite subset has a joint Gaussian distribution [31]. It is completely specified by a mean function $m(\mathbf{x})$ and a covariance function $k(\mathbf{x},\mathbf{x}^{\prime})$ where the covariance function typically depends on a set of hyperparameters $\varphi$ . The GP is usually used as a prior over latent function $f(\mathbf{x})\sim \mathcal{GP}(m,k)$ . We can combine the GP prior with observed data to give a posterior over desired latent function.
53
+
54
+ Consider a regression problem, where the training data is $\mathcal{D} = \{(\mathbf{x}_i,y_i)\}_{i = 1}^n$ . Each $y_{i}$ is a noisy realization of the latent function at location $\mathbf{x}_i$ such that $y_{i} = f_{i} + \epsilon_{i}$ , where $\epsilon_{i}\sim \mathcal{N}(\epsilon_{i}|0,\sigma^{2})$ is typically assumed to be an i.i.d. Gaussian noise and $f_{i} = f(\mathbf{x}_{i})$ . Denoting all training inputs as $\mathbf{X}$ , all training outputs $\mathbf{y}$ , and all latent function values at $\mathbf{X}$ as $\mathbf{f}$ . Placing a GP prior on the latent function $f$ , we obtain
55
+
56
+ $$
57
+ p (\mathbf {f} | \mathbf {X}, \varphi) \sim \mathcal {N} (\mathbf {f} | m (\mathbf {X}), \mathbf {K} _ {n n}), \tag {1}
58
+ $$
59
+
60
+ where $\mathbf{K}_{nn}$ is the covariance function evaluated between all the training points. For simplicity of exposition, the mean function is set to zero $m(\mathbf{x}) = 0$ . It does not reduce the generalizability of the GP as long as the problem is correctly transformed, e.g., by subtracting the mean of $\mathbf{y}$ . One can induce a posterior distribution on $\mathbf{f}$ according to Bayes' rule
61
+
62
+ $$
63
+ p (\mathbf {f} | \mathcal {D}, \sigma^ {2}, \varphi) \sim \mathcal {N} (\mathbf {f} | \mathbf {K} _ {n n} (\mathbf {K} _ {n n} ^ {\epsilon}) ^ {- 1} \mathbf {y}, \boldsymbol {\Sigma}), \tag {2}
64
+ $$
65
+
66
+ which is also a multivariate normal distribution, where $\mathbf{K}_{nn}^{\epsilon} = \mathbf{K}_{nn} + \sigma^{2}\mathbf{I}$ and $\boldsymbol {\Sigma} = (\mathbf{K}_{nn}^{-1} + \sigma^{-2}\mathbf{I})^{-1}$ . The posterior distribution of $\mathbf{f}$ can help to compute the posterior predictive distribution of $\mathbf{f}_*$ at any test location $\mathbf{X}_{*}$ :
67
+
68
+ $$
69
+ \begin{array}{l} p \left(\mathbf {f} _ {*} \mid \mathcal {D}, \mathbf {X} _ {*}\right) = \int p \left(\mathbf {f} _ {*} \mid \mathbf {f}, \mathbf {X} _ {*}\right) p (\mathbf {f} \mid \mathcal {D}) d \mathbf {f} \tag {3} \\ = \mathcal {N} \left(\mathbf {f} _ {*} | \mathbf {K} _ {* n} \left(\mathbf {K} _ {n n} ^ {\epsilon}\right) ^ {- 1} \mathbf {y}, \mathbf {K} _ {* *} - \mathbf {K} _ {* n} \left(\mathbf {K} _ {n n} ^ {\epsilon}\right) ^ {- 1} \mathbf {K} _ {n *}\right), \\ \end{array}
70
+ $$
71
+
72
+ where, $\sigma^2$ and $\varphi$ are omitted for brevity. The above posterior depends on the values of $\sigma^2$ and $\varphi$ , which can be inferred by maximizing the log marginal likelihood:
73
+
74
+ $$
75
+ \log p (\mathbf {y} \mid \sigma^ {2}, \varphi) = \log \mathcal {N} (\mathbf {y} \mid \mathbf {0}, \mathbf {K} _ {n n} ^ {\epsilon}), \tag {4}
76
+ $$
77
+
78
+ which automatically achieves the bias-variance trade-off.
79
+
80
+ # 4. Robust Gaussian Process Regression Model
81
+
82
+ In real-world scenarios, a huge discrepancy between the model and data-generating process often occurs, leading to outliers. However, a critical flaw of Gaussian observation model is its non-robustness. This poses a great challenge to GPR in practice. To this end, this section introduces a robust GPR, which identifies the mode of data and simultaneously approximates the posterior of the desired model.
83
+
84
+ # 4.1. Problem Formulation
85
+
86
+ We explicitly consider the generation process of outliers. Since outliers are independent of the latent function, a reasonable assumption is that they obey a uniform distribution $\frac{1}{a}$ , where the output space is a bounded region that covers the output $\mathbf{y}$ and $a$ is a constant denoting the volume of the region. The inliers follow a regular Gaussian noise $\mathcal{N}(y_i|f_i,\sigma^2)$ . We then associate each data with a latent variable $z_{i} \in \{0,1\}$ , where $z_{i} = 0$ indicates that the observation is generated by outlier distribution and $z_{i} = 1$ inlier. The joint distribution of $(y_{i},z_{i})$ is then a mixture of two models:
87
+
88
+ $$
89
+ p \left(y _ {i}, z _ {i} \mid f _ {i}, \boldsymbol {\theta}\right) = \left\{\left(1 - \gamma\right) \frac {1}{a} \right\} ^ {1 - z _ {i}} \left\{\gamma \mathcal {N} \left(y _ {i} \mid f _ {i}, \sigma^ {2}\right) \right\} ^ {z _ {i}}, \tag {5}
90
+ $$
91
+
92
+ where $\theta = (\sigma^2, \gamma, \varphi)$ denotes the latent variables set and $\gamma = p(z_i = 1)$ is the mixing coefficient.
93
+
94
+ Prior distributions describe beliefs about the model variables before the inference. As a robust GPR problem, we place a GP prior over the latent function, i.e., Eq. (1). The variable $\gamma$ controls the probability of occurrence of two models. We suppose it follows a Beta distribution
95
+
96
+ $$
97
+ p (\gamma) \sim \operatorname {B e t a} (\gamma | B _ {a}, B _ {b}), \tag {6}
98
+ $$
99
+
100
+ where $B_{a}$ and $B_{b}$ are hyperparameters. As we will see, $p(\gamma)$ is conjugate to the posterior and hence easier to handle. For other variables $(\sigma^2,\varphi)$ , we introduce noninformative prior.
101
+
102
+ Full joint distribution is obtained by incorporating the above prior distributions as follows:
103
+
104
+ $$
105
+ p (\mathcal {D}, \mathbf {f}, \boldsymbol {\theta}) \propto p (\mathbf {f} | \mathbf {X}) p (\gamma) \prod_ {i = 1} ^ {n} p \left(y _ {i}, z _ {i} \mid f _ {i}, \sigma^ {2}, \gamma\right), \tag {7}
106
+ $$
107
+
108
+ where $\pmb{\theta} = (\sigma^2, \gamma, \mathbf{Z}, \varphi)$ and $\mathbf{Z} \in \{0,1\}^n$ is the indicator variable vector. Our goal is to obtain the posterior $p(\mathbf{f}, \pmb{\theta}|\mathcal{D})$ , which can be induced from $p(\mathcal{D}, \mathbf{f}, \pmb{\theta})$ . Nevertheless, as there are two possibilities for each observation, i.e., inlier and outlier, it will generate $2^n$ combinations to be evaluated. Thus, the exact estimation is analytically intractable and approximation is needed. Next, we use the variational inference to approximate the exact posterior. Once the variational distribution is obtained, the posterior predictive distribution can be determined.
109
+
110
+ # 4.2. Variational Inference
111
+
112
+ We wish to directly approximate the posterior $p(\mathbf{f},\boldsymbol {\theta}|\mathcal{D})$ using a variational posterior $q(\mathbf{f},\boldsymbol {\theta})$ such that the KL divergence between $p(\mathbf{f},\boldsymbol {\theta}|\mathcal{D})$ and $q(\mathbf{f},\boldsymbol {\theta})$ is minimized. The minimization is equivalently expressed as maximization of lower bound for the true log marginal likelihood:
113
+
114
+ $$
115
+ \mathcal {L} (\mathbf {f}, \boldsymbol {\theta}) = \int q (\mathbf {f}, \boldsymbol {\theta}) \log \left(\frac {p (\mathbf {f} , \boldsymbol {\theta} , \mathcal {D})}{q (\mathbf {f} , \boldsymbol {\theta})}\right) d \mathbf {f} d \boldsymbol {\theta}. \tag {8}
116
+ $$
117
+
118
+ The maximization is, however, intractable since the optimal solution is $\hat{q} (\mathbf{f},\pmb {\theta}) = p(\mathbf{f},\pmb {\theta}|\mathcal{D})$ . To this end, we consider instead a restricted distribution of $q(\mathbf{f},\pmb {\theta})$
119
+
120
+ Mean-field factorization supposes $q(\mathbf{f},\pmb {\theta})$ can be partitioned into disjoint groups. Considering the ease of calculation, we factorize $q(\mathbf{f},\pmb {\theta}) = q_{1}(\mathbf{f})q_{2}(\gamma)q_{3}(\mathbf{Z})q_{4}(\sigma^{2},\pmb {\varphi})$ . Besides, we assume $q_{4}$ is Dirac delta function, i.e., it has a point mass at $(\sigma^2,\varphi)$ . With this factorization, the maximization of Eq. (8) is obtained by optimizing each of the factors. For $q_{i}$ , the minimization becomes
121
+
122
+ $$
123
+ \hat {q} _ {i} = \underset {q _ {i}} {\arg \min } \int q _ {i} \log \left(\frac {\exp \mathbb {E} _ {j \neq i} [ \log p (\mathbf {f} , \boldsymbol {\theta} , \mathcal {D}) ]}{q _ {i}}\right) d \boldsymbol {\theta} _ {i}, \tag {9}
124
+ $$
125
+
126
+ where $\mathbb{E}_{j\neq i}[\cdot ]$ denotes the expectation over $q_{j}$ for $j\neq i$ Note that Eq. (9) is the negative KL divergence between $q_{i}$ and $\mathbb{E}_{j\neq i}\left[\log p(\mathbf{f},\boldsymbol {\theta},\mathcal{D})\right]$ . Thus, the optimal $q_{i}$ is given by
127
+
128
+ $$
129
+ \log \hat {q} _ {i} = \mathbb {E} _ {j \neq i} [ \log p (\mathbf {f}, \boldsymbol {\theta}, \mathcal {D}) ] + \text {c o n s t .} \tag {10}
130
+ $$
131
+
132
+ # 4.3. Sparse Inducing Variable Approximation
133
+
134
+ The variational posterior approximation presented above can easily integrate sparse inducing variables [37], which allows the reduction of time complexity of GPR from $\mathcal{O}(n^3)$ to $\mathcal{O}(nm^2)$ . Introducing $m$ pseudo-inputs $\mathbf{X}_m$ with inducing variables $\mathbf{f}_m$ , our goal is to use $\mathbf{f}_m$ to approximate $q(\mathbf{f})$ , where $\mathbf{f}_m$ akin to $\mathbf{f}$ follows the same GP prior that $p(\mathbf{f}_m) \sim \mathcal{N}(\mathbf{0}, \mathbf{K}_{mm})$ . We suppose $\mathbf{f}_m$ is a sufficient statistic for $\mathbf{f}$ , i.e., for any value $u$ , $p(u|\mathbf{f}, \mathbf{f}_m) = p(u|\mathbf{f}_m)$ holds [37]. In this setting, the augmented variational posterior $q(\mathbf{f}, \mathbf{f}_m) = p(\mathbf{f}|\mathbf{f}_m)\phi(\mathbf{f}_m)$ and augmented latent variables set $\boldsymbol{\theta} = (\gamma, \sigma^2, \mathbf{Z}, \varphi, \mathbf{X}_m)$ with factorization $q(\mathbf{f}, \mathbf{f}_m, \boldsymbol{\theta}) = q_1(\mathbf{f}, \mathbf{f}_m)q_2(\gamma)q_3(\mathbf{Z})q_4(\sigma^2, \varphi, \mathbf{X}_m)$ . The bound (8) becomes
135
+
136
+ $$
137
+ \mathcal {L} = \int q (\mathbf {f}, \mathbf {f} _ {m}, \boldsymbol {\theta}) \log \left(\frac {p \left(\mathbf {f} _ {m}\right) p (\gamma) p (\mathbf {y} , \mathbf {Z} \mid \mathbf {f})}{\phi \left(\mathbf {f} _ {m}\right) q _ {2 - 4}}\right) d \mathbf {f} d \mathbf {f} _ {m} d \boldsymbol {\theta}, \tag {11}
138
+ $$
139
+
140
+ where the term $p(\mathbf{f}|\mathbf{f}_m)$ inside the log is eliminated. In what follows, we maximize the bound analytically by solving the optimal choice of the variational distribution $q(\mathbf{f},\mathbf{f}_m,\boldsymbol {\theta})$
141
+
142
+ # 4.4. Maximize Lower Bound
143
+
144
+ To optimize the bound (11) in mean-field factorization, a common approach is to use the coordinate ascent technique,
145
+
146
+ i.e., optimizing one factor $q_{i}$ while keeping the others $q_{j\neq i}$ fixed, and cycling through these factors in turn. Convergence is guaranteed because the bound is convex and monotonically increasing with respect to each factor [5]. Next, we give closed-form expressions for factor updates without detailed derivations, which can be found in suppl. material.
147
+
148
+ Updating $q_{1}(\mathbf{f},\mathbf{f}_{m})$ means finding the optimal posterior GP given $q_{2}$ $q_{3}$ , and $q_{4}$ . We denote $p_i = \mathbb{E}[z_i]$ and $\mathbf{P} =$ diag $(p_1,p_2,\dots ,p_n)$ . Integrating over f, (11) becomes
149
+
150
+ $$
151
+ \mathcal {L} = \int \phi (\mathbf {f} _ {m}) \log \left(\frac {p \left(\mathbf {f} _ {m}\right) Q \left(\mathbf {f} _ {m} , \mathbf {y}\right)}{\phi \left(\mathbf {f} _ {m}\right)}\right) d \mathbf {f} _ {m} + \text {c o n s t .}, \tag {12}
152
+ $$
153
+
154
+ where
155
+
156
+ $$
157
+ \log Q \left(\mathbf {f} _ {m}, \mathbf {y}\right) = \log \mathcal {N} \left(\mathbf {y} \mid \boldsymbol {\mu} _ {\mathbf {f} \mid \mathbf {f} _ {m}}, \sigma^ {2} \mathbf {P} ^ {- 1}\right) - \frac {1}{2 \sigma^ {2}} \operatorname {t r} \left(\mathbf {P K} _ {\mathbf {f} \mid \mathbf {f} _ {m}}\right), \tag {13}
158
+ $$
159
+
160
+ $\mu_{\mathbf{f}|\mathbf{f}_m} = \mathbf{K}_{nm}\mathbf{K}_{mm}^{-1}\mathbf{f}_m,$ $\mathbf{K}_{\mathbf{f}|\mathbf{f}_m} = \mathbf{K}_{nn} - \mathbf{K}_{nm}\mathbf{K}_{mm}^{-1}\mathbf{K}_{mn}$ are the expectation and covariance of $p(\mathbf{f}|\mathbf{f}_m)$ , respectively. With some linear algebraic manipulation, the optimal $\hat{\phi} (\mathbf{f}_m)$ is given by a multivariate Gaussian distribution
161
+
162
+ $$
163
+ \begin{array}{l} \hat {\phi} (\mathbf {f} _ {m}) \propto p (\mathbf {f} _ {m}) Q (\mathbf {f} _ {m}, \mathbf {y}) \tag {14} \\ = \mathcal {N} (\mathbf {f} _ {m} | \boldsymbol {\mu} _ {m}, \mathbf {A} _ {m}), \\ \end{array}
164
+ $$
165
+
166
+ where $\pmb{\mu}_{m} = \sigma^{-2}\mathbf{K}_{mm}\pmb{\Sigma}\mathbf{K}_{mn}\mathbf{P}\mathbf{y},\mathbf{A}_{m} = \mathbf{K}_{mm}\pmb{\Sigma}\mathbf{K}_{mm},$ and $\pmb {\Sigma} = (\mathbf{K}_{mm} + \sigma^{-2}\mathbf{K}_{mn}\mathbf{P}\mathbf{K}_{nm})^{-1}$ Note that we can always recover $q(\mathbf{f})$ from $q(\mathbf{f},\mathbf{f}_m) = p(\mathbf{f}|\mathbf{f}_m)\phi (\mathbf{f}_m)$ by marginalizing out $\mathbf{f}_m$ , which gives optimal $\hat{q} (\mathbf{f}) =$ $\mathcal{N}(\mathbf{f}|\boldsymbol {\mu}_{\mathbf{f}},\mathbf{A})$ , where $\pmb {\mu_{\mathrm{f}}} = \mathbf{K}_{nm}\mathbf{K}_{mm}^{-1}\pmb {\mu}_m$ and $\mathbf{A} =$ $\mathbf{K}_{\mathbf{f}|\mathbf{f}_m} + \mathbf{K}_{nm}\pmb{\Sigma}\mathbf{K}_{mn}$
167
+
168
+ Remark 1 With the optimal $\hat{\phi}(\mathbf{f}_m)$ , we obtain $\langle \mathcal{N}_i \rangle = \exp(\mathbb{E}[\log \mathcal{N}_i])$ as follows
169
+
170
+ $$
171
+ \langle \mathcal {N} _ {i} \rangle = \mathcal {N} \left(y _ {i} \mid \boldsymbol {\mu} _ {\mathbf {f} i}, \sigma^ {2}\right) \exp \left(- \frac {1}{2 \sigma^ {2}} \mathbf {A} _ {i i}\right), \tag {15}
172
+ $$
173
+
174
+ where $\mathcal{N}_i$ denotes $\mathcal{N}(y_i|f_i,\sigma^2)$ , $\pmb{\mu}_{\mathbf{fi}}$ is the $i$ -th element of $\pmb{\mu}_{\mathbf{f}}$ , and $\mathbf{A}_{ii}$ is the $i$ -th diagonal element of $\mathbf{A}$ .
175
+
176
+ Updating $q_{2}(\gamma)$ indicates updating the inlier ratio or mixture. Suppose $q_{1}, q_{3}$ , and $q_{4}$ are given and note $\hat{n} = \operatorname{tr}(\mathbf{P})$ . According to Eq. (10), we conclude that $\hat{q}_{2}(\gamma)$ again follows a Beta distribution
177
+
178
+ $$
179
+ \hat {q} _ {2} (\gamma) = \operatorname {B e t a} (\gamma | B _ {a} + \hat {n}, B _ {b} + n - \hat {n}). \tag {16}
180
+ $$
181
+
182
+ Remark 2 With the optimal $\hat{q}_2(\gamma)$ , according to the properties of Beta distribution, we obtain $\langle \gamma \rangle = \exp (\mathbb{E}[\log \gamma ])$ and $\langle 1 - \gamma \rangle = \exp (\mathbb{E}[\log (1 - \gamma)])$ as follows
183
+
184
+ $$
185
+ \langle \gamma \rangle = \exp (\psi (B _ {a} + \hat {n}) - \psi (B _ {a} + B _ {b} + n)), \tag {17}
186
+ $$
187
+
188
+ $$
189
+ \langle 1 - \gamma \rangle = \exp (\psi (B _ {b} + n - \hat {n}) - \psi (B _ {a} + B _ {b} + n)), \tag {17}
190
+ $$
191
+
192
+ where $\psi (\cdot)$ is the digamma function.
193
+
194
+ Updating $q_{3}(\mathbf{Z})$ represents determining the probability of each data being an inlier given $q_{1}, q_{2}$ , and $q_{4}$ . It encodes the mode of each data point. From Eq. (10), it leads to a
195
+
196
+ further factorization $\hat{q}_3(\mathbf{Z}) = \prod_{i=1}^n \hat{q}_3^{[i]}(z_i)$ and $\hat{q}_3^{[i]}(z_i)$ is shown to follow a Bernoulli distribution:
197
+
198
+ $$
199
+ \hat {q} _ {3} ^ {[ i ]} \left(z _ {i}\right) = \left(1 - p _ {i}\right) ^ {1 - z _ {i}} p _ {i} ^ {z _ {i}}, \tag {18}
200
+ $$
201
+
202
+ $$
203
+ \text {w h e r e} \quad p _ {i} = \frac {\langle \gamma \rangle \left\langle \mathcal {N} _ {i} \right\rangle}{\left. \langle 1 - \gamma \rangle / a + \langle \gamma \rangle \left\langle \mathcal {N} _ {i} \right\rangle \right.}. \tag {19}
204
+ $$
205
+
206
+ Remark 3 $p_i$ shows the posterior probability of the $i$ -th data being an inlier. It depends on both the deviation between $y_i$ and the posterior mean $\mu_{\mathrm{fi}}$ , and the term containing the posterior variance $\mathbf{A}_{ii}$ . The importance of this variance term is that it encodes the uncertainty of the latent function at $\mathbf{x}_i$ , which means if we are uncertain about that data, then we cannot identify it as an inlier, even though $y_i$ and $\mu_{\mathrm{fi}}$ are close. This is helpful to avoid overfitting.
207
+
208
+ Updating $q_{4}(\sigma^{2},\varphi ,\mathbf{X}_{m})$ involves the optimization of hyperparameters. Specifically, $\sigma^2$ controls the noise level, $\varphi$ represents the shape of the kernel function, and $\mathbf{X}_m$ is the location of the pseudo-inputs. As assumed previously, $q_{4}(\sigma^{2},\varphi ,\mathbf{X}_{m})$ obeys the Dirac delta distribution, which means we directly maximize the lower bound (11) with respect to $(\sigma^2,\varphi ,\mathbf{X}_m)$ , rather than using Eq. (10).
209
+
210
+ Given $q_{1}$ , $q_{2}$ , and $q_{3}$ , taking derivative of (11) over $\sigma^2$ and setting to zero, we obtain a closed-form expression
211
+
212
+ $$
213
+ \hat {\sigma} ^ {2} = \frac {1}{\hat {n}} \left(\mathbf {y} - \boldsymbol {\mu} _ {\mathbf {f}}\right) ^ {\top} \mathbf {P} \left(\mathbf {y} - \boldsymbol {\mu} _ {\mathbf {f}}\right) + \frac {1}{\hat {n}} \operatorname {t r} (\mathbf {P A}), \tag {20}
214
+ $$
215
+
216
+ where $\operatorname{tr}(\cdot)$ is the trace. The lower bound (11) is, however, difficult to optimize for the remaining hyperparameters $(\varphi, \mathbf{X}_m)$ since the integral is intractable. To this end, we use the reverse Jensen's inequality and obtain
217
+
218
+ $$
219
+ \mathcal {L} _ {2} := \log \mathcal {N} (\mathbf {y} | 0, \mathbf {K} _ {\mathbf {y}}) - \frac {1}{2 \sigma^ {2}} \operatorname {t r} \left(\mathbf {P K} _ {\mathbf {f} | \mathbf {f} _ {m}}\right) + \text {c o n s t .} \leq \mathcal {L}, \tag {21}
220
+ $$
221
+
222
+ where $\mathbf{K_y} = \sigma^2\mathbf{P}^{-1} + \mathbf{K}_{nm}\mathbf{K}_{mm}^{-1}\mathbf{K}_{mn}$ . To set the optimized hyperparameters $(\varphi, \mathbf{X}_m)$ , we seek partial derivatives of $\mathcal{L}_2$ with respect to them. Thereafter, the standard gradient descent algorithm provides the value that maximizes $\mathcal{L}_2$ .
223
+
224
+ # 4.5. Stochastic Variational Inference
225
+
226
+ Modern applications often involve handling massive data. However, the GPR equipped with variational inference does not easily scale to big data, even with inducing variables. This is because we need to go through the entire training data at each coordinate update. As the training data size grows, the computational overhead becomes larger.
227
+
228
+ An alternative to coordinate ascent is to use SVI [20], which is similar to stochastic gradient descent (SGD), that performs stochastic optimization from the noisy but cheap-to-compute estimates of the gradient. It iterates between subsampling a subset of data and adjusting the hidden structure solely based on the subset. Thus, SVI is much more efficient than conventional variational inference. To apply SVI to the probabilistic model, a set of global variables is
229
+
230
+ needed [18]. In our robust GPR setting, the inducing variables $\mathbf{f}_m$ and mixing coefficient variable $\gamma$ fill this role.
231
+
232
+ Natural Gradient. In gradient-based optimization, the natural gradient accounts for the geometric structure of probability parameters and is used in SVI to replace the Euclidean gradient. In exponential families, it is given by premultiplying the usual gradient by the inverse Fisher information matrix $G(\eta)^{-1}$ , where $\eta$ is the canonical parameter of the exponential families. Since both $\phi(\mathbf{f}_m)$ and $q(\gamma)$ are exponential family distributions, they enjoy simple natural gradients [19] of the bound (11) such that $g(\eta) = \mathbb{E}[\eta] - \eta$ , where $g(\cdot)$ denotes the natural gradient. For Gaussian distribution $\phi(\mathbf{f}_m)$ , its canonical parameter is $[\mathbf{A}_m^{-1}\boldsymbol{\mu}_m, -\frac{1}{2}\mathbf{A}_m^{-1}]$ . For Beta distribution $q(\gamma)$ , its canonical parameter is $[B_a + \hat{n}, B_b + n - \hat{n}]$ . Thereafter, we obtain the update rule using natural gradient with a step size $\epsilon_t$
233
+
234
+ $$
235
+ \boldsymbol {\eta} ^ {t} = \boldsymbol {\eta} ^ {t - 1} + \varepsilon_ {t} g (\boldsymbol {\eta} ^ {t - 1}) = (1 - \varepsilon_ {t}) \boldsymbol {\eta} ^ {t - 1} + \varepsilon_ {t} \mathbb {E} [ \boldsymbol {\eta} ]. \tag {22}
236
+ $$
237
+
238
+ Note that when the step size $\varepsilon_{t} = 1$ , we recover the original variational update in Eq. (14) and Eq. (16). Now, the noisy natural gradient can be easily computed by sampling either individual or mini-batch of the data.
239
+
240
+ # 4.6. Making Predictions
241
+
242
+ So far we have described how to infer the approximate posterior of the latent variables. To make predictions, we use the predictive distribution similar to Eq. (3):
243
+
244
+ $$
245
+ \begin{array}{l} p \left(\mathbf {f} _ {*} \mid \mathcal {D}, \mathbf {X} _ {*}\right) = \int p \left(\mathbf {f} _ {*} \mid \mathbf {f} _ {m}\right) p \left(\mathbf {f} \mid \mathbf {f} _ {m}\right) \phi \left(\mathbf {f} _ {m}\right) d \mathbf {f} _ {m} d \mathbf {f} \tag {23} \\ = \mathcal {N} \left(\mathbf {f} _ {*} | \mathbf {K} _ {* m} \mathbf {K} _ {m m} ^ {- 1} \boldsymbol {\mu} _ {m}, \mathbf {K} _ {\mathbf {f} _ {*} | \mathbf {f} _ {m}} + \mathbf {K} _ {* m} \boldsymbol {\Sigma} \mathbf {K} _ {m *}\right), \\ \end{array}
246
+ $$
247
+
248
+ where $\mathbf{K}_{\mathbf{f}_{*}|\mathbf{f}_{m}} = \mathbf{K}_{**} - \mathbf{K}_{*m}\mathbf{K}_{mm}^{-1}\mathbf{K}_{m*}$ .
249
+
250
+ # 5. Applications
251
+
252
+ In this section, we describe how to use the proposed GPR model to deal with real-world tasks, including feature matching and dense gene expression imputation.
253
+
254
+ # 5.1. Feature Matching
255
+
256
+ Feature matching aims at establishing reliable feature point correspondences. To solve the matching problem in a regression manner, we first construct a putative correspondence set $\mathcal{P} = \{(\mathbf{x}_i,\mathbf{x}_i')\}_{i = 1}^n$ by nearest neighbor (NN) matching, where $\mathbf{x}_i$ and $\mathbf{x}_i'$ represent the coordinates of feature points. The putative set is then converted to motion vector set, i.e., $\mathcal{D} = \{(\mathbf{x}_i,\mathbf{y}_i)\}_{i = 1}^n$ , where $\mathbf{y}_i = \mathbf{x}_i' - \mathbf{x}_i$ . The motion of an object projected on the image plane is known to be slow-and-smooth [38], which is well described by GP. Thus, the motion vectors of inliers can be seen as sampling on the underlying model $\mathbf{f}$ , while the outliers are spuriously distributed. Our goal is to determine the correctness of each data. The task, however, is non-trivial as there are often
257
+
258
+ massive outliers (even up to $90\%$ ). Moreover, the feature points in each image can reach several thousands, which requires high computational efficiency. As we will see, our model is robust and fast enough to handle such cases.
259
+
260
+ Inference Setting. To apply our model to feature matching, some special considerations need to be set. The feature coordinates are first normalized. For the covariance kernel, we choose the squared exponential (SE) kernel $k(\mathbf{x},\mathbf{x}^{\prime}) = \lambda^{-1}\exp (-\beta \| \mathbf{x} - \mathbf{x}^{\prime}\|^{2})$ . We find that fixed hyperparameters work well for most situations, which are empirically set to $\lambda = 4$ and $\beta = 0.1$ . The positions of the inducing variables are randomly selected from the training data and kept fixed during inference. The ratio information is also considered as a prior. The mini-batch size is set to $\max \left\{\frac{n}{8},200\right\}$ . After inference convergences, the $i$ -th data is decided to be inlier if $p_i > \tau$ , where $\tau$ is a threshold and is set to $\tau = 0.75$ .
261
+
262
+ Predictions. One of the attractive properties of our model for this task is that when the training is completed, the resulting posterior GP is able to predict the motion vector of each feature point. Combined with the descriptors' information, we can yield much more correct correspondences rather than being limited to putative set. More correct correspondences are crucial for downstream tasks, as they can significantly improve the accuracy.
263
+
264
+ # 5.2. Dense Gene Expression Imputation
265
+
266
+ Recently, technological advances have made it possible to measure spatially resolved transcriptome-wide mRNA expressions with spatial information in tissue samples. GPR is thus utilized to infer gene expression in discrete samples, generating dense gene expression in 2D, 3D, and even spatiotemporal [3]. Unfortunately, due to equipment malfunctions and process disturbances, the readout data may contain bad spots. In this case, a vanilla GPR may lead to completely wrong results and thus mislead the subsequent analysis. By comparison, the robust GPR developed in this paper copes well with outliers and is able to give meaningful imputation results. For this application, the implementation is simple and requires little extra consideration.
267
+
268
+ # 6. Experiments
269
+
270
+ In this section, we first conduct numerical experiments to evaluate the performance in a fully controlled environment. After that, we apply our model to tackle feature matching problem and compare it with many state-of-the-art methods tailored for feature matching. In the end, the evaluation of gene expression imputation is performed. All experiments are conducted on a desktop with Intel i7-10700 CPU, 16 GB memory with MATLAB except deep learning methods.
271
+
272
+ # 6.1. Numerical Experiments
273
+
274
+ We study two datasets, i.e., Neal dataset [28] and Friedman dataset [14]. The Neal dataset is one-dimensional data,
275
+
276
+ ![](images/bf6975106d089e2f181b198558389f9a26f504bb91fd7013d7378b6516549c3a.jpg)
277
+ Figure 2. Visual examples on Neal data. Our model approximates the exact posterior GP using only inliers, even with massive outliers. The legend is shown in the bottom.
278
+
279
+ where the input $x$ is drawn from a uniform distribution within $[-2.5, 2.5]$ . The target value is calculated as:
280
+
281
+ $$
282
+ f (x) = 0. 3 + 0. 4 x + 0. 5 \sin (2. 7 x) + 1. 1 / \left(1 + x ^ {2}\right).
283
+ $$
284
+
285
+ The Gaussian noise is added to the target value with zero mean and variance of 0.1. Next, outliers are injected into the training data, where the value is drawn from uniform distribution within $[-5, 5]$ . The Friedman is 10-dimensional data and the function value depends on the first five dimensions
286
+
287
+ $$
288
+ f (\mathbf {x}) = 1 0 \sin \left(\pi x _ {1} x _ {2}\right) + 2 0 \left(x _ {3} - 0. 5\right) ^ {2} + 1 0 x _ {4} + 5 x _ {5},
289
+ $$
290
+
291
+ while the remaining dimensions complicate the task by adding a feature selection process. The input $\mathbf{x}$ is sampled uniformly within $[0,1]^{10}$ . Gaussian noise $\mathcal{N}(0,0.01)$ is added and random outliers are distributed among [5, 25].
292
+
293
+ We fix the inlier number to 100 and change the outlier ratio. GPR with Student's- $t$ (St) [32], Laplace (Lap) [32], and Gaussian mixture distribution (MEM) [8], as well as variational free energy (VEF) [37], are chosen as competitors. Some intuitive examples on Neal data are shown in Fig. 2, where outlier ratio increases from 0 to $80\%$ . We see that the prediction by our model almost exactly reproduces the full GP prediction using only inliers. In contrast, GPR-St and -Lap give poor approximations. GPR-MEM is able to give a better answer, yet still not as good as ours, especially at extremely high outlier ratios.
294
+
295
+ To statistically verify performance, we select mean absolute error (MAE, m), root mean square error (RMSE, r), and negative log of predictive probability (NLP, n) as metrics. 1000 testing points are generated in equal intervals. The statistical results on Neal and Friedman in Table 1 suggest that our method outperforms other alternatives in terms of robustness and accuracy. More results including KL and statistical variance are provided in suppl. material.
296
+
297
+ # 6.2. Feature Matching
298
+
299
+ We collect three large and diverse feature matching datasets with three different vision tasks, say YFCC100M
300
+
301
+ Table 1. Quantitative results on Neal and Friedman data with different outlier ratios. Bold indicates the best.
302
+
303
+ <table><tr><td colspan="10">Neal [28]</td></tr><tr><td rowspan="2">Method</td><td colspan="3">10%</td><td colspan="3">50%</td><td colspan="3">80%</td></tr><tr><td>m</td><td>r</td><td>n</td><td>m</td><td>r</td><td>n</td><td>m</td><td>r</td><td>n</td></tr><tr><td>GPR-VFE [37]</td><td>0.18</td><td>0.21</td><td>0.73</td><td>0.58</td><td>0.77</td><td>1.75</td><td>0.76</td><td>1.08</td><td>1.96</td></tr><tr><td>GPR-St [32]</td><td>0.03</td><td>0.03</td><td>-0.41</td><td>0.09</td><td>0.19</td><td>0.39</td><td>0.77</td><td>1.15</td><td>1.66</td></tr><tr><td>GPR-Lap [32]</td><td>0.03</td><td>0.07</td><td>-1.17</td><td>0.42</td><td>0.56</td><td>1.62</td><td>0.35</td><td>0.72</td><td>2.03</td></tr><tr><td>GPR-MEM [8]</td><td>0.04</td><td>0.08</td><td>-1.36</td><td>0.06</td><td>0.12</td><td>-0.66</td><td>0.77</td><td>1.52</td><td>2.04</td></tr><tr><td>Ours</td><td>0.01</td><td>0.02</td><td>-1.38</td><td>0.03</td><td>0.04</td><td>-1.25</td><td>0.03</td><td>0.03</td><td>-0.89</td></tr><tr><td colspan="10">Friedman [14]</td></tr><tr><td rowspan="2">Method</td><td colspan="3">20%</td><td colspan="3">30%</td><td colspan="3">50%</td></tr><tr><td>m</td><td>r</td><td>n</td><td>m</td><td>r</td><td>n</td><td>m</td><td>r</td><td>n</td></tr><tr><td>GPR-VFE [37]</td><td>1.12</td><td>1.24</td><td>2.28</td><td>1.47</td><td>1.62</td><td>3.08</td><td>2.88</td><td>3.16</td><td>6.93</td></tr><tr><td>GPR-St [32]</td><td>0.04</td><td>0.04</td><td>-0.08</td><td>0.91</td><td>1.10</td><td>3.59</td><td>1.47</td><td>2.18</td><td>7.16</td></tr><tr><td>GPR-Lap [32]</td><td>0.23</td><td>0.29</td><td>0.37</td><td>0.46</td><td>0.61</td><td>0.94</td><td>2.03</td><td>2.23</td><td>3.29</td></tr><tr><td>GPR-MEM [8]</td><td>0.11</td><td>0.21</td><td>-0.99</td><td>0.13</td><td>0.24</td><td>-0.66</td><td>2.13</td><td>2.32</td><td>2.98</td></tr><tr><td>Ours</td><td>0.04</td><td>0.06</td><td>-1.46</td><td>0.06</td><td>0.08</td><td>-0.35</td><td>0.10</td><td>0.14</td><td>0.55</td></tr></table>
304
+
305
+ ![](images/d63bae63fad5a96af15bc558e42713b63d1df2c55e121e76a9bfee1978ae7202.jpg)
306
+ Figure 3. Qualitative illustration of relative pose estimation. We show the true positive in blue and false positive in red. Best viewed in color with $200\%$ zoom in.
307
+
308
+ dataset [36] (4000 pairs) for relative pose estimation, CPC dataset [42] (1000 pairs) for fundamental matrix estimation, and HPatches dataset [1] (580 pairs) for homography estimation. Six state-of-the-arts from handcrafted to deep learning are chosen for comparison, namely, MAGSAC++ [2], LPM [26], MCDM [25], CRC [12], PointCN [43], and OANet [44], with baseline ratio test. Briefly, MAGSAC++ is resampling-based method, LPM is a heuristic method that preserves local structure of feature points, CRC interpolates a smooth vector field using Fourier bases, PointCN uses deep neural networks with context normalization to encode global contextual information, OANet further captures local contextual information by clustering unordered correspondences. Note that we assess currently other robust GPR methods are not capable to handle the feature matching problem, thus none of them are chosen for comparison.
309
+
310
+ # 6.2.1 Relative Pose Estimation
311
+
312
+ Following [44], Yahoo's YFCC100M dataset [36] generates 4000 testing image pairs. SIFT [24] with NN matching is adopted to establish putative correspondences. The maximum number of keypoints for each image is limited to 4000. RANSAC [13] is used to estimate the relative camera pose.
313
+
314
+ Table 2. Quantitative comparison for relative pose estimation. Bold indicates the best and underline ranks the second.
315
+
316
+ <table><tr><td colspan="7">YFCC100M [36]</td></tr><tr><td rowspan="2">Method</td><td colspan="3">AUC</td><td rowspan="2">P</td><td rowspan="2">R</td><td rowspan="2">F</td></tr><tr><td>@5°</td><td>@10°</td><td>@20°</td></tr><tr><td>Ratio test</td><td>28.8</td><td>39.2</td><td>52.1</td><td>24.3</td><td>55.3</td><td>32.8</td></tr><tr><td>MSC++ [2]</td><td>24.9</td><td>33.7</td><td>45.7</td><td>24.3</td><td>55.3</td><td>32.8</td></tr><tr><td>PointCN [43]</td><td>45.1</td><td>54.6</td><td>65.0</td><td>47.2</td><td>73.4</td><td>54.9</td></tr><tr><td>OANet [44]</td><td>53.6</td><td>64.1</td><td>75.8</td><td>51.2</td><td>87.0</td><td>61.7</td></tr><tr><td>LPM [26]</td><td>38.6</td><td>48.8</td><td>59.7</td><td>51.4</td><td>43.6</td><td>45.9</td></tr><tr><td>MCDM [25]</td><td>47.8</td><td>59.4</td><td>70.7</td><td>63.0</td><td>46.5</td><td>52.7</td></tr><tr><td>CRC [12]</td><td>38.1</td><td>48.2</td><td>58.6</td><td>41.2</td><td>57.8</td><td>44.5</td></tr><tr><td>Ours</td><td>53.8</td><td>63.7</td><td>73.8</td><td>85.3</td><td>54.0</td><td>63.7</td></tr></table>
317
+
318
+ ![](images/dc04c4bb8db41b27359a91b479060ac3afea0f3337e0a7442160a15dd2c9426b.jpg)
319
+ Figure 4. Visual examples of homography and fundamental matrix estimation. We show the true positive in blue and false positive in red. Best viewed in color with $200\%$ zoom in.
320
+
321
+ We measure the maximum of rotation and translation errors in degree and report the approximate Area Under the Curve (AUC) with thresholds 5, 10, and 20 degrees. Precision, recall, and F1-score before RANSAC with respect to the ground-truth inliers are also presented. See suppl. material for a detailed description of the dataset and the metrics.
322
+
323
+ Some visualized examples are demonstrated in Fig. 3. Statistics results are comprehensively reported in Table 2. Since deep learning methods such as PointCN and OANet are trained in such dataset with geometric loss, they can capture geometric information and therefore have higher AUC compared to other handcrafted methods. By comparison, our method achieves comparable results to these deep learning methods with much higher precision and F1-score.
324
+
325
+ # 6.2.2 Homography & Fundamental Matrix Estimation
326
+
327
+ Homography and fundamental matrix estimation are critical parts in computer vision. The maximum keypoints for HPatches and CPC are limited to 4000. For evaluation metrics, we choose the precision, recall, and F1-score as well as the accuracy of the estimation. For homography estimation, we adopt homography error defined in [11], and for fundamental matrix, we follow [4] and use normalized symmetric geometry distance (NSGD) as the metrics. An estimate is classified as accurate or not by a threshold, which is set to 4 pixels for homography error and 0.05 for the NSGD [4].
328
+
329
+ Table 3. Statistical results for homography and fundamental matrix estimation. Bold ranks the first, underline the second.
330
+
331
+ <table><tr><td rowspan="2">Method</td><td colspan="4">HPatches [1]</td><td colspan="4">CPC [42]</td></tr><tr><td>Acc.</td><td>P</td><td>R</td><td>F</td><td>Acc.</td><td>P</td><td>R</td><td>F</td></tr><tr><td>Ratio test</td><td>71.7</td><td>71.6</td><td>51.6</td><td>63.0</td><td>17.0</td><td>12.7</td><td>63.9</td><td>20.5</td></tr><tr><td>MSC++ [2]</td><td>74.4</td><td>87.3</td><td>92.4</td><td>86.0</td><td>6.9</td><td>19.7</td><td>29.6</td><td>19.4</td></tr><tr><td>PointCN [43]</td><td>73.1</td><td>76.9</td><td>85.7</td><td>78.6</td><td>29.3</td><td>19.8</td><td>70.2</td><td>28.8</td></tr><tr><td>OANet [44]</td><td>70.0</td><td>73.5</td><td>77.2</td><td>72.1</td><td>47.0</td><td>34.6</td><td>81.0</td><td>45.4</td></tr><tr><td>LPM [26]</td><td>60.6</td><td>72.5</td><td>50.5</td><td>52.3</td><td>23.8</td><td>21.9</td><td>48.9</td><td>26.1</td></tr><tr><td>MCDM [25]</td><td>77.7</td><td>62.9</td><td>91.0</td><td>73.0</td><td>34.8</td><td>23.0</td><td>61.0</td><td>32.5</td></tr><tr><td>CRC [12]</td><td>71.7</td><td>75.5</td><td>92.5</td><td>80.4</td><td>28.7</td><td>24.8</td><td>47.2</td><td>30.4</td></tr><tr><td>Ours</td><td>78.1</td><td>89.9</td><td>94.3</td><td>91.5</td><td>49.6</td><td>61.5</td><td>59.3</td><td>57.2</td></tr><tr><td>Ours*</td><td>82.4</td><td>73.9</td><td>109.6</td><td>-</td><td>51.9</td><td>39.2</td><td>97.3</td><td>-</td></tr></table>
332
+
333
+ Table 4. Ablation studies on feature matching. Bold is the best.
334
+
335
+ <table><tr><td colspan="2">Method</td><td>Prec.</td><td>Recall</td><td>F1</td><td>Time</td></tr><tr><td colspan="2">CRC [12]</td><td>50.7</td><td>67.8</td><td>54.9</td><td>68.1</td></tr><tr><td colspan="2">OANet (on GPU) [44]</td><td>47.9</td><td>75.2</td><td>56.0</td><td>41.2</td></tr><tr><td rowspan="5">RS-GPR</td><td>No SVI</td><td>76.3</td><td>66.1</td><td>70.0</td><td>47.1</td></tr><tr><td>No inducing variables</td><td>72.6</td><td>69.0</td><td>69.5</td><td>1.6e5</td></tr><tr><td>Opt. Hyperparameters</td><td>85.0</td><td>61.2</td><td>67.7</td><td>1556</td></tr><tr><td>EM optimization</td><td>67.6</td><td>69.8</td><td>65.5</td><td>43.7</td></tr><tr><td>Default</td><td>88.0</td><td>68.2</td><td>75.6</td><td>9.8</td></tr></table>
336
+
337
+ Several visual examples are provided in Fig. 4. Note that our predictive model is also presented and indicated by the superscript ${}^{*}\ast$ . It establishes significantly more correct correspondences compared to other methods. Table 3 reports the statistical results. As it indicates, our method outperforms all other methods in the two tasks. When using the predictive model, it can even produce more correct correspondences than the putative set (e.g., the recall in HPatches is higher than 100), thus further improving accuracy.
338
+
339
+ # 6.2.3 Ablation Studies of Feature Matching
340
+
341
+ To measure performance, especially runtime, in different settings, we conduct ablation studies on 100 image pairs from YFCC100M. It mainly includes four choices, i.e., not using SVI, not using inducing variables, optimizing the kernel hyperparameters, and using EM optimization instead of variational inference. We also include the CRC and OANet to compare the runtime. The results are reported in Table 4, where we see that our method performs the best with the default setting described in Sect. 5.1. In particular, the running time is much faster compared to other settings and methods.
342
+
343
+ # 6.3. Dense Gene Expression Imputation
344
+
345
+ We collect two spatial transcriptomics datasets from mouse olfactory bulb and human breast cancer biopsies [34], respectively. Briefly, the spatial transcriptomics gene expression data is measured from thin tissue sections placed on an array with ploy probes and spatially resolved DNA barcodes. These create a grid of circular spots where the mRNA abundance of each spot is measured using probes with the barcodes encoding the spatial locations. The mouse olfactory bulb and human breast cancer biopsies datasets
346
+
347
+ ![](images/f18ddbf02269356ae12c989aebb22acf2f900a111dfd7aa45aba705fdda444c2.jpg)
348
+ Figure 5. Visualization of dense spatial gene expression imputation. The spatial transcriptomics data is plotted on the top row. The bottom row shows the imputation results of different methods. We stack the spatial transcriptomics (shown as dots) on the imputed results to show the quality of the imputation.
349
+
350
+ ![](images/55745084f69e7e663fc4106324058dc6db9f1fa989348d7b22a8e582325efa9a.jpg)
351
+
352
+ Table 5. Statistical results of dense gene expression imputation. Bold indicates the best.
353
+
354
+ <table><tr><td colspan="7">Mouse olfactory bulb</td></tr><tr><td rowspan="2">Method</td><td colspan="3">Pattern 1</td><td colspan="3">Pattern 2</td></tr><tr><td>m</td><td>r</td><td>n</td><td>m</td><td>r</td><td>n</td></tr><tr><td>GPR-Exact</td><td>0.46</td><td>0.59</td><td>1.25</td><td>0.33</td><td>0.40</td><td>1.10</td></tr><tr><td>GPR-St [32]</td><td>0.07</td><td>0.14</td><td>-0.57</td><td>0.02</td><td>0.04</td><td>-1.33</td></tr><tr><td>GPR-Lap [32]</td><td>0.20</td><td>0.27</td><td>0.33</td><td>0.48</td><td>0.54</td><td>0.94</td></tr><tr><td>Ours</td><td>0.03</td><td>0.05</td><td>-1.99</td><td>0.02</td><td>0.04</td><td>-2.21</td></tr><tr><td colspan="7">Breast cancer tissue</td></tr><tr><td rowspan="2">Method</td><td colspan="3">Pattern 1</td><td colspan="3">Pattern 2</td></tr><tr><td>m</td><td>r</td><td>n</td><td>m</td><td>r</td><td>n</td></tr><tr><td>GPR-Exact</td><td>0.24</td><td>0.35</td><td>1.44</td><td>0.23</td><td>0.33</td><td>1.15</td></tr><tr><td>GPR-St [32]</td><td>0.19</td><td>0.32</td><td>0.37</td><td>0.18</td><td>0.30</td><td>0.22</td></tr><tr><td>GPR-Lap [32]</td><td>0.25</td><td>0.36</td><td>0.56</td><td>0.25</td><td>0.37</td><td>0.53</td></tr><tr><td>Ours</td><td>0.12</td><td>0.16</td><td>-0.54</td><td>0.11</td><td>0.17</td><td>-0.39</td></tr></table>
355
+
356
+ contain 14859 and 12856 measured genes, respectively. To identify spatially variable genes, we adopt SpatialDE [35] and select two patterns for each dataset. We next add 5 random outliers to each identified pattern to simulate the outlier scenario. In the end, different GPR methods are applied to impute dense spatial gene expression.
357
+
358
+ Two visual examples are presented in Fig. 5, where in each example, the second row shows the imputed results. As can be seen, our method accurately infers the correct gene expression regardless of the outliers. To give quantitative results, we randomly split each dataset into $80\%$ training data and $20\%$ test data. As mentioned above, we use the MAE, RMSE, and NLP to measure the performance, and report the statistical results in Table 5. We see that our method always has the lowest error.
359
+
360
+ # 7. Analysis
361
+
362
+ Optimization Process. We visualize the intermediate results of the optimization process on Neal data with $50\%$ outliers and present it in the top row of Fig. 6. We see that the field of $p$ in Eq. (19) gradually converges to the inliers. After convergence, the posterior mean and variance functions are very close to the ground truth, and the inducing variables are restricted to the interval of the inliers.
363
+
364
+ The Role of Term $\mathbf{A}_{ii}$ . We remove the term containing $\mathbf{A}_{ii}$ in Eq. (15) and show the results in the bottom of Fig. 6.
365
+
366
+ ![](images/63607d253c4cdef0168a9b0993d4d1c20ff65bda2af6dc79d4cd66053cea2219.jpg)
367
+ Figure 6. Intermediate results of the optimization process. The top row shows the full model while the bottom removes term $\mathbf{A}_{ii}$ . We visualize the field of $p$ , which indicates the inlier probability at any location. The deeper the blue color, the higher the probability.
368
+
369
+ ![](images/7675747a1a2dd41bde5fa61c995b2e69b4a3152c8469d6feffc58aea63ae31f3.jpg)
370
+ Figure 7. Our model deal with different non-uniform outliers.
371
+
372
+ ![](images/62e6027f2054a4ce502523dbd7253a07b2680fd0a94db254e8cfea899846273d.jpg)
373
+ Figure 8. The inference time of each GPR method. Our model with SVI can easily handle big data.
374
+
375
+ Without variance term, the regression tends to overfit.
376
+
377
+ Non-uniform Outliers. Different types of non-uniform outliers are tested in Fig. 7. More detailed results are given in suppl. material. Our model exhibits strong generalization and robustness. The reason is that uniform distribution is a very weak assumption with high flexibility. It accommodates many outlier distributions.
378
+
379
+ **Scalability.** We test the runtime towards different problem sizes on Neal data with $50\%$ outliers. Other methods as well as ours w. and w/o. SVI are included. The number of inducing variables is set to 15 for our method in both two settings. As plotted in Fig. 8, our method can easily scale to large-scale data, especially when equipped with SVI.
380
+
381
+ # 8. Conclusion
382
+
383
+ This paper presents a robust and scalable GPR using variational learning. It uses an outlier robust mixture likelihood model, where the uniform distribution accounts for the outliers. A variational formulation is introduced to learn the mode of data and hyperparameters by minimizing the KL divergence between the true posterior distribution and the approximate one. Its most attractive property is that it can rigorously approximate the exact posterior. Inducing variable approximation and SVI further extend our model to big data. We apply our model to feature matching and dense gene expression imputation. Extensive results show the significant improvement over existing robust GPRs.
384
+
385
+ # References
386
+
387
+ [1] Vassileios Balntas, Karel Lenc, Andrea Vedaldi, and Krystian Mikolajczyk. Hpatches: A benchmark and evaluation of handcrafted and learned local descriptors. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5173-5182, 2017. 6, 7
388
+ [2] Daniel Barath, Jana Noskova, Maksym Ivashechkin, and Jiri Matas. Magsac++, a fast, reliable and accurate robust estimator. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1304-1312, 2020. 6, 7
389
+ [3] Sophie Bergmann, Christopher A Penfold, Erin Slatery, Dylan Siriwardena, Charis Drummer, Stephen Clark, Stanley E Strawbridge, Keiko Kishimoto, Alice Vickers, Mukul Tewary, et al. Spatial profiling of early primate gastrulation in utero. Nature, 609(7925):136-143, 2022. 1, 5
390
+ [4] Jia-Wang Bian, Yu-Huan Wu, Ji Zhao, Yun Liu, Le Zhang, Ming-Ming Cheng, and Ian Reid. An evaluation of feature matchers for fundamental matrix estimation. In Proceedings of the British Machine Vision Conference, 2019. 7
391
+ [5] Christopher M Bishop and Nasser M Nasrabadi. Pattern Recognition and Machine Learning. Springer, 2006. 4
392
+ [6] Michael J Black and Anand Rangarajan. On the unification of line processes, outlier rejection, and robust statistics with applications in early vision. International Journal of Computer Vision, 19(1):57-91, 1996. 1
393
+ [7] Lehel Csato and Manfred Opper. Sparse on-line gaussian processes. Neural Computation, 14(3):641-668, 2002. 2
394
+ [8] Atefeh Daemi, Hariprasad Kodamana, and Biao Huang. Gaussian process modelling with gaussian mixture likelihood. Journal of Process Control, 81:209-220, 2019. 1, 2, 6
395
+ [9] A Philip Dawid. Posterior expectations for large observations. Biometrika, 60(3):664-667, 1973. 1
396
+ [10] Bruno De Finetti. The bayesian approach to the rejection of outliers. In Proceedings of the fourth Berkeley Symposium on Probability and Statistics, pages 199–210, 1961. 1
397
+ [11] Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superpoint: Self-supervised interest point detection and description. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops, pages 224-236, 2018. 7
398
+ [12] Aoxiang Fan, Xingyu Jiang, Yong Ma, Xiaoguang Mei, and Jiayi Ma. Smoothness-driven consensus based on compact representation for robust feature matching. IEEE Transactions on Neural Networks and Learning Systems, 2021. 6, 7
399
+ [13] Martin A Fischler and Robert C Bolles. Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. Communications of the ACM, 24(6):381-395, 1981. 6
400
+ [14] Jerome H Friedman. Multivariate adaptive regression splines. The Annals of Statistics, 19(1):1-67, 1991. 5, 6
401
+ [15] Yarin Gal and Richard Turner. Improving the gaussian process sparse spectrum approximation by representing uncertainty in frequency inputs. In Proceedings of the Interna
402
+
403
+ tional Conference on Machine Learning, pages 655-664, 2015. 2
404
+ [16] John Geweke. Bayesian treatment of the independent student-t linear model. Journal of Applied Econometrics, 8(S1):S19-S40, 1993. 1
405
+ [17] Paul Goldberg, Christopher Williams, and Christopher Bishop. Regression with input-dependent noise: A gaussian process treatment. Advances in Neural Information Processing Systems, 10, 1997. 1
406
+ [18] James Hensman, Nicolo Fusi, and Neil D Lawrence. Gaussian processes for big data. arXiv preprint arXiv:1309.6835, 2013. 2, 5
407
+ [19] James Hensman, Magnus Rattray, and Neil Lawrence. Fast variational inference in the conjugate exponential family. Advances in Neural Information Processing Systems, 25, 2012. 5
408
+ [20] Matthew D Hoffman, David M Blei, Chong Wang, and John Paisley. Stochastic variational inference. Journal of Machine Learning Research, 14:1303-1347, 2013. 2, 4
409
+ [21] Pasi Jylänki, Jarno Vanhatalo, and Aki Vehtari. Robust gaussian process regression with a student-t likelihood. Journal of Machine Learning Research, 12(11):3227-3257, 2011. 1
410
+ [22] Malte Kuss. Gaussian process models for robust regression, classification, and reinforcement learning. PhD thesis, ech-nische Universität Darmstadt Darmstadt, Germany, 2006. 1, 2
411
+ [23] Miguel Lázaro-Gredilla, Joaquin Quinonero-Candela, Carl Edward Rasmussen, and Aníbal R Figueiras-Vidal. Sparse spectrum gaussian process regression. The Journal of Machine Learning Research, 11:1865-1881, 2010. 2
412
+ [24] David G Lowe. Distinctive image features from scale-invariant keypoints. International Journal of Computer Vision, 60(2):91-110, 2004. 6
413
+ [25] Jiayi Ma, Aoxiang Fan, Xingyu Jiang, and Guobao Xiao. Feature matching via motion-consistency driven probabilistic graphical model. International Journal of Computer Vision, 130(9):2249-2264, 2022. 6, 7
414
+ [26] Jiayi Ma, Ji Zhao, Junjun Jiang, Huabing Zhou, and Xiaojie Guo. Locality preserving matching. International Journal of Computer Vision, 127(5):512-531, 2019. 6, 7
415
+ [27] Andrew Naish-Guzman and Sean Holden. Robust regression with twinned gaussian processes. Advances in Neural Information Processing Systems, 20, 2007. 1
416
+ [28] Radford M Neal. Monte carlo implementation of gaussian process models for bayesian regression and classification. arXiv preprint physics/9701026, 1997. 1, 5, 6
417
+ [29] Anthony O'Hagan. On outlier rejection phenomena in bayes inference. Journal of the Royal Statistical Society: Series B (Methodological), 41(3):358-367, 1979. 1
418
+ [30] Rishik Ranjan, Biao Huang, and Alireza Fatehi. Robust gaussian process modeling using em algorithm. Journal of Process Control, 42:125-136, 2016. 1, 2
419
+ [31] Carl Edward Rasmussen. Gaussian processes in machine learning. In Summer School on Machine Learning, pages 63-71. Springer, 2003. 1, 2
420
+ [32] Carl Edward Rasmussen and Hannes Nickisch. Gaussian processes for machine learning (gpml) toolbox. The Jour-
421
+
422
+ nal of Machine Learning Research, 11:3011-3015, 2010. 6, 8
423
+ [33] Carl Edward Rasmussen and Joaquin Quinonero-Candela. Healing the relevance vector machine through augmentation. In Proceedings of the International Conference on Machine learning, pages 689-696, 2005. 2
424
+ [34] Patrik L Stahl, Fredrik Salmen, Sanja Vickovic, Anna Lundmark, José Fernández Navarro, Jens Magnusson, Stefania Giacomello, Michaela Asp, Jakub O Westholm, Mikael Huss, et al. Visualization and analysis of gene expression in tissue sections by spatial transcriptomics. Science, 353(6294):78-82, 2016. 7
425
+ [35] Valentine Svensson, Sarah A Teichmann, and Oliver Stegle. Spatialde: identification of spatially variable genes. Nature Methods, 15(5):343-346, 2018. 8
426
+ [36] Bart Thomee, David A Shamma, Gerald Friedland, Benjamin Elizalde, Karl Ni, Douglas Poland, Damian Borth, and Li-Jia Li. Yfcc100m: The new data in multimedia research. Communications of the ACM, 59(2):64-73, 2016. 6, 7
427
+ [37] Michalis Titsias. Variational learning of inducing variables in sparse gaussian processes. In Proceedings of the Artificial Intelligence and Statistics, pages 567-574, 2009. 2, 3, 6
428
+ [38] Shimon Ullman. The interpretation of structure from motion. Proceedings of the Royal Society of London. Series B. Biological Sciences, 203(1153):405-426, 1979. 5
429
+ [39] Jarno Vanhatalo, Pasi Jylänk, and Aki Vehtari. Gaussian process regression with student-t likelihood. Advances in Neural Information Processing Systems, 22, 2009. 2
430
+ [40] Christopher Williams and Matthias Seeger. Using the nystrom method to speed up kernel machines. Advances in Neural Information Processing Systems, 13, 2000. 2
431
+ [41] Oliver Williams and Andrew Fitzgibbon. Gaussian process implicit surfaces. In *Gaussian Processes in Practice*, 2006. 1
432
+ [42] Kyle Wilson and Noah Snavely. Robust global translations with 1dsfm. In Proceedings of the European Conference on Computer Vision, pages 61-75, 2014. 6, 7
433
+ [43] Kwang Moo Yi, Eduard Trulls, Yuki Ono, Vincent Lepetit, Mathieu Salzmann, and Pascal Fua. Learning to find good correspondences. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2666-2674, 2018. 6, 7
434
+ [44] Jiahui Zhang, Dawei Sun, Zixin Luo, Anbang Yao, Lei Zhou, Tianwei Shen, Yurong Chen, Long Quan, and Honggen Liao. Learning two-view correspondences and geometry using order-aware network. In Proceedings of the IEEE International Conference on Computer Vision, pages 5845-5854, 2019. 6, 7
2023/Robust and Scalable Gaussian Process Regression and Its Applications/images.zip ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:5f71e0c112cf5b64baf016e0842cdc72ab7335e31648465db65f9b9c27025e4c
3
+ size 726771
2023/Robust and Scalable Gaussian Process Regression and Its Applications/layout.json ADDED
The diff for this file is too large to render. See raw diff
 
2023/RobustNeRF_ Ignoring Distractors With Robust Losses/2f329418-58fe-4ad7-b462-6d0d0ce341f1_content_list.json ADDED
@@ -0,0 +1,2145 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ [
2
+ {
3
+ "type": "text",
4
+ "text": "RobustNeRF: Ignoring Distractors with Robust Losses<sup>4</sup>",
5
+ "text_level": 1,
6
+ "bbox": [
7
+ 204,
8
+ 128,
9
+ 764,
10
+ 152
11
+ ],
12
+ "page_idx": 0
13
+ },
14
+ {
15
+ "type": "text",
16
+ "text": "Sara Sabour $^{1,2}$ Suhani Vora $^{1}$ Daniel Duckworth $^{1}$ Ivan Krasin $^{1}$ David J. Fleet $^{1,2}$ Andrea Tagliafasacchi $^{1,2,3}$",
17
+ "bbox": [
18
+ 210,
19
+ 179,
20
+ 756,
21
+ 215
22
+ ],
23
+ "page_idx": 0
24
+ },
25
+ {
26
+ "type": "text",
27
+ "text": "<sup>1</sup>Google Research, Brain Team <sup>2</sup>University of Toronto <sup>3</sup>Simon Fraser University",
28
+ "bbox": [
29
+ 228,
30
+ 217,
31
+ 736,
32
+ 233
33
+ ],
34
+ "page_idx": 0
35
+ },
36
+ {
37
+ "type": "text",
38
+ "text": "Abstract",
39
+ "text_level": 1,
40
+ "bbox": [
41
+ 233,
42
+ 268,
43
+ 313,
44
+ 282
45
+ ],
46
+ "page_idx": 0
47
+ },
48
+ {
49
+ "type": "text",
50
+ "text": "Neural radiance fields (NeRF) excel at synthesizing new views given multi-view, calibrated images of a static scene. When scenes include distractors, which are not persistent during image capture (moving objects, lighting variations, shadows), artifacts appear as view-dependent effects or 'floaters'. To cope with distractors, we advocate a form of robust estimation for NeRF training, modeling distractors in training data as outliers of an optimization problem. Our method successfully removes outliers from a scene and improves upon our baselines, on synthetic and real-world scenes. Our technique is simple to incorporate in modern NeRF frameworks, with few hyper-parameters. It does not assume a priori knowledge of the types of distractors, and is instead focused on the optimization problem rather than pre-processing or modeling transient objects. More results at https://robustnerf.github.io/public.",
51
+ "bbox": [
52
+ 75,
53
+ 291,
54
+ 473,
55
+ 532
56
+ ],
57
+ "page_idx": 0
58
+ },
59
+ {
60
+ "type": "text",
61
+ "text": "1. Introduction",
62
+ "text_level": 1,
63
+ "bbox": [
64
+ 76,
65
+ 554,
66
+ 209,
67
+ 569
68
+ ],
69
+ "page_idx": 0
70
+ },
71
+ {
72
+ "type": "text",
73
+ "text": "The ability to understand the structure of a static 3D scene from 2D images alone is a fundamental problem in computer vision [44]. It finds applications in AR/VR for mapping virtual environments [6, 36, 61], in autonomous robotics for action planning [1], and in photogrammetry to create digital copies of real-world objects [34].",
74
+ "bbox": [
75
+ 75,
76
+ 575,
77
+ 468,
78
+ 666
79
+ ],
80
+ "page_idx": 0
81
+ },
82
+ {
83
+ "type": "text",
84
+ "text": "Neural fields [55] have recently revolutionized this classical task, by storing 3D representations within the weights of a neural network [39]. These representations are optimized by back-propagating image differences. When the fields store view-dependent radiance and volumetric rendering is employed [21], we can capture 3D scenes with photo-realistic accuracy, and we refer to the generated representation as Neural Radiance Fields, or NeRF [25]).",
85
+ "bbox": [
86
+ 75,
87
+ 667,
88
+ 468,
89
+ 787
90
+ ],
91
+ "page_idx": 0
92
+ },
93
+ {
94
+ "type": "text",
95
+ "text": "Training of NeRF models generally requires a large collection of images equipped with accurate camera calibration, which can often be recovered via structure-from-motion [37]. Behind its simplicity, NeRF hides several assumptions. As models are typically trained to minimize error in RGB color space, it is of paramount importance",
96
+ "bbox": [
97
+ 75,
98
+ 787,
99
+ 468,
100
+ 878
101
+ ],
102
+ "page_idx": 0
103
+ },
104
+ {
105
+ "type": "image",
106
+ "img_path": "images/1f94b3ae1ab59c57b87c5e4884be69f5c0048ed2a2ee83bf4879413ecd792bef.jpg",
107
+ "image_caption": [
108
+ "Figure 1. NeRF assumes photometric consistency in the observed images of a scene. Violations of this assumption, as with the images in the top row, yield reconstructed scenes with inconsistent content in the form of \"floaters\" (highlighted with ellipses). We introduce a simple technique that produces clean reconstruction by automatically ignoring distractors without explicit supervision."
109
+ ],
110
+ "image_footnote": [],
111
+ "bbox": [
112
+ 503,
113
+ 267,
114
+ 890,
115
+ 433
116
+ ],
117
+ "page_idx": 0
118
+ },
119
+ {
120
+ "type": "text",
121
+ "text": "that images are photometrically consistent - two photos taken from the same vantage point should be identical up to noise. Unless one employs a method explicitly accounting for it [35], one should manually hold a camera's focus, exposure, white-balance, and ISO fixed.",
122
+ "bbox": [
123
+ 496,
124
+ 537,
125
+ 892,
126
+ 611
127
+ ],
128
+ "page_idx": 0
129
+ },
130
+ {
131
+ "type": "text",
132
+ "text": "However, properly configuring one's camera is not all that is required to capture high-quality NeRFs – it is also important to avoid distractors: anything that isn't persistent throughout the entire capture session. Distractors come in many shapes and forms, from the hard-shadows cast by the operators as they explore the scene to a pet or child casually walking within the camera's field of view. Distractors are tedious to remove manually, as this would require pixel-by-pixel labeling. They are also tedious to detect, as typical NeRF scenes are trained from hundreds of input images, and the types of distractors are not known a priori. If distractors are ignored, the quality of the reconstruction scene suffers significantly; see Figure 1.",
133
+ "bbox": [
134
+ 496,
135
+ 613,
136
+ 893,
137
+ 808
138
+ ],
139
+ "page_idx": 0
140
+ },
141
+ {
142
+ "type": "text",
143
+ "text": "In a typical capture session, it is difficult to capture multiple images of the same scene from the same viewpoint, rendering distractors challenging to model mathematically. As such, while view-dependent effects are what give NeRF their realistic look, how can the model tell the difference between a distractor and a view-dependent effect?",
144
+ "bbox": [
145
+ 496,
146
+ 810,
147
+ 893,
148
+ 900
149
+ ],
150
+ "page_idx": 0
151
+ },
152
+ {
153
+ "type": "header",
154
+ "text": "CVF",
155
+ "bbox": [
156
+ 106,
157
+ 2,
158
+ 181,
159
+ 42
160
+ ],
161
+ "page_idx": 0
162
+ },
163
+ {
164
+ "type": "header",
165
+ "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.",
166
+ "bbox": [
167
+ 236,
168
+ 0,
169
+ 807,
170
+ 46
171
+ ],
172
+ "page_idx": 0
173
+ },
174
+ {
175
+ "type": "page_footnote",
176
+ "text": "4Work done at Google Research.",
177
+ "bbox": [
178
+ 94,
179
+ 886,
180
+ 271,
181
+ 898
182
+ ],
183
+ "page_idx": 0
184
+ },
185
+ {
186
+ "type": "page_number",
187
+ "text": "20626",
188
+ "bbox": [
189
+ 478,
190
+ 944,
191
+ 519,
192
+ 957
193
+ ],
194
+ "page_idx": 0
195
+ },
196
+ {
197
+ "type": "text",
198
+ "text": "Despite the challenges, the research community has devised several approaches to overcome this issue:",
199
+ "bbox": [
200
+ 76,
201
+ 90,
202
+ 467,
203
+ 119
204
+ ],
205
+ "page_idx": 1
206
+ },
207
+ {
208
+ "type": "list",
209
+ "sub_type": "text",
210
+ "list_items": [
211
+ "- If distractors are known to belong to a specific class (e.g., people), one can remove them with a pre-trained semantic segmentation model [35, 43] – this process does not generalize to “unexpected” distractors such as shadows.",
212
+ "- One can model distractors as per-image transient phenomena, and control the balance of transient/persistent modeling [23] - however, it is difficult to tune the losses that control this Pareto-optimal objective.",
213
+ "- One can model data in time (i.e., high-framerate video) and decompose the scene into static and dynamic (i.e., distractor) components [53] – but this clearly only applies to video rather than photo collection captures."
214
+ ],
215
+ "bbox": [
216
+ 76,
217
+ 121,
218
+ 467,
219
+ 301
220
+ ],
221
+ "page_idx": 1
222
+ },
223
+ {
224
+ "type": "text",
225
+ "text": "Conversely, we approach the problem of distractors by modeling them as outliers in NeRF optimization.",
226
+ "bbox": [
227
+ 76,
228
+ 301,
229
+ 467,
230
+ 332
231
+ ],
232
+ "page_idx": 1
233
+ },
234
+ {
235
+ "type": "text",
236
+ "text": "We analyze the aforementioned techniques through the lens of robust estimation, allowing us to understand their behavior, and to design a method that is not only simpler to implement but also more effective (see Figure 1). As a result, we obtain a method that is straightforward to implement, requires minimal-to-no hyper-parameter tuning, and achieves state-of-the-art performance. We evaluate our method:",
237
+ "bbox": [
238
+ 75,
239
+ 333,
240
+ 467,
241
+ 436
242
+ ],
243
+ "page_idx": 1
244
+ },
245
+ {
246
+ "type": "list",
247
+ "sub_type": "text",
248
+ "list_items": [
249
+ "- quantitatively, in terms of reconstruction with synthetically, yet photo-realistically, rendered data;",
250
+ "- qualitatively on publicly available datasets (often finetuned to work effectively with previous methods);",
251
+ "- on a new collection of natural and synthetic scenes, including those autonomously acquired by a robot, allowing us to demonstrate the sensitivity of previous methods to hyper-parameter tuning."
252
+ ],
253
+ "bbox": [
254
+ 76,
255
+ 438,
256
+ 467,
257
+ 558
258
+ ],
259
+ "page_idx": 1
260
+ },
261
+ {
262
+ "type": "text",
263
+ "text": "2. Related Work",
264
+ "text_level": 1,
265
+ "bbox": [
266
+ 76,
267
+ 571,
268
+ 217,
269
+ 585
270
+ ],
271
+ "page_idx": 1
272
+ },
273
+ {
274
+ "type": "text",
275
+ "text": "We briefly review the basics and notation of Neural Radiance Fields. We then describe recent progress in NeRF research, paying particular attention to techniques for modeling of static/dynamic scenes.",
276
+ "bbox": [
277
+ 76,
278
+ 593,
279
+ 467,
280
+ 654
281
+ ],
282
+ "page_idx": 1
283
+ },
284
+ {
285
+ "type": "text",
286
+ "text": "Neural Radiance Fields. A neural radiance field (NeRF) is a continuous volumetric representation of a 3D scene, stored within the parameters of a neural network $\\theta$ . The representation maps a position $\\mathbf{x}$ and view direction $\\mathbf{d}$ to a view-dependent RGB color and view-independent density:",
287
+ "bbox": [
288
+ 76,
289
+ 659,
290
+ 467,
291
+ 734
292
+ ],
293
+ "page_idx": 1
294
+ },
295
+ {
296
+ "type": "equation",
297
+ "text": "\n$$\n\\left. \\begin{array}{l} \\mathbf {c} (\\mathbf {x}, \\mathbf {d}) \\\\ \\sigma (\\mathbf {x}) \\end{array} \\right\\} f (\\mathbf {x}, \\mathbf {d}; \\boldsymbol {\\theta}) \\tag {1}\n$$\n",
298
+ "text_format": "latex",
299
+ "bbox": [
300
+ 204,
301
+ 744,
302
+ 467,
303
+ 785
304
+ ],
305
+ "page_idx": 1
306
+ },
307
+ {
308
+ "type": "text",
309
+ "text": "This representation is trained from a collection, $\\{(\\mathbf{C}_i,\\mathbf{T}_i)\\}$ of images $\\mathbf{C}_i$ with corresponding calibration parameters $\\mathbf{T}_i$ (camera extrinsic and intrinsic).",
310
+ "bbox": [
311
+ 76,
312
+ 794,
313
+ 467,
314
+ 839
315
+ ],
316
+ "page_idx": 1
317
+ },
318
+ {
319
+ "type": "text",
320
+ "text": "During training the calibration information is employed to convert each pixel of the image into a ray $\\mathbf{r} = (\\mathbf{o},\\mathbf{d})$ , and rays are drawn randomly from input images to form a training mini-batch $(\\mathbf{r}\\sim \\mathbf{C}_i)$ . The parameters $\\theta$ are optimized to",
321
+ "bbox": [
322
+ 76,
323
+ 839,
324
+ 467,
325
+ 900
326
+ ],
327
+ "page_idx": 1
328
+ },
329
+ {
330
+ "type": "text",
331
+ "text": "correctly predict the colors of the pixels in the batch via the L2 photometric-reconstruction loss:",
332
+ "bbox": [
333
+ 498,
334
+ 90,
335
+ 890,
336
+ 119
337
+ ],
338
+ "page_idx": 1
339
+ },
340
+ {
341
+ "type": "equation",
342
+ "text": "\n$$\n\\mathcal {L} _ {\\mathrm {r g b}} (\\boldsymbol {\\theta}) = \\sum_ {i} \\mathbb {E} _ {\\mathbf {r} \\sim \\mathbf {C} _ {i}} \\left[ \\mathcal {L} _ {\\mathrm {r g b}} ^ {\\mathbf {r}, i} (\\boldsymbol {\\theta}) \\right] \\tag {2}\n$$\n",
343
+ "text_format": "latex",
344
+ "bbox": [
345
+ 586,
346
+ 133,
347
+ 890,
348
+ 167
349
+ ],
350
+ "page_idx": 1
351
+ },
352
+ {
353
+ "type": "equation",
354
+ "text": "\n$$\n\\mathcal {L} _ {\\mathrm {r g b}} ^ {\\mathbf {r}, i} (\\boldsymbol {\\theta}) = \\left| \\left| \\mathbf {C} (\\mathbf {r}; \\boldsymbol {\\theta}) - \\mathbf {C} _ {i} (\\mathbf {r}) \\right| \\right| _ {2} ^ {2} \\tag {3}\n$$\n",
355
+ "text_format": "latex",
356
+ "bbox": [
357
+ 588,
358
+ 170,
359
+ 890,
360
+ 191
361
+ ],
362
+ "page_idx": 1
363
+ },
364
+ {
365
+ "type": "text",
366
+ "text": "Parameterizing the ray as $\\mathbf{r}(t) = \\mathbf{o} + t\\mathbf{d}$ , the NeRF model image $\\mathbf{C}(\\mathbf{r};\\boldsymbol{\\theta})$ is generated pixel-by-pixel volumetric rendering based on $\\sigma(\\cdot)$ and $\\mathbf{c}(\\cdot)$ (e.g., see [25, 42]).",
367
+ "bbox": [
368
+ 498,
369
+ 203,
370
+ 890,
371
+ 250
372
+ ],
373
+ "page_idx": 1
374
+ },
375
+ {
376
+ "type": "text",
377
+ "text": "Recent progress on NeRF models. NeRF models have recently been extended in several ways. A major thread has been the speedup of training [15, 27] and inference [6, 13], enabling today's models to be trained in minutes [27], and rendered on mobile in real-time [6]. While initially restricted to forward-facing scenes, researchers quickly found ways to model real-world $360^{\\circ}$ scenes [4, 59], and to reduce the required number of images, via sensor fusion [35] or hand-designed priors [28]. We can now deal with image artifacts such as motion blur [22], exposure [24], and lens distortion [14]. And the requirement of (precise) camera calibrations is quickly being relaxed with the introduction of techniques for local camera refinement [8, 19], or direct inference [58]. While a NeRF typically represents geometry via volumetric density, there exist models custom-tailored to predict surfaces [29, 51], which can be extended to use predicted normals to significantly improve reconstruction quality [50, 57]. Given high-quality normals [47], inferring the (rendering) structure of a scene becomes a possibility [5]. We also note recent papers about additional applications to generalization [56], semantic understanding [48], generative modeling [33], robotics [1], and text-to-3D [31].",
378
+ "bbox": [
379
+ 496,
380
+ 256,
381
+ 890,
382
+ 589
383
+ ],
384
+ "page_idx": 1
385
+ },
386
+ {
387
+ "type": "text",
388
+ "text": "Modeling non-static scenes. For unstructured scenes like those considered here, the community has focused on reconstructing both static and non-static elements from video. The most direct approach, treating time as an auxiliary input, leads to cloudy geometry and a lack of fine detail [11, 54]. Directly optimizing per-frame latent codes as an auxiliary input has proved more effective [17, 30, 53]. The most widely-adopted approach is to fit a time-conditioned deformation field mapping 3D points between pairs of frames [18, 49] or to a canonical coordinate frame [9, 10, 20, 32, 45]. Given how sparsely spacetime is sampled, all methods require careful regularization, optimization, or additional training signals to achieve acceptable results.",
389
+ "bbox": [
390
+ 496,
391
+ 597,
392
+ 890,
393
+ 806
394
+ ],
395
+ "page_idx": 1
396
+ },
397
+ {
398
+ "type": "text",
399
+ "text": "Relatively little attention has been given to removing non-static elements. One common approach is to segment and ignore pixels which are likely to be distractors [35, 43]. While this eliminates larger objects, it fails to account for secondary effects like shadows. Prior attempts to model distractors as outliers still leave residual cloudy geometry [23].",
400
+ "bbox": [
401
+ 496,
402
+ 809,
403
+ 890,
404
+ 900
405
+ ],
406
+ "page_idx": 1
407
+ },
408
+ {
409
+ "type": "page_number",
410
+ "text": "20627",
411
+ "bbox": [
412
+ 478,
413
+ 944,
414
+ 517,
415
+ 955
416
+ ],
417
+ "page_idx": 1
418
+ },
419
+ {
420
+ "type": "image",
421
+ "img_path": "images/fd427e53b06992ad7e2399ed5219e2db17f6d6b5aac0badd013f1d3667e33e83.jpg",
422
+ "image_caption": [
423
+ "Figure 2. Ambiguity - A simple 2D scene where a static object (blue) is captured by three cameras. During the first and third capture the scene is not photo-consistent as a distractor was within the field of view. Not photo-consistent portions of the scene can end up being encoded as view-dependent effects - even when we assume ground truth geometry."
424
+ ],
425
+ "image_footnote": [],
426
+ "bbox": [
427
+ 86,
428
+ 85,
429
+ 566,
430
+ 223
431
+ ],
432
+ "page_idx": 2
433
+ },
434
+ {
435
+ "type": "image",
436
+ "img_path": "images/ff166e4f12987f76838b6513f6e0bb8ec9a53ef5607f55d7408084a762a7996d.jpg",
437
+ "image_caption": [],
438
+ "image_footnote": [],
439
+ "bbox": [
440
+ 573,
441
+ 85,
442
+ 875,
443
+ 223
444
+ ],
445
+ "page_idx": 2
446
+ },
447
+ {
448
+ "type": "text",
449
+ "text": "3. Method",
450
+ "text_level": 1,
451
+ "bbox": [
452
+ 76,
453
+ 282,
454
+ 168,
455
+ 299
456
+ ],
457
+ "page_idx": 2
458
+ },
459
+ {
460
+ "type": "text",
461
+ "text": "The classical NeRF training losses (3) are effective for capturing scenes that are photometrically consistent, leading to the photo-realistic novel-view synthesis that we are now accustomed to seeing in recent research. However, \"what happens when there are elements of the scene that are not persistent throughout the entire capture session?\" Simple examples of such scenes include those in which an object is only present in some fraction of the observed images, or may not remain in the same position in all observed images. For example, Figure 2 depicts a 2D scene comprising a persistent object (the truck), along with several transient objects (e.g., people and a dog). While rays in blue from the three cameras intersect the truck, the green and orange rays from cameras 1 and 3 intersect transient objects. For video capture and spatio-temporal NeRF models, the persistent objects comprise the \"static\" portion of the scene, while the rest would be called the \"dynamic\".",
462
+ "bbox": [
463
+ 75,
464
+ 306,
465
+ 473,
466
+ 564
467
+ ],
468
+ "page_idx": 2
469
+ },
470
+ {
471
+ "type": "text",
472
+ "text": "3.1. Sensitivity to outliers",
473
+ "text_level": 1,
474
+ "bbox": [
475
+ 76,
476
+ 575,
477
+ 274,
478
+ 590
479
+ ],
480
+ "page_idx": 2
481
+ },
482
+ {
483
+ "type": "text",
484
+ "text": "For Lambertian scenes, photo-consistent structure is view independent, as scene radiance only depends on the incident light [16]. For such scenes, view-dependent NeRF models like (1), trained by minimizing (3), admit local optima in which transient objects are explained by view-dependent terms. Figure 2 depicts this, with the outgoing color corresponding to the memorized color of the outlier - i.e. view-dependent radiance. Such models exploit the view-dependent capacity of the model to over-fit observations, effectively memorizing the transient objects. One can alter the model to remove dependence on $\\mathbf{d}$ but the L2 loss remains problematic as least-squares (LS) estimators are sensitive to outliers, or heavy-tailed noise distributions.",
485
+ "bbox": [
486
+ 75,
487
+ 597,
488
+ 468,
489
+ 792
490
+ ],
491
+ "page_idx": 2
492
+ },
493
+ {
494
+ "type": "text",
495
+ "text": "Under more natural conditions, dropping the Lambertian assumption, the problem becomes more complex as both non-Lambertian reflectance phenomena and outliers can be explained as view-dependent radiance. While we want the models to capture photo-consistent view-dependent radiance, outliers and other transient phenomena should ideally be ignored. And in such cases, optimization with an L2",
496
+ "bbox": [
497
+ 75,
498
+ 795,
499
+ 470,
500
+ 901
501
+ ],
502
+ "page_idx": 2
503
+ },
504
+ {
505
+ "type": "text",
506
+ "text": "loss (3) yields significant errors in reconstruction; see Figure 1. Problems like these are pervasive in NeRF model fitting, especially in uncontrolled environments with complex reflectance, non-rigidity, or independently moving objects.",
507
+ "bbox": [
508
+ 496,
509
+ 285,
510
+ 890,
511
+ 347
512
+ ],
513
+ "page_idx": 2
514
+ },
515
+ {
516
+ "type": "text",
517
+ "text": "3.2. Robustness to outliers",
518
+ "text_level": 1,
519
+ "bbox": [
520
+ 498,
521
+ 353,
522
+ 705,
523
+ 368
524
+ ],
525
+ "page_idx": 2
526
+ },
527
+ {
528
+ "type": "text",
529
+ "text": "Robustness via semantic segmentation. One way to reduce outlier contamination during NeRF model optimization is to rely on an oracle $\\mathbf{S}$ that specifies whether a given pixel $\\mathbf{r}$ from image $i$ is an outlier, and should therefore be excluded from the empirical loss, replacing (3) with:",
530
+ "bbox": [
531
+ 496,
532
+ 376,
533
+ 892,
534
+ 450
535
+ ],
536
+ "page_idx": 2
537
+ },
538
+ {
539
+ "type": "equation",
540
+ "text": "\n$$\n\\mathcal {L} _ {\\text {o r a c l e}} ^ {\\mathbf {r}, i} (\\boldsymbol {\\theta}) = \\mathbf {S} _ {i} (\\mathbf {r}) \\cdot | | \\mathbf {C} (\\mathbf {r}; \\boldsymbol {\\theta}) - \\mathbf {C} _ {i} (\\mathbf {r}) | | _ {2} ^ {2} \\tag {4}\n$$\n",
541
+ "text_format": "latex",
542
+ "bbox": [
543
+ 558,
544
+ 458,
545
+ 890,
546
+ 477
547
+ ],
548
+ "page_idx": 2
549
+ },
550
+ {
551
+ "type": "text",
552
+ "text": "In practice, a pre-trained (semantic) segmentation network $S$ might serve as an oracle, $\\mathbf{S}_i = S(\\mathbf{C}_i)$ . E.g., Nerf-in-the-wild [23] employed a semantic segmenter to remove pixels occupied by people, as they are outliers in the context of photo-tourism. Urban Radiance Fields [35] segmented out sky pixels, while LOL-NeRF [33] ignored pixels not belonging to faces. The obvious problem with this approach is the need for an oracle to detect arbitrary distractors.",
553
+ "bbox": [
554
+ 496,
555
+ 484,
556
+ 892,
557
+ 604
558
+ ],
559
+ "page_idx": 2
560
+ },
561
+ {
562
+ "type": "text",
563
+ "text": "Robust estimators. Another way to reduce sensitivity to outliers is to replace the conventional L2 loss (3) with a robust loss (e.g., [2, 41]), so that photometrically-inconsistent observations can be down-weighted during optimization. Given a robust kernel $\\kappa(\\cdot)$ , we rewrite our training loss as:",
564
+ "bbox": [
565
+ 496,
566
+ 611,
567
+ 890,
568
+ 686
569
+ ],
570
+ "page_idx": 2
571
+ },
572
+ {
573
+ "type": "equation",
574
+ "text": "\n$$\n\\mathcal {L} _ {\\text {r o b u s t}} ^ {\\mathbf {r}, i} (\\boldsymbol {\\theta}) = \\kappa \\left(\\left| \\left| \\mathbf {C} (\\mathbf {r}; \\boldsymbol {\\theta}) - \\mathbf {C} _ {i} (\\mathbf {r}) \\right| \\right| _ {2}\\right) \\tag {5}\n$$\n",
575
+ "text_format": "latex",
576
+ "bbox": [
577
+ 571,
578
+ 693,
579
+ 890,
580
+ 712
581
+ ],
582
+ "page_idx": 2
583
+ },
584
+ {
585
+ "type": "text",
586
+ "text": "where $\\kappa(\\cdot)$ is positive and monotonically increasing. MipNeRF [3], for example, employs an L1 loss $\\kappa(\\epsilon) = |\\epsilon|$ , which provides some degree of robustness to outliers during NeRF training. Given our analysis, a valid question is whether we can straightforwardly employ a robust kernel to approach our problem, and if so, given the large variety of robust kernels [2], which is the kernel of choice.",
587
+ "bbox": [
588
+ 496,
589
+ 719,
590
+ 890,
591
+ 824
592
+ ],
593
+ "page_idx": 2
594
+ },
595
+ {
596
+ "type": "text",
597
+ "text": "Unfortunately, as discussed above, outliers and non-Lambertian effects can both be modelled as view-dependent effects (see Figure 3). As a consequence, with simple application of robust estimators it can be difficult to separate signal from noise. Figure 4 shows examples in which outliers",
598
+ "bbox": [
599
+ 496,
600
+ 825,
601
+ 890,
602
+ 900
603
+ ],
604
+ "page_idx": 2
605
+ },
606
+ {
607
+ "type": "page_number",
608
+ "text": "20628",
609
+ "bbox": [
610
+ 478,
611
+ 944,
612
+ 519,
613
+ 957
614
+ ],
615
+ "page_idx": 2
616
+ },
617
+ {
618
+ "type": "image",
619
+ "img_path": "images/11ad7dc1cc45fb13dab8eb420d7cc8ad66432e0546b9533578c9214db3f57180.jpg",
620
+ "image_caption": [],
621
+ "image_footnote": [],
622
+ "bbox": [
623
+ 83,
624
+ 87,
625
+ 271,
626
+ 167
627
+ ],
628
+ "page_idx": 3
629
+ },
630
+ {
631
+ "type": "image",
632
+ "img_path": "images/b2eb3620a00498884b091da8a7002ce0ca793164b22359e7226635d9c8a96708.jpg",
633
+ "image_caption": [],
634
+ "image_footnote": [],
635
+ "bbox": [
636
+ 274,
637
+ 88,
638
+ 460,
639
+ 167
640
+ ],
641
+ "page_idx": 3
642
+ },
643
+ {
644
+ "type": "image",
645
+ "img_path": "images/d79c260859e20c9b34c744b290b5c9725455d80ee2d5fcfb5fbf68d85187829b.jpg",
646
+ "image_caption": [
647
+ "Figure 3. Histograms - Robust estimators perform well when the distribution of residuals agrees with the one implied by the estimator (e.g., Gaussian for L2, Laplacian for L1). Here we visualize the ground-truth distribution of residuals (bottom-left), which is hardly a good match with any simple parametric distribution."
648
+ ],
649
+ "image_footnote": [],
650
+ "bbox": [
651
+ 81,
652
+ 167,
653
+ 271,
654
+ 253
655
+ ],
656
+ "page_idx": 3
657
+ },
658
+ {
659
+ "type": "image",
660
+ "img_path": "images/bc0cfa8000d411cb3d2e6833322a1287d9b33fa4e89c2466baecdf67d05404a1.jpg",
661
+ "image_caption": [],
662
+ "image_footnote": [],
663
+ "bbox": [
664
+ 274,
665
+ 172,
666
+ 460,
667
+ 252
668
+ ],
669
+ "page_idx": 3
670
+ },
671
+ {
672
+ "type": "text",
673
+ "text": "are removed, but fine-grained texture and view-dependent details are also lost, or conversely, fine-grained details are preserved, but outliers cause artifacts in the reconstructed scene. One can also observe mixtures of these cases in which details are not captured well, nor are outliers fully removed. We find that this behaviour occurs consistently for many different robust estimators and parameter settings.",
674
+ "bbox": [
675
+ 75,
676
+ 345,
677
+ 467,
678
+ 450
679
+ ],
680
+ "page_idx": 3
681
+ },
682
+ {
683
+ "type": "text",
684
+ "text": "Training time can also be problematic. The robust estimator gradient w.r.t. model parameters can be expressed using the chain rule as",
685
+ "bbox": [
686
+ 75,
687
+ 452,
688
+ 468,
689
+ 498
690
+ ],
691
+ "page_idx": 3
692
+ },
693
+ {
694
+ "type": "equation",
695
+ "text": "\n$$\n\\left. \\frac {\\partial \\kappa (\\epsilon (\\boldsymbol {\\theta}))}{\\partial \\boldsymbol {\\theta}} \\right| _ {\\boldsymbol {\\theta} ^ {(t)}} = \\left. \\frac {\\partial \\kappa (\\epsilon)}{\\partial \\epsilon} \\right| _ {\\epsilon (\\boldsymbol {\\theta} ^ {(t)})} \\cdot \\left. \\frac {\\partial \\epsilon (\\boldsymbol {\\theta})}{\\partial \\boldsymbol {\\theta}} \\right| _ {\\boldsymbol {\\theta} ^ {(t)}} \\tag {6}\n$$\n",
696
+ "text_format": "latex",
697
+ "bbox": [
698
+ 129,
699
+ 510,
700
+ 468,
701
+ 547
702
+ ],
703
+ "page_idx": 3
704
+ },
705
+ {
706
+ "type": "text",
707
+ "text": "The second factor is the classical NeRF gradient. The first factor is the kernel gradient evaluated at the current error residual $\\epsilon (\\pmb{\\theta}^{(t)})$ . During training, large residuals can equivalently come from high-frequency details that have not yet been learnt, or they may arise from outliers (see Figure 4 (bottom)). This explain why robust optimization, implemented as (5), should not be expected to decouple high-frequency details from outliers. Further, when strongly robust kernels are employed, like redescending estimators, this also explains the loss of visual fidelity. That is, because the gradient of (large) residuals get down-weighted by the (small) gradients of the kernel, slowing down the learning of these fine-grained details (see Figure 4 (top)).",
708
+ "bbox": [
709
+ 75,
710
+ 559,
711
+ 468,
712
+ 755
713
+ ],
714
+ "page_idx": 3
715
+ },
716
+ {
717
+ "type": "text",
718
+ "text": "3.3. Robustness via Trimmed Least Squares",
719
+ "text_level": 1,
720
+ "bbox": [
721
+ 76,
722
+ 768,
723
+ 415,
724
+ 784
725
+ ],
726
+ "page_idx": 3
727
+ },
728
+ {
729
+ "type": "text",
730
+ "text": "In what follows we advocate a form of iteratively reweighted least-squares (IRLS) with a Trimmed least squares (LS) loss for NeRF model fitting.",
731
+ "bbox": [
732
+ 75,
733
+ 789,
734
+ 468,
735
+ 834
736
+ ],
737
+ "page_idx": 3
738
+ },
739
+ {
740
+ "type": "text",
741
+ "text": "Iteratively Reweighted least Squares. IRLS is a widely used method for robust estimation that involves solving a sequence of weighted LS problems, the weights of which are adapted to reduce the influence of outliers. To that end,",
742
+ "bbox": [
743
+ 75,
744
+ 839,
745
+ 468,
746
+ 898
747
+ ],
748
+ "page_idx": 3
749
+ },
750
+ {
751
+ "type": "image",
752
+ "img_path": "images/7938b39f31acfdc501893b4a8f3350a2f3af6e6029632895dc97faf118a27321.jpg",
753
+ "image_caption": [],
754
+ "image_footnote": [],
755
+ "bbox": [
756
+ 501,
757
+ 88,
758
+ 689,
759
+ 169
760
+ ],
761
+ "page_idx": 3
762
+ },
763
+ {
764
+ "type": "image",
765
+ "img_path": "images/8bee22d7570c63c3c8dcb2a2fb19c6b6aacc77cd83387cf79c109fb8235f552f.jpg",
766
+ "image_caption": [],
767
+ "image_footnote": [],
768
+ "bbox": [
769
+ 702,
770
+ 90,
771
+ 888,
772
+ 162
773
+ ],
774
+ "page_idx": 3
775
+ },
776
+ {
777
+ "type": "image",
778
+ "img_path": "images/1632681e93a49b53846aba35cb9a9a030cc78ccb806d3756c22ec532e55ff874.jpg",
779
+ "image_caption": [
780
+ "Figure 4. Kernels - (top-left) Family of robust kernels [2], including L2 ( $\\alpha = 2$ ), Charbonnier ( $\\alpha = 1$ ) and Geman-McClure ( $\\alpha = -2$ ). (top-right) Mid-training, residual magnitudes are similar for distractors and fine-grained details, and pixels with large residuals are learned more slowly, as the gradient of re-descending kernels flattens out. (bottom-right) A too aggressive Geman-McClure in down-weighting large residuals removes both outliers and high-frequency detail. (bottom-left) A less aggressive Geman-McClure does not effectively remove outliers."
781
+ ],
782
+ "image_footnote": [],
783
+ "bbox": [
784
+ 501,
785
+ 171,
786
+ 689,
787
+ 244
788
+ ],
789
+ "page_idx": 3
790
+ },
791
+ {
792
+ "type": "image",
793
+ "img_path": "images/eef272506fab0a7293f190e56385e2b5f26c2fe42d91278bc7e3ba201f641252.jpg",
794
+ "image_caption": [],
795
+ "image_footnote": [],
796
+ "bbox": [
797
+ 702,
798
+ 171,
799
+ 888,
800
+ 244
801
+ ],
802
+ "page_idx": 3
803
+ },
804
+ {
805
+ "type": "text",
806
+ "text": "at iteration $t$ , one can write the loss as",
807
+ "bbox": [
808
+ 500,
809
+ 397,
810
+ 751,
811
+ 411
812
+ ],
813
+ "page_idx": 3
814
+ },
815
+ {
816
+ "type": "equation",
817
+ "text": "\n$$\n\\mathcal {L} _ {\\text {r o b u s t}} ^ {\\mathbf {r}, i} \\left(\\boldsymbol {\\theta} ^ {(t)}\\right) = \\omega \\left(\\boldsymbol {\\epsilon} ^ {(t - 1)} (\\mathbf {r})\\right) \\cdot | | \\mathbf {C} (\\mathbf {r}; \\boldsymbol {\\theta} ^ {(t)}) - \\mathbf {C} _ {i} (\\mathbf {r}) | | _ {2} ^ {2}\n$$\n",
818
+ "text_format": "latex",
819
+ "bbox": [
820
+ 519,
821
+ 419,
822
+ 870,
823
+ 440
824
+ ],
825
+ "page_idx": 3
826
+ },
827
+ {
828
+ "type": "equation",
829
+ "text": "\n$$\n\\boldsymbol {\\epsilon} ^ {(t - 1)} (\\mathbf {r}) = \\left| \\left| \\mathbf {C} (\\mathbf {r}; \\boldsymbol {\\theta} ^ {(t - 1)}) - \\mathbf {C} _ {i} (\\mathbf {r}) \\right| \\right| _ {2} \\tag {7}\n$$\n",
830
+ "text_format": "latex",
831
+ "bbox": [
832
+ 537,
833
+ 441,
834
+ 890,
835
+ 460
836
+ ],
837
+ "page_idx": 3
838
+ },
839
+ {
840
+ "type": "text",
841
+ "text": "For weight functions given by $\\omega (\\epsilon) = \\epsilon^{-1}\\cdot \\partial \\kappa (\\epsilon) / \\partial \\epsilon$ one can show that, under suitable conditions, the iteration converges to a local minima of (5) (see [41, Sec. 3]).",
842
+ "bbox": [
843
+ 500,
844
+ 470,
845
+ 890,
846
+ 516
847
+ ],
848
+ "page_idx": 3
849
+ },
850
+ {
851
+ "type": "text",
852
+ "text": "This framework admits a broad family of losses, including maximum likelihood estimators for heavy-tailed noise processes. Examples in Figure 4 include the Charbonnier loss (smoothed L1), and more aggressive redescending estimators such as the Lorentzian or Geman-McClure [2]. The objective in (4) can also be viewed as a weighted LS objective, the binary weights of which are provided by an oracle. And, as discussed at length below, one can also view several recent methods like NeRFW [23] and $\\mathrm{D}^2\\mathrm{NeRF}$ [53] through the lens of IRLS and weighted LS.",
853
+ "bbox": [
854
+ 496,
855
+ 517,
856
+ 890,
857
+ 667
858
+ ],
859
+ "page_idx": 3
860
+ },
861
+ {
862
+ "type": "text",
863
+ "text": "Nevertheless, choosing a suitable weight function $\\omega(\\epsilon)$ for NeRF optimization is non-trivial, due in large part to the intrinsic ambiguity between view-dependent radiance phenomena and outliers. One might try to solve this problem by learning a neural weight function [40], although generating enough annotated training data might be prohibitive. Instead, the approach taken below is to exploit inductive biases in the structure of outliers, combined with the simplicity of a robust, trimmed LS estimator.",
864
+ "bbox": [
865
+ 496,
866
+ 667,
867
+ 890,
868
+ 803
869
+ ],
870
+ "page_idx": 3
871
+ },
872
+ {
873
+ "type": "text",
874
+ "text": "Trimmed Robust Kernels. Our goal is to develop a weight function for use in iteratively weighted LS optimization that is simple and captures useful inductive biases for NeRF optimization. For simplicity we opt for a binary weight function with intuitive parameters that adapts naturally through model fitting so that fine-grained image details that are not",
875
+ "bbox": [
876
+ 496,
877
+ 809,
878
+ 890,
879
+ 898
880
+ ],
881
+ "page_idx": 3
882
+ },
883
+ {
884
+ "type": "page_number",
885
+ "text": "20629",
886
+ "bbox": [
887
+ 478,
888
+ 944,
889
+ 517,
890
+ 955
891
+ ],
892
+ "page_idx": 3
893
+ },
894
+ {
895
+ "type": "image",
896
+ "img_path": "images/c4d7aa68fc1cf767be6cc6e9d208ab8c015e53701c307328f83104d47b5e8f5c.jpg",
897
+ "image_caption": [],
898
+ "image_footnote": [],
899
+ "bbox": [
900
+ 112,
901
+ 89,
902
+ 297,
903
+ 170
904
+ ],
905
+ "page_idx": 4
906
+ },
907
+ {
908
+ "type": "image",
909
+ "img_path": "images/f91fb7f5b893bb47da572b0df940fc0b094b10279f430799e1478a8e794435e8.jpg",
910
+ "image_caption": [
911
+ "residuals - $\\epsilon (\\mathbf{r})$"
912
+ ],
913
+ "image_footnote": [],
914
+ "bbox": [
915
+ 112,
916
+ 171,
917
+ 297,
918
+ 196
919
+ ],
920
+ "page_idx": 4
921
+ },
922
+ {
923
+ "type": "image",
924
+ "img_path": "images/d9a7c3c9248e6b0d2e49b80e508445785b7faf30297405fd92a356c86778a670.jpg",
925
+ "image_caption": [],
926
+ "image_footnote": [],
927
+ "bbox": [
928
+ 297,
929
+ 89,
930
+ 482,
931
+ 170
932
+ ],
933
+ "page_idx": 4
934
+ },
935
+ {
936
+ "type": "image",
937
+ "img_path": "images/873f6abe2d818eeb1542748a04be93455f2e67da076891565998d6dc7678c5f5.jpg",
938
+ "image_caption": [
939
+ "inliers - $\\tilde{\\omega} (\\mathbf{r})$",
940
+ "Figure 5. Algorithm - We visualize our weight function computed by residuals on two examples: (top) the residuals of a (mid-training) NeRF rendered from a training viewpoint, (bottom) a toy residual image containing residual of small spatial extent (dot, line) and residuals of large spatial extent (squares). Notice residuals with large magnitude but small spatial extent (texture of the box, dot, line) are included in the optimization, while weaker residuals with larger spatial extent are excluded. Note that while we operate on patches, we visualize the weight function on the whole image to facilitate visualization."
941
+ ],
942
+ "image_footnote": [],
943
+ "bbox": [
944
+ 297,
945
+ 171,
946
+ 483,
947
+ 196
948
+ ],
949
+ "page_idx": 4
950
+ },
951
+ {
952
+ "type": "image",
953
+ "img_path": "images/d39ccfe40e70ee86c4959d6925556f5b4f52cee785752e345afb7026b573e088.jpg",
954
+ "image_caption": [],
955
+ "image_footnote": [],
956
+ "bbox": [
957
+ 482,
958
+ 89,
959
+ 668,
960
+ 170
961
+ ],
962
+ "page_idx": 4
963
+ },
964
+ {
965
+ "type": "image",
966
+ "img_path": "images/3a9ee80c514580cfc6f5f57876a041719846aed5e201e272d35434e954f86f15.jpg",
967
+ "image_caption": [
968
+ "diffusion - $\\tilde{\\omega}(\\mathbf{r}) \\oplus \\mathcal{B}_{3 \\times 3}$"
969
+ ],
970
+ "image_footnote": [],
971
+ "bbox": [
972
+ 482,
973
+ 171,
974
+ 668,
975
+ 196
976
+ ],
977
+ "page_idx": 4
978
+ },
979
+ {
980
+ "type": "image",
981
+ "img_path": "images/09be3306f8248a3d9f4bd6c55d6918e93896a53fa487fe67237b06357342877a.jpg",
982
+ "image_caption": [],
983
+ "image_footnote": [],
984
+ "bbox": [
985
+ 668,
986
+ 89,
987
+ 857,
988
+ 170
989
+ ],
990
+ "page_idx": 4
991
+ },
992
+ {
993
+ "type": "image",
994
+ "img_path": "images/a3df0d1d6602f3b8faab9d6badefac0d455997cc96da692275f316984e52437b.jpg",
995
+ "image_caption": [
996
+ "(IRLS) weights - $\\mathcal{W}(\\mathbf{r})$"
997
+ ],
998
+ "image_footnote": [],
999
+ "bbox": [
1000
+ 669,
1001
+ 171,
1002
+ 857,
1003
+ 196
1004
+ ],
1005
+ "page_idx": 4
1006
+ },
1007
+ {
1008
+ "type": "text",
1009
+ "text": "outliers can be learned quickly. It is also important to capture the structured nature of typical outliers, contrary to the typical i.i.d. assumption in most robust estimator formulations. To this end, the weight function should capture spatial smoothness of the outlier process, recognizing that objects typically have continuous local support, and hence outliers are expected to occupy large, connected regions of an image (e.g., the silhouette of a person to be segmented out from a photo-tourism dataset).",
1010
+ "bbox": [
1011
+ 75,
1012
+ 300,
1013
+ 468,
1014
+ 435
1015
+ ],
1016
+ "page_idx": 4
1017
+ },
1018
+ {
1019
+ "type": "text",
1020
+ "text": "Surprisingly, a relatively simple weight function embodies these properties and performs extremely well in practice. The weight function is based on so-called trimmed estimators that are used in trimmed least-squares, like that used in trimmed ICP [7]. We first sort residuals, and assume that residuals below a certain percentile are inliers. Picking the $50\\%$ percentile for convenience (i.e., median), we define",
1021
+ "bbox": [
1022
+ 75,
1023
+ 436,
1024
+ 468,
1025
+ 542
1026
+ ],
1027
+ "page_idx": 4
1028
+ },
1029
+ {
1030
+ "type": "equation",
1031
+ "text": "\n$$\n\\tilde {\\omega} (\\mathbf {r}) = \\epsilon (\\mathbf {r}) \\leq \\mathcal {T} _ {\\epsilon}, \\quad \\mathcal {T} _ {\\epsilon} = \\operatorname {M e d i a n} _ {\\mathbf {r}} \\left\\{\\epsilon (\\mathbf {r}) \\right\\}. \\tag {8}\n$$\n",
1032
+ "text_format": "latex",
1033
+ "bbox": [
1034
+ 125,
1035
+ 549,
1036
+ 468,
1037
+ 566
1038
+ ],
1039
+ "page_idx": 4
1040
+ },
1041
+ {
1042
+ "type": "text",
1043
+ "text": "To capture spatial smoothness of outliers we spatially diffuse inlier/outlier labels $\\omega$ with a $3\\times 3$ box kernel $\\mathcal{B}_{3\\times 3}$ . Formally, we define",
1044
+ "bbox": [
1045
+ 76,
1046
+ 571,
1047
+ 468,
1048
+ 616
1049
+ ],
1050
+ "page_idx": 4
1051
+ },
1052
+ {
1053
+ "type": "equation",
1054
+ "text": "\n$$\n\\mathcal {W} (\\mathbf {r}) = \\left(\\tilde {\\omega} (\\mathbf {r}) * \\mathcal {B} _ {3 \\times 3}\\right) \\geq \\mathcal {T} _ {*}, \\quad \\mathcal {T} _ {*} = 0. 5. \\tag {9}\n$$\n",
1055
+ "text_format": "latex",
1056
+ "bbox": [
1057
+ 122,
1058
+ 625,
1059
+ 468,
1060
+ 641
1061
+ ],
1062
+ "page_idx": 4
1063
+ },
1064
+ {
1065
+ "type": "text",
1066
+ "text": "This helps to avoid classifying high-frequency details as outliers, allowing them to be captured by the NeRF model during optimization (see Figure 5).",
1067
+ "bbox": [
1068
+ 75,
1069
+ 647,
1070
+ 468,
1071
+ 691
1072
+ ],
1073
+ "page_idx": 4
1074
+ },
1075
+ {
1076
+ "type": "text",
1077
+ "text": "While the trimmed weight function (9) improves the robustness of model fitting, it sometimes misclassifies fine-grained image details early in training where the NeRF model first captures coarse-grained structure. These localized texture elements may emerge but only after very long training times. We find that stronger inductive bias to spatially coherence allows fine-grained details to be learned more quickly. To that end, we aggregate the detection of outliers on $16 \\times 16$ neighborhoods; i.e., we label entire $8 \\times 8$ patches as outliers or inliers based on the behavior of $\\mathcal{W}$ in the $16 \\times 16$ neighborhood of the patch. Denoting the $N \\times N$ neighborhood of pixels around $\\mathbf{r}$ as $\\mathcal{R}_N(\\mathbf{r})$ , we define",
1078
+ "bbox": [
1079
+ 75,
1080
+ 691,
1081
+ 468,
1082
+ 873
1083
+ ],
1084
+ "page_idx": 4
1085
+ },
1086
+ {
1087
+ "type": "equation",
1088
+ "text": "\n$$\n\\omega \\left(\\mathcal {R} _ {8} (\\mathbf {r})\\right) = \\mathbb {E} _ {\\mathbf {s} \\sim \\mathcal {R} _ {1 6} (\\mathbf {r})} \\left[ \\mathcal {W} (\\mathbf {s}) \\right] \\geq \\mathcal {T} _ {\\mathcal {R}}, \\quad \\mathcal {T} _ {\\mathcal {R}} = 0. 6. \\tag {10}\n$$\n",
1089
+ "text_format": "latex",
1090
+ "bbox": [
1091
+ 86,
1092
+ 880,
1093
+ 468,
1094
+ 897
1095
+ ],
1096
+ "page_idx": 4
1097
+ },
1098
+ {
1099
+ "type": "image",
1100
+ "img_path": "images/bf26a6902d7a41ed48389fa1f5b89d4469c7a3ae868ee3ba2d6b6dd6db473a18.jpg",
1101
+ "image_caption": [
1102
+ "Ground Truth"
1103
+ ],
1104
+ "image_footnote": [],
1105
+ "bbox": [
1106
+ 504,
1107
+ 300,
1108
+ 586,
1109
+ 363
1110
+ ],
1111
+ "page_idx": 4
1112
+ },
1113
+ {
1114
+ "type": "image",
1115
+ "img_path": "images/4416c7398ccd94ebc9d367fa80dd34745d1f0b22821109b958a335b77eb5ef87.jpg",
1116
+ "image_caption": [
1117
+ "RobustNeRF"
1118
+ ],
1119
+ "image_footnote": [],
1120
+ "bbox": [
1121
+ 604,
1122
+ 300,
1123
+ 686,
1124
+ 362
1125
+ ],
1126
+ "page_idx": 4
1127
+ },
1128
+ {
1129
+ "type": "image",
1130
+ "img_path": "images/0b6c7d1202a0081f52befd0f8113b7eaa05410296dc88312c5fd84db8c93517f.jpg",
1131
+ "image_caption": [
1132
+ "Train View"
1133
+ ],
1134
+ "image_footnote": [],
1135
+ "bbox": [
1136
+ 705,
1137
+ 300,
1138
+ 787,
1139
+ 362
1140
+ ],
1141
+ "page_idx": 4
1142
+ },
1143
+ {
1144
+ "type": "image",
1145
+ "img_path": "images/3a02c34a52cb7d88bc69b2089f4a9245d1437487a0c9d1e3882be254e4ca87fc.jpg",
1146
+ "image_caption": [
1147
+ "Distractors"
1148
+ ],
1149
+ "image_footnote": [],
1150
+ "bbox": [
1151
+ 808,
1152
+ 300,
1153
+ 888,
1154
+ 362
1155
+ ],
1156
+ "page_idx": 4
1157
+ },
1158
+ {
1159
+ "type": "image",
1160
+ "img_path": "images/61d593844bc4764698a50b4773d9444124bb85159f02cb5638f6aa39ecafd740.jpg",
1161
+ "image_caption": [
1162
+ "Figure 6. Residuals - For the dataset shown in the top row, we visualize the dynamics of the RobustNeRF training residuals, which show how over time the estimated distractor weights go from being random $(t / T) = 0.5\\%)$ to identify distractor pixels $(t / T) = 100\\%$ without any explicit supervision."
1163
+ ],
1164
+ "image_footnote": [],
1165
+ "bbox": [
1166
+ 506,
1167
+ 380,
1168
+ 887,
1169
+ 518
1170
+ ],
1171
+ "page_idx": 4
1172
+ },
1173
+ {
1174
+ "type": "text",
1175
+ "text": "The final weight function is the union of the three masks in Eqns. 8 -10. This robust weight function evolves during optimization, as one expects with IRLS where the weights are a function of the residuals at the previous iteration. That is, the labeling of pixels as inliers/outliers changes during training, and settles around masks similar to the one an oracle would provide as training converges (see Figure 6).",
1176
+ "bbox": [
1177
+ 496,
1178
+ 609,
1179
+ 890,
1180
+ 715
1181
+ ],
1182
+ "page_idx": 4
1183
+ },
1184
+ {
1185
+ "type": "text",
1186
+ "text": "4. Experiments",
1187
+ "text_level": 1,
1188
+ "bbox": [
1189
+ 500,
1190
+ 727,
1191
+ 632,
1192
+ 744
1193
+ ],
1194
+ "page_idx": 4
1195
+ },
1196
+ {
1197
+ "type": "text",
1198
+ "text": "We implement our robust loss function in the MultiNeRF codebase [26] and apply it to mip-NeRF 360 [4]. We dub this method \"RobustNeRF\". To evaluate RobustNeRF, we compare against baselines on several scenes containing different types of distractors. Where possible, we quantitatively compare reconstructions to held-out, distraction-free images; we report three metrics, averaged across held-out frames, namely, PSNR, SSIM [52], and LPIPS [60].",
1199
+ "bbox": [
1200
+ 496,
1201
+ 750,
1202
+ 890,
1203
+ 869
1204
+ ],
1205
+ "page_idx": 4
1206
+ },
1207
+ {
1208
+ "type": "text",
1209
+ "text": "We compare different methods on two collections of scenes, i.e., those provided by the authors of $\\mathrm{D}^2\\mathrm{NeRF}$ , and",
1210
+ "bbox": [
1211
+ 500,
1212
+ 869,
1213
+ 890,
1214
+ 900
1215
+ ],
1216
+ "page_idx": 4
1217
+ },
1218
+ {
1219
+ "type": "page_number",
1220
+ "text": "20630",
1221
+ "bbox": [
1222
+ 478,
1223
+ 944,
1224
+ 517,
1225
+ 955
1226
+ ],
1227
+ "page_idx": 4
1228
+ },
1229
+ {
1230
+ "type": "image",
1231
+ "img_path": "images/fd0b180c1026e2be7ccdee830cae294216bd8c940336f0220309076f21627f21.jpg",
1232
+ "image_caption": [
1233
+ "··"
1234
+ ],
1235
+ "image_footnote": [],
1236
+ "bbox": [
1237
+ 83,
1238
+ 90,
1239
+ 176,
1240
+ 202
1241
+ ],
1242
+ "page_idx": 5
1243
+ },
1244
+ {
1245
+ "type": "image",
1246
+ "img_path": "images/22cb10d1fc054634cd88f8e4e16a0ab3bf0cb02ffedc48c416b29e5cde255850.jpg",
1247
+ "image_caption": [
1248
+ "··"
1249
+ ],
1250
+ "image_footnote": [],
1251
+ "bbox": [
1252
+ 179,
1253
+ 90,
1254
+ 272,
1255
+ 200
1256
+ ],
1257
+ "page_idx": 5
1258
+ },
1259
+ {
1260
+ "type": "image",
1261
+ "img_path": "images/7a8807948cc6eeba13b5d92f78a16697ee949ec41f3a5424f1ade3c127bec8b5.jpg",
1262
+ "image_caption": [
1263
+ "···"
1264
+ ],
1265
+ "image_footnote": [],
1266
+ "bbox": [
1267
+ 274,
1268
+ 90,
1269
+ 367,
1270
+ 200
1271
+ ],
1272
+ "page_idx": 5
1273
+ },
1274
+ {
1275
+ "type": "image",
1276
+ "img_path": "images/c1697cdb8d34afb169bc52ec894798f33b8f724db8ce82cf41b82abed8e3d25a.jpg",
1277
+ "image_caption": [
1278
+ "···"
1279
+ ],
1280
+ "image_footnote": [],
1281
+ "bbox": [
1282
+ 369,
1283
+ 90,
1284
+ 462,
1285
+ 200
1286
+ ],
1287
+ "page_idx": 5
1288
+ },
1289
+ {
1290
+ "type": "image",
1291
+ "img_path": "images/762874ad94fc2fd98f1240146b07723da08dad632c699a0932c7aa82b5de6a09.jpg",
1292
+ "image_caption": [
1293
+ "Statue"
1294
+ ],
1295
+ "image_footnote": [],
1296
+ "bbox": [
1297
+ 83,
1298
+ 215,
1299
+ 176,
1300
+ 268
1301
+ ],
1302
+ "page_idx": 5
1303
+ },
1304
+ {
1305
+ "type": "image",
1306
+ "img_path": "images/dbcea64e120747e1ed4cea2e373652528276341671ecec1e4f4912acb22390c0.jpg",
1307
+ "image_caption": [
1308
+ "Android"
1309
+ ],
1310
+ "image_footnote": [],
1311
+ "bbox": [
1312
+ 179,
1313
+ 215,
1314
+ 272,
1315
+ 268
1316
+ ],
1317
+ "page_idx": 5
1318
+ },
1319
+ {
1320
+ "type": "image",
1321
+ "img_path": "images/66cc33a619f3416b49a8d2b6c643ed33b59e00a0f8f22684e90cfb6ed0e62fa7.jpg",
1322
+ "image_caption": [
1323
+ "Crab",
1324
+ "Figure 7. Dataset - Sample training images showing the distractors in each scene. Statue and Android were acquired manually, and the others with a robotic arm. In the robotic setting we have pixel-perfect alignment of distractor vs. distractor-free images."
1325
+ ],
1326
+ "image_footnote": [],
1327
+ "bbox": [
1328
+ 274,
1329
+ 215,
1330
+ 367,
1331
+ 268
1332
+ ],
1333
+ "page_idx": 5
1334
+ },
1335
+ {
1336
+ "type": "image",
1337
+ "img_path": "images/d4665e589c43ddd6ae0a880b9b65793acf9e80ac3bc7d18a3a2c1524afd65028.jpg",
1338
+ "image_caption": [
1339
+ "BabyYoda"
1340
+ ],
1341
+ "image_footnote": [],
1342
+ "bbox": [
1343
+ 369,
1344
+ 215,
1345
+ 462,
1346
+ 268
1347
+ ],
1348
+ "page_idx": 5
1349
+ },
1350
+ {
1351
+ "type": "text",
1352
+ "text": "novel datasets described below. We also present a series of illustrative experiments on synthetic scenes, shedding light on RobustNeRF's efficacy and inner workings.",
1353
+ "bbox": [
1354
+ 75,
1355
+ 359,
1356
+ 468,
1357
+ 406
1358
+ ],
1359
+ "page_idx": 5
1360
+ },
1361
+ {
1362
+ "type": "text",
1363
+ "text": "4.1. Baselines",
1364
+ "text_level": 1,
1365
+ "bbox": [
1366
+ 76,
1367
+ 417,
1368
+ 184,
1369
+ 431
1370
+ ],
1371
+ "page_idx": 5
1372
+ },
1373
+ {
1374
+ "type": "text",
1375
+ "text": "We compare RobustNeRF to variants of mip-NeRF 360 optimized with different loss functions $(L_{2}, L_{1}$ , and Carbonnier). These variants serve as natural baselines for models with limited or no robustness to outliers. We also compare to $\\mathrm{D}^2\\mathrm{NeRF}$ , a recent method for reconstructing dynamic scenes from monocular video. Unlike our method, $\\mathrm{D}^2\\mathrm{NeRF}$ is designed to reconstruct distractors rather than discard them. While $\\mathrm{D}^2\\mathrm{NeRF}$ is presented as a method for monocular video, it does not presuppose temporal continuity, and can be directly applied to unordered images. We omit additional comparisons to NeRF-W as its performance falls short of $\\mathrm{D}^2\\mathrm{NeRF}$ [53]. For more details on model training, see the supplementary material.",
1376
+ "bbox": [
1377
+ 75,
1378
+ 438,
1379
+ 468,
1380
+ 635
1381
+ ],
1382
+ "page_idx": 5
1383
+ },
1384
+ {
1385
+ "type": "text",
1386
+ "text": "4.2. Datasets - Figure 7",
1387
+ "text_level": 1,
1388
+ "bbox": [
1389
+ 76,
1390
+ 648,
1391
+ 261,
1392
+ 664
1393
+ ],
1394
+ "page_idx": 5
1395
+ },
1396
+ {
1397
+ "type": "text",
1398
+ "text": "In addition to scenes from $\\mathrm{D}^2\\mathrm{NeRF}$ , we introduce a set of natural and synthetic scenes. They facilitate the evaluation of RobustNeRF's effectiveness on illustrative use cases, and they enable empirical analysis under controlled conditions.",
1399
+ "bbox": [
1400
+ 75,
1401
+ 667,
1402
+ 468,
1403
+ 729
1404
+ ],
1405
+ "page_idx": 5
1406
+ },
1407
+ {
1408
+ "type": "text",
1409
+ "text": "Natural scenes. We capture seven natural scenes exemplifying different types of distractors. Scenes are captured in three settings, on the street, in an apartment and in a robotics lab. Distractor objects are moved, or are allowed to move, between frames to simulate capture over extended periods of time. We vary the number of unique distractors from 1 (Statue) to 150 (BabyYoda), and their movements. Unlike prior work on monocular video, frames are captured without a clear temporal ordering (see Figure 7). The other three (i.e., Street1, Street2, and Gloss) include view-dependence effects, the results of which are shown in the supplementary",
1410
+ "bbox": [
1411
+ 75,
1412
+ 734,
1413
+ 467,
1414
+ 900
1415
+ ],
1416
+ "page_idx": 5
1417
+ },
1418
+ {
1419
+ "type": "text",
1420
+ "text": "material. We also capture additional frames without distractors to enable quantitative evaluations. Camera poses are estimated using COLMAP [38]. A full description of each scene in the supplementary material.",
1421
+ "bbox": [
1422
+ 496,
1423
+ 90,
1424
+ 890,
1425
+ 151
1426
+ ],
1427
+ "page_idx": 5
1428
+ },
1429
+ {
1430
+ "type": "text",
1431
+ "text": "Synthetic scenes. To further evaluate RobustNeRF, we generate synthetic scenes using the Kubric dataset generator [12]. Each scene is constructed by placing a set of simple geometries in an empty, texture-less room. In each scene, a subset of objects remain fixed while the other objects (i.e., distractors) change position from frame to frame. By varying the number of objects, their size, and the way they move, we control the level of distraction in each scene. We use these scenes to examine RobustNeRF's sensitivity to its hyperparameters, see supplementary material.",
1432
+ "bbox": [
1433
+ 496,
1434
+ 154,
1435
+ 890,
1436
+ 305
1437
+ ],
1438
+ "page_idx": 5
1439
+ },
1440
+ {
1441
+ "type": "text",
1442
+ "text": "4.3. Evaluation",
1443
+ "text_level": 1,
1444
+ "bbox": [
1445
+ 500,
1446
+ 314,
1447
+ 619,
1448
+ 329
1449
+ ],
1450
+ "page_idx": 5
1451
+ },
1452
+ {
1453
+ "type": "text",
1454
+ "text": "We evaluate RobustNeRF on its ability to ignore distractors while accurately reconstructing the static elements of a scene. We train RobustNeRF, $\\mathrm{D}^2\\mathrm{NeRF}$ , and variants of mip-NeRF 360 on scenes where distraction-free frames are available. Models are trained on frames with distractors and evaluated on distractor-free frames.",
1455
+ "bbox": [
1456
+ 496,
1457
+ 334,
1458
+ 890,
1459
+ 422
1460
+ ],
1461
+ "page_idx": 5
1462
+ },
1463
+ {
1464
+ "type": "text",
1465
+ "text": "Comparison to mip-NeRF 360 - Figure 8. On natural scenes, RobustNeRF generally outperforms variants of mip-NeRF 360 by 1.3 to $4.7\\mathrm{dB}$ in PSNR. As $L_{2}, L_{1}$ , and Charbonnier losses weigh all pixels equally, the model is forced to represent, rather than ignore, distractors as \"clouds\" with view-dependent appearance. We find clouds to be most apparent when distractors remain stationary for multiple frames. In contrast, RobustNeRF's loss isolates distractor pixels and assigns them a weight of zero (see Figure 6). To establish an upper bound on reconstruction accuracy, we train mip-NeRF 360 with Charbonnier loss on distraction-free versions of each scene, the images for which are taken from (approximately) the same viewpoints. Reassuringly, RobustNeRF when trained on distraction-free frames, achieves nearly identical accuracy; see Figure 11.",
1466
+ "bbox": [
1467
+ 496,
1468
+ 428,
1469
+ 890,
1470
+ 654
1471
+ ],
1472
+ "page_idx": 5
1473
+ },
1474
+ {
1475
+ "type": "text",
1476
+ "text": "While RobustNeRF consistently outperforms mip-NeRF 360, the gap is smaller in the Apartment scenes (Statue, Android) than the Robotics Lab scenes (Crab, BabyYoda). This can be explained by challenging background geometry, errors in camera parameter estimation, and imperceptible changes to scene appearance. For further discussion, see the supplementary material.",
1477
+ "bbox": [
1478
+ 496,
1479
+ 654,
1480
+ 890,
1481
+ 760
1482
+ ],
1483
+ "page_idx": 5
1484
+ },
1485
+ {
1486
+ "type": "text",
1487
+ "text": "Comparison to $\\mathbf{D}^2\\mathrm{NeRF}$ - Figure 9. Quantitatively, RobustNeRF matches or outperforms $\\mathrm{D}^2\\mathrm{NeRF}$ by as much as $12\\mathrm{dB}$ PSNR depending on the number of unique outlier objects in the capture. Results on $\\mathrm{D}^2\\mathrm{NeRF}$ real scenes are provided in the supplementary material for qualitative comparison. In Statue and Android, 1 and 3 non-rigid objects are moved around the scene, respectively. $\\mathrm{D}^2\\mathrm{NeRF}$ is able to model these objects and thus separate them from the scenes' static content. In the remaining scenes, a much",
1488
+ "bbox": [
1489
+ 496,
1490
+ 763,
1491
+ 890,
1492
+ 900
1493
+ ],
1494
+ "page_idx": 5
1495
+ },
1496
+ {
1497
+ "type": "page_number",
1498
+ "text": "20631",
1499
+ "bbox": [
1500
+ 478,
1501
+ 944,
1502
+ 517,
1503
+ 955
1504
+ ],
1505
+ "page_idx": 5
1506
+ },
1507
+ {
1508
+ "type": "table",
1509
+ "img_path": "images/9c0d942be36ab8e476ae45e835cd0345cdf7b4d7694878a0c1c937a0c4711672.jpg",
1510
+ "table_caption": [],
1511
+ "table_footnote": [],
1512
+ "table_body": "<table><tr><td rowspan=\"2\"></td><td colspan=\"3\">Statue</td><td colspan=\"3\">Android</td><td colspan=\"3\">Crab</td><td colspan=\"3\">BabyYoda</td></tr><tr><td>LPIPS↓</td><td>SSIM↑</td><td>PSNR↑</td><td>LPIPS↓</td><td>SSIM↑</td><td>PSNR↑</td><td>LPIPS↓</td><td>SSIM↑</td><td>PSNR↑</td><td>LPIPS↓</td><td>SSIM↑</td><td>PSNR↑</td></tr><tr><td>mip-NeRF 360 (L2)</td><td>0.36</td><td>0.66</td><td>19.09</td><td>0.40</td><td>0.65</td><td>19.35</td><td>0.27</td><td>0.77</td><td>25.73</td><td>0.31</td><td>0.75</td><td>22.97</td></tr><tr><td>mip-NeRF 360 (L1)</td><td>0.30</td><td>0.72</td><td>19.55</td><td>0.40</td><td>0.66</td><td>19.38</td><td>0.22</td><td>0.79</td><td>26.69</td><td>0.22</td><td>0.80</td><td>26.15</td></tr><tr><td>mip-NeRF 360 (Ch.)</td><td>0.30</td><td>0.73</td><td>19.64</td><td>0.40</td><td>0.66</td><td>19.53</td><td>0.21</td><td>0.80</td><td>27.72</td><td>0.23</td><td>0.80</td><td>25.22</td></tr><tr><td>D2NeRF</td><td>0.48</td><td>0.49</td><td>19.09</td><td>0.43</td><td>0.57</td><td>20.61</td><td>0.42</td><td>0.68</td><td>21.18</td><td>0.44</td><td>0.65</td><td>17.32</td></tr><tr><td>RobustNeRF</td><td>0.28</td><td>0.75</td><td>20.89</td><td>0.31</td><td>0.65</td><td>21.72</td><td>0.21</td><td>0.81</td><td>30.75</td><td>0.20</td><td>0.83</td><td>30.87</td></tr><tr><td>mip-NeRF 360 (clean)</td><td>0.19</td><td>0.80</td><td>23.57</td><td>0.31</td><td>0.71</td><td>23.10</td><td>0.16</td><td>0.84</td><td>32.55</td><td>0.16</td><td>0.84</td><td>32.63</td></tr></table>",
1513
+ "bbox": [
1514
+ 78,
1515
+ 85,
1516
+ 890,
1517
+ 220
1518
+ ],
1519
+ "page_idx": 6
1520
+ },
1521
+ {
1522
+ "type": "image",
1523
+ "img_path": "images/2472bef537acf5ecbf70fd1ccd4e8fdd4925aac7be2d02f5078b3cadb5f67ca5.jpg",
1524
+ "image_caption": [
1525
+ "MipNeRF360 (L2)"
1526
+ ],
1527
+ "image_footnote": [],
1528
+ "bbox": [
1529
+ 81,
1530
+ 239,
1531
+ 238,
1532
+ 332
1533
+ ],
1534
+ "page_idx": 6
1535
+ },
1536
+ {
1537
+ "type": "image",
1538
+ "img_path": "images/af7d9acea775d2155ed10237b7999d99195bb134366ad25a6c1e360d937f9f42.jpg",
1539
+ "image_caption": [
1540
+ "MipNeRF360 (L1)"
1541
+ ],
1542
+ "image_footnote": [],
1543
+ "bbox": [
1544
+ 243,
1545
+ 239,
1546
+ 401,
1547
+ 332
1548
+ ],
1549
+ "page_idx": 6
1550
+ },
1551
+ {
1552
+ "type": "image",
1553
+ "img_path": "images/13a1dc7479797324164b7080843547f38b6e2fb62ce563031af7a75ba9138338.jpg",
1554
+ "image_caption": [
1555
+ "MipNeRF360 (Ch)"
1556
+ ],
1557
+ "image_footnote": [],
1558
+ "bbox": [
1559
+ 406,
1560
+ 239,
1561
+ 563,
1562
+ 332
1563
+ ],
1564
+ "page_idx": 6
1565
+ },
1566
+ {
1567
+ "type": "image",
1568
+ "img_path": "images/714501e76ffdf113f080c6f7b7789dd195653e9a529af0ee71a92d27ba2a2214.jpg",
1569
+ "image_caption": [
1570
+ "$\\mathrm{D}^2\\mathrm{NeRF}$"
1571
+ ],
1572
+ "image_footnote": [],
1573
+ "bbox": [
1574
+ 566,
1575
+ 239,
1576
+ 725,
1577
+ 332
1578
+ ],
1579
+ "page_idx": 6
1580
+ },
1581
+ {
1582
+ "type": "image",
1583
+ "img_path": "images/3ddc66a93c631c646317fe0aa2640bfdf35c5cc4d28f9c65636ca20f691b07c4.jpg",
1584
+ "image_caption": [
1585
+ "RobustNeRF"
1586
+ ],
1587
+ "image_footnote": [],
1588
+ "bbox": [
1589
+ 730,
1590
+ 239,
1591
+ 888,
1592
+ 332
1593
+ ],
1594
+ "page_idx": 6
1595
+ },
1596
+ {
1597
+ "type": "image",
1598
+ "img_path": "images/7fd91bf56dc9e30d6dcc2265a52d415b47fdfc66e4518b1befa598acb895655d.jpg",
1599
+ "image_caption": [],
1600
+ "image_footnote": [],
1601
+ "bbox": [
1602
+ 83,
1603
+ 335,
1604
+ 238,
1605
+ 428
1606
+ ],
1607
+ "page_idx": 6
1608
+ },
1609
+ {
1610
+ "type": "image",
1611
+ "img_path": "images/0647871967bd735d4b498a761950e451bffe8a160ccfff0372e0513819fbc28f.jpg",
1612
+ "image_caption": [],
1613
+ "image_footnote": [],
1614
+ "bbox": [
1615
+ 243,
1616
+ 335,
1617
+ 401,
1618
+ 428
1619
+ ],
1620
+ "page_idx": 6
1621
+ },
1622
+ {
1623
+ "type": "image",
1624
+ "img_path": "images/b658f6f97a6edd10f1d7a6cdd644fdd1ce388b732f1914e6aaeda8051c38167f.jpg",
1625
+ "image_caption": [],
1626
+ "image_footnote": [],
1627
+ "bbox": [
1628
+ 406,
1629
+ 335,
1630
+ 563,
1631
+ 428
1632
+ ],
1633
+ "page_idx": 6
1634
+ },
1635
+ {
1636
+ "type": "image",
1637
+ "img_path": "images/a779e3c1ee430366f17af15c093cba5a31cacd675bc5a43f4a67457a2f945046.jpg",
1638
+ "image_caption": [],
1639
+ "image_footnote": [],
1640
+ "bbox": [
1641
+ 568,
1642
+ 335,
1643
+ 725,
1644
+ 428
1645
+ ],
1646
+ "page_idx": 6
1647
+ },
1648
+ {
1649
+ "type": "image",
1650
+ "img_path": "images/5d6a9d369db5377343f4d05f840cff309f2bcdbfec3a29c11b9f81bb5d6e35dc.jpg",
1651
+ "image_caption": [],
1652
+ "image_footnote": [],
1653
+ "bbox": [
1654
+ 730,
1655
+ 335,
1656
+ 888,
1657
+ 428
1658
+ ],
1659
+ "page_idx": 6
1660
+ },
1661
+ {
1662
+ "type": "image",
1663
+ "img_path": "images/908fbde1d517ea610c37fcfc2d9fa26961946c3ac07c40f3516176ce479013fe.jpg",
1664
+ "image_caption": [],
1665
+ "image_footnote": [],
1666
+ "bbox": [
1667
+ 83,
1668
+ 431,
1669
+ 238,
1670
+ 523
1671
+ ],
1672
+ "page_idx": 6
1673
+ },
1674
+ {
1675
+ "type": "image",
1676
+ "img_path": "images/941d667498108edd1f131e7b2e9577b587c75b45666ebec28ffe5295b8276fac.jpg",
1677
+ "image_caption": [],
1678
+ "image_footnote": [],
1679
+ "bbox": [
1680
+ 243,
1681
+ 431,
1682
+ 401,
1683
+ 523
1684
+ ],
1685
+ "page_idx": 6
1686
+ },
1687
+ {
1688
+ "type": "image",
1689
+ "img_path": "images/7cdaf75c778f195b955d2151692a4b2e4dcfd3b2cccad382ba65da930cc8d33b.jpg",
1690
+ "image_caption": [],
1691
+ "image_footnote": [],
1692
+ "bbox": [
1693
+ 406,
1694
+ 431,
1695
+ 563,
1696
+ 523
1697
+ ],
1698
+ "page_idx": 6
1699
+ },
1700
+ {
1701
+ "type": "image",
1702
+ "img_path": "images/0c66e136417c9ec76a7ce9cd70ea66232e91904c114038bd3185ee2df4289495.jpg",
1703
+ "image_caption": [],
1704
+ "image_footnote": [],
1705
+ "bbox": [
1706
+ 568,
1707
+ 431,
1708
+ 725,
1709
+ 523
1710
+ ],
1711
+ "page_idx": 6
1712
+ },
1713
+ {
1714
+ "type": "image",
1715
+ "img_path": "images/e32af9c9ac8a43c2649ac21e09d064db9b60e09650a3536d662beeace1b62791.jpg",
1716
+ "image_caption": [],
1717
+ "image_footnote": [],
1718
+ "bbox": [
1719
+ 730,
1720
+ 431,
1721
+ 888,
1722
+ 523
1723
+ ],
1724
+ "page_idx": 6
1725
+ },
1726
+ {
1727
+ "type": "image",
1728
+ "img_path": "images/84ef9a524a3021812a8c9f681d2f616deddd59904b35a59c6139c3e949778e33.jpg",
1729
+ "image_caption": [
1730
+ "Figure 8. Evaluation on Natural Scenes - RobustNeRF outperforms baselines and $\\mathrm{D}^2\\mathrm{NeRF}$ [53] on novel view synthesis with real-world captures. The table provides a quantitative comparison of RobustNeRF, $\\mathrm{D}^2\\mathrm{NeRF}$ and mip-NeRF 360 using different reconstruction losses. The last row reports mip-NeRF 360 trained on a distractor-free version of each dataset, giving an upperbound for RobustNeRF performance. We also visualize samples from each scene rendered with each of the methods. See Supplementary Material for more samples."
1731
+ ],
1732
+ "image_footnote": [],
1733
+ "bbox": [
1734
+ 83,
1735
+ 527,
1736
+ 238,
1737
+ 619
1738
+ ],
1739
+ "page_idx": 6
1740
+ },
1741
+ {
1742
+ "type": "image",
1743
+ "img_path": "images/fbff17edcd0c79df32fd429ad15ec1ef06b80d4dca1fab87e2456bd3826010ad.jpg",
1744
+ "image_caption": [],
1745
+ "image_footnote": [],
1746
+ "bbox": [
1747
+ 243,
1748
+ 527,
1749
+ 401,
1750
+ 619
1751
+ ],
1752
+ "page_idx": 6
1753
+ },
1754
+ {
1755
+ "type": "image",
1756
+ "img_path": "images/e9215e4e4f884ac6d10f101a9563ca8e96bf72151a128a1751def30b93e22c45.jpg",
1757
+ "image_caption": [],
1758
+ "image_footnote": [],
1759
+ "bbox": [
1760
+ 406,
1761
+ 527,
1762
+ 563,
1763
+ 619
1764
+ ],
1765
+ "page_idx": 6
1766
+ },
1767
+ {
1768
+ "type": "image",
1769
+ "img_path": "images/b2ab403dd8bd35e27f24f173505838e737fc32feaeaff8ed4698dd5bbd5f6794.jpg",
1770
+ "image_caption": [],
1771
+ "image_footnote": [],
1772
+ "bbox": [
1773
+ 568,
1774
+ 527,
1775
+ 725,
1776
+ 619
1777
+ ],
1778
+ "page_idx": 6
1779
+ },
1780
+ {
1781
+ "type": "image",
1782
+ "img_path": "images/06c7d9010e371fb2c3acf11e25eeec78573aa9900a82f80795f0ec317f4a5b25.jpg",
1783
+ "image_caption": [],
1784
+ "image_footnote": [],
1785
+ "bbox": [
1786
+ 730,
1787
+ 527,
1788
+ 888,
1789
+ 619
1790
+ ],
1791
+ "page_idx": 6
1792
+ },
1793
+ {
1794
+ "type": "text",
1795
+ "text": "larger pool of 100 to 150 unique, non-static objects are used – too many for $\\mathrm{D}^2\\mathrm{NeRF}$ to model effectively. As a result, “cloud” artifacts appear in its static representation, similar to those produced by mip-NeRF 360. In contrast, Robust-NeRF identifies non-static content as outliers and omits it during reconstruction. Although both methods use a similar number of parameters, $\\mathrm{D}^2\\mathrm{NeRF}$ 's peak memory usage is $2.3\\mathrm{x}$ higher than RobustNeRF and $37\\mathrm{x}$ higher when normalizing for batch size. This is a direct consequence of model architecture: $\\mathrm{D}^2\\mathrm{NeRF}$ is tailored to simultaneously modeling static and dynamic content and thus merits higher complexity. To remain comparable, we limit image resolution to 0.2 megapixels for all experiments.",
1796
+ "bbox": [
1797
+ 75,
1798
+ 699,
1799
+ 472,
1800
+ 896
1801
+ ],
1802
+ "page_idx": 6
1803
+ },
1804
+ {
1805
+ "type": "text",
1806
+ "text": "Ablations - Figure 10. We ablate elements of the Robust-NeRF loss on the crab scene, comparing to an upper bound on the reconstruction accuracy of mip-NeRF 360 trained on distractor-free (clean) images from identical viewpoints. Our trimmed estimator (8) successfully eliminates distractors at the expense of high frequency texture and a lower PSNR. With smoothing (9), fine details are recovered, at the cost of longer training times. With the spatial window (10), RobustNeRF training time is on-par with mip-NeRF 360. We also ablate patch size and the trimming threshold (see Supplementary Material); we find that RobustNeRF is insensitive to trimming threshold, and that reducing the patch size offsets the gains from smoothing and patching.",
1807
+ "bbox": [
1808
+ 496,
1809
+ 699,
1810
+ 890,
1811
+ 896
1812
+ ],
1813
+ "page_idx": 6
1814
+ },
1815
+ {
1816
+ "type": "page_number",
1817
+ "text": "20632",
1818
+ "bbox": [
1819
+ 478,
1820
+ 945,
1821
+ 517,
1822
+ 955
1823
+ ],
1824
+ "page_idx": 6
1825
+ },
1826
+ {
1827
+ "type": "table",
1828
+ "img_path": "images/81c326f99a2d9f37889109b6a36918676e3c948c687fa65f8e8763fe160cc0f8.jpg",
1829
+ "table_caption": [],
1830
+ "table_footnote": [],
1831
+ "table_body": "<table><tr><td rowspan=\"2\"></td><td colspan=\"3\">Car</td><td colspan=\"3\">Cars</td><td colspan=\"3\">Bag</td><td colspan=\"3\">Chairs</td><td colspan=\"3\">Pillow</td><td></td></tr><tr><td>LPIPS↓</td><td>MS-SSIM↑</td><td>PSNR↑</td><td>LPIPS↓</td><td>MS-SSIM↑</td><td>PSNR↑</td><td>LPIPS↓</td><td>MS-SSIM↑</td><td>PSNR↑</td><td>LPIPS↓</td><td>MS-SSIM↑</td><td>PSNR↑</td><td>LPIPS↓</td><td>MS-SSIM↑</td><td>PSNR↑</td><td></td></tr><tr><td>NeRF-W [23]</td><td>.218</td><td>.814</td><td>24.23</td><td>.243</td><td>.873</td><td>24.51</td><td>.139</td><td>.791</td><td>20.65</td><td>.150</td><td>.681</td><td>23.77</td><td>.088</td><td>.935</td><td>28.24</td><td></td></tr><tr><td>NSFF [18]</td><td>.200</td><td>.806</td><td>24.90</td><td>.620</td><td>.376</td><td>10.29</td><td>.108</td><td>.892</td><td>25.62</td><td>.682</td><td>.284</td><td>12.82</td><td>.782</td><td>.343</td><td>4.55</td><td></td></tr><tr><td>NeuralDiff [46]</td><td>.065</td><td>.952</td><td>31.89</td><td>.098</td><td>.921</td><td>25.93</td><td>.117</td><td>.910</td><td>29.02</td><td>.112</td><td>.722</td><td>24.42</td><td>.565</td><td>.652</td><td>20.09</td><td></td></tr><tr><td>D2NeRF [53]</td><td>.062</td><td>.975</td><td>34.27</td><td>.090</td><td>.953</td><td>26.27</td><td>.076</td><td>.979</td><td>34.14</td><td>.095</td><td>.707</td><td>24.63</td><td>.076</td><td>.979</td><td>36.58</td><td></td></tr><tr><td>RobustNeRF</td><td>.013</td><td>.988</td><td>37.73</td><td>.063</td><td>.957</td><td>26.31</td><td>.006</td><td>.995</td><td>41.82</td><td>.007</td><td>.992</td><td>41.23</td><td>.018</td><td>.990</td><td>38.95</td><td></td></tr><tr><td>Train View</td><td colspan=\"3\">Test View</td><td colspan=\"3\">NeRF-W</td><td colspan=\"3\">NSFF</td><td colspan=\"3\">NeuralDiff</td><td colspan=\"3\">D2NeRF</td><td>RobustNeRF</td></tr></table>",
1832
+ "bbox": [
1833
+ 81,
1834
+ 87,
1835
+ 890,
1836
+ 277
1837
+ ],
1838
+ "page_idx": 7
1839
+ },
1840
+ {
1841
+ "type": "table",
1842
+ "img_path": "images/31ab415c63a78f3acd67bea5122ed8b8332933561ca566bba81396917a5ac763.jpg",
1843
+ "table_caption": [
1844
+ "Figure 9. Evaluations on $\\mathbf{D}^2\\mathrm{NeRF}$ Synthetic Scenes - Quantitative and qualitative evaluations on the Kubric synthetic dataset introduced by $\\mathbf{D}^2\\mathrm{NeRF}$ , consisting of 200 training frames (with distractor) and 100 novel views for evaluation (without distractor)."
1845
+ ],
1846
+ "table_footnote": [],
1847
+ "table_body": "<table><tr><td></td><td>LPIPS↓</td><td>SSIM↑</td><td>PSNR↑</td><td>Updates to PSNR=30</td></tr><tr><td>mip-NeRF 360 (L2)</td><td>0.31</td><td>0.75</td><td>22.97</td><td>-</td></tr><tr><td>+ robust (8)</td><td>0.39</td><td>0.60</td><td>18.21</td><td>-</td></tr><tr><td>+ smoothing (9)</td><td>0.22</td><td>0.81</td><td>30.01</td><td>250K</td></tr><tr><td>+ patching (10)</td><td>0.21</td><td>0.81</td><td>30.75</td><td>70K</td></tr><tr><td>oracle (clean)</td><td>0.16</td><td>0.84</td><td>32.55</td><td>25K</td></tr><tr><td>+ robust</td><td colspan=\"3\">+ smoothing</td><td>+ patching</td></tr></table>",
1848
+ "bbox": [
1849
+ 80,
1850
+ 321,
1851
+ 467,
1852
+ 496
1853
+ ],
1854
+ "page_idx": 7
1855
+ },
1856
+ {
1857
+ "type": "text",
1858
+ "text": "Figure 10. Ablations - Blindly trimming the loss causes details to be lost. Smoothing recovers fine-grained detail, while patch-based evaluation speeds up training and adds more detail. Patching enables the model to reach PSNR of 30, almost $4 \\times$ faster.",
1859
+ "bbox": [
1860
+ 75,
1861
+ 497,
1862
+ 468,
1863
+ 551
1864
+ ],
1865
+ "page_idx": 7
1866
+ },
1867
+ {
1868
+ "type": "text",
1869
+ "text": "Sensitivity - Figure 11. We find that RobustNeRF is remarkably robust to the amount of clutter in a dataset. We define an image as \"cluttered\" if it contains some number of distractor pixels. The figure shows how the reconstruction accuracy of RobustNeRF and mip-NeRF 360 depends on the fraction of training images with distractors, keeping the training set size constant. As the fraction increases, mip-NeRF 360's accuracy steadily drops from 33 to $25\\mathrm{dB}$ , while RobustNeRF's remains steadily above 31 dB throughout. In the distraction-free regime, we find that RobustNeRF mildly under-performs mip-NeRF 360, both in reconstruction quality and the time needed for training. This follows from the statistical inefficiency induced by the trimmed estimator (8), for which a percentage of pixels will be discarded even if they do not correspond to distractors.",
1870
+ "bbox": [
1871
+ 75,
1872
+ 574,
1873
+ 468,
1874
+ 801
1875
+ ],
1876
+ "page_idx": 7
1877
+ },
1878
+ {
1879
+ "type": "text",
1880
+ "text": "5. Conclusions",
1881
+ "text_level": 1,
1882
+ "bbox": [
1883
+ 76,
1884
+ 816,
1885
+ 204,
1886
+ 832
1887
+ ],
1888
+ "page_idx": 7
1889
+ },
1890
+ {
1891
+ "type": "text",
1892
+ "text": "We address a central problem in training NeRF models, namely, optimization in the presence of distractors, such as transient or moving objects and photometric phenomena that are not persistent throughout the capture session.",
1893
+ "bbox": [
1894
+ 75,
1895
+ 839,
1896
+ 468,
1897
+ 902
1898
+ ],
1899
+ "page_idx": 7
1900
+ },
1901
+ {
1902
+ "type": "image",
1903
+ "img_path": "images/8d20921cfffe3ec11df060f2cc54bfdaeaa6671ce873faf1eb8f2439c641d3b5.jpg",
1904
+ "image_caption": [
1905
+ "Figure 11. Sensitivity and Limitations - (left) Reconstruction accuracy for BabyYoda as we increase the fraction of train images with distractors. (right) Accuracy vs training time on clean BabyYoda images (distractor-free)."
1906
+ ],
1907
+ "image_footnote": [],
1908
+ "bbox": [
1909
+ 504,
1910
+ 318,
1911
+ 697,
1912
+ 426
1913
+ ],
1914
+ "page_idx": 7
1915
+ },
1916
+ {
1917
+ "type": "image",
1918
+ "img_path": "images/6aefd32abf79c23dfc1f2bc20cbe57a82060ff522497877b96c668389f8568a1.jpg",
1919
+ "image_caption": [],
1920
+ "image_footnote": [],
1921
+ "bbox": [
1922
+ 709,
1923
+ 318,
1924
+ 888,
1925
+ 426
1926
+ ],
1927
+ "page_idx": 7
1928
+ },
1929
+ {
1930
+ "type": "text",
1931
+ "text": "Viewed through the lens of robust estimation, we formulate training as a form of iteratively re-weighted least squares, with a variant of trimmed LS, and an inductive bias on the smoothness of the outlier process. RobustNeRF is surprisingly simple, yet effective on a wide range of datasets. RobustNeRF is shown to outperform recent state-of-the-art methods [4, 53], qualitatively and quantitatively, on a suite of synthetic datasets, common benchmark datasets, and new datasets captured by a robot, allowing fine-grained control over distractors for comparison with previous methods. While our experiments explore robust estimation in the context of mip-NeRF 360, the RobustNeRF loss can be incorporated within other NeRF models.",
1932
+ "bbox": [
1933
+ 496,
1934
+ 502,
1935
+ 892,
1936
+ 698
1937
+ ],
1938
+ "page_idx": 7
1939
+ },
1940
+ {
1941
+ "type": "text",
1942
+ "text": "Limitations. While RobustNeRF performs well on scenes with distractors, the loss entails some statistical inefficiency. On clean data, this yields somewhat poorer reconstructions, often taking longer to train (see Figure 11). Future work will consider very small distractors, which may require adaptation of the spatial support used for outlier/inlier decisions. It would also be interesting to learn a neural weight function, further improving RobustNeRF; active learning may be useful in this context. Finally, it would be interesting to include our robust loss in other NeRF frameworks.",
1943
+ "bbox": [
1944
+ 496,
1945
+ 700,
1946
+ 890,
1947
+ 851
1948
+ ],
1949
+ "page_idx": 7
1950
+ },
1951
+ {
1952
+ "type": "text",
1953
+ "text": "Acknowledgements We thank Pete Florence and Konstantinos Rematas for helpful feedback, and Tianhao Wu for help with $\\mathrm{D}^2\\mathrm{NeRF}$ experiments.",
1954
+ "bbox": [
1955
+ 496,
1956
+ 854,
1957
+ 890,
1958
+ 900
1959
+ ],
1960
+ "page_idx": 7
1961
+ },
1962
+ {
1963
+ "type": "page_number",
1964
+ "text": "20633",
1965
+ "bbox": [
1966
+ 478,
1967
+ 944,
1968
+ 517,
1969
+ 955
1970
+ ],
1971
+ "page_idx": 7
1972
+ },
1973
+ {
1974
+ "type": "text",
1975
+ "text": "References",
1976
+ "text_level": 1,
1977
+ "bbox": [
1978
+ 78,
1979
+ 89,
1980
+ 174,
1981
+ 106
1982
+ ],
1983
+ "page_idx": 8
1984
+ },
1985
+ {
1986
+ "type": "list",
1987
+ "sub_type": "ref_text",
1988
+ "list_items": [
1989
+ "[1] Michal Adamkiewicz, Timothy Chen, Adam Caccavale, Rachel Gardner, Preston Culbertson, Jeannette Bohg, and Mac Schwager. Vision-only robot navigation in a neural radiance world. IEEE Robotics and Automation Letters, 2022. 1, 2",
1990
+ "[2] Jonathan T. Barron. A general and adaptive robust loss function. Proc. CVPR, 2019. 3, 4",
1991
+ "[3] Jonathan T. Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P. Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. ICCV, 2021. 3",
1992
+ "[4] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In Proc. CVPR, 2022. 2, 5, 8",
1993
+ "[5] Mark Boss, Andreas Engelhardt, Abhishek Kar, Yuanzhen Li, Deqing Sun, Jonathan T. Barron, Hendrik P.A. Lensch, and Varun Jampani. SAMURAI: Shape And Material from Unconstrained Real-world Arbitrary Image collections. In Proc. NeurIPS, 2022. 2",
1994
+ "[6] Zhiqin Chen, Thomas Funkhouser, Peter Hedman, and Andrea Tagliasacchi. Mobilenerf: Exploiting the polygon rasterization pipeline for efficient neural field rendering on mobile architectures. arXiv preprint arXiv:2208.00277, 2022. 1, 2",
1995
+ "[7] Dmitry Chetverikov, Dmitry Svirko, Dmitry Stepanov, and Pavel Krsek. The trimmed iterative closest point algorithm. In International Conference on Pattern Recognition, 2002. 5",
1996
+ "[8] Shin-Fang Chng, Sameera Ramasinghe, Jamie Sherrah, and Simon Lucey. Garf: Gaussian activated radiance fields for high fidelity reconstruction and pose estimation. arXiv eprints, 2022. 2",
1997
+ "[9] Yilun Du, Yinan Zhang, Hong-Xing Yu, Joshua B Tenenbaum, and Jiajun Wu. Neural radiance flow for 4d view synthesis and video processing. In Proc. ICCV. IEEE Computer Society, 2021. 2",
1998
+ "[10] Jiemin Fang, Taoran Yi, Xinggang Wang, Lingxi Xie, Xiaopeng Zhang, Wenyu Liu, Matthias Nießner, and Qi Tian. Fast dynamic radiance fields with time-aware neural voxels. arXiv preprint arXiv:2205.15285, 2022. 2",
1999
+ "[11] Chen Gao, Ayush Saraf, Johannes Kopf, and Jia-Bin Huang. Dynamic view synthesis from dynamic monocular video. In Proc. ICCV, 2021. 2",
2000
+ "[12] Klaus Greff, Francois Belletti, Lucas Beyer, Carl Doersch, Yilun Du, Daniel Duckworth, David J Fleet, Dan Gnanapragasam, Florian Golemo, Charles Herrmann, Thomas Kipf, Abhijit Kundu, Dmitry Lagun, Issam Laradji, Hsueh-Ti (Derek) Liu, Henning Meyer, Yishu Miao, Derek Nowrouzezahrai, Cengiz Oztireli, Etienne Pot, Noha Radwan, Daniel Rebain, Sara Sabour, Mehdi S. M. Sajjadi, Matan Sela, Vincent Sitzmann, Austin Stone, Deqing Sun, Suhani Vora, Ziyu Wang, Tianhao Wu, Kwang Moo Yi, Fangcheng Zhong, and Andrea Tagliausacchi. Kubric: a scalable dataset generator. In Proc. CVPR, 2022. 6",
2001
+ "[13] Peter Hedman, Pratul P. Srinivasan, Ben Mildenhall, Jonathan T. Barron, and Paul Debevec. Baking neural radi"
2002
+ ],
2003
+ "bbox": [
2004
+ 78,
2005
+ 114,
2006
+ 470,
2007
+ 902
2008
+ ],
2009
+ "page_idx": 8
2010
+ },
2011
+ {
2012
+ "type": "list",
2013
+ "sub_type": "ref_text",
2014
+ "list_items": [
2015
+ "ance fields for real-time view synthesis. Proc. ICCV, 2021. 2",
2016
+ "[14] Yoonwoo Jeong, Seokjun Ahn, Christopher Choy, Anima Anandkumar, Minsu Cho, and Jaesik Park. Self-calibrating neural radiance fields. In Proc. ICCV, 2021. 2",
2017
+ "[15] Animesh Karnewar, Tobias Ritschel, Oliver Wang, and Niloy Mitra. Relu fields: The little non-linearity that could. TOG (Proc. SIGGRAPH), 2022. 2",
2018
+ "[16] KN Kutulakos and SM Seitz. A theory of shape by space carving. *IJCV*, 2000. 3",
2019
+ "[17] Tianye Li, Mira Slavcheva, Michael Zollhoefer, Simon Green, Christoph Lassner, Changil Kim, Tanner Schmidt, Steven Lovegrove, Michael Goesele, Richard Newcombe, et al. Neural 3d video synthesis from multi-view video. In Proc. CVPR, 2022. 2",
2020
+ "[18] Zhengqi Li, Simon Niklaus, Noah Snavely, and Oliver Wang. Neural scene flow fields for space-time view synthesis of dynamic scenes. In Proc. CVPR, 2021. 2, 8",
2021
+ "[19] Chen-Hsuan Lin, Wei-Chiu Ma, Antonio Torralba, and Simon Lucey. Barf: Bundle-adjusting neural radiance fields. In Proc. CVPR, 2021. 2",
2022
+ "[20] Jia-Wei Liu, Yan-Pei Cao, Weijia Mao, Wenqiao Zhang, David Junhao Zhang, Jussi Keppo, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. Devrf: Fast deformable voxel radiance fields for dynamic scenes. arXiv preprint arXiv:2205.15723, 2022. 2",
2023
+ "[21] Stephen Lombardi, Tomas Simon, Jason Saragih, Gabriel Schwartz, Andreas Lehrmann, and Yaser Sheikh. Neural volumes: Learning dynamic renderable volumes from images. ACM TOG, 2019. 1",
2024
+ "[22] Li Ma, Xiaoyu Li, Jing Liao, Qi Zhang, Xuan Wang, Jue Wang, and Pedro V. Sander. Deblur-nerf: Neural radiance fields from blurry images. In Proc. CVPR, 2022. 2",
2025
+ "[23] Ricardo Martin-Brualla, Noha Radwan, Mehdi SM Sajjadi, Jonathan T Barron, Alexey Dosovitskiy, and Daniel Duckworth. Nerf in the wild: Neural radiance fields for unconstrained photo collections. In Proc. CVPR, 2021. 2, 3, 4, 8",
2026
+ "[24] Ben Mildenhall, Peter Hedman, Ricardo Martin-Brualla, Pratul P. Srinivasan, and Jonathan T. Barron. NeRF in the dark: High dynamic range view synthesis from noisy raw images. In Proc. CVPR, 2021. 2",
2027
+ "[25] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In Proc. ECCV, 2020. 1, 2",
2028
+ "[26] Ben Mildenhall, Dor Verbin, Pratul P. Srinivasan, Peter Hedman, Ricardo Martin-Brualla, and Jonathan T. Barron. Multinerf: a code release for Mip-NeRF 360, Ref-NeRF, and RawNeRF, 2022. 5",
2029
+ "[27] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. TOG (Proc. SIGGRAPH), 2022. 2",
2030
+ "[28] Michael Niemeyer, Jonathan T. Barron, Ben Mildenhall, Mehdi S. M. Sajjadi, Andreas Geiger, and Noha Radwan. Regnerf: Regularizing neural radiance fields for view synthesis from sparse inputs. In Proc. CVPR, 2022. 2"
2031
+ ],
2032
+ "bbox": [
2033
+ 501,
2034
+ 92,
2035
+ 890,
2036
+ 900
2037
+ ],
2038
+ "page_idx": 8
2039
+ },
2040
+ {
2041
+ "type": "page_number",
2042
+ "text": "20634",
2043
+ "bbox": [
2044
+ 478,
2045
+ 944,
2046
+ 519,
2047
+ 955
2048
+ ],
2049
+ "page_idx": 8
2050
+ },
2051
+ {
2052
+ "type": "list",
2053
+ "sub_type": "ref_text",
2054
+ "list_items": [
2055
+ "[29] Michael Oechsle, Songyou Peng, and Andreas Geiger. Unisurf: Unifying neural implicit surfaces and radiance fields for multi-view reconstruction. In Proc. ICCV, 2021. 2",
2056
+ "[30] Keunhong Park, Utkarsh Sinha, Peter Hedman, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Ricardo MartinBrualla, and Steven M Seitz. Hypernerf: a higher-dimensional representation for topologically varying neural radiance fields. ACM Transactions on Graphics (TOG), 2021. 2",
2057
+ "[31] Ben Poole, Ajay Jain, Jonathan T. Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint 2209.14988, 2022. 2",
2058
+ "[32] Albert Pumarola, Enric Corona, Gerard Pons-Moll, and Francesc Moreno-Noguer. D-nerf: Neural radiance fields for dynamic scenes. In Proc. CVPR, 2021. 2",
2059
+ "[33] Daniel Rebain, Mark Matthews, Kwang Moo Yi, Dmitry Lagun, and Andrea Tagliasacchi. LOLNeRF: Learn from One Look. In Proc. CVPR, 2022. 2, 3",
2060
+ "[34] Jeremy Reizenstein, Roman Shapovalov, Philipp Henzler, Luca Sbordone, Patrick Labatut, and David Novotny. Common objects in 3d: Large-scale learning and evaluation of real-life 3d category reconstruction. In Proc. ICCV, 2021. 1",
2061
+ "[35] Konstantinos Rematas, Andrew Liu, Pratul P. Srinivasan, Jonathan T. Barron, Andrea Tagliasacchi, Tom Funkhouser, and Vittorio Ferrari. Urban radiance fields. Proc. CVPR, 2022. 1, 2, 3",
2062
+ "[36] Antoni Rosinol, John J Leonard, and Luca Carlone. Nerf-slam: Real-time dense monocular slam with neural radiance fields. arXiv preprint arXiv:2210.13641, 2022. 1",
2063
+ "[37] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Proc. CVPR, 2016. 1",
2064
+ "[38] Johannes Lutz Schonberger, Enliang Zheng, Marc Pollefeys, and Jan-Michael Frahm. Pixelwise view selection for unstructured multi-view stereo. In Proc. ECCV, 2016. 6",
2065
+ "[39] Vincent Sitzmann, Michael Zollhöfer, and Gordon Wetzstein. Scene representation networks: Continuous 3d-structure-aware neural scene representations. In Proc. NeurIPS, 2019. 1",
2066
+ "[40] Weiwei Sun, Wei Jiang, Andrea Tagliasacchi, Eduard Trulls, and Kwang Moo Yi. ACNe: Attentive Context Normalization for Robust Permutation-Equivariant Learning. Proc. CVPR, 2020. 4",
2067
+ "[41] Andrea Tagliasacchi and Hao Li. Modern techniques and applications for real-time non-rigid registration. In Proc. SIGGRAPH Asia (Technical Course Notes), 2016. 3, 4",
2068
+ "[42] Andrea Tagliasacchi and Ben Mildenhall. Volume Rendering Digest (for NeRF), 2022. 2",
2069
+ "[43] Matthew Tancik, Vincent Casser, Xinchen Yan, Sabeek Pradhan, Ben Mildenhall, Pratul P Srinivasan, Jonathan T Barron, and Henrik Kretzschmar. Block-nerf: Scalable large scene neural view synthesis. In Proc. CVPR, 2022. 2",
2070
+ "[44] Ayush Tewari, Justus Thies, Ben Mildenhall, Pratul Srinivasan, Edgar Tretschk, W Yifan, Christoph Lassner, Vincent Sitzmann, Ricardo Martin-Brualla, Stephen Lombardi, et al. Advances in neural rendering. In Computer Graphics Forum, 2022. 1"
2071
+ ],
2072
+ "bbox": [
2073
+ 78,
2074
+ 90,
2075
+ 468,
2076
+ 898
2077
+ ],
2078
+ "page_idx": 9
2079
+ },
2080
+ {
2081
+ "type": "list",
2082
+ "sub_type": "ref_text",
2083
+ "list_items": [
2084
+ "[45] Edgar Tretschk, Ayush Tewari, Vladislav Golyanik, Michael Zollhöfer, Christoph Lassner, and Christian Theobalt. Non-rigid neural radiance fields: Reconstruction and novel view synthesis of a dynamic scene from monocular video. In Proc. ICCV, 2021. 2",
2085
+ "[46] Vadim Tschernezki, Diane Larlus, and Andrea Vedaldi. NeuralDiff: Segmenting 3D objects that move in egocentric videos. In Proc. 3DV, 2021. 8",
2086
+ "[47] Dor Verbin, Peter Hedman, Ben Mildenhall, Todd Zickler, Jonathan T. Barron, and Pratul P. Srinivasan. Ref-NeRF: Structured view-dependent appearance for neural radiance fields. Proc. CVPR, 2022. 2",
2087
+ "[48] Suhani Vora, Noha Radwan, Klaus Greff, Henning Meyer, Kyle Genova, Mehdi SM Sajjadi, Etienne Pot, Andrea Tagliasacchi, and Daniel Duckworth. Nesf: Neural semantic fields for generalizable semantic segmentation of 3d scenes. TMLR, 2021. 2",
2088
+ "[49] Chaoyang Wang, Ben Eckart, Simon Lucey, and Orazio Gallo. Neural trajectory fields for dynamic novel view synthesis. arXiv preprint arXiv:2105.05994, 2021. 2",
2089
+ "[50] Jiepeng Wang, Peng Wang, Xiaoxiao Long, Christian Theobalt, Taku Komura, Lingjie Liu, and Wenping Wang. Neuris: Neural reconstruction of indoor scenes using normal priors. arXiv preprint, 2022. 2",
2090
+ "[51] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. Proc. NeurIPS, 2021. 2",
2091
+ "[52] Z. Wang, A.C. Bovik, H.R. Sheikh, and E.P. Simoncelli. Image quality assessment: From error visibility to structural similarity. IEEE Transactions on Image Processing, 2004. 5",
2092
+ "[53] Tianhao Wu, Fangcheng Zhong, Forrester Cole, Andrea Tagliasacchi, and Cengiz Oztireli. D2nerf: Self-supervised decoupling of dynamic and static objects from a monocular video. In Proc. NeurIPS, 2022. 2, 4, 6, 7, 8",
2093
+ "[54] Wenqi Xian, Jia-Bin Huang, Johannes Kopf, and Changil Kim. Space-time neural irradiance fields for free-viewpoint video. In Proc. CVPR, 2021. 2",
2094
+ "[55] Yiheng Xie, Towaki Takikawa, Shunsuke Saito, Or Litany, Shiqin Yan, Numair Khan, Federico Tombari, James Tompkin, Vincent Sitzmann, and Srinath Sridhar. Neural fields in visual computing and beyond. Comput. Graph. Forum, 2022. 1",
2095
+ "[56] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. pixelnerf: Neural radiance fields from one or few images. In Proc. CVPR, 2021. 2",
2096
+ "[57] Zehao Yu, Songyou Peng, Michael Niemeyer, Torsten Sattler, and Andreas Geiger. Monosdf: Exploring monocular geometric cues for neural implicit surface reconstruction. In Proc. NeurIPS, 2022. 2",
2097
+ "[58] Jason Y Zhang, Deva Ramanan, and Shubham Tulsiani. Relpose: Predicting probabilistic relative rotation for single objects in the wild. In Proc. ECCV, 2022. 2",
2098
+ "[59] Kai Zhang, Gernot Riegler, Noah Snavely, and Vladlen Koltun. Nerf++: Analyzing and improving neural radiance fields. arXiv preprint arXiv:2010.07492, 2020. 2"
2099
+ ],
2100
+ "bbox": [
2101
+ 501,
2102
+ 92,
2103
+ 890,
2104
+ 900
2105
+ ],
2106
+ "page_idx": 9
2107
+ },
2108
+ {
2109
+ "type": "page_number",
2110
+ "text": "20635",
2111
+ "bbox": [
2112
+ 478,
2113
+ 945,
2114
+ 517,
2115
+ 955
2116
+ ],
2117
+ "page_idx": 9
2118
+ },
2119
+ {
2120
+ "type": "list",
2121
+ "sub_type": "ref_text",
2122
+ "list_items": [
2123
+ "[60] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proc. CVPR, 2018. 5",
2124
+ "[61] Zihan Zhu, Songyou Peng, Viktor Larsson, Weiwei Xu, Hujun Bao, Zhaopeng Cui, Martin R Oswald, and Marc Pollefeys. Nice-slam: Neural implicit scalable encoding for slam. In Proc. CVPR, 2022. 1"
2125
+ ],
2126
+ "bbox": [
2127
+ 78,
2128
+ 90,
2129
+ 468,
2130
+ 189
2131
+ ],
2132
+ "page_idx": 10
2133
+ },
2134
+ {
2135
+ "type": "page_number",
2136
+ "text": "20636",
2137
+ "bbox": [
2138
+ 478,
2139
+ 945,
2140
+ 519,
2141
+ 955
2142
+ ],
2143
+ "page_idx": 10
2144
+ }
2145
+ ]
2023/RobustNeRF_ Ignoring Distractors With Robust Losses/2f329418-58fe-4ad7-b462-6d0d0ce341f1_model.json ADDED
The diff for this file is too large to render. See raw diff
 
2023/RobustNeRF_ Ignoring Distractors With Robust Losses/2f329418-58fe-4ad7-b462-6d0d0ce341f1_origin.pdf ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:570f108a49794a550628444b971624584343962b0f6b98372b20e04595baf6ad
3
+ size 9855441
2023/RobustNeRF_ Ignoring Distractors With Robust Losses/full.md ADDED
@@ -0,0 +1,424 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # RobustNeRF: Ignoring Distractors with Robust Losses<sup>4</sup>
2
+
3
+ Sara Sabour $^{1,2}$ Suhani Vora $^{1}$ Daniel Duckworth $^{1}$ Ivan Krasin $^{1}$ David J. Fleet $^{1,2}$ Andrea Tagliafasacchi $^{1,2,3}$
4
+
5
+ <sup>1</sup>Google Research, Brain Team <sup>2</sup>University of Toronto <sup>3</sup>Simon Fraser University
6
+
7
+ # Abstract
8
+
9
+ Neural radiance fields (NeRF) excel at synthesizing new views given multi-view, calibrated images of a static scene. When scenes include distractors, which are not persistent during image capture (moving objects, lighting variations, shadows), artifacts appear as view-dependent effects or 'floaters'. To cope with distractors, we advocate a form of robust estimation for NeRF training, modeling distractors in training data as outliers of an optimization problem. Our method successfully removes outliers from a scene and improves upon our baselines, on synthetic and real-world scenes. Our technique is simple to incorporate in modern NeRF frameworks, with few hyper-parameters. It does not assume a priori knowledge of the types of distractors, and is instead focused on the optimization problem rather than pre-processing or modeling transient objects. More results at https://robustnerf.github.io/public.
10
+
11
+ # 1. Introduction
12
+
13
+ The ability to understand the structure of a static 3D scene from 2D images alone is a fundamental problem in computer vision [44]. It finds applications in AR/VR for mapping virtual environments [6, 36, 61], in autonomous robotics for action planning [1], and in photogrammetry to create digital copies of real-world objects [34].
14
+
15
+ Neural fields [55] have recently revolutionized this classical task, by storing 3D representations within the weights of a neural network [39]. These representations are optimized by back-propagating image differences. When the fields store view-dependent radiance and volumetric rendering is employed [21], we can capture 3D scenes with photo-realistic accuracy, and we refer to the generated representation as Neural Radiance Fields, or NeRF [25]).
16
+
17
+ Training of NeRF models generally requires a large collection of images equipped with accurate camera calibration, which can often be recovered via structure-from-motion [37]. Behind its simplicity, NeRF hides several assumptions. As models are typically trained to minimize error in RGB color space, it is of paramount importance
18
+
19
+ ![](images/1f94b3ae1ab59c57b87c5e4884be69f5c0048ed2a2ee83bf4879413ecd792bef.jpg)
20
+ Figure 1. NeRF assumes photometric consistency in the observed images of a scene. Violations of this assumption, as with the images in the top row, yield reconstructed scenes with inconsistent content in the form of "floaters" (highlighted with ellipses). We introduce a simple technique that produces clean reconstruction by automatically ignoring distractors without explicit supervision.
21
+
22
+ that images are photometrically consistent - two photos taken from the same vantage point should be identical up to noise. Unless one employs a method explicitly accounting for it [35], one should manually hold a camera's focus, exposure, white-balance, and ISO fixed.
23
+
24
+ However, properly configuring one's camera is not all that is required to capture high-quality NeRFs – it is also important to avoid distractors: anything that isn't persistent throughout the entire capture session. Distractors come in many shapes and forms, from the hard-shadows cast by the operators as they explore the scene to a pet or child casually walking within the camera's field of view. Distractors are tedious to remove manually, as this would require pixel-by-pixel labeling. They are also tedious to detect, as typical NeRF scenes are trained from hundreds of input images, and the types of distractors are not known a priori. If distractors are ignored, the quality of the reconstruction scene suffers significantly; see Figure 1.
25
+
26
+ In a typical capture session, it is difficult to capture multiple images of the same scene from the same viewpoint, rendering distractors challenging to model mathematically. As such, while view-dependent effects are what give NeRF their realistic look, how can the model tell the difference between a distractor and a view-dependent effect?
27
+
28
+ Despite the challenges, the research community has devised several approaches to overcome this issue:
29
+
30
+ - If distractors are known to belong to a specific class (e.g., people), one can remove them with a pre-trained semantic segmentation model [35, 43] – this process does not generalize to “unexpected” distractors such as shadows.
31
+ - One can model distractors as per-image transient phenomena, and control the balance of transient/persistent modeling [23] - however, it is difficult to tune the losses that control this Pareto-optimal objective.
32
+ - One can model data in time (i.e., high-framerate video) and decompose the scene into static and dynamic (i.e., distractor) components [53] – but this clearly only applies to video rather than photo collection captures.
33
+
34
+ Conversely, we approach the problem of distractors by modeling them as outliers in NeRF optimization.
35
+
36
+ We analyze the aforementioned techniques through the lens of robust estimation, allowing us to understand their behavior, and to design a method that is not only simpler to implement but also more effective (see Figure 1). As a result, we obtain a method that is straightforward to implement, requires minimal-to-no hyper-parameter tuning, and achieves state-of-the-art performance. We evaluate our method:
37
+
38
+ - quantitatively, in terms of reconstruction with synthetically, yet photo-realistically, rendered data;
39
+ - qualitatively on publicly available datasets (often finetuned to work effectively with previous methods);
40
+ - on a new collection of natural and synthetic scenes, including those autonomously acquired by a robot, allowing us to demonstrate the sensitivity of previous methods to hyper-parameter tuning.
41
+
42
+ # 2. Related Work
43
+
44
+ We briefly review the basics and notation of Neural Radiance Fields. We then describe recent progress in NeRF research, paying particular attention to techniques for modeling of static/dynamic scenes.
45
+
46
+ Neural Radiance Fields. A neural radiance field (NeRF) is a continuous volumetric representation of a 3D scene, stored within the parameters of a neural network $\theta$ . The representation maps a position $\mathbf{x}$ and view direction $\mathbf{d}$ to a view-dependent RGB color and view-independent density:
47
+
48
+ $$
49
+ \left. \begin{array}{l} \mathbf {c} (\mathbf {x}, \mathbf {d}) \\ \sigma (\mathbf {x}) \end{array} \right\} f (\mathbf {x}, \mathbf {d}; \boldsymbol {\theta}) \tag {1}
50
+ $$
51
+
52
+ This representation is trained from a collection, $\{(\mathbf{C}_i,\mathbf{T}_i)\}$ of images $\mathbf{C}_i$ with corresponding calibration parameters $\mathbf{T}_i$ (camera extrinsic and intrinsic).
53
+
54
+ During training the calibration information is employed to convert each pixel of the image into a ray $\mathbf{r} = (\mathbf{o},\mathbf{d})$ , and rays are drawn randomly from input images to form a training mini-batch $(\mathbf{r}\sim \mathbf{C}_i)$ . The parameters $\theta$ are optimized to
55
+
56
+ correctly predict the colors of the pixels in the batch via the L2 photometric-reconstruction loss:
57
+
58
+ $$
59
+ \mathcal {L} _ {\mathrm {r g b}} (\boldsymbol {\theta}) = \sum_ {i} \mathbb {E} _ {\mathbf {r} \sim \mathbf {C} _ {i}} \left[ \mathcal {L} _ {\mathrm {r g b}} ^ {\mathbf {r}, i} (\boldsymbol {\theta}) \right] \tag {2}
60
+ $$
61
+
62
+ $$
63
+ \mathcal {L} _ {\mathrm {r g b}} ^ {\mathbf {r}, i} (\boldsymbol {\theta}) = \left| \left| \mathbf {C} (\mathbf {r}; \boldsymbol {\theta}) - \mathbf {C} _ {i} (\mathbf {r}) \right| \right| _ {2} ^ {2} \tag {3}
64
+ $$
65
+
66
+ Parameterizing the ray as $\mathbf{r}(t) = \mathbf{o} + t\mathbf{d}$ , the NeRF model image $\mathbf{C}(\mathbf{r};\boldsymbol{\theta})$ is generated pixel-by-pixel volumetric rendering based on $\sigma(\cdot)$ and $\mathbf{c}(\cdot)$ (e.g., see [25, 42]).
67
+
68
+ Recent progress on NeRF models. NeRF models have recently been extended in several ways. A major thread has been the speedup of training [15, 27] and inference [6, 13], enabling today's models to be trained in minutes [27], and rendered on mobile in real-time [6]. While initially restricted to forward-facing scenes, researchers quickly found ways to model real-world $360^{\circ}$ scenes [4, 59], and to reduce the required number of images, via sensor fusion [35] or hand-designed priors [28]. We can now deal with image artifacts such as motion blur [22], exposure [24], and lens distortion [14]. And the requirement of (precise) camera calibrations is quickly being relaxed with the introduction of techniques for local camera refinement [8, 19], or direct inference [58]. While a NeRF typically represents geometry via volumetric density, there exist models custom-tailored to predict surfaces [29, 51], which can be extended to use predicted normals to significantly improve reconstruction quality [50, 57]. Given high-quality normals [47], inferring the (rendering) structure of a scene becomes a possibility [5]. We also note recent papers about additional applications to generalization [56], semantic understanding [48], generative modeling [33], robotics [1], and text-to-3D [31].
69
+
70
+ Modeling non-static scenes. For unstructured scenes like those considered here, the community has focused on reconstructing both static and non-static elements from video. The most direct approach, treating time as an auxiliary input, leads to cloudy geometry and a lack of fine detail [11, 54]. Directly optimizing per-frame latent codes as an auxiliary input has proved more effective [17, 30, 53]. The most widely-adopted approach is to fit a time-conditioned deformation field mapping 3D points between pairs of frames [18, 49] or to a canonical coordinate frame [9, 10, 20, 32, 45]. Given how sparsely spacetime is sampled, all methods require careful regularization, optimization, or additional training signals to achieve acceptable results.
71
+
72
+ Relatively little attention has been given to removing non-static elements. One common approach is to segment and ignore pixels which are likely to be distractors [35, 43]. While this eliminates larger objects, it fails to account for secondary effects like shadows. Prior attempts to model distractors as outliers still leave residual cloudy geometry [23].
73
+
74
+ ![](images/fd427e53b06992ad7e2399ed5219e2db17f6d6b5aac0badd013f1d3667e33e83.jpg)
75
+ Figure 2. Ambiguity - A simple 2D scene where a static object (blue) is captured by three cameras. During the first and third capture the scene is not photo-consistent as a distractor was within the field of view. Not photo-consistent portions of the scene can end up being encoded as view-dependent effects - even when we assume ground truth geometry.
76
+
77
+ ![](images/ff166e4f12987f76838b6513f6e0bb8ec9a53ef5607f55d7408084a762a7996d.jpg)
78
+
79
+ # 3. Method
80
+
81
+ The classical NeRF training losses (3) are effective for capturing scenes that are photometrically consistent, leading to the photo-realistic novel-view synthesis that we are now accustomed to seeing in recent research. However, "what happens when there are elements of the scene that are not persistent throughout the entire capture session?" Simple examples of such scenes include those in which an object is only present in some fraction of the observed images, or may not remain in the same position in all observed images. For example, Figure 2 depicts a 2D scene comprising a persistent object (the truck), along with several transient objects (e.g., people and a dog). While rays in blue from the three cameras intersect the truck, the green and orange rays from cameras 1 and 3 intersect transient objects. For video capture and spatio-temporal NeRF models, the persistent objects comprise the "static" portion of the scene, while the rest would be called the "dynamic".
82
+
83
+ # 3.1. Sensitivity to outliers
84
+
85
+ For Lambertian scenes, photo-consistent structure is view independent, as scene radiance only depends on the incident light [16]. For such scenes, view-dependent NeRF models like (1), trained by minimizing (3), admit local optima in which transient objects are explained by view-dependent terms. Figure 2 depicts this, with the outgoing color corresponding to the memorized color of the outlier - i.e. view-dependent radiance. Such models exploit the view-dependent capacity of the model to over-fit observations, effectively memorizing the transient objects. One can alter the model to remove dependence on $\mathbf{d}$ but the L2 loss remains problematic as least-squares (LS) estimators are sensitive to outliers, or heavy-tailed noise distributions.
86
+
87
+ Under more natural conditions, dropping the Lambertian assumption, the problem becomes more complex as both non-Lambertian reflectance phenomena and outliers can be explained as view-dependent radiance. While we want the models to capture photo-consistent view-dependent radiance, outliers and other transient phenomena should ideally be ignored. And in such cases, optimization with an L2
88
+
89
+ loss (3) yields significant errors in reconstruction; see Figure 1. Problems like these are pervasive in NeRF model fitting, especially in uncontrolled environments with complex reflectance, non-rigidity, or independently moving objects.
90
+
91
+ # 3.2. Robustness to outliers
92
+
93
+ Robustness via semantic segmentation. One way to reduce outlier contamination during NeRF model optimization is to rely on an oracle $\mathbf{S}$ that specifies whether a given pixel $\mathbf{r}$ from image $i$ is an outlier, and should therefore be excluded from the empirical loss, replacing (3) with:
94
+
95
+ $$
96
+ \mathcal {L} _ {\text {o r a c l e}} ^ {\mathbf {r}, i} (\boldsymbol {\theta}) = \mathbf {S} _ {i} (\mathbf {r}) \cdot | | \mathbf {C} (\mathbf {r}; \boldsymbol {\theta}) - \mathbf {C} _ {i} (\mathbf {r}) | | _ {2} ^ {2} \tag {4}
97
+ $$
98
+
99
+ In practice, a pre-trained (semantic) segmentation network $S$ might serve as an oracle, $\mathbf{S}_i = S(\mathbf{C}_i)$ . E.g., Nerf-in-the-wild [23] employed a semantic segmenter to remove pixels occupied by people, as they are outliers in the context of photo-tourism. Urban Radiance Fields [35] segmented out sky pixels, while LOL-NeRF [33] ignored pixels not belonging to faces. The obvious problem with this approach is the need for an oracle to detect arbitrary distractors.
100
+
101
+ Robust estimators. Another way to reduce sensitivity to outliers is to replace the conventional L2 loss (3) with a robust loss (e.g., [2, 41]), so that photometrically-inconsistent observations can be down-weighted during optimization. Given a robust kernel $\kappa(\cdot)$ , we rewrite our training loss as:
102
+
103
+ $$
104
+ \mathcal {L} _ {\text {r o b u s t}} ^ {\mathbf {r}, i} (\boldsymbol {\theta}) = \kappa \left(\left| \left| \mathbf {C} (\mathbf {r}; \boldsymbol {\theta}) - \mathbf {C} _ {i} (\mathbf {r}) \right| \right| _ {2}\right) \tag {5}
105
+ $$
106
+
107
+ where $\kappa(\cdot)$ is positive and monotonically increasing. MipNeRF [3], for example, employs an L1 loss $\kappa(\epsilon) = |\epsilon|$ , which provides some degree of robustness to outliers during NeRF training. Given our analysis, a valid question is whether we can straightforwardly employ a robust kernel to approach our problem, and if so, given the large variety of robust kernels [2], which is the kernel of choice.
108
+
109
+ Unfortunately, as discussed above, outliers and non-Lambertian effects can both be modelled as view-dependent effects (see Figure 3). As a consequence, with simple application of robust estimators it can be difficult to separate signal from noise. Figure 4 shows examples in which outliers
110
+
111
+ ![](images/11ad7dc1cc45fb13dab8eb420d7cc8ad66432e0546b9533578c9214db3f57180.jpg)
112
+
113
+ ![](images/b2eb3620a00498884b091da8a7002ce0ca793164b22359e7226635d9c8a96708.jpg)
114
+
115
+ ![](images/d79c260859e20c9b34c744b290b5c9725455d80ee2d5fcfb5fbf68d85187829b.jpg)
116
+ Figure 3. Histograms - Robust estimators perform well when the distribution of residuals agrees with the one implied by the estimator (e.g., Gaussian for L2, Laplacian for L1). Here we visualize the ground-truth distribution of residuals (bottom-left), which is hardly a good match with any simple parametric distribution.
117
+
118
+ ![](images/bc0cfa8000d411cb3d2e6833322a1287d9b33fa4e89c2466baecdf67d05404a1.jpg)
119
+
120
+ are removed, but fine-grained texture and view-dependent details are also lost, or conversely, fine-grained details are preserved, but outliers cause artifacts in the reconstructed scene. One can also observe mixtures of these cases in which details are not captured well, nor are outliers fully removed. We find that this behaviour occurs consistently for many different robust estimators and parameter settings.
121
+
122
+ Training time can also be problematic. The robust estimator gradient w.r.t. model parameters can be expressed using the chain rule as
123
+
124
+ $$
125
+ \left. \frac {\partial \kappa (\epsilon (\boldsymbol {\theta}))}{\partial \boldsymbol {\theta}} \right| _ {\boldsymbol {\theta} ^ {(t)}} = \left. \frac {\partial \kappa (\epsilon)}{\partial \epsilon} \right| _ {\epsilon (\boldsymbol {\theta} ^ {(t)})} \cdot \left. \frac {\partial \epsilon (\boldsymbol {\theta})}{\partial \boldsymbol {\theta}} \right| _ {\boldsymbol {\theta} ^ {(t)}} \tag {6}
126
+ $$
127
+
128
+ The second factor is the classical NeRF gradient. The first factor is the kernel gradient evaluated at the current error residual $\epsilon (\pmb{\theta}^{(t)})$ . During training, large residuals can equivalently come from high-frequency details that have not yet been learnt, or they may arise from outliers (see Figure 4 (bottom)). This explain why robust optimization, implemented as (5), should not be expected to decouple high-frequency details from outliers. Further, when strongly robust kernels are employed, like redescending estimators, this also explains the loss of visual fidelity. That is, because the gradient of (large) residuals get down-weighted by the (small) gradients of the kernel, slowing down the learning of these fine-grained details (see Figure 4 (top)).
129
+
130
+ # 3.3. Robustness via Trimmed Least Squares
131
+
132
+ In what follows we advocate a form of iteratively reweighted least-squares (IRLS) with a Trimmed least squares (LS) loss for NeRF model fitting.
133
+
134
+ Iteratively Reweighted least Squares. IRLS is a widely used method for robust estimation that involves solving a sequence of weighted LS problems, the weights of which are adapted to reduce the influence of outliers. To that end,
135
+
136
+ ![](images/7938b39f31acfdc501893b4a8f3350a2f3af6e6029632895dc97faf118a27321.jpg)
137
+
138
+ ![](images/8bee22d7570c63c3c8dcb2a2fb19c6b6aacc77cd83387cf79c109fb8235f552f.jpg)
139
+
140
+ ![](images/1632681e93a49b53846aba35cb9a9a030cc78ccb806d3756c22ec532e55ff874.jpg)
141
+ Figure 4. Kernels - (top-left) Family of robust kernels [2], including L2 ( $\alpha = 2$ ), Charbonnier ( $\alpha = 1$ ) and Geman-McClure ( $\alpha = -2$ ). (top-right) Mid-training, residual magnitudes are similar for distractors and fine-grained details, and pixels with large residuals are learned more slowly, as the gradient of re-descending kernels flattens out. (bottom-right) A too aggressive Geman-McClure in down-weighting large residuals removes both outliers and high-frequency detail. (bottom-left) A less aggressive Geman-McClure does not effectively remove outliers.
142
+
143
+ ![](images/eef272506fab0a7293f190e56385e2b5f26c2fe42d91278bc7e3ba201f641252.jpg)
144
+
145
+ at iteration $t$ , one can write the loss as
146
+
147
+ $$
148
+ \mathcal {L} _ {\text {r o b u s t}} ^ {\mathbf {r}, i} \left(\boldsymbol {\theta} ^ {(t)}\right) = \omega \left(\boldsymbol {\epsilon} ^ {(t - 1)} (\mathbf {r})\right) \cdot | | \mathbf {C} (\mathbf {r}; \boldsymbol {\theta} ^ {(t)}) - \mathbf {C} _ {i} (\mathbf {r}) | | _ {2} ^ {2}
149
+ $$
150
+
151
+ $$
152
+ \boldsymbol {\epsilon} ^ {(t - 1)} (\mathbf {r}) = \left| \left| \mathbf {C} (\mathbf {r}; \boldsymbol {\theta} ^ {(t - 1)}) - \mathbf {C} _ {i} (\mathbf {r}) \right| \right| _ {2} \tag {7}
153
+ $$
154
+
155
+ For weight functions given by $\omega (\epsilon) = \epsilon^{-1}\cdot \partial \kappa (\epsilon) / \partial \epsilon$ one can show that, under suitable conditions, the iteration converges to a local minima of (5) (see [41, Sec. 3]).
156
+
157
+ This framework admits a broad family of losses, including maximum likelihood estimators for heavy-tailed noise processes. Examples in Figure 4 include the Charbonnier loss (smoothed L1), and more aggressive redescending estimators such as the Lorentzian or Geman-McClure [2]. The objective in (4) can also be viewed as a weighted LS objective, the binary weights of which are provided by an oracle. And, as discussed at length below, one can also view several recent methods like NeRFW [23] and $\mathrm{D}^2\mathrm{NeRF}$ [53] through the lens of IRLS and weighted LS.
158
+
159
+ Nevertheless, choosing a suitable weight function $\omega(\epsilon)$ for NeRF optimization is non-trivial, due in large part to the intrinsic ambiguity between view-dependent radiance phenomena and outliers. One might try to solve this problem by learning a neural weight function [40], although generating enough annotated training data might be prohibitive. Instead, the approach taken below is to exploit inductive biases in the structure of outliers, combined with the simplicity of a robust, trimmed LS estimator.
160
+
161
+ Trimmed Robust Kernels. Our goal is to develop a weight function for use in iteratively weighted LS optimization that is simple and captures useful inductive biases for NeRF optimization. For simplicity we opt for a binary weight function with intuitive parameters that adapts naturally through model fitting so that fine-grained image details that are not
162
+
163
+ ![](images/c4d7aa68fc1cf767be6cc6e9d208ab8c015e53701c307328f83104d47b5e8f5c.jpg)
164
+
165
+ ![](images/f91fb7f5b893bb47da572b0df940fc0b094b10279f430799e1478a8e794435e8.jpg)
166
+ residuals - $\epsilon (\mathbf{r})$
167
+
168
+ ![](images/d9a7c3c9248e6b0d2e49b80e508445785b7faf30297405fd92a356c86778a670.jpg)
169
+
170
+ ![](images/873f6abe2d818eeb1542748a04be93455f2e67da076891565998d6dc7678c5f5.jpg)
171
+ inliers - $\tilde{\omega} (\mathbf{r})$
172
+ Figure 5. Algorithm - We visualize our weight function computed by residuals on two examples: (top) the residuals of a (mid-training) NeRF rendered from a training viewpoint, (bottom) a toy residual image containing residual of small spatial extent (dot, line) and residuals of large spatial extent (squares). Notice residuals with large magnitude but small spatial extent (texture of the box, dot, line) are included in the optimization, while weaker residuals with larger spatial extent are excluded. Note that while we operate on patches, we visualize the weight function on the whole image to facilitate visualization.
173
+
174
+ ![](images/d39ccfe40e70ee86c4959d6925556f5b4f52cee785752e345afb7026b573e088.jpg)
175
+
176
+ ![](images/3a9ee80c514580cfc6f5f57876a041719846aed5e201e272d35434e954f86f15.jpg)
177
+ diffusion - $\tilde{\omega}(\mathbf{r}) \oplus \mathcal{B}_{3 \times 3}$
178
+
179
+ ![](images/09be3306f8248a3d9f4bd6c55d6918e93896a53fa487fe67237b06357342877a.jpg)
180
+
181
+ ![](images/a3df0d1d6602f3b8faab9d6badefac0d455997cc96da692275f316984e52437b.jpg)
182
+ (IRLS) weights - $\mathcal{W}(\mathbf{r})$
183
+
184
+ outliers can be learned quickly. It is also important to capture the structured nature of typical outliers, contrary to the typical i.i.d. assumption in most robust estimator formulations. To this end, the weight function should capture spatial smoothness of the outlier process, recognizing that objects typically have continuous local support, and hence outliers are expected to occupy large, connected regions of an image (e.g., the silhouette of a person to be segmented out from a photo-tourism dataset).
185
+
186
+ Surprisingly, a relatively simple weight function embodies these properties and performs extremely well in practice. The weight function is based on so-called trimmed estimators that are used in trimmed least-squares, like that used in trimmed ICP [7]. We first sort residuals, and assume that residuals below a certain percentile are inliers. Picking the $50\%$ percentile for convenience (i.e., median), we define
187
+
188
+ $$
189
+ \tilde {\omega} (\mathbf {r}) = \epsilon (\mathbf {r}) \leq \mathcal {T} _ {\epsilon}, \quad \mathcal {T} _ {\epsilon} = \operatorname {M e d i a n} _ {\mathbf {r}} \left\{\epsilon (\mathbf {r}) \right\}. \tag {8}
190
+ $$
191
+
192
+ To capture spatial smoothness of outliers we spatially diffuse inlier/outlier labels $\omega$ with a $3\times 3$ box kernel $\mathcal{B}_{3\times 3}$ . Formally, we define
193
+
194
+ $$
195
+ \mathcal {W} (\mathbf {r}) = \left(\tilde {\omega} (\mathbf {r}) * \mathcal {B} _ {3 \times 3}\right) \geq \mathcal {T} _ {*}, \quad \mathcal {T} _ {*} = 0. 5. \tag {9}
196
+ $$
197
+
198
+ This helps to avoid classifying high-frequency details as outliers, allowing them to be captured by the NeRF model during optimization (see Figure 5).
199
+
200
+ While the trimmed weight function (9) improves the robustness of model fitting, it sometimes misclassifies fine-grained image details early in training where the NeRF model first captures coarse-grained structure. These localized texture elements may emerge but only after very long training times. We find that stronger inductive bias to spatially coherence allows fine-grained details to be learned more quickly. To that end, we aggregate the detection of outliers on $16 \times 16$ neighborhoods; i.e., we label entire $8 \times 8$ patches as outliers or inliers based on the behavior of $\mathcal{W}$ in the $16 \times 16$ neighborhood of the patch. Denoting the $N \times N$ neighborhood of pixels around $\mathbf{r}$ as $\mathcal{R}_N(\mathbf{r})$ , we define
201
+
202
+ $$
203
+ \omega \left(\mathcal {R} _ {8} (\mathbf {r})\right) = \mathbb {E} _ {\mathbf {s} \sim \mathcal {R} _ {1 6} (\mathbf {r})} \left[ \mathcal {W} (\mathbf {s}) \right] \geq \mathcal {T} _ {\mathcal {R}}, \quad \mathcal {T} _ {\mathcal {R}} = 0. 6. \tag {10}
204
+ $$
205
+
206
+ ![](images/bf26a6902d7a41ed48389fa1f5b89d4469c7a3ae868ee3ba2d6b6dd6db473a18.jpg)
207
+ Ground Truth
208
+
209
+ ![](images/4416c7398ccd94ebc9d367fa80dd34745d1f0b22821109b958a335b77eb5ef87.jpg)
210
+ RobustNeRF
211
+
212
+ ![](images/0b6c7d1202a0081f52befd0f8113b7eaa05410296dc88312c5fd84db8c93517f.jpg)
213
+ Train View
214
+
215
+ ![](images/3a02c34a52cb7d88bc69b2089f4a9245d1437487a0c9d1e3882be254e4ca87fc.jpg)
216
+ Distractors
217
+
218
+ ![](images/61d593844bc4764698a50b4773d9444124bb85159f02cb5638f6aa39ecafd740.jpg)
219
+ Figure 6. Residuals - For the dataset shown in the top row, we visualize the dynamics of the RobustNeRF training residuals, which show how over time the estimated distractor weights go from being random $(t / T) = 0.5\%)$ to identify distractor pixels $(t / T) = 100\%$ without any explicit supervision.
220
+
221
+ The final weight function is the union of the three masks in Eqns. 8 -10. This robust weight function evolves during optimization, as one expects with IRLS where the weights are a function of the residuals at the previous iteration. That is, the labeling of pixels as inliers/outliers changes during training, and settles around masks similar to the one an oracle would provide as training converges (see Figure 6).
222
+
223
+ # 4. Experiments
224
+
225
+ We implement our robust loss function in the MultiNeRF codebase [26] and apply it to mip-NeRF 360 [4]. We dub this method "RobustNeRF". To evaluate RobustNeRF, we compare against baselines on several scenes containing different types of distractors. Where possible, we quantitatively compare reconstructions to held-out, distraction-free images; we report three metrics, averaged across held-out frames, namely, PSNR, SSIM [52], and LPIPS [60].
226
+
227
+ We compare different methods on two collections of scenes, i.e., those provided by the authors of $\mathrm{D}^2\mathrm{NeRF}$ , and
228
+
229
+ ![](images/fd0b180c1026e2be7ccdee830cae294216bd8c940336f0220309076f21627f21.jpg)
230
+ ··
231
+
232
+ ![](images/22cb10d1fc054634cd88f8e4e16a0ab3bf0cb02ffedc48c416b29e5cde255850.jpg)
233
+ ··
234
+
235
+ ![](images/7a8807948cc6eeba13b5d92f78a16697ee949ec41f3a5424f1ade3c127bec8b5.jpg)
236
+ ···
237
+
238
+ ![](images/c1697cdb8d34afb169bc52ec894798f33b8f724db8ce82cf41b82abed8e3d25a.jpg)
239
+ ···
240
+
241
+ ![](images/762874ad94fc2fd98f1240146b07723da08dad632c699a0932c7aa82b5de6a09.jpg)
242
+ Statue
243
+
244
+ ![](images/dbcea64e120747e1ed4cea2e373652528276341671ecec1e4f4912acb22390c0.jpg)
245
+ Android
246
+
247
+ ![](images/66cc33a619f3416b49a8d2b6c643ed33b59e00a0f8f22684e90cfb6ed0e62fa7.jpg)
248
+ Crab
249
+ Figure 7. Dataset - Sample training images showing the distractors in each scene. Statue and Android were acquired manually, and the others with a robotic arm. In the robotic setting we have pixel-perfect alignment of distractor vs. distractor-free images.
250
+
251
+ ![](images/d4665e589c43ddd6ae0a880b9b65793acf9e80ac3bc7d18a3a2c1524afd65028.jpg)
252
+ BabyYoda
253
+
254
+ novel datasets described below. We also present a series of illustrative experiments on synthetic scenes, shedding light on RobustNeRF's efficacy and inner workings.
255
+
256
+ # 4.1. Baselines
257
+
258
+ We compare RobustNeRF to variants of mip-NeRF 360 optimized with different loss functions $(L_{2}, L_{1}$ , and Carbonnier). These variants serve as natural baselines for models with limited or no robustness to outliers. We also compare to $\mathrm{D}^2\mathrm{NeRF}$ , a recent method for reconstructing dynamic scenes from monocular video. Unlike our method, $\mathrm{D}^2\mathrm{NeRF}$ is designed to reconstruct distractors rather than discard them. While $\mathrm{D}^2\mathrm{NeRF}$ is presented as a method for monocular video, it does not presuppose temporal continuity, and can be directly applied to unordered images. We omit additional comparisons to NeRF-W as its performance falls short of $\mathrm{D}^2\mathrm{NeRF}$ [53]. For more details on model training, see the supplementary material.
259
+
260
+ # 4.2. Datasets - Figure 7
261
+
262
+ In addition to scenes from $\mathrm{D}^2\mathrm{NeRF}$ , we introduce a set of natural and synthetic scenes. They facilitate the evaluation of RobustNeRF's effectiveness on illustrative use cases, and they enable empirical analysis under controlled conditions.
263
+
264
+ Natural scenes. We capture seven natural scenes exemplifying different types of distractors. Scenes are captured in three settings, on the street, in an apartment and in a robotics lab. Distractor objects are moved, or are allowed to move, between frames to simulate capture over extended periods of time. We vary the number of unique distractors from 1 (Statue) to 150 (BabyYoda), and their movements. Unlike prior work on monocular video, frames are captured without a clear temporal ordering (see Figure 7). The other three (i.e., Street1, Street2, and Gloss) include view-dependence effects, the results of which are shown in the supplementary
265
+
266
+ material. We also capture additional frames without distractors to enable quantitative evaluations. Camera poses are estimated using COLMAP [38]. A full description of each scene in the supplementary material.
267
+
268
+ Synthetic scenes. To further evaluate RobustNeRF, we generate synthetic scenes using the Kubric dataset generator [12]. Each scene is constructed by placing a set of simple geometries in an empty, texture-less room. In each scene, a subset of objects remain fixed while the other objects (i.e., distractors) change position from frame to frame. By varying the number of objects, their size, and the way they move, we control the level of distraction in each scene. We use these scenes to examine RobustNeRF's sensitivity to its hyperparameters, see supplementary material.
269
+
270
+ # 4.3. Evaluation
271
+
272
+ We evaluate RobustNeRF on its ability to ignore distractors while accurately reconstructing the static elements of a scene. We train RobustNeRF, $\mathrm{D}^2\mathrm{NeRF}$ , and variants of mip-NeRF 360 on scenes where distraction-free frames are available. Models are trained on frames with distractors and evaluated on distractor-free frames.
273
+
274
+ Comparison to mip-NeRF 360 - Figure 8. On natural scenes, RobustNeRF generally outperforms variants of mip-NeRF 360 by 1.3 to $4.7\mathrm{dB}$ in PSNR. As $L_{2}, L_{1}$ , and Charbonnier losses weigh all pixels equally, the model is forced to represent, rather than ignore, distractors as "clouds" with view-dependent appearance. We find clouds to be most apparent when distractors remain stationary for multiple frames. In contrast, RobustNeRF's loss isolates distractor pixels and assigns them a weight of zero (see Figure 6). To establish an upper bound on reconstruction accuracy, we train mip-NeRF 360 with Charbonnier loss on distraction-free versions of each scene, the images for which are taken from (approximately) the same viewpoints. Reassuringly, RobustNeRF when trained on distraction-free frames, achieves nearly identical accuracy; see Figure 11.
275
+
276
+ While RobustNeRF consistently outperforms mip-NeRF 360, the gap is smaller in the Apartment scenes (Statue, Android) than the Robotics Lab scenes (Crab, BabyYoda). This can be explained by challenging background geometry, errors in camera parameter estimation, and imperceptible changes to scene appearance. For further discussion, see the supplementary material.
277
+
278
+ Comparison to $\mathbf{D}^2\mathrm{NeRF}$ - Figure 9. Quantitatively, RobustNeRF matches or outperforms $\mathrm{D}^2\mathrm{NeRF}$ by as much as $12\mathrm{dB}$ PSNR depending on the number of unique outlier objects in the capture. Results on $\mathrm{D}^2\mathrm{NeRF}$ real scenes are provided in the supplementary material for qualitative comparison. In Statue and Android, 1 and 3 non-rigid objects are moved around the scene, respectively. $\mathrm{D}^2\mathrm{NeRF}$ is able to model these objects and thus separate them from the scenes' static content. In the remaining scenes, a much
279
+
280
+ <table><tr><td rowspan="2"></td><td colspan="3">Statue</td><td colspan="3">Android</td><td colspan="3">Crab</td><td colspan="3">BabyYoda</td></tr><tr><td>LPIPS↓</td><td>SSIM↑</td><td>PSNR↑</td><td>LPIPS↓</td><td>SSIM↑</td><td>PSNR↑</td><td>LPIPS↓</td><td>SSIM↑</td><td>PSNR↑</td><td>LPIPS↓</td><td>SSIM↑</td><td>PSNR↑</td></tr><tr><td>mip-NeRF 360 (L2)</td><td>0.36</td><td>0.66</td><td>19.09</td><td>0.40</td><td>0.65</td><td>19.35</td><td>0.27</td><td>0.77</td><td>25.73</td><td>0.31</td><td>0.75</td><td>22.97</td></tr><tr><td>mip-NeRF 360 (L1)</td><td>0.30</td><td>0.72</td><td>19.55</td><td>0.40</td><td>0.66</td><td>19.38</td><td>0.22</td><td>0.79</td><td>26.69</td><td>0.22</td><td>0.80</td><td>26.15</td></tr><tr><td>mip-NeRF 360 (Ch.)</td><td>0.30</td><td>0.73</td><td>19.64</td><td>0.40</td><td>0.66</td><td>19.53</td><td>0.21</td><td>0.80</td><td>27.72</td><td>0.23</td><td>0.80</td><td>25.22</td></tr><tr><td>D2NeRF</td><td>0.48</td><td>0.49</td><td>19.09</td><td>0.43</td><td>0.57</td><td>20.61</td><td>0.42</td><td>0.68</td><td>21.18</td><td>0.44</td><td>0.65</td><td>17.32</td></tr><tr><td>RobustNeRF</td><td>0.28</td><td>0.75</td><td>20.89</td><td>0.31</td><td>0.65</td><td>21.72</td><td>0.21</td><td>0.81</td><td>30.75</td><td>0.20</td><td>0.83</td><td>30.87</td></tr><tr><td>mip-NeRF 360 (clean)</td><td>0.19</td><td>0.80</td><td>23.57</td><td>0.31</td><td>0.71</td><td>23.10</td><td>0.16</td><td>0.84</td><td>32.55</td><td>0.16</td><td>0.84</td><td>32.63</td></tr></table>
281
+
282
+ ![](images/2472bef537acf5ecbf70fd1ccd4e8fdd4925aac7be2d02f5078b3cadb5f67ca5.jpg)
283
+ MipNeRF360 (L2)
284
+
285
+ ![](images/af7d9acea775d2155ed10237b7999d99195bb134366ad25a6c1e360d937f9f42.jpg)
286
+ MipNeRF360 (L1)
287
+
288
+ ![](images/13a1dc7479797324164b7080843547f38b6e2fb62ce563031af7a75ba9138338.jpg)
289
+ MipNeRF360 (Ch)
290
+
291
+ ![](images/714501e76ffdf113f080c6f7b7789dd195653e9a529af0ee71a92d27ba2a2214.jpg)
292
+ $\mathrm{D}^2\mathrm{NeRF}$
293
+
294
+ ![](images/3ddc66a93c631c646317fe0aa2640bfdf35c5cc4d28f9c65636ca20f691b07c4.jpg)
295
+ RobustNeRF
296
+
297
+ ![](images/7fd91bf56dc9e30d6dcc2265a52d415b47fdfc66e4518b1befa598acb895655d.jpg)
298
+
299
+ ![](images/0647871967bd735d4b498a761950e451bffe8a160ccfff0372e0513819fbc28f.jpg)
300
+
301
+ ![](images/b658f6f97a6edd10f1d7a6cdd644fdd1ce388b732f1914e6aaeda8051c38167f.jpg)
302
+
303
+ ![](images/a779e3c1ee430366f17af15c093cba5a31cacd675bc5a43f4a67457a2f945046.jpg)
304
+
305
+ ![](images/5d6a9d369db5377343f4d05f840cff309f2bcdbfec3a29c11b9f81bb5d6e35dc.jpg)
306
+
307
+ ![](images/908fbde1d517ea610c37fcfc2d9fa26961946c3ac07c40f3516176ce479013fe.jpg)
308
+
309
+ ![](images/941d667498108edd1f131e7b2e9577b587c75b45666ebec28ffe5295b8276fac.jpg)
310
+
311
+ ![](images/7cdaf75c778f195b955d2151692a4b2e4dcfd3b2cccad382ba65da930cc8d33b.jpg)
312
+
313
+ ![](images/0c66e136417c9ec76a7ce9cd70ea66232e91904c114038bd3185ee2df4289495.jpg)
314
+
315
+ ![](images/e32af9c9ac8a43c2649ac21e09d064db9b60e09650a3536d662beeace1b62791.jpg)
316
+
317
+ ![](images/84ef9a524a3021812a8c9f681d2f616deddd59904b35a59c6139c3e949778e33.jpg)
318
+ Figure 8. Evaluation on Natural Scenes - RobustNeRF outperforms baselines and $\mathrm{D}^2\mathrm{NeRF}$ [53] on novel view synthesis with real-world captures. The table provides a quantitative comparison of RobustNeRF, $\mathrm{D}^2\mathrm{NeRF}$ and mip-NeRF 360 using different reconstruction losses. The last row reports mip-NeRF 360 trained on a distractor-free version of each dataset, giving an upperbound for RobustNeRF performance. We also visualize samples from each scene rendered with each of the methods. See Supplementary Material for more samples.
319
+
320
+ ![](images/fbff17edcd0c79df32fd429ad15ec1ef06b80d4dca1fab87e2456bd3826010ad.jpg)
321
+
322
+ ![](images/e9215e4e4f884ac6d10f101a9563ca8e96bf72151a128a1751def30b93e22c45.jpg)
323
+
324
+ ![](images/b2ab403dd8bd35e27f24f173505838e737fc32feaeaff8ed4698dd5bbd5f6794.jpg)
325
+
326
+ ![](images/06c7d9010e371fb2c3acf11e25eeec78573aa9900a82f80795f0ec317f4a5b25.jpg)
327
+
328
+ larger pool of 100 to 150 unique, non-static objects are used – too many for $\mathrm{D}^2\mathrm{NeRF}$ to model effectively. As a result, “cloud” artifacts appear in its static representation, similar to those produced by mip-NeRF 360. In contrast, Robust-NeRF identifies non-static content as outliers and omits it during reconstruction. Although both methods use a similar number of parameters, $\mathrm{D}^2\mathrm{NeRF}$ 's peak memory usage is $2.3\mathrm{x}$ higher than RobustNeRF and $37\mathrm{x}$ higher when normalizing for batch size. This is a direct consequence of model architecture: $\mathrm{D}^2\mathrm{NeRF}$ is tailored to simultaneously modeling static and dynamic content and thus merits higher complexity. To remain comparable, we limit image resolution to 0.2 megapixels for all experiments.
329
+
330
+ Ablations - Figure 10. We ablate elements of the Robust-NeRF loss on the crab scene, comparing to an upper bound on the reconstruction accuracy of mip-NeRF 360 trained on distractor-free (clean) images from identical viewpoints. Our trimmed estimator (8) successfully eliminates distractors at the expense of high frequency texture and a lower PSNR. With smoothing (9), fine details are recovered, at the cost of longer training times. With the spatial window (10), RobustNeRF training time is on-par with mip-NeRF 360. We also ablate patch size and the trimming threshold (see Supplementary Material); we find that RobustNeRF is insensitive to trimming threshold, and that reducing the patch size offsets the gains from smoothing and patching.
331
+
332
+ <table><tr><td rowspan="2"></td><td colspan="3">Car</td><td colspan="3">Cars</td><td colspan="3">Bag</td><td colspan="3">Chairs</td><td colspan="3">Pillow</td><td></td></tr><tr><td>LPIPS↓</td><td>MS-SSIM↑</td><td>PSNR↑</td><td>LPIPS↓</td><td>MS-SSIM↑</td><td>PSNR↑</td><td>LPIPS↓</td><td>MS-SSIM↑</td><td>PSNR↑</td><td>LPIPS↓</td><td>MS-SSIM↑</td><td>PSNR↑</td><td>LPIPS↓</td><td>MS-SSIM↑</td><td>PSNR↑</td><td></td></tr><tr><td>NeRF-W [23]</td><td>.218</td><td>.814</td><td>24.23</td><td>.243</td><td>.873</td><td>24.51</td><td>.139</td><td>.791</td><td>20.65</td><td>.150</td><td>.681</td><td>23.77</td><td>.088</td><td>.935</td><td>28.24</td><td></td></tr><tr><td>NSFF [18]</td><td>.200</td><td>.806</td><td>24.90</td><td>.620</td><td>.376</td><td>10.29</td><td>.108</td><td>.892</td><td>25.62</td><td>.682</td><td>.284</td><td>12.82</td><td>.782</td><td>.343</td><td>4.55</td><td></td></tr><tr><td>NeuralDiff [46]</td><td>.065</td><td>.952</td><td>31.89</td><td>.098</td><td>.921</td><td>25.93</td><td>.117</td><td>.910</td><td>29.02</td><td>.112</td><td>.722</td><td>24.42</td><td>.565</td><td>.652</td><td>20.09</td><td></td></tr><tr><td>D2NeRF [53]</td><td>.062</td><td>.975</td><td>34.27</td><td>.090</td><td>.953</td><td>26.27</td><td>.076</td><td>.979</td><td>34.14</td><td>.095</td><td>.707</td><td>24.63</td><td>.076</td><td>.979</td><td>36.58</td><td></td></tr><tr><td>RobustNeRF</td><td>.013</td><td>.988</td><td>37.73</td><td>.063</td><td>.957</td><td>26.31</td><td>.006</td><td>.995</td><td>41.82</td><td>.007</td><td>.992</td><td>41.23</td><td>.018</td><td>.990</td><td>38.95</td><td></td></tr><tr><td>Train View</td><td colspan="3">Test View</td><td colspan="3">NeRF-W</td><td colspan="3">NSFF</td><td colspan="3">NeuralDiff</td><td colspan="3">D2NeRF</td><td>RobustNeRF</td></tr></table>
333
+
334
+ Figure 9. Evaluations on $\mathbf{D}^2\mathrm{NeRF}$ Synthetic Scenes - Quantitative and qualitative evaluations on the Kubric synthetic dataset introduced by $\mathbf{D}^2\mathrm{NeRF}$ , consisting of 200 training frames (with distractor) and 100 novel views for evaluation (without distractor).
335
+
336
+ <table><tr><td></td><td>LPIPS↓</td><td>SSIM↑</td><td>PSNR↑</td><td>Updates to PSNR=30</td></tr><tr><td>mip-NeRF 360 (L2)</td><td>0.31</td><td>0.75</td><td>22.97</td><td>-</td></tr><tr><td>+ robust (8)</td><td>0.39</td><td>0.60</td><td>18.21</td><td>-</td></tr><tr><td>+ smoothing (9)</td><td>0.22</td><td>0.81</td><td>30.01</td><td>250K</td></tr><tr><td>+ patching (10)</td><td>0.21</td><td>0.81</td><td>30.75</td><td>70K</td></tr><tr><td>oracle (clean)</td><td>0.16</td><td>0.84</td><td>32.55</td><td>25K</td></tr><tr><td>+ robust</td><td colspan="3">+ smoothing</td><td>+ patching</td></tr></table>
337
+
338
+ Figure 10. Ablations - Blindly trimming the loss causes details to be lost. Smoothing recovers fine-grained detail, while patch-based evaluation speeds up training and adds more detail. Patching enables the model to reach PSNR of 30, almost $4 \times$ faster.
339
+
340
+ Sensitivity - Figure 11. We find that RobustNeRF is remarkably robust to the amount of clutter in a dataset. We define an image as "cluttered" if it contains some number of distractor pixels. The figure shows how the reconstruction accuracy of RobustNeRF and mip-NeRF 360 depends on the fraction of training images with distractors, keeping the training set size constant. As the fraction increases, mip-NeRF 360's accuracy steadily drops from 33 to $25\mathrm{dB}$ , while RobustNeRF's remains steadily above 31 dB throughout. In the distraction-free regime, we find that RobustNeRF mildly under-performs mip-NeRF 360, both in reconstruction quality and the time needed for training. This follows from the statistical inefficiency induced by the trimmed estimator (8), for which a percentage of pixels will be discarded even if they do not correspond to distractors.
341
+
342
+ # 5. Conclusions
343
+
344
+ We address a central problem in training NeRF models, namely, optimization in the presence of distractors, such as transient or moving objects and photometric phenomena that are not persistent throughout the capture session.
345
+
346
+ ![](images/8d20921cfffe3ec11df060f2cc54bfdaeaa6671ce873faf1eb8f2439c641d3b5.jpg)
347
+ Figure 11. Sensitivity and Limitations - (left) Reconstruction accuracy for BabyYoda as we increase the fraction of train images with distractors. (right) Accuracy vs training time on clean BabyYoda images (distractor-free).
348
+
349
+ ![](images/6aefd32abf79c23dfc1f2bc20cbe57a82060ff522497877b96c668389f8568a1.jpg)
350
+
351
+ Viewed through the lens of robust estimation, we formulate training as a form of iteratively re-weighted least squares, with a variant of trimmed LS, and an inductive bias on the smoothness of the outlier process. RobustNeRF is surprisingly simple, yet effective on a wide range of datasets. RobustNeRF is shown to outperform recent state-of-the-art methods [4, 53], qualitatively and quantitatively, on a suite of synthetic datasets, common benchmark datasets, and new datasets captured by a robot, allowing fine-grained control over distractors for comparison with previous methods. While our experiments explore robust estimation in the context of mip-NeRF 360, the RobustNeRF loss can be incorporated within other NeRF models.
352
+
353
+ Limitations. While RobustNeRF performs well on scenes with distractors, the loss entails some statistical inefficiency. On clean data, this yields somewhat poorer reconstructions, often taking longer to train (see Figure 11). Future work will consider very small distractors, which may require adaptation of the spatial support used for outlier/inlier decisions. It would also be interesting to learn a neural weight function, further improving RobustNeRF; active learning may be useful in this context. Finally, it would be interesting to include our robust loss in other NeRF frameworks.
354
+
355
+ Acknowledgements We thank Pete Florence and Konstantinos Rematas for helpful feedback, and Tianhao Wu for help with $\mathrm{D}^2\mathrm{NeRF}$ experiments.
356
+
357
+ # References
358
+
359
+ [1] Michal Adamkiewicz, Timothy Chen, Adam Caccavale, Rachel Gardner, Preston Culbertson, Jeannette Bohg, and Mac Schwager. Vision-only robot navigation in a neural radiance world. IEEE Robotics and Automation Letters, 2022. 1, 2
360
+ [2] Jonathan T. Barron. A general and adaptive robust loss function. Proc. CVPR, 2019. 3, 4
361
+ [3] Jonathan T. Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P. Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. ICCV, 2021. 3
362
+ [4] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In Proc. CVPR, 2022. 2, 5, 8
363
+ [5] Mark Boss, Andreas Engelhardt, Abhishek Kar, Yuanzhen Li, Deqing Sun, Jonathan T. Barron, Hendrik P.A. Lensch, and Varun Jampani. SAMURAI: Shape And Material from Unconstrained Real-world Arbitrary Image collections. In Proc. NeurIPS, 2022. 2
364
+ [6] Zhiqin Chen, Thomas Funkhouser, Peter Hedman, and Andrea Tagliasacchi. Mobilenerf: Exploiting the polygon rasterization pipeline for efficient neural field rendering on mobile architectures. arXiv preprint arXiv:2208.00277, 2022. 1, 2
365
+ [7] Dmitry Chetverikov, Dmitry Svirko, Dmitry Stepanov, and Pavel Krsek. The trimmed iterative closest point algorithm. In International Conference on Pattern Recognition, 2002. 5
366
+ [8] Shin-Fang Chng, Sameera Ramasinghe, Jamie Sherrah, and Simon Lucey. Garf: Gaussian activated radiance fields for high fidelity reconstruction and pose estimation. arXiv eprints, 2022. 2
367
+ [9] Yilun Du, Yinan Zhang, Hong-Xing Yu, Joshua B Tenenbaum, and Jiajun Wu. Neural radiance flow for 4d view synthesis and video processing. In Proc. ICCV. IEEE Computer Society, 2021. 2
368
+ [10] Jiemin Fang, Taoran Yi, Xinggang Wang, Lingxi Xie, Xiaopeng Zhang, Wenyu Liu, Matthias Nießner, and Qi Tian. Fast dynamic radiance fields with time-aware neural voxels. arXiv preprint arXiv:2205.15285, 2022. 2
369
+ [11] Chen Gao, Ayush Saraf, Johannes Kopf, and Jia-Bin Huang. Dynamic view synthesis from dynamic monocular video. In Proc. ICCV, 2021. 2
370
+ [12] Klaus Greff, Francois Belletti, Lucas Beyer, Carl Doersch, Yilun Du, Daniel Duckworth, David J Fleet, Dan Gnanapragasam, Florian Golemo, Charles Herrmann, Thomas Kipf, Abhijit Kundu, Dmitry Lagun, Issam Laradji, Hsueh-Ti (Derek) Liu, Henning Meyer, Yishu Miao, Derek Nowrouzezahrai, Cengiz Oztireli, Etienne Pot, Noha Radwan, Daniel Rebain, Sara Sabour, Mehdi S. M. Sajjadi, Matan Sela, Vincent Sitzmann, Austin Stone, Deqing Sun, Suhani Vora, Ziyu Wang, Tianhao Wu, Kwang Moo Yi, Fangcheng Zhong, and Andrea Tagliausacchi. Kubric: a scalable dataset generator. In Proc. CVPR, 2022. 6
371
+ [13] Peter Hedman, Pratul P. Srinivasan, Ben Mildenhall, Jonathan T. Barron, and Paul Debevec. Baking neural radi
372
+
373
+ ance fields for real-time view synthesis. Proc. ICCV, 2021. 2
374
+ [14] Yoonwoo Jeong, Seokjun Ahn, Christopher Choy, Anima Anandkumar, Minsu Cho, and Jaesik Park. Self-calibrating neural radiance fields. In Proc. ICCV, 2021. 2
375
+ [15] Animesh Karnewar, Tobias Ritschel, Oliver Wang, and Niloy Mitra. Relu fields: The little non-linearity that could. TOG (Proc. SIGGRAPH), 2022. 2
376
+ [16] KN Kutulakos and SM Seitz. A theory of shape by space carving. *IJCV*, 2000. 3
377
+ [17] Tianye Li, Mira Slavcheva, Michael Zollhoefer, Simon Green, Christoph Lassner, Changil Kim, Tanner Schmidt, Steven Lovegrove, Michael Goesele, Richard Newcombe, et al. Neural 3d video synthesis from multi-view video. In Proc. CVPR, 2022. 2
378
+ [18] Zhengqi Li, Simon Niklaus, Noah Snavely, and Oliver Wang. Neural scene flow fields for space-time view synthesis of dynamic scenes. In Proc. CVPR, 2021. 2, 8
379
+ [19] Chen-Hsuan Lin, Wei-Chiu Ma, Antonio Torralba, and Simon Lucey. Barf: Bundle-adjusting neural radiance fields. In Proc. CVPR, 2021. 2
380
+ [20] Jia-Wei Liu, Yan-Pei Cao, Weijia Mao, Wenqiao Zhang, David Junhao Zhang, Jussi Keppo, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. Devrf: Fast deformable voxel radiance fields for dynamic scenes. arXiv preprint arXiv:2205.15723, 2022. 2
381
+ [21] Stephen Lombardi, Tomas Simon, Jason Saragih, Gabriel Schwartz, Andreas Lehrmann, and Yaser Sheikh. Neural volumes: Learning dynamic renderable volumes from images. ACM TOG, 2019. 1
382
+ [22] Li Ma, Xiaoyu Li, Jing Liao, Qi Zhang, Xuan Wang, Jue Wang, and Pedro V. Sander. Deblur-nerf: Neural radiance fields from blurry images. In Proc. CVPR, 2022. 2
383
+ [23] Ricardo Martin-Brualla, Noha Radwan, Mehdi SM Sajjadi, Jonathan T Barron, Alexey Dosovitskiy, and Daniel Duckworth. Nerf in the wild: Neural radiance fields for unconstrained photo collections. In Proc. CVPR, 2021. 2, 3, 4, 8
384
+ [24] Ben Mildenhall, Peter Hedman, Ricardo Martin-Brualla, Pratul P. Srinivasan, and Jonathan T. Barron. NeRF in the dark: High dynamic range view synthesis from noisy raw images. In Proc. CVPR, 2021. 2
385
+ [25] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In Proc. ECCV, 2020. 1, 2
386
+ [26] Ben Mildenhall, Dor Verbin, Pratul P. Srinivasan, Peter Hedman, Ricardo Martin-Brualla, and Jonathan T. Barron. Multinerf: a code release for Mip-NeRF 360, Ref-NeRF, and RawNeRF, 2022. 5
387
+ [27] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. TOG (Proc. SIGGRAPH), 2022. 2
388
+ [28] Michael Niemeyer, Jonathan T. Barron, Ben Mildenhall, Mehdi S. M. Sajjadi, Andreas Geiger, and Noha Radwan. Regnerf: Regularizing neural radiance fields for view synthesis from sparse inputs. In Proc. CVPR, 2022. 2
389
+
390
+ [29] Michael Oechsle, Songyou Peng, and Andreas Geiger. Unisurf: Unifying neural implicit surfaces and radiance fields for multi-view reconstruction. In Proc. ICCV, 2021. 2
391
+ [30] Keunhong Park, Utkarsh Sinha, Peter Hedman, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Ricardo MartinBrualla, and Steven M Seitz. Hypernerf: a higher-dimensional representation for topologically varying neural radiance fields. ACM Transactions on Graphics (TOG), 2021. 2
392
+ [31] Ben Poole, Ajay Jain, Jonathan T. Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint 2209.14988, 2022. 2
393
+ [32] Albert Pumarola, Enric Corona, Gerard Pons-Moll, and Francesc Moreno-Noguer. D-nerf: Neural radiance fields for dynamic scenes. In Proc. CVPR, 2021. 2
394
+ [33] Daniel Rebain, Mark Matthews, Kwang Moo Yi, Dmitry Lagun, and Andrea Tagliasacchi. LOLNeRF: Learn from One Look. In Proc. CVPR, 2022. 2, 3
395
+ [34] Jeremy Reizenstein, Roman Shapovalov, Philipp Henzler, Luca Sbordone, Patrick Labatut, and David Novotny. Common objects in 3d: Large-scale learning and evaluation of real-life 3d category reconstruction. In Proc. ICCV, 2021. 1
396
+ [35] Konstantinos Rematas, Andrew Liu, Pratul P. Srinivasan, Jonathan T. Barron, Andrea Tagliasacchi, Tom Funkhouser, and Vittorio Ferrari. Urban radiance fields. Proc. CVPR, 2022. 1, 2, 3
397
+ [36] Antoni Rosinol, John J Leonard, and Luca Carlone. Nerf-slam: Real-time dense monocular slam with neural radiance fields. arXiv preprint arXiv:2210.13641, 2022. 1
398
+ [37] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Proc. CVPR, 2016. 1
399
+ [38] Johannes Lutz Schonberger, Enliang Zheng, Marc Pollefeys, and Jan-Michael Frahm. Pixelwise view selection for unstructured multi-view stereo. In Proc. ECCV, 2016. 6
400
+ [39] Vincent Sitzmann, Michael Zollhöfer, and Gordon Wetzstein. Scene representation networks: Continuous 3d-structure-aware neural scene representations. In Proc. NeurIPS, 2019. 1
401
+ [40] Weiwei Sun, Wei Jiang, Andrea Tagliasacchi, Eduard Trulls, and Kwang Moo Yi. ACNe: Attentive Context Normalization for Robust Permutation-Equivariant Learning. Proc. CVPR, 2020. 4
402
+ [41] Andrea Tagliasacchi and Hao Li. Modern techniques and applications for real-time non-rigid registration. In Proc. SIGGRAPH Asia (Technical Course Notes), 2016. 3, 4
403
+ [42] Andrea Tagliasacchi and Ben Mildenhall. Volume Rendering Digest (for NeRF), 2022. 2
404
+ [43] Matthew Tancik, Vincent Casser, Xinchen Yan, Sabeek Pradhan, Ben Mildenhall, Pratul P Srinivasan, Jonathan T Barron, and Henrik Kretzschmar. Block-nerf: Scalable large scene neural view synthesis. In Proc. CVPR, 2022. 2
405
+ [44] Ayush Tewari, Justus Thies, Ben Mildenhall, Pratul Srinivasan, Edgar Tretschk, W Yifan, Christoph Lassner, Vincent Sitzmann, Ricardo Martin-Brualla, Stephen Lombardi, et al. Advances in neural rendering. In Computer Graphics Forum, 2022. 1
406
+
407
+ [45] Edgar Tretschk, Ayush Tewari, Vladislav Golyanik, Michael Zollhöfer, Christoph Lassner, and Christian Theobalt. Non-rigid neural radiance fields: Reconstruction and novel view synthesis of a dynamic scene from monocular video. In Proc. ICCV, 2021. 2
408
+ [46] Vadim Tschernezki, Diane Larlus, and Andrea Vedaldi. NeuralDiff: Segmenting 3D objects that move in egocentric videos. In Proc. 3DV, 2021. 8
409
+ [47] Dor Verbin, Peter Hedman, Ben Mildenhall, Todd Zickler, Jonathan T. Barron, and Pratul P. Srinivasan. Ref-NeRF: Structured view-dependent appearance for neural radiance fields. Proc. CVPR, 2022. 2
410
+ [48] Suhani Vora, Noha Radwan, Klaus Greff, Henning Meyer, Kyle Genova, Mehdi SM Sajjadi, Etienne Pot, Andrea Tagliasacchi, and Daniel Duckworth. Nesf: Neural semantic fields for generalizable semantic segmentation of 3d scenes. TMLR, 2021. 2
411
+ [49] Chaoyang Wang, Ben Eckart, Simon Lucey, and Orazio Gallo. Neural trajectory fields for dynamic novel view synthesis. arXiv preprint arXiv:2105.05994, 2021. 2
412
+ [50] Jiepeng Wang, Peng Wang, Xiaoxiao Long, Christian Theobalt, Taku Komura, Lingjie Liu, and Wenping Wang. Neuris: Neural reconstruction of indoor scenes using normal priors. arXiv preprint, 2022. 2
413
+ [51] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. Proc. NeurIPS, 2021. 2
414
+ [52] Z. Wang, A.C. Bovik, H.R. Sheikh, and E.P. Simoncelli. Image quality assessment: From error visibility to structural similarity. IEEE Transactions on Image Processing, 2004. 5
415
+ [53] Tianhao Wu, Fangcheng Zhong, Forrester Cole, Andrea Tagliasacchi, and Cengiz Oztireli. D2nerf: Self-supervised decoupling of dynamic and static objects from a monocular video. In Proc. NeurIPS, 2022. 2, 4, 6, 7, 8
416
+ [54] Wenqi Xian, Jia-Bin Huang, Johannes Kopf, and Changil Kim. Space-time neural irradiance fields for free-viewpoint video. In Proc. CVPR, 2021. 2
417
+ [55] Yiheng Xie, Towaki Takikawa, Shunsuke Saito, Or Litany, Shiqin Yan, Numair Khan, Federico Tombari, James Tompkin, Vincent Sitzmann, and Srinath Sridhar. Neural fields in visual computing and beyond. Comput. Graph. Forum, 2022. 1
418
+ [56] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. pixelnerf: Neural radiance fields from one or few images. In Proc. CVPR, 2021. 2
419
+ [57] Zehao Yu, Songyou Peng, Michael Niemeyer, Torsten Sattler, and Andreas Geiger. Monosdf: Exploring monocular geometric cues for neural implicit surface reconstruction. In Proc. NeurIPS, 2022. 2
420
+ [58] Jason Y Zhang, Deva Ramanan, and Shubham Tulsiani. Relpose: Predicting probabilistic relative rotation for single objects in the wild. In Proc. ECCV, 2022. 2
421
+ [59] Kai Zhang, Gernot Riegler, Noah Snavely, and Vladlen Koltun. Nerf++: Analyzing and improving neural radiance fields. arXiv preprint arXiv:2010.07492, 2020. 2
422
+
423
+ [60] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proc. CVPR, 2018. 5
424
+ [61] Zihan Zhu, Songyou Peng, Viktor Larsson, Weiwei Xu, Hujun Bao, Zhaopeng Cui, Martin R Oswald, and Marc Pollefeys. Nice-slam: Neural implicit scalable encoding for slam. In Proc. CVPR, 2022. 1
2023/RobustNeRF_ Ignoring Distractors With Robust Losses/images.zip ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:31a5a89794b65d4efe3631c87d64a96724ef42cbe51d9e1b266bf9834edb78c5
3
+ size 741443
2023/RobustNeRF_ Ignoring Distractors With Robust Losses/layout.json ADDED
The diff for this file is too large to render. See raw diff
 
2023/Role of Transients in Two-Bounce Non-Line-of-Sight Imaging/dd8b3d32-1f20-46bb-9a36-d731b9d46ed2_content_list.json ADDED
@@ -0,0 +1,1589 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ [
2
+ {
3
+ "type": "text",
4
+ "text": "Role of Transients in Two-Bounce Non-Line-of-Sight Imaging",
5
+ "text_level": 1,
6
+ "bbox": [
7
+ 174,
8
+ 130,
9
+ 795,
10
+ 154
11
+ ],
12
+ "page_idx": 0
13
+ },
14
+ {
15
+ "type": "text",
16
+ "text": "Siddharth Somasundaram<sup>1</sup> Akshit Dave<sup>2</sup> Connor Henley<sup>1</sup> Ashok Veeraraghavan<sup>2</sup> Ramesh Raskar<sup>1</sup>",
17
+ "bbox": [
18
+ 240,
19
+ 179,
20
+ 720,
21
+ 215
22
+ ],
23
+ "page_idx": 0
24
+ },
25
+ {
26
+ "type": "text",
27
+ "text": "<sup>1</sup>Massachusetts Institute of Technology <sup>2</sup>Rice University",
28
+ "bbox": [
29
+ 248,
30
+ 215,
31
+ 722,
32
+ 234
33
+ ],
34
+ "page_idx": 0
35
+ },
36
+ {
37
+ "type": "text",
38
+ "text": "{sidsoma,co24401,raskar}@mit.edu,{akshat,vashok}@rice.edu",
39
+ "bbox": [
40
+ 228,
41
+ 236,
42
+ 736,
43
+ 252
44
+ ],
45
+ "page_idx": 0
46
+ },
47
+ {
48
+ "type": "image",
49
+ "img_path": "images/7aeab3f60d5fffa2bd2ad3467548e8845c54e83e6be1cc168a60191a04250c76.jpg",
50
+ "image_caption": [
51
+ "(a) Two-Bounce NLOS"
52
+ ],
53
+ "image_footnote": [],
54
+ "bbox": [
55
+ 86,
56
+ 286,
57
+ 276,
58
+ 472
59
+ ],
60
+ "page_idx": 0
61
+ },
62
+ {
63
+ "type": "image",
64
+ "img_path": "images/118345b65ad62ac1f6b1a164a33c258b22fb20ad950318959ea43658f57ef8d0.jpg",
65
+ "image_caption": [
66
+ "(b) Application Settings"
67
+ ],
68
+ "image_footnote": [],
69
+ "bbox": [
70
+ 279,
71
+ 286,
72
+ 447,
73
+ 472
74
+ ],
75
+ "page_idx": 0
76
+ },
77
+ {
78
+ "type": "image",
79
+ "img_path": "images/52c0223be557335de100b98aeb25d634c86691168774718254326334a72d2710.jpg",
80
+ "image_caption": [
81
+ "(c) Experimental Setup",
82
+ "Figure 1. Two-bounce Transient NLOS Imaging. (a) We propose using two-bounce transients for NLOS imaging. Two-bounce signals capture information about the shadows in a scene, and use of transient information can reduce the number of measurements needed for reconstruction by using multiplexed illumination. (b) Our work can be applicable in many scenarios, such as tunnels, corridors, streets, and forests. (c)-(d) We experimentally validate the use of two-bounce transients on real objects. We show that time-of-flight information enables more robust reconstructions of occluded objects while using fewer measurements."
83
+ ],
84
+ "image_footnote": [],
85
+ "bbox": [
86
+ 452,
87
+ 286,
88
+ 602,
89
+ 472
90
+ ],
91
+ "page_idx": 0
92
+ },
93
+ {
94
+ "type": "image",
95
+ "img_path": "images/c2882cdcc836f088655b91a89b156cfbbe64c7a1a2d68e2e585d6f6782bc3ac9.jpg",
96
+ "image_caption": [
97
+ "(d) Reconstruction with 10 Measurements"
98
+ ],
99
+ "image_footnote": [],
100
+ "bbox": [
101
+ 604,
102
+ 286,
103
+ 880,
104
+ 472
105
+ ],
106
+ "page_idx": 0
107
+ },
108
+ {
109
+ "type": "text",
110
+ "text": "Abstract",
111
+ "text_level": 1,
112
+ "bbox": [
113
+ 233,
114
+ 583,
115
+ 313,
116
+ 599
117
+ ],
118
+ "page_idx": 0
119
+ },
120
+ {
121
+ "type": "text",
122
+ "text": "The goal of non-line-of-sight (NLOS) imaging is to image objects occluded from the camera's field of view using multiply scattered light. Recent works have demonstrated the feasibility of two-bounce (2B) NLOS imaging by scanning a laser and measuring cast shadows of occluded objects in scenes with two relay surfaces. In this work, we study the role of time-of-flight (ToF) measurements, i.e. transients, in 2B-NLOS under multiplexed illumination. Specifically, we study how ToF information can reduce the number of measurements and spatial resolution needed for shape reconstruction. We present our findings with respect to trade-offs in (1) temporal resolution, (2) spatial resolution, and (3) number of image captures by studying SNR and recoverability as functions of system parameters. This leads to a formal definition of the mathematical constraints for 2B lidar. We believe that our work lays an analytical groundwork for design of future NLOS imaging systems, especially as ToF sensors become increasingly ubiquitous.",
123
+ "bbox": [
124
+ 75,
125
+ 611,
126
+ 472,
127
+ 883
128
+ ],
129
+ "page_idx": 0
130
+ },
131
+ {
132
+ "type": "text",
133
+ "text": "1. Introduction",
134
+ "text_level": 1,
135
+ "bbox": [
136
+ 501,
137
+ 583,
138
+ 630,
139
+ 599
140
+ ],
141
+ "page_idx": 0
142
+ },
143
+ {
144
+ "type": "text",
145
+ "text": "Non-line-of-sight (NLOS) imaging aims to reconstruct objects occluded from direct line of sight and has the potential to be transformative in numerous applications across autonomous driving, search and rescue, and non-invasive medical imaging [24]. The key approach is to measure light that has undergone multiple surface scattering events and computationally invert these measurements to estimate hidden geometries. Recent work used two-bounce light, as shown in Fig. 2a, to reconstruct high quality shapes behind occluders [15]. The key idea is that two-bounce light captures information about the shadows of the occluded object. By scanning the laser source at different points 1 on a relay surface and measuring multiple shadow images, it is possible to reconstruct the hidden object by computing the visual hull [20] of the measured shadows.",
146
+ "bbox": [
147
+ 496,
148
+ 608,
149
+ 890,
150
+ 835
151
+ ],
152
+ "page_idx": 0
153
+ },
154
+ {
155
+ "type": "text",
156
+ "text": "Benefits of Two-Bounce Two-bounce (2B) light can be captured using two relay surfaces on opposite sides of the hidden scene (Fig. 1a). This can occur in a variety of real-world settings such as tunnels, hallways, streets, and clit",
157
+ "bbox": [
158
+ 496,
159
+ 839,
160
+ 893,
161
+ 900
162
+ ],
163
+ "page_idx": 0
164
+ },
165
+ {
166
+ "type": "header",
167
+ "text": "CVF",
168
+ "bbox": [
169
+ 106,
170
+ 2,
171
+ 181,
172
+ 42
173
+ ],
174
+ "page_idx": 0
175
+ },
176
+ {
177
+ "type": "header",
178
+ "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.",
179
+ "bbox": [
180
+ 236,
181
+ 0,
182
+ 810,
183
+ 46
184
+ ],
185
+ "page_idx": 0
186
+ },
187
+ {
188
+ "type": "page_number",
189
+ "text": "9192",
190
+ "bbox": [
191
+ 480,
192
+ 944,
193
+ 514,
194
+ 955
195
+ ],
196
+ "page_idx": 0
197
+ },
198
+ {
199
+ "type": "image",
200
+ "img_path": "images/dac7c60efc57b3d2780b5a921c14a32785da739cb4a0e855d406b94c62a09000.jpg",
201
+ "image_caption": [
202
+ "(a) Two-Bounce NLOS Imaging"
203
+ ],
204
+ "image_footnote": [],
205
+ "bbox": [
206
+ 80,
207
+ 88,
208
+ 267,
209
+ 200
210
+ ],
211
+ "page_idx": 1
212
+ },
213
+ {
214
+ "type": "image",
215
+ "img_path": "images/afdd89e2ce005569ede0fd9f3f24ea328c4a24daef53b9947c7063d6f5e8f52d.jpg",
216
+ "image_caption": [
217
+ "(b) Two-Bounce Transient NLOS",
218
+ "Figure 2. Two-Bounce NLOS Imaging. (a) Two-bounce NLOS imaging performs 3D reconstruction by using information contained in the shadows of the hidden object when illuminated by one virtual source 1 at a time. (b) In this work, we study the benefits of using transient information for two-bounce NLOS imaging under the presence of multiplexed illumination."
219
+ ],
220
+ "image_footnote": [],
221
+ "bbox": [
222
+ 269,
223
+ 88,
224
+ 467,
225
+ 200
226
+ ],
227
+ "page_idx": 1
228
+ },
229
+ {
230
+ "type": "text",
231
+ "text": "tered environments like forests (Fig. 1b). For example, imagine an autonomous vehicle in a tunnel being able to see ahead of the car in front of it by using two-bounce signals from the sides of the tunnel. Two-bounce light also helps alleviate the signal-to-noise ratio (SNR) issue in 3B-NLOS imaging, since 3B signals experience significant signal attenuation at each surface scatter event. Furthermore, 2B light captures information about the cast shadows of the hidden object instead of reflectance, as in the 3B case. As a result, the SNR of 2B signals is independent of the object's albedo and can therefore robustly image darker objects.",
232
+ "bbox": [
233
+ 75,
234
+ 339,
235
+ 467,
236
+ 503
237
+ ],
238
+ "page_idx": 1
239
+ },
240
+ {
241
+ "type": "text",
242
+ "text": "Contributions In this work, we propose using two-bounce transients for NLOS imaging to enable the use of multiplexed illumination. 3B-NLOS transient imaging systems have been analyzed previously, but we believe we are the first to analyze the use of multiplexed illumination and transient measurements for 2B-NLOS (Fig. 2b). Our key insight is that multiplexed illumination enables measurement of occluded objects with fewer image captures, and 2B transients enable demultiplexing shadows, as shown in Fig. 3. The robust SNR and promise of few-shot capture afforded by 2B transients makes the idea an important direction for few-shot NLOS imaging in general environments. We summarize our contributions below.",
243
+ "bbox": [
244
+ 75,
245
+ 511,
246
+ 467,
247
+ 705
248
+ ],
249
+ "page_idx": 1
250
+ },
251
+ {
252
+ "type": "list",
253
+ "sub_type": "text",
254
+ "list_items": [
255
+ "- Model: We present new insights into the algebraic formulations of two-bounce lidar for NLOS imaging.",
256
+ "- Analysis: We analyze tradeoffs with spatial resolution, temporal resolution, and number of image captures by using quantitative metrics such as SNR and recoverability, as well as qualitative validation from real and simulated results."
257
+ ],
258
+ "bbox": [
259
+ 96,
260
+ 720,
261
+ 467,
262
+ 839
263
+ ],
264
+ "page_idx": 1
265
+ },
266
+ {
267
+ "type": "text",
268
+ "text": "Although we envision the ideas introduced here to inspire future works in single-shot NLOS imaging, we do not claim this as a contribution in this paper.",
269
+ "bbox": [
270
+ 76,
271
+ 854,
272
+ 467,
273
+ 900
274
+ ],
275
+ "page_idx": 1
276
+ },
277
+ {
278
+ "type": "text",
279
+ "text": "Scope of This Work All data used for this work is from simulation and a single-pixel SPAD sensor. We do not physically realize few-shot results in this paper due to limited availability of high-resolution SPAD arrays presently. Instead, we emulate SPAD array measurements by scanning a single-pixel SPAD across the field of view and and emulate multiplexed illumination by scanning a laser spot, acquiring individual transient images, and summing the images in post-processing. These measurements, however, are more conducive to our goal of performing analysis that explores the landscape of possibilities in NLOS imaging with advances in ToF sensors. We believe that the ideas and analysis introduced in this work will be increasingly relevant as SPAD array technology matures [19,26,44].",
280
+ "bbox": [
281
+ 496,
282
+ 90,
283
+ 890,
284
+ 303
285
+ ],
286
+ "page_idx": 1
287
+ },
288
+ {
289
+ "type": "text",
290
+ "text": "2. Related Work",
291
+ "text_level": 1,
292
+ "bbox": [
293
+ 500,
294
+ 316,
295
+ 640,
296
+ 330
297
+ ],
298
+ "page_idx": 1
299
+ },
300
+ {
301
+ "type": "text",
302
+ "text": "Non-Line-of-Sight Imaging",
303
+ "text_level": 1,
304
+ "bbox": [
305
+ 500,
306
+ 340,
307
+ 714,
308
+ 357
309
+ ],
310
+ "page_idx": 1
311
+ },
312
+ {
313
+ "type": "text",
314
+ "text": "ToF Methods: The non-line-of-sight imaging (NLOS) problem aims to see around a corner by measuring light reflected off intermediary surfaces. A large body of work models the response of a hidden scene to a temporal pulse of light and generally perform a variant of ellipsoidal tomography or solve a linear inverse problem [3, 5, 11, 12, 18, 39]. The temporal profiles can probe features of the hidden space, such as Fermat paths [42], scene impulse response [23], and individual voxels [28]. Other works derive efficient algorithms using a confocal setup [2, 22, 27] and data-driven techniques [8]. NLOS has also been demonstrated with amplitude modulated continuous wave (AMCW) ToF by exploiting sparsity and fitting measurements to a Lambertian reflectance model [13, 16].",
315
+ "bbox": [
316
+ 496,
317
+ 363,
318
+ 890,
319
+ 575
320
+ ],
321
+ "page_idx": 1
322
+ },
323
+ {
324
+ "type": "text",
325
+ "text": "Other Active Methods: Chen et al. use steady-state (i.e. time-integrated) intensity information by leveraging partial directionality of diffuse reflectors [7]. Radio wavelengths can also be used to image through walls and/or around them [1, 32, 43]. There has also been significant progress in exploiting speckle correlations [9, 17, 25, 34] and synthetic wave holography [40, 41] for NLOS imaging. Henley et al. demonstrate the use of two-bounce light (i.e. shadows) for NLOS using intensity sensors [15]. An extension of this work uses ToF constraints on two-bounce pathlengths to isolate shadows from ambient background and higher-order inter-reflections [14]. However multiplexing was not used.",
326
+ "bbox": [
327
+ 496,
328
+ 579,
329
+ 890,
330
+ 760
331
+ ],
332
+ "page_idx": 1
333
+ },
334
+ {
335
+ "type": "text",
336
+ "text": "Passive Methods: Passive methods make use of optical information already present in the scene without an active source. Many passive approaches exploit polarization cues [10,36], ambient thermal light [21,36], shadows [35], multiview reflections [38], and neural networks [33,37] to reconstruct or make inferences about the scene outside the camera's field of view. Occlusion-assisted imaging makes use of information present at sharp discontinuities in LOS and NLOS scenes by turning them into virtual pinholes [4,30] or",
337
+ "bbox": [
338
+ 496,
339
+ 763,
340
+ 890,
341
+ 900
342
+ ],
343
+ "page_idx": 1
344
+ },
345
+ {
346
+ "type": "page_number",
347
+ "text": "9193",
348
+ "bbox": [
349
+ 480,
350
+ 944,
351
+ 514,
352
+ 955
353
+ ],
354
+ "page_idx": 1
355
+ },
356
+ {
357
+ "type": "table",
358
+ "img_path": "images/ff5b994feed9ce68024242b3f89d30db7eb411fd529b7290592fedf6504f3d38.jpg",
359
+ "table_caption": [],
360
+ "table_footnote": [],
361
+ "table_body": "<table><tr><td rowspan=\"2\">NLOS Imaging Domain</td><td colspan=\"3\">Key Components</td><td colspan=\"3\">Key Features</td></tr><tr><td>Assisting Surfaces</td><td>Sensor Type</td><td>Light Source</td><td>Number of Captures</td><td>Recovered Dimensions</td><td>Signal Quality</td></tr><tr><td rowspan=\"2\">Passive Occlusion-Aided [4,31,33]</td><td>Relay Wall/Floor</td><td rowspan=\"2\">Intensity</td><td rowspan=\"2\">Passive</td><td rowspan=\"2\">Low</td><td rowspan=\"2\">2D</td><td rowspan=\"2\">Low (strong background)</td></tr><tr><td>Edge Occluders</td></tr><tr><td rowspan=\"2\">Three-Bounce Transient [18,27,30,42]</td><td>Relay Wall/Floor</td><td rowspan=\"2\">ToF</td><td rowspan=\"2\">Active (Scan)</td><td rowspan=\"2\">High</td><td rowspan=\"2\">3D</td><td rowspan=\"2\">Medium (3-bounce)</td></tr><tr><td>Edge Occluders</td></tr><tr><td>Two-Bounce Transient (Ours)</td><td>Two LOS Walls</td><td>ToF</td><td>Active (Flash)</td><td>Low</td><td>3D</td><td>High (2-bounce)</td></tr></table>",
362
+ "bbox": [
363
+ 104,
364
+ 88,
365
+ 867,
366
+ 272
367
+ ],
368
+ "page_idx": 2
369
+ },
370
+ {
371
+ "type": "text",
372
+ "text": "Table 1. Comparison of NLOS Methods. Passive approaches require fewer measurements but can either recover up to two dimensions or are strongly affected by ambient noise. Works exploiting three-bounce light paths demonstrate 3D imaging around the corner but require a bounce on the hidden surface, reducing the signal quality. We analyze exploiting two-bounce transients for NLOS imaging.",
373
+ "bbox": [
374
+ 75,
375
+ 280,
376
+ 893,
377
+ 324
378
+ ],
379
+ "page_idx": 2
380
+ },
381
+ {
382
+ "type": "text",
383
+ "text": "virtual pinspecks [31]. Intensity-only methods are difficult to scale to scenarios with arbitrary lighting conditions.",
384
+ "bbox": [
385
+ 76,
386
+ 349,
387
+ 470,
388
+ 380
389
+ ],
390
+ "page_idx": 2
391
+ },
392
+ {
393
+ "type": "text",
394
+ "text": "3. Forward and Adjoint Operators",
395
+ "text_level": 1,
396
+ "bbox": [
397
+ 76,
398
+ 404,
399
+ 372,
400
+ 422
401
+ ],
402
+ "page_idx": 2
403
+ },
404
+ {
405
+ "type": "text",
406
+ "text": "3.1. Problem Setup",
407
+ "text_level": 1,
408
+ "bbox": [
409
+ 76,
410
+ 436,
411
+ 228,
412
+ 454
413
+ ],
414
+ "page_idx": 2
415
+ },
416
+ {
417
+ "type": "text",
418
+ "text": "We first introduce terminology and notation for our problem formulation. The object of interest lies behind an occluder and is flanked on two sides by planar Lambertian relay surfaces. Without loss of generality, the right wall is the illumination wall $\\mathcal{L}$ and the left wall is the observation wall $\\mathcal{O}$ . Light incident on the illumination wall at a point $l$ is a virtual source. Each virtual source diffusely radiates light towards the observation wall. Light will arrive at a virtual detector $s$ from $l$ if the object doesn't lie along the ray connecting $s$ and $l$ , as shown in Fig. 2a. If the object obstructs this ray, $s$ is said to lie in the shadow of the object. We assume that every virtual source and virtual detector is visible to the laser and camera field of view, respectively. The hidden scene is discretized into a voxel grid described by a binary occupancy (or transparency) function. This implicitly assumes that we are reconstructing opaque surfaces. We neglect the effect of higher order bounces (e.g. interreflections, 3B light) because SPADs can temporally gate out the longer ToF of 3B light. Furthermore, the intensity of higher order bounces will be much lower than that of two-bounce light because all surfaces are assumed to be diffuse.",
419
+ "bbox": [
420
+ 75,
421
+ 469,
422
+ 472,
423
+ 786
424
+ ],
425
+ "page_idx": 2
426
+ },
427
+ {
428
+ "type": "text",
429
+ "text": "We denote light transients as ToF measurements for light that arrives at the virtual detector and shadow transients as ToF measurements for light that didn't arrive at the virtual detector because of the object's shadow. Both of these are two-bounce transients. We show that modeling the problem in terms of shadow transients enables reconstruction of the hidden objects shape by solving a linear inverse problem.",
430
+ "bbox": [
431
+ 75,
432
+ 794,
433
+ 468,
434
+ 902
435
+ ],
436
+ "page_idx": 2
437
+ },
438
+ {
439
+ "type": "image",
440
+ "img_path": "images/6f41010c71b9b4e84377d5a0bb51f56fd62c5842531823a5568b6a7bf5682fd8.jpg",
441
+ "image_caption": [
442
+ "Figure 3. Time-of-Flight Information for Demultiplexing Shadows. Assume the point $s$ lies in shadow. When using multiplexed illumination, there exists an ambiguity as to whether $s$ lies in the shadow of $l_1$ , $l_2$ , or both. However, the paths $L - l_1 - s - C$ and $L - l_2 - s - C$ have different pathlengths, suggesting the usefulness of ToF information for demultiplexing shadows."
443
+ ],
444
+ "image_footnote": [],
445
+ "bbox": [
446
+ 545,
447
+ 347,
448
+ 844,
449
+ 536
450
+ ],
451
+ "page_idx": 2
452
+ },
453
+ {
454
+ "type": "text",
455
+ "text": "3.2. Linear Forward Model",
456
+ "text_level": 1,
457
+ "bbox": [
458
+ 500,
459
+ 650,
460
+ 714,
461
+ 666
462
+ ],
463
+ "page_idx": 2
464
+ },
465
+ {
466
+ "type": "text",
467
+ "text": "Light Transients Consider the case where the occluded scene is empty. The measured transient is then given by",
468
+ "bbox": [
469
+ 500,
470
+ 672,
471
+ 890,
472
+ 705
473
+ ],
474
+ "page_idx": 2
475
+ },
476
+ {
477
+ "type": "equation",
478
+ "text": "\n$$\ni _ {0} (t; \\mathbf {s}) = \\alpha \\iint_ {\\mathcal {L}} \\frac {\\delta (\\| \\mathbf {l} - \\mathbf {g} \\| + \\| \\mathbf {l} - \\mathbf {s} \\| - c t)}{\\| \\mathbf {l} - \\mathbf {s} \\| ^ {2}} d \\mathbf {l}, \\tag {1}\n$$\n",
479
+ "text_format": "latex",
480
+ "bbox": [
481
+ 537,
482
+ 709,
483
+ 890,
484
+ 743
485
+ ],
486
+ "page_idx": 2
487
+ },
488
+ {
489
+ "type": "text",
490
+ "text": "where $\\mathbf{g}$ is the location of the laser, $c$ is the speed of light, and $\\alpha$ is the intensity of the laser. We refer to $i_0$ as the empty transient. For simplicity, we ignore $\\alpha$ for the rest of the derivations in this section and assume that the length between the virtual detector and camera is calibrated or known. The numerator of the integrand",
491
+ "bbox": [
492
+ 496,
493
+ 748,
494
+ 893,
495
+ 839
496
+ ],
497
+ "page_idx": 2
498
+ },
499
+ {
500
+ "type": "equation",
501
+ "text": "\n$$\ni (t; \\mathbf {s}, \\mathbf {l}) = \\delta (\\| \\mathbf {l} - \\mathbf {g} \\| + \\| \\mathbf {l} - \\mathbf {s} \\| - c t) \\tag {2}\n$$\n",
502
+ "text_format": "latex",
503
+ "bbox": [
504
+ 568,
505
+ 845,
506
+ 890,
507
+ 864
508
+ ],
509
+ "page_idx": 2
510
+ },
511
+ {
512
+ "type": "text",
513
+ "text": "describes the equation of an ellipsoid, with foci at $\\mathbf{g}$ and s and major axis length $ct$ . When an object is present in the",
514
+ "bbox": [
515
+ 498,
516
+ 869,
517
+ 890,
518
+ 901
519
+ ],
520
+ "page_idx": 2
521
+ },
522
+ {
523
+ "type": "page_number",
524
+ "text": "9194",
525
+ "bbox": [
526
+ 480,
527
+ 944,
528
+ 514,
529
+ 955
530
+ ],
531
+ "page_idx": 2
532
+ },
533
+ {
534
+ "type": "image",
535
+ "img_path": "images/dc5705168e811c254e22b1d68ae3ba330bd0192dfe1a920cfa8f9180350463ca.jpg",
536
+ "image_caption": [
537
+ "Figure 4. Light and Shadow Transients. Three spacetime curves are depicted. (left) Empty transient $i_0$ with no hidden object, (middle) light transient $i_l$ with an object, and (right) shadow transient $i_s = i_0 - i_l$ . Switching from light transients $i_l$ to shadow transients $i_s$ enables derivation of a linear forward model."
538
+ ],
539
+ "image_footnote": [],
540
+ "bbox": [
541
+ 89,
542
+ 89,
543
+ 459,
544
+ 154
545
+ ],
546
+ "page_idx": 3
547
+ },
548
+ {
549
+ "type": "text",
550
+ "text": "occluded scene, the measured light transient is given by",
551
+ "bbox": [
552
+ 76,
553
+ 261,
554
+ 444,
555
+ 276
556
+ ],
557
+ "page_idx": 3
558
+ },
559
+ {
560
+ "type": "equation",
561
+ "text": "\n$$\ni _ {l} (t; \\mathbf {s}) = \\iint_ {\\mathcal {L}} V _ {\\mathbf {s}, \\mathbf {l}} \\cdot \\frac {i (t ; \\mathbf {s} , \\mathbf {l})}{\\| \\mathbf {l} - \\mathbf {s} \\| ^ {2}} d \\mathbf {l}, \\tag {3}\n$$\n",
562
+ "text_format": "latex",
563
+ "bbox": [
564
+ 163,
565
+ 287,
566
+ 468,
567
+ 321
568
+ ],
569
+ "page_idx": 3
570
+ },
571
+ {
572
+ "type": "text",
573
+ "text": "where $V_{\\mathbf{s},1}$ is the visibility between s and l.",
574
+ "bbox": [
575
+ 76,
576
+ 332,
577
+ 359,
578
+ 348
579
+ ],
580
+ "page_idx": 3
581
+ },
582
+ {
583
+ "type": "equation",
584
+ "text": "\n$$\nV _ {\\mathbf {s}, \\mathbf {l}} = \\exp \\left(\\oint_ {\\mathbf {l}} ^ {\\mathbf {s}} \\ln \\left(f _ {t} (\\mathbf {x})\\right) d \\mathbf {x}\\right) \\tag {4}\n$$\n",
585
+ "text_format": "latex",
586
+ "bbox": [
587
+ 166,
588
+ 353,
589
+ 468,
590
+ 387
591
+ ],
592
+ "page_idx": 3
593
+ },
594
+ {
595
+ "type": "text",
596
+ "text": "where Eq. (4) computes the product integral along the line connecting $\\mathbf{l}$ and $\\mathbf{s}$ and $f_{t}(\\mathbf{x})\\in \\{0,1\\}$ is the transparency function of voxel $\\mathbf{x}$ in the occluded region.",
597
+ "bbox": [
598
+ 76,
599
+ 398,
600
+ 468,
601
+ 444
602
+ ],
603
+ "page_idx": 3
604
+ },
605
+ {
606
+ "type": "text",
607
+ "text": "Shadow Transients The non-linearity of Eq. (4) makes model inversion challenging with light transients. Casting the formulation from light space-time to shadow space-time enables a linear formulation of the problem. A shadow transient at $s$ is given by",
608
+ "bbox": [
609
+ 76,
610
+ 450,
611
+ 468,
612
+ 527
613
+ ],
614
+ "page_idx": 3
615
+ },
616
+ {
617
+ "type": "equation",
618
+ "text": "\n$$\ni _ {s} (t; \\mathbf {s}) = i _ {0} (t; \\mathbf {s}) - i _ {l} (t; \\mathbf {s}). \\tag {5}\n$$\n",
619
+ "text_format": "latex",
620
+ "bbox": [
621
+ 179,
622
+ 540,
623
+ 468,
624
+ 556
625
+ ],
626
+ "page_idx": 3
627
+ },
628
+ {
629
+ "type": "text",
630
+ "text": "Fig. 4 illustrates the difference between light transients and shadow transients. Shadow transients directly probe voxel occupancy, whereas light transients probe for voxel transparency. For example, if the occluded scene is empty except for voxel $\\mathbf{x}$ , the shadow transient is given by",
631
+ "bbox": [
632
+ 76,
633
+ 569,
634
+ 468,
635
+ 645
636
+ ],
637
+ "page_idx": 3
638
+ },
639
+ {
640
+ "type": "equation",
641
+ "text": "\n$$\ni _ {s} (t; \\mathbf {s}, \\mathbf {x}) = \\frac {\\delta (\\| \\mathbf {l} (\\mathbf {x} , \\mathbf {s}) - \\mathbf {g} \\| + \\| \\mathbf {l} (\\mathbf {x} , \\mathbf {s}) - \\mathbf {s} \\| - c t)}{\\| \\mathbf {l} (\\mathbf {x} , \\mathbf {s}) - \\mathbf {s} \\| ^ {2}}, \\tag {6}\n$$\n",
642
+ "text_format": "latex",
643
+ "bbox": [
644
+ 89,
645
+ 657,
646
+ 468,
647
+ 690
648
+ ],
649
+ "page_idx": 3
650
+ },
651
+ {
652
+ "type": "text",
653
+ "text": "where $\\mathbf{l}(\\mathbf{x},\\mathbf{s})$ is a point on the illumination wall that lies along the ray connecting $\\mathbf{x}$ and $\\mathbf{s}$ . If the wall doesn't intersect this ray, $i_s(t;\\mathbf{x},\\mathbf{s}) = 0$ .",
654
+ "bbox": [
655
+ 76,
656
+ 702,
657
+ 467,
658
+ 748
659
+ ],
660
+ "page_idx": 3
661
+ },
662
+ {
663
+ "type": "text",
664
+ "text": "Linear Relaxation Now, consider the general case where a single opaque object lies in the hidden scene. We can describe the shadow transient as",
665
+ "bbox": [
666
+ 76,
667
+ 755,
668
+ 468,
669
+ 799
670
+ ],
671
+ "page_idx": 3
672
+ },
673
+ {
674
+ "type": "equation",
675
+ "text": "\n$$\ni _ {s} (t; \\mathbf {s}) = \\iint_ {\\mathcal {L}} O _ {\\mathbf {s}, \\mathbf {l}} \\cdot \\frac {i (t ; \\mathbf {s} , \\mathbf {l})}{\\| \\mathbf {l} - \\mathbf {s} \\| ^ {2}} d \\mathbf {l}, \\tag {7}\n$$\n",
676
+ "text_format": "latex",
677
+ "bbox": [
678
+ 160,
679
+ 809,
680
+ 468,
681
+ 843
682
+ ],
683
+ "page_idx": 3
684
+ },
685
+ {
686
+ "type": "text",
687
+ "text": "where $O_{\\mathbf{s},1} \\in \\{0,1\\}$ is the occlusion function (denotes if the ray between 1 and s is occluded). It can be expressed as the indicator function",
688
+ "bbox": [
689
+ 76,
690
+ 854,
691
+ 468,
692
+ 898
693
+ ],
694
+ "page_idx": 3
695
+ },
696
+ {
697
+ "type": "equation",
698
+ "text": "\n$$\n\\begin{array}{l} O _ {\\mathbf {s}, \\mathrm {l}} = \\left\\{ \\begin{array}{l l} 1, & \\text {i f} \\oint_ {\\mathrm {l}} ^ {\\mathbf {s}} f _ {o} (\\mathbf {x}) d \\mathbf {x} > 0 \\\\ 0, & \\text {o t h e r w i s e} \\end{array} \\right. \\tag {8} \\\\ \\approx \\oint_ {1} ^ {\\mathbf {s}} f _ {o} (\\mathbf {x}) d \\mathbf {x}, \\\\ \\end{array}\n$$\n",
699
+ "text_format": "latex",
700
+ "bbox": [
701
+ 584,
702
+ 84,
703
+ 890,
704
+ 159
705
+ ],
706
+ "page_idx": 3
707
+ },
708
+ {
709
+ "type": "text",
710
+ "text": "where $f_{o}(\\mathbf{x})$ is the voxel occupancy function of the hidden scene. The linear relaxation in Eq. (8) is possible because NLOS reconstructions typically only consider ToF information, but not radiometric information. The above formulation will preserve the structure of the streak images in spacetime, even if the intensities are overestimated. Combining Eq. (5), (7), and (8), we obtain",
711
+ "bbox": [
712
+ 498,
713
+ 170,
714
+ 890,
715
+ 275
716
+ ],
717
+ "page_idx": 3
718
+ },
719
+ {
720
+ "type": "equation",
721
+ "text": "\n$$\ni _ {0} (t; \\mathbf {s}) - i _ {l} (t; \\mathbf {s}) = \\iint_ {\\mathcal {L}} \\frac {i (t ; \\mathbf {s} , \\mathbf {l})}{\\| \\mathbf {l} - \\mathbf {s} \\| ^ {2}} \\cdot \\left(\\oint_ {\\mathbf {l}} ^ {\\mathbf {s}} f _ {o} (\\mathbf {x}) d \\mathbf {x}\\right) d \\mathbf {l}, \\tag {9}\n$$\n",
722
+ "text_format": "latex",
723
+ "bbox": [
724
+ 506,
725
+ 286,
726
+ 890,
727
+ 320
728
+ ],
729
+ "page_idx": 3
730
+ },
731
+ {
732
+ "type": "text",
733
+ "text": "where we solve for $f_{o}(\\mathbf{x})$ . We can estimate $i_0$ from the scene geometry and $i_l$ is the measured transient.",
734
+ "bbox": [
735
+ 498,
736
+ 330,
737
+ 890,
738
+ 361
739
+ ],
740
+ "page_idx": 3
741
+ },
742
+ {
743
+ "type": "text",
744
+ "text": "Discrete Beam Illumination Discretizing $\\mathcal{L}$ as $K$ discrete laser spots and the hidden space into a voxel grid, we can write Eq. (9) as",
745
+ "bbox": [
746
+ 498,
747
+ 364,
748
+ 890,
749
+ 411
750
+ ],
751
+ "page_idx": 3
752
+ },
753
+ {
754
+ "type": "equation",
755
+ "text": "\n$$\ni _ {s} [ t; \\mathbf {s} ] = \\sum_ {k = 1} ^ {K} \\frac {\\delta [ \\| \\mathbf {l} _ {k} - \\mathbf {g} \\| + \\| \\mathbf {l} _ {k} - \\mathbf {s} \\| - c t ]}{\\| \\mathbf {l} _ {k} - \\mathbf {s} \\| ^ {2}} \\cdot \\left(\\sum_ {\\mathbf {x} \\in \\mathcal {P} _ {\\mathbf {l} _ {k}, \\mathbf {s}}} f _ {o} [ \\mathbf {x} ]\\right), \\tag {10}\n$$\n",
756
+ "text_format": "latex",
757
+ "bbox": [
758
+ 500,
759
+ 422,
760
+ 900,
761
+ 479
762
+ ],
763
+ "page_idx": 3
764
+ },
765
+ {
766
+ "type": "text",
767
+ "text": "where $\\mathcal{P}_{1,\\mathrm{s}}$ denotes the set of voxels lying along the path between 1 and s. Eq. (10) can be expressed in matrix form",
768
+ "bbox": [
769
+ 500,
770
+ 479,
771
+ 888,
772
+ 510
773
+ ],
774
+ "page_idx": 3
775
+ },
776
+ {
777
+ "type": "equation",
778
+ "text": "\n$$\n\\mathbf {I} _ {s} = \\mathbf {P B f} = \\mathbf {A f}, \\tag {11}\n$$\n",
779
+ "text_format": "latex",
780
+ "bbox": [
781
+ 633,
782
+ 523,
783
+ 890,
784
+ 539
785
+ ],
786
+ "page_idx": 3
787
+ },
788
+ {
789
+ "type": "text",
790
+ "text": "where $\\mathbf{I}_s\\in \\mathbb{R}^{n_u n_v n_t}$ is the vectorized space-time measurements, $\\mathbf{f}\\in \\mathbb{R}^{n_x n_y n_z}$ is the voxel occupancies, and $\\mathbf{A}\\in \\mathbb{R}^{n_u n_v n_t\\times n_x n_y n_z}$ is the measurement operator mapping voxel occupancies to space-time measurements. $n_u$ and $n_v$ are the number of pixels along the $u$ and $v$ direction, and $n_t$ is the number of timing bins. $n_x,n_y,$ and $n_z$ are the number of voxels in the $x,y,$ and $z$ directions.",
791
+ "bbox": [
792
+ 498,
793
+ 551,
794
+ 890,
795
+ 655
796
+ ],
797
+ "page_idx": 3
798
+ },
799
+ {
800
+ "type": "text",
801
+ "text": "$\\mathbf{P}$ and $\\mathbf{B}$ model the temporal and spatial relationships, respectively, between virtual sources, virtual detectors, and voxels. $\\mathbf{P} \\in \\mathbb{R}^{n_u n_v n_t \\times n_u n_v K}$ is a block diagonal matrix, with the column space of the $i$ th block $\\mathbf{P}^i \\in \\mathbb{R}^{n_t \\times K}$ containing the corresponding $\\delta(\\cdot)$ functions from Eq. (2) for virtual detector $\\mathbf{s}_i$ and all $K$ virtual sources. $\\mathbf{B} \\in \\mathbb{R}^{n_u n_v K \\times n_x n_y n_z}$ is composed of $n_u n_v$ vertical sub-blocks. Each entry $\\mathbf{B}_{kj}^i \\in \\{0,1\\}$ of the $i$ th sub-block $\\mathbf{B}^i \\in \\mathbb{R}^{K \\times n_x n_y n_z}$ of sub-block $i$ indicates whether voxel $j$ lies along the line connecting virtual detector $\\mathbf{s}_i$ and virtual source $\\mathbf{l}_k$ . The multiplication $\\mathbf{Bf}$ performs the line integral in Eq. (8). The $k$ th column vector of $\\mathbf{P}$ is the space-time impulse response (STIR) to the $k$ th illumination source when the hidden scene is completely occupied (note that we are dealing with shadow transients, not light transients). The $i$ th column of $\\mathbf{A}$ forms the expected space-time measurements",
802
+ "bbox": [
803
+ 498,
804
+ 656,
805
+ 890,
806
+ 900
807
+ ],
808
+ "page_idx": 3
809
+ },
810
+ {
811
+ "type": "page_number",
812
+ "text": "9195",
813
+ "bbox": [
814
+ 480,
815
+ 944,
816
+ 514,
817
+ 955
818
+ ],
819
+ "page_idx": 3
820
+ },
821
+ {
822
+ "type": "text",
823
+ "text": "if only the $i$ th voxel is occupied (i.e. the STIR to a single voxel). We neglect the $r^2$ falloff term in our reconstructions. Additional explanation of the A matrix is provided in the supplementary. The linear model in Eq. (11) resembles that of 3B-bounce NLOS, where $\\mathbf{A}$ is a measurement operator that maps voxels $\\mathbf{f}$ to measurements $\\mathbf{I}$ . However, in 3B-NLOS, the information is encoded into the reflectance of the object, whereas in 2B-NLOS, the information is encoded into the shadows. Therefore, the former is linear with respects to light transients, and the latter is linear with respect to shadow transients.",
824
+ "bbox": [
825
+ 75,
826
+ 90,
827
+ 472,
828
+ 257
829
+ ],
830
+ "page_idx": 4
831
+ },
832
+ {
833
+ "type": "text",
834
+ "text": "3.3. Reconstruction Algorithm",
835
+ "text_level": 1,
836
+ "bbox": [
837
+ 76,
838
+ 268,
839
+ 316,
840
+ 286
841
+ ],
842
+ "page_idx": 4
843
+ },
844
+ {
845
+ "type": "text",
846
+ "text": "Reconstruction of $\\mathbf{f}$ can be performed by backprojection",
847
+ "bbox": [
848
+ 96,
849
+ 294,
850
+ 468,
851
+ 309
852
+ ],
853
+ "page_idx": 4
854
+ },
855
+ {
856
+ "type": "equation",
857
+ "text": "\n$$\n\\mathbf {f} _ {b p} = \\mathbf {A} ^ {\\top} \\mathbf {I} _ {s}. \\tag {12}\n$$\n",
858
+ "text_format": "latex",
859
+ "bbox": [
860
+ 227,
861
+ 316,
862
+ 468,
863
+ 335
864
+ ],
865
+ "page_idx": 4
866
+ },
867
+ {
868
+ "type": "text",
869
+ "text": "Backprojection produces a low-frequency reconstruction of $\\mathbf{f}$ . In practice, filtering $\\mathbf{f}_{bp}$ with a Laplacian filter",
870
+ "bbox": [
871
+ 76,
872
+ 345,
873
+ 470,
874
+ 378
875
+ ],
876
+ "page_idx": 4
877
+ },
878
+ {
879
+ "type": "equation",
880
+ "text": "\n$$\n\\mathbf {f} _ {f b p} = \\left(\\frac {\\partial^ {2}}{\\partial x ^ {2}} + \\frac {\\partial^ {2}}{\\partial y ^ {2}} + \\frac {\\partial^ {2}}{\\partial z ^ {2}}\\right) \\mathbf {f} _ {b p} \\tag {13}\n$$\n",
881
+ "text_format": "latex",
882
+ "bbox": [
883
+ 158,
884
+ 390,
885
+ 468,
886
+ 431
887
+ ],
888
+ "page_idx": 4
889
+ },
890
+ {
891
+ "type": "text",
892
+ "text": "has been shown to yield reasonably good results in ellipsoidal tomography in 3-bounce NLOS and line-based tomography in medical imaging. Our linear formulation also enables reconstruction with priors. We recognize that these optimization-based methods would improve reconstruction over unfiltered backprojection, which is equivalent to a low-frequency version of voxel carving in our setup. However, we restrict our focus to unfiltered backprojection because of its convenient relationship with the Gram matrix, which enables further analysis of our method in Sec. 4.2.",
893
+ "bbox": [
894
+ 75,
895
+ 443,
896
+ 468,
897
+ 594
898
+ ],
899
+ "page_idx": 4
900
+ },
901
+ {
902
+ "type": "text",
903
+ "text": "4. Design and Analysis",
904
+ "text_level": 1,
905
+ "bbox": [
906
+ 76,
907
+ 611,
908
+ 272,
909
+ 628
910
+ ],
911
+ "page_idx": 4
912
+ },
913
+ {
914
+ "type": "text",
915
+ "text": "In this section, we analyze how the performance of the proposed two-bounce transient system varies with system parameters and the geometry of LOS relay surfaces. We also analyze the impact of temporal resolution, spatial resolution, and number of measurements on recoverability. The amount of multiplexing in the scene is inversely related to the number of measurements. For example, if 60 laser spots are used and 15 measurements are captured, then each image contains 4 multiplexed shadows. We use SNR, coherence, and FWHM of the PSF to quantify performance.",
916
+ "bbox": [
917
+ 75,
918
+ 637,
919
+ 468,
920
+ 789
921
+ ],
922
+ "page_idx": 4
923
+ },
924
+ {
925
+ "type": "text",
926
+ "text": "4.1. Voxel SNR Analysis",
927
+ "text_level": 1,
928
+ "bbox": [
929
+ 76,
930
+ 800,
931
+ 266,
932
+ 816
933
+ ],
934
+ "page_idx": 4
935
+ },
936
+ {
937
+ "type": "text",
938
+ "text": "We quantify the noise characteristics of the measured shadow two-bounce transients using per-voxel SNR. Consider the case where only voxel $\\mathbf{x}$ is occupied in the hidden region. We derive the SNR of the number of \"shadow\" photons received in the measured shadow transient $\\mathbf{I}_s^{\\mathbf{x}}$ . In vector",
939
+ "bbox": [
940
+ 75,
941
+ 824,
942
+ 470,
943
+ 902
944
+ ],
945
+ "page_idx": 4
946
+ },
947
+ {
948
+ "type": "image",
949
+ "img_path": "images/4f3db07091b483a62612cdaf18a59324896d0484893c3c68aaf647b579b2ea8a.jpg",
950
+ "image_caption": [
951
+ "Figure 5. SNR Analysis. The goal of this analysis is to determine how scene geometry affects the SNR of the two-bounce measurement at each voxel in the hidden scene. The first two cases consider the orientations of the relay surfaces, and the third considers the field of view of the camera. SNR is highest in the middle of the hidden scene for parallel surfaces, orienting the walls closer together increases SNR due to reduced $r^2$ falloff, and SNR is highest closer to the virtual detectors for small FOV cameras."
952
+ ],
953
+ "image_footnote": [],
954
+ "bbox": [
955
+ 511,
956
+ 89,
957
+ 877,
958
+ 220
959
+ ],
960
+ "page_idx": 4
961
+ },
962
+ {
963
+ "type": "text",
964
+ "text": "form, the occupancy function can be described as $\\delta^{\\mathbf{x}}$ . From Eq. (11), the shadow transient is given by",
965
+ "bbox": [
966
+ 498,
967
+ 364,
968
+ 892,
969
+ 396
970
+ ],
971
+ "page_idx": 4
972
+ },
973
+ {
974
+ "type": "equation",
975
+ "text": "\n$$\n\\mathbf {I} _ {s} ^ {\\mathbf {x}} = \\mathbf {A} \\delta^ {\\mathbf {x}} \\tag {14}\n$$\n",
976
+ "text_format": "latex",
977
+ "bbox": [
978
+ 656,
979
+ 404,
980
+ 890,
981
+ 421
982
+ ],
983
+ "page_idx": 4
984
+ },
985
+ {
986
+ "type": "text",
987
+ "text": "and the light transients measured by the SPAD sensor are given by $\\mathbf{I}_l^{\\mathbf{x}}$ . For this analysis, we account for $r^2$ falloff because it is a physically accurate phenomenon that impacts SNR. From Eq. (5),",
988
+ "bbox": [
989
+ 498,
990
+ 429,
991
+ 892,
992
+ 491
993
+ ],
994
+ "page_idx": 4
995
+ },
996
+ {
997
+ "type": "equation",
998
+ "text": "\n$$\n\\mathbf {I} _ {l} ^ {\\mathbf {x}} = \\mathbf {I} _ {0} - \\mathbf {A} \\delta^ {\\mathbf {x}} \\tag {15}\n$$\n",
999
+ "text_format": "latex",
1000
+ "bbox": [
1001
+ 640,
1002
+ 498,
1003
+ 890,
1004
+ 515
1005
+ ],
1006
+ "page_idx": 4
1007
+ },
1008
+ {
1009
+ "type": "text",
1010
+ "text": "where $\\mathbf{I}_0$ is the empty transient and assumed to be calibrated. Assuming no SPAD pile-up distortion, the number of photons arriving at the sensor for each time bin of $\\mathbf{I}_l^{\\mathbf{x}}$ follows an independent Poisson distribution [29]. Thus, the total number of photons $P^{\\mathbf{x}}$ measured follows a Poisson distribution with the rate parameter being the sum of all the photon counts for all time bins for all pixels.",
1011
+ "bbox": [
1012
+ 496,
1013
+ 523,
1014
+ 892,
1015
+ 630
1016
+ ],
1017
+ "page_idx": 4
1018
+ },
1019
+ {
1020
+ "type": "equation",
1021
+ "text": "\n$$\nP ^ {\\mathbf {x}} \\sim \\operatorname {P o i s s o n} \\left(\\mathbf {1} ^ {T} \\left(\\mathbf {I} _ {0} - \\mathbf {A} \\delta^ {\\mathbf {x}}\\right)\\right) \\tag {16}\n$$\n",
1022
+ "text_format": "latex",
1023
+ "bbox": [
1024
+ 589,
1025
+ 637,
1026
+ 890,
1027
+ 655
1028
+ ],
1029
+ "page_idx": 4
1030
+ },
1031
+ {
1032
+ "type": "text",
1033
+ "text": "where $\\mathbf{1} \\in \\mathbb{R}^{n_u n_v n_t}$ is the one vector. Using the fact that SNR of a Poisson distribution is $\\mu / \\sigma = \\sqrt{\\lambda}$ with rate parameter $\\lambda$ , the per-voxel SNR $S^{\\mathbf{x}}$ is given by",
1034
+ "bbox": [
1035
+ 496,
1036
+ 662,
1037
+ 892,
1038
+ 709
1039
+ ],
1040
+ "page_idx": 4
1041
+ },
1042
+ {
1043
+ "type": "equation",
1044
+ "text": "\n$$\nS ^ {\\mathbf {x}} = \\sqrt {\\mathbf {1} ^ {T} \\left(\\mathbf {I} _ {0} - \\mathbf {A} \\boldsymbol {\\delta} ^ {\\mathbf {x}}\\right)} \\tag {17}\n$$\n",
1045
+ "text_format": "latex",
1046
+ "bbox": [
1047
+ 614,
1048
+ 718,
1049
+ 890,
1050
+ 742
1051
+ ],
1052
+ "page_idx": 4
1053
+ },
1054
+ {
1055
+ "type": "text",
1056
+ "text": "The per-voxel SNR is plotted in Fig. 5 for three different scenarios. The first two scenarios deal with the orientation of the walls, and the third considers the field of view of the camera. In Fig. 5a, we can see that the SNR is higher for center voxels, as more rays connecting the illumination wall and observation wall pass through the center voxels. When the relay walls are oriented towards each other (Fig. 5b), the SNR increases due to reduced $r^2$ falloff. When the camera's field of view is small compared to the laser (Fig. 5c), regions closer to the observation wall have the highest SNR.",
1057
+ "bbox": [
1058
+ 496,
1059
+ 750,
1060
+ 893,
1061
+ 901
1062
+ ],
1063
+ "page_idx": 4
1064
+ },
1065
+ {
1066
+ "type": "page_number",
1067
+ "text": "9196",
1068
+ "bbox": [
1069
+ 480,
1070
+ 944,
1071
+ 514,
1072
+ 955
1073
+ ],
1074
+ "page_idx": 4
1075
+ },
1076
+ {
1077
+ "type": "image",
1078
+ "img_path": "images/5e8c362db0744c8c9755964b4fda3be0b9a90703ab8cbc9ccf84fdd8847089e7.jpg",
1079
+ "image_caption": [
1080
+ "Figure 6. Spatially Varying PSF. In two-bounce lidar, the PSF has a double cone shape. The orientation of the double cone varies strongly as a function of the orientation and length of the relay surfaces (i.e. the virtual apertures) and the voxel position."
1081
+ ],
1082
+ "image_footnote": [],
1083
+ "bbox": [
1084
+ 102,
1085
+ 99,
1086
+ 436,
1087
+ 251
1088
+ ],
1089
+ "page_idx": 5
1090
+ },
1091
+ {
1092
+ "type": "text",
1093
+ "text": "4.2. Recoverability Analysis",
1094
+ "text_level": 1,
1095
+ "bbox": [
1096
+ 76,
1097
+ 351,
1098
+ 294,
1099
+ 368
1100
+ ],
1101
+ "page_idx": 5
1102
+ },
1103
+ {
1104
+ "type": "text",
1105
+ "text": "Resolution We now study the resolution limits of our method. In particular, we consider two key metrics in our measurement operator: mutual coherence and full width at half maximum (FWHM) of the PSF. We are interested in how (1) temporal resolution, (2) spatial resolution $N$ , and (3) number of measurements $M$ impact each of our metrics. From Eq. (12), we know that the backprojected occupancy can be obtained from the occupancy function by",
1106
+ "bbox": [
1107
+ 75,
1108
+ 375,
1109
+ 468,
1110
+ 496
1111
+ ],
1112
+ "page_idx": 5
1113
+ },
1114
+ {
1115
+ "type": "equation",
1116
+ "text": "\n$$\n\\mathbf {f} _ {b p} = \\left(\\mathbf {A} ^ {\\top} \\mathbf {A}\\right) \\mathbf {f}, \\tag {18}\n$$\n",
1117
+ "text_format": "latex",
1118
+ "bbox": [
1119
+ 217,
1120
+ 507,
1121
+ 468,
1122
+ 526
1123
+ ],
1124
+ "page_idx": 5
1125
+ },
1126
+ {
1127
+ "type": "text",
1128
+ "text": "where $\\mathbf{G} = \\mathbf{A}^{\\top}\\mathbf{A}$ is the Gram matrix. We can interpret the columns (or rows) of the Gram matrix to be the spatially varying point spread function (PSF) of the imaging system, as shown in Fig. 6. Each column $\\mathbf{G}_i$ describes how a voxel $\\mathbf{x}_i$ is blurred during backprojection reconstruction.",
1129
+ "bbox": [
1130
+ 75,
1131
+ 537,
1132
+ 468,
1133
+ 613
1134
+ ],
1135
+ "page_idx": 5
1136
+ },
1137
+ {
1138
+ "type": "text",
1139
+ "text": "We can see the effect of different imaging parameters on the PSF in Fig. 7, where we sweep different values of (1) $M$ and (2) $N$ . In particular, we note that the lobes of the PSF are suppressed when using temporal information, even with low spatial resolution and few number of image measurements. When using no temporal information, the PSF is still roughly centered at the correct location, but has large FWHM. The sharpness of the PSF when using temporal information indicates that it is better suited to applications that require finer voxel resolution with few-shot capture. In case (1) without transients, we see that the FWHM of the PSF is large along both the $x$ and $y$ direction for small values of $M$ . As $M$ approaches the number of illumination spots, the FWHM approaches the theoretical lower bound. In case (2), increased spatial resolution helps better localize the center of the PSF, but the FWHM does not improve. The transient cases are quite robust to both low number of measurements and low spatial resolution. Intuitively, higher values of $N$ in the transient case slightly help accentuate the center voxel",
1140
+ "bbox": [
1141
+ 75,
1142
+ 614,
1143
+ 472,
1144
+ 901
1145
+ ],
1146
+ "page_idx": 5
1147
+ },
1148
+ {
1149
+ "type": "text",
1150
+ "text": "by projecting more diverse rays through the center voxel.",
1151
+ "bbox": [
1152
+ 498,
1153
+ 90,
1154
+ 877,
1155
+ 107
1156
+ ],
1157
+ "page_idx": 5
1158
+ },
1159
+ {
1160
+ "type": "text",
1161
+ "text": "Mutual Coherence We wish to further quantify the impact of spatio-temporal resolution and number of measurements on the recoverability of the hidden scene. We know from compressive sensing theory that sparse recovery of the hidden scene is possible when the mutual coherence of $\\mathbf{A}$ is sufficiently small. The mutual coherence is defined as",
1162
+ "bbox": [
1163
+ 498,
1164
+ 109,
1165
+ 890,
1166
+ 200
1167
+ ],
1168
+ "page_idx": 5
1169
+ },
1170
+ {
1171
+ "type": "equation",
1172
+ "text": "\n$$\n\\mu = \\max _ {1 \\leq i \\neq j \\leq n} \\frac {\\left| \\mathbf {A} _ {i} ^ {\\top} \\mathbf {A} _ {j} \\right|}{\\left| \\mathbf {A} _ {i} \\right| \\left| \\mathbf {A} _ {j} \\right|}. \\tag {19}\n$$\n",
1173
+ "text_format": "latex",
1174
+ "bbox": [
1175
+ 612,
1176
+ 212,
1177
+ 890,
1178
+ 247
1179
+ ],
1180
+ "page_idx": 5
1181
+ },
1182
+ {
1183
+ "type": "text",
1184
+ "text": "Intuitively, the coherence measures the largest \"similarity\" or correlation between any two column vectors in measurement matrix $\\mathbf{A}$ [6]. Physically, this quantifies how similar the STIRs of two different voxels are. If $\\mu = 1$ , then two voxels have the exact same STIR, which means that there exists an ambiguity as to which voxel resulted in a certain spatio-temporal response. Alternatively, $\\mu$ can be defined as the maximum off-diagonal entry of the Gram matrix. Fig. 8 illustrates how the Gram matrix (and subsequently the mutual coherence) is affected by coarser timing resolutions. From compressive sensing theory, we know that a $K$ -sparse scene can be recovered if $K \\leq 0.5(1 + 1 / \\mu)$ , which motivates our empirical analysis of the effect of spatial resolution and temporal resolution on $\\mu$ .",
1185
+ "bbox": [
1186
+ 496,
1187
+ 256,
1188
+ 892,
1189
+ 468
1190
+ ],
1191
+ "page_idx": 5
1192
+ },
1193
+ {
1194
+ "type": "text",
1195
+ "text": "From Fig. 9, we can see the impact that spatial and temporal resolution have on mutual coherence in a single-shot setup with multiplexed illumination. Consider a scene similar to that of Fig. 6, where the walls are $2\\mathrm{m}$ apart, the walls are $3\\mathrm{m}$ long, the axial distance to the walls is $2\\mathrm{m}$ , and the voxel resolution is discretized to $2\\mathrm{cm}$ . Increasing the spatial resolution at a fixed temporal resolution appears to have little to no effect on the mutual coherence. As Kadambi et al. also point out, improvements of coherence through higher spatial resolution are limited by the size of the virtual aperture (i.e. the wall) [16]. On the other hand, the difference between $100~\\mathrm{ps}$ resolution and $10~\\mathrm{ps}$ resolution reduces the coherence from 0.9 to 0.3 for a fixed spatial resolution. Any timing resolution coarser than $\\sim 120$ ps results in a coherence of 1 for multiplexed measurements because pathlength differences for different shadows are no longer resolvable. This analysis motivates the importance of temporal resolution when using multiplexed illumination.",
1196
+ "bbox": [
1197
+ 496,
1198
+ 468,
1199
+ 892,
1200
+ 741
1201
+ ],
1202
+ "page_idx": 5
1203
+ },
1204
+ {
1205
+ "type": "text",
1206
+ "text": "5. Experiments and Results",
1207
+ "text_level": 1,
1208
+ "bbox": [
1209
+ 500,
1210
+ 753,
1211
+ 733,
1212
+ 771
1213
+ ],
1214
+ "page_idx": 5
1215
+ },
1216
+ {
1217
+ "type": "text",
1218
+ "text": "We scan a single-pixel SPAD across the field of view on the left wall to mimic the output of a SPAD array. We scan a single laser spot across the right wall, and sum up the corresponding transient images in post-processing to mimic the effect of multiplexing. Further details are provided in the supplementary. We first validate our model using simulated data in Fig. 10, where we demonstrate that we can localize four letters and reconstruct their 2D shape with just a single",
1219
+ "bbox": [
1220
+ 496,
1221
+ 779,
1222
+ 890,
1223
+ 900
1224
+ ],
1225
+ "page_idx": 5
1226
+ },
1227
+ {
1228
+ "type": "page_number",
1229
+ "text": "9197",
1230
+ "bbox": [
1231
+ 480,
1232
+ 944,
1233
+ 514,
1234
+ 955
1235
+ ],
1236
+ "page_idx": 5
1237
+ },
1238
+ {
1239
+ "type": "image",
1240
+ "img_path": "images/8917c0c3376b274e4c1aec23ce917105bd1496bf57c63b2ca47e202dafefab2f.jpg",
1241
+ "image_caption": [
1242
+ "Figure 7. PSF as a Function of Number of Measurements and Spatial Resolution. The PSF of the center voxel is shown in the transient and intensity-only case when multiplexed with 60 laser spots. (1) the number of pixels $N = 40$ are kept fixed with the number of measurements $M$ being swept (left) and (2) $M = 10$ is kept fixed and $N$ is swept (right). Transient measurements result in PSFs with narrower FWHMs, even with low spatial resolution and low number of measurements."
1243
+ ],
1244
+ "image_footnote": [],
1245
+ "bbox": [
1246
+ 83,
1247
+ 89,
1248
+ 496,
1249
+ 340
1250
+ ],
1251
+ "page_idx": 6
1252
+ },
1253
+ {
1254
+ "type": "image",
1255
+ "img_path": "images/1edc73563e372c28f865e5f33e9f5ed0763c3157ca12399ed9b489b4d511ba4e.jpg",
1256
+ "image_caption": [],
1257
+ "image_footnote": [],
1258
+ "bbox": [
1259
+ 500,
1260
+ 90,
1261
+ 890,
1262
+ 340
1263
+ ],
1264
+ "page_idx": 6
1265
+ },
1266
+ {
1267
+ "type": "image",
1268
+ "img_path": "images/16f0d107255a743cc7c7b3edbfebd1230b98d899f7b505bddff4798abb3c252b.jpg",
1269
+ "image_caption": [
1270
+ "Figure 8. Gram Matrix as a Function of Temporal Resolution. For a defined voxel resolution and fixed spatial resolution. The columns (or rows) of the Gram matrix form the spatially varying PSFs shown in Fig. 6. The off-diagonal streaks describe the coherence between different voxels. As expected, higher temporal resolution decreases the coherence of the Gram matrix, which increases the likelihood of guaranteed sparse recovery."
1271
+ ],
1272
+ "image_footnote": [],
1273
+ "bbox": [
1274
+ 89,
1275
+ 433,
1276
+ 455,
1277
+ 513
1278
+ ],
1279
+ "page_idx": 6
1280
+ },
1281
+ {
1282
+ "type": "text",
1283
+ "text": "capture. We render $i_{o}$ and $i_{l}$ to obtain a noiseless version of the shadow transient for this experiment. Then, we reconstruct the shapes of real mannequins and must estimate $i_{0}$ . Fig. 11 shows that ToF measurements provide a noticeable improvement in image reconstruction. When there's no multiplexing (i.e., the number of laser spots is the same as number of measurements), the intensity-based method is similar to the ToF-based method. However, in the intensity-only case, the object becomes increasingly blurry until it is no longer visible as the amount of multiplexing increases. In the ToF case for single-shot, we can see a blurry image roughly localized where the center of the mannequin is as well as some high-frequency detail. Furthermore, the top view of the ToF reconstructions show that the high-frequency shape of the arms can be reconstructed even with multiplexed illumination, indicating the promise of single",
1284
+ "bbox": [
1285
+ 75,
1286
+ 659,
1287
+ 472,
1288
+ 901
1289
+ ],
1290
+ "page_idx": 6
1291
+ },
1292
+ {
1293
+ "type": "image",
1294
+ "img_path": "images/a939948ccd7e2441a4107bbde6b02c4135b2b8f32bbefa8083f6280b2129e310.jpg",
1295
+ "image_caption": [
1296
+ "Figure 9. Tradeoffs Between Spatial Resolution and Temporal Resolution. We quantify the tradeoff using mutual coherence. We use the same setup as Fig. 6, but the walls have length $3\\mathrm{m}$ , the walls are space $2\\mathrm{m}$ apart, and the distance to the walls is $2\\mathrm{m}$ . We discretize the voxel resolution to $2\\mathrm{cm}(\\sim 67\\mathrm{ps})$ . The bottom right of the figure, where spatial and temporal resolution are highest is where mutual coherence is the lowest. For a fixed temporal resolution, improvements in coherence quickly plateau. Improvements in temporal resolution dominate improvements in spatial resolution when reducing coherence, suggesting the importance of ToF for demultiplexing illumination in two-bounce NLOS imaging."
1297
+ ],
1298
+ "image_footnote": [],
1299
+ "bbox": [
1300
+ 568,
1301
+ 434,
1302
+ 823,
1303
+ 601
1304
+ ],
1305
+ "page_idx": 6
1306
+ },
1307
+ {
1308
+ "type": "text",
1309
+ "text": "shot capture with two-bounce ToF information. Fig. 12 shows that poorer temporal resolution results in aliasing artifacts (repeated arms and legs), whereas the rough shape of the object can still be seen with lower spatial resolution. Finally, we compare our method to the probabilistic voxel carving method used in [15] in Fig. 13. Voxel carving results in sharper reconstructions than backprojection.",
1310
+ "bbox": [
1311
+ 496,
1312
+ 794,
1313
+ 890,
1314
+ 900
1315
+ ],
1316
+ "page_idx": 6
1317
+ },
1318
+ {
1319
+ "type": "page_number",
1320
+ "text": "9198",
1321
+ "bbox": [
1322
+ 482,
1323
+ 944,
1324
+ 514,
1325
+ 955
1326
+ ],
1327
+ "page_idx": 6
1328
+ },
1329
+ {
1330
+ "type": "image",
1331
+ "img_path": "images/73d56e3e1d2c94eac3c110d2cccf8f917bb7ef872fe2dbc3535b42a6c8e81187.jpg",
1332
+ "image_caption": [
1333
+ "Figure 10. Single-Shot Localization. We obtain single-shot reconstructions of the \"N\", \"L\", \"O\", and \"S\" letters in the hidden scene with and without transient information. We see that using transient information, we are able to effectively localize the hidden object and even get a sharp outline of the letters, demonstrating the promise of single-shot capture with two-bounce transients."
1334
+ ],
1335
+ "image_footnote": [],
1336
+ "bbox": [
1337
+ 91,
1338
+ 88,
1339
+ 457,
1340
+ 258
1341
+ ],
1342
+ "page_idx": 7
1343
+ },
1344
+ {
1345
+ "type": "image",
1346
+ "img_path": "images/052e98ac960dfbd0dda5d08d4c729a58a63172b1586062dfa28ba96a9705a46b.jpg",
1347
+ "image_caption": [
1348
+ "Figure 11. Transients Enable Scene Recovery from Multiplexed Illumination. We perform backprojection of measurements with varying levels of multiplexing. Plotted in the top two rows are $x$ slices of the reconstructed volume. Under no multiplexing (right), intensity is sufficient to carve the shadow along detected pixel. But reconstruction quality sharply deteriorates as we move towards the multiplexing regime (left). When measuring with transients, time-of-flight information allows us to disambiguate between different shadows and recover a rough silhouette of the object even with just a single measurement."
1349
+ ],
1350
+ "image_footnote": [],
1351
+ "bbox": [
1352
+ 91,
1353
+ 391,
1354
+ 459,
1355
+ 651
1356
+ ],
1357
+ "page_idx": 7
1358
+ },
1359
+ {
1360
+ "type": "text",
1361
+ "text": "However, our method can be improved by filtering or incorporating priors, and is also generalizable to multiplexed illumination by utilizing ToF information.",
1362
+ "bbox": [
1363
+ 75,
1364
+ 854,
1365
+ 470,
1366
+ 902
1367
+ ],
1368
+ "page_idx": 7
1369
+ },
1370
+ {
1371
+ "type": "image",
1372
+ "img_path": "images/ad76127c3b9f4afd934c1107a982ed0d2ca3477e6dc88bdfb5454de7fd317d7e.jpg",
1373
+ "image_caption": [
1374
+ "Figure 12. Spatial Resolution vs. Timing Resolution. We plot reconstructions with different spatio-temporal resolutions. Lower timing resolution results in aliasing artifacts (i.e. repeated arms and legs). Lower spatial resolution introduces hole artifacts."
1375
+ ],
1376
+ "image_footnote": [],
1377
+ "bbox": [
1378
+ 517,
1379
+ 92,
1380
+ 851,
1381
+ 280
1382
+ ],
1383
+ "page_idx": 7
1384
+ },
1385
+ {
1386
+ "type": "image",
1387
+ "img_path": "images/fd216f0dc7a5bd6663190c42867f44743f539f7e6a813ffe1cf67a6a302b7c89.jpg",
1388
+ "image_caption": [
1389
+ "Figure 13. Scanning Baseline Comparison. We compare to the voxel carving method used in [15] (a) before and (b) after thresholding when scanning 1 laser spot. Our method is more general, but voxel carving yields sharper results with laser scanning."
1390
+ ],
1391
+ "image_footnote": [],
1392
+ "bbox": [
1393
+ 511,
1394
+ 358,
1395
+ 879,
1396
+ 501
1397
+ ],
1398
+ "page_idx": 7
1399
+ },
1400
+ {
1401
+ "type": "text",
1402
+ "text": "6. Discussion",
1403
+ "text_level": 1,
1404
+ "bbox": [
1405
+ 500,
1406
+ 590,
1407
+ 612,
1408
+ 607
1409
+ ],
1410
+ "page_idx": 7
1411
+ },
1412
+ {
1413
+ "type": "text",
1414
+ "text": "In this paper, we introduced the use of two-bounce transients for NLOS imaging. Compared to 3B-NLOS, 2B-NLOS enables more robust imaging of occluded objects in the presence of scenes with arbitrary lighting conditions and low albedo objects. Transient 2B-NLOS is promising for few-shot capture of occluded objects by using multiplexed illumination, especially as SPAD array technology matures. When comparing tradeoffs in temporal resolution, spatial resolution, and number of measurements, we qualitatively and quantitatively verify with real and simulated results that temporal resolution is most important for demultiplexing shadows. We anticipate future work to extend the analysis in this paper to SPAD arrays for real-time NLOS imaging.",
1415
+ "bbox": [
1416
+ 496,
1417
+ 617,
1418
+ 893,
1419
+ 815
1420
+ ],
1421
+ "page_idx": 7
1422
+ },
1423
+ {
1424
+ "type": "text",
1425
+ "text": "Acknowledgements",
1426
+ "text_level": 1,
1427
+ "bbox": [
1428
+ 500,
1429
+ 829,
1430
+ 668,
1431
+ 847
1432
+ ],
1433
+ "page_idx": 7
1434
+ },
1435
+ {
1436
+ "type": "text",
1437
+ "text": "AD and AV are supported by NSF Expeditions Award #IIS-1730574 and NSF CAREER Award #IIS-1652633. CH was supported by a Draper Scholarship.",
1438
+ "bbox": [
1439
+ 498,
1440
+ 854,
1441
+ 893,
1442
+ 902
1443
+ ],
1444
+ "page_idx": 7
1445
+ },
1446
+ {
1447
+ "type": "page_number",
1448
+ "text": "9199",
1449
+ "bbox": [
1450
+ 482,
1451
+ 944,
1452
+ 514,
1453
+ 955
1454
+ ],
1455
+ "page_idx": 7
1456
+ },
1457
+ {
1458
+ "type": "text",
1459
+ "text": "References",
1460
+ "text_level": 1,
1461
+ "bbox": [
1462
+ 78,
1463
+ 89,
1464
+ 173,
1465
+ 104
1466
+ ],
1467
+ "page_idx": 8
1468
+ },
1469
+ {
1470
+ "type": "list",
1471
+ "sub_type": "ref_text",
1472
+ "list_items": [
1473
+ "[1] Fadel Adib and Dina Katabi. See through walls with wifi! In Proceedings of the ACM SIGCOMM 2013 conference on SIGCOMM, pages 75-86, 2013. 2",
1474
+ "[2] Byeongjoo Ahn, Akshat Dave, Ashok Veeraraghavan, Ioannis Gkioulekas, and Aswin C Sankaranarayanan. Convolutional approximations to the general non-line-of-sight imaging operator. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7889-7899, 2019. 2",
1475
+ "[3] Victor Arellano, Diego Gutierrez, and Adrian Jarabo. Fast back-projection for non-line of sight reconstruction. In ACM SIGGRAPH 2017 Posters, pages 1-2. 2017. 2",
1476
+ "[4] Katherine L Bouman, Vickie Ye, Adam B Yedidia, Frédo Durand, Gregory W Wornell, Antonio Torralba, and William T Freeman. Turning corners into cameras: Principles and methods. In Proceedings of the IEEE International Conference on Computer Vision, pages 2270-2278, 2017. 2, 3",
1477
+ "[5] Mauro Buttafava, Jessica Zeman, Alberto Tosi, Kevin Eliceiri, and Andreas Velten. Non-line-of-sight imaging using a time-gated single photon avalanche diode. Optics express, 23(16):20997-21011, 2015. 2",
1478
+ "[6] Emmanuel J Candès and Michael B Wakin. An introduction to compressive sampling. IEEE signal processing magazine, 25(2):21-30, 2008. 6",
1479
+ "[7] Wenzheng Chen, Simon Daneau, Fahim Mannan, and Felix Heide. Steady-state non-line-of-sight imaging. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6790-6799, 2019. 2",
1480
+ "[8] Wenzheng Chen, Fangyin Wei, Kiriakos N Kutulakos, Szymon Rusinkiewicz, and Felix Heide. Learned feature embeddings for non-line-of-sight imaging and recognition. ACM Transactions on Graphics (ToG), 39(6):1-18, 2020. 2",
1481
+ "[9] Akshit Dave, Muralidhar Madabhushi Balaji, Prasanna Rangarajan, Ashok Veeraraghavan, and Marc P Christensen. Foveated non-line-of-sight imaging. In Computational Optical Sensing and Imaging, pages CTh5C-6. Optica Publishing Group, 2020. 2",
1482
+ "[10] Akshit Dave, Yongyi Zhao, and Ashok Veeraraghavan. Pandora: Polarization-aided neural decomposition of radiance. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23–27, 2022, Proceedings, Part VII, pages 538–556. Springer, 2022. 2",
1483
+ "[11] Genevieve Gariepy, Francesco Tonolini, Robert Henderson, Jonathan Leach, and Daniele Faccio. Detection and tracking of moving objects hidden from view. Nature Photonics, 10(1):23-26, 2016. 2",
1484
+ "[12] Felix Heide, Matthew O'Toole, Kai Zang, David B Lindell, Steven Diamond, and Gordon Wetzstein. Non-line-of-sight imaging with partial occluders and surface normals. ACM Transactions on Graphics (ToG), 38(3):1-10, 2019. 2",
1485
+ "[13] Felix Heide, Lei Xiao, Wolfgang Heidrich, and Matthias B Hullin. Diffuse mirrors: 3d reconstruction from diffuse indirect illumination using inexpensive time-of-flight sensors. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3222-3229, 2014. 2"
1486
+ ],
1487
+ "bbox": [
1488
+ 78,
1489
+ 114,
1490
+ 468,
1491
+ 900
1492
+ ],
1493
+ "page_idx": 8
1494
+ },
1495
+ {
1496
+ "type": "list",
1497
+ "sub_type": "ref_text",
1498
+ "list_items": [
1499
+ "[14] Connor Henley, Joseph Hollmann, and Ramesh Raskar. Bounce-flash lidar. IEEE Transactions on Computational Imaging, 8:411-424, 2022. 2",
1500
+ "[15] Connor Henley, Tomohiro Maeda, Tristan Swedish, and Ramesh Raskar. Imaging behind occluders using two-bounce light. In ECCV, pages 573-588. Springer, 2020. 1, 2, 7, 8",
1501
+ "[16] Achuta Kadambi, Hang Zhao, Boxin Shi, and Ramesh Raskar. Occluded imaging with time-of-flight sensors. ACM Transactions on Graphics (ToG), 35(2):1-12, 2016. 2, 6",
1502
+ "[17] Ori Katz, Pierre Heidmann, Mathias Fink, and Sylvain Gigan. Non-invasive single-shot imaging through scattering layers and around corners via speckle correlations. Nature photonics, 8(10):784-790, 2014. 2",
1503
+ "[18] Ahmed Kirmani, Tyler Hutchison, James Davis, and Ramesh Raskar. Looking around the corner using transient imaging. In 2009 IEEE 12th International Conference on Computer Vision, pages 159-166. IEEE, 2009. 2, 3",
1504
+ "[19] Oichi Kumagai, Junichi Ohmachi, Masao Matsumura, Shinichiro Yagi, Kenichi Tayu, Keitaro Amagawa, Tomohiro Matsukawa, Osamu Ozawa, Daisuke Hirono, Yasuhiro Shinozuka, et al. 7.3 a $189 \\times 600$ back-illuminated stacked spad direct time-of-flight depth sensor for automotive lidar systems. In 2021 IEEE International Solid-State Circuits Conference (ISSCC), volume 64, pages 110-112. IEEE, 2021. 2",
1505
+ "[20] Aldo Laurentini. The visual hull concept for silhouette-based image understanding. PAMI, 16(2):150-162, 1994. 1",
1506
+ "[21] Di Lin, Connor Hashemi, and James R Leger. Passive non-line-of-sight imaging using plenoptic information. JOSA A, 37(4):540-551, 2020. 2",
1507
+ "[22] David B Lindell, Gordon Wetzstein, and Matthew O'Toole. Wave-based non-line-of-sight imaging using fast fkr migration. ACM Transactions on Graphics (ToG), 38(4):1-13, 2019. 2",
1508
+ "[23] Xiaochun Liu, Ibón Guillén, Marco La Manna, Ji Hyun Nam, Syed Azer Reza, Toan Huu Le, Adrian Jarabo, Diego Gutierrez, and Andreas Velten. Non-line-of-sight imaging using phasor-field virtual wave optics. Nature, 572(7771):620-623, 2019. 2",
1509
+ "[24] Tomohiro Maeda, Guy Satat, Tristan Swedish, Lagnojita Sinha, and Ramesh Raskar. Recent advances in imaging around corners. arXiv preprint arXiv:1910.05613, 2019. 1",
1510
+ "[25] Christopher A Metzler, Felix Heide, Prasana Rangarajan, Muralidhar Madabhushi Balaji, Aparna Viswanath, Ashok Veeraraghavan, and Richard G Baraniuk. Deep-inverse correlography: towards real-time high-resolution non-line-of-sight imaging. Optica, 7(1):63-71, 2020. 2",
1511
+ "[26] Kazuhiro Morimoto, Andrei Ardelean, Ming-Lo Wu, Arin Can Ulku, Ivan Michel Antolovic, Claudio Bruschini, and Edoardo Charbon. Megapixel time-gated spad image sensor for 2d and 3d imaging applications. Optica, 7(4):346–354, 2020. 2",
1512
+ "[27] Matthew O'Toole, David B Lindell, and Gordon Wetzstein. Confocal non-line-of-sight imaging based on the light-cone transform. Nature, 555(7696):338-341, 2018. 2, 3",
1513
+ "[28] Adithya Pediredla, Akshit Dave, and Ashok Veeraraghavan. Snlos: Non-line-of-sight scanning through temporal focus"
1514
+ ],
1515
+ "bbox": [
1516
+ 501,
1517
+ 92,
1518
+ 890,
1519
+ 900
1520
+ ],
1521
+ "page_idx": 8
1522
+ },
1523
+ {
1524
+ "type": "page_number",
1525
+ "text": "9200",
1526
+ "bbox": [
1527
+ 482,
1528
+ 945,
1529
+ 514,
1530
+ 955
1531
+ ],
1532
+ "page_idx": 8
1533
+ },
1534
+ {
1535
+ "type": "list",
1536
+ "sub_type": "ref_text",
1537
+ "list_items": [
1538
+ "ing. In 2019 IEEE International Conference on Computational Photography (ICCP), pages 1-13. IEEE, 2019. 2",
1539
+ "[29] Adithya K Pediredla, Aswin C Sankaranarayanan, Mauro Buttafava, Alberto Tosi, and Ashok Veeraraghavan. Signal processing based pile-up compensation for gated single-photon avalanche diodes. arXiv preprint arXiv:1806.07437, 2018.5",
1540
+ "[30] Joshua Rapp, Charles Saunders, Julián Tachella, John Murray-Bruce, Yoann Altmann, Jean-Yves Tourneret, Stephen McLaughlin, Robin Dawson, Franco NC Wong, and Vivek K Goyal. Seeing around corners with edge-resolved transient imaging. Nature communications, 11(1):1-10, 2020. 2, 3",
1541
+ "[31] Charles Saunders, John Murray-Bruce, and Vivek K Goyal. Computational periscopy with an ordinary digital camera. Nature, 565(7740):472-475, 2019. 3",
1542
+ "[32] Nicolas Scheiner, Florian Kraus, Fangyin Wei, Buu Phan, Fahim Mannan, Nils Appenrodt, Werner Ritter, Jurgen Dickmann, Klaus Dietmayer, Bernhard Sick, et al. Seeing around street corners: Non-line-of-sight detection and tracking in-the-wild using doppler radar. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2068-2077, 2020. 2",
1543
+ "[33] Prafull Sharma, Miika Aittala, Yoav Y Schechner, Antonio Torralba, Gregory W Wornell, William T Freeman, and Frédo Durand. What you can learn by staring at a blank wall. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2330-2339, 2021. 2, 3",
1544
+ "[34] Brandon M Smith, Matthew O'Toole, and Mohit Gupta. Tracking multiple objects outside the line of sight using speckle imaging. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6258-6266, 2018. 2",
1545
+ "[35] Tristan Swedish, Connor Henley, and Ramesh Raskar. Objects as cameras: Estimating high-frequency illumination from shadows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2593-2602, 2021. 2",
1546
+ "[36] Kenichiro Tanaka, Yasuhiro Mukaigawa, and Achuta Kadambi. Polarized non-line-of-sight imaging. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2136-2145, 2020. 2",
1547
+ "[37] Matthew Tancik, Guy Satat, and Ramesh Raskar. Flash photography for data-driven hidden scene recovery. arXiv preprint arXiv:1810.11710, 2018. 2",
1548
+ "[38] Kushagra Tiwary, Askhat Dave, Nikhil Behari, Tzofi Klinghoffer, Ashok Veeraraghavan, and Ramesh Raskar. Orca: Glossy objects as radiance field cameras. arXiv preprint arXiv:2212.04531, 2022. 2",
1549
+ "[39] Andreas Velten, Thomas Willwacher, Otkrist Gupta, Ashok Veeraraghavan, Moungi G Bawendi, and Ramesh Raskar. Recovering three-dimensional shape around a corner using ultrafast time-of-flight imaging. Nature communications, 3(1):1-8, 2012. 2",
1550
+ "[40] Florian Willomitzer, Fengqiang Li, Muralidhar Madabhushi Balaji, Prasanna Ranganajan, and Oliver Cossairt. High resolution non-line-of-sight imaging with superheterodyne re"
1551
+ ],
1552
+ "bbox": [
1553
+ 78,
1554
+ 92,
1555
+ 468,
1556
+ 900
1557
+ ],
1558
+ "page_idx": 9
1559
+ },
1560
+ {
1561
+ "type": "list",
1562
+ "sub_type": "ref_text",
1563
+ "list_items": [
1564
+ "mote digital holography. In Computational Optical Sensing and Imaging, pages CM2A-2. Optica Publishing Group, 2019. 2",
1565
+ "[41] Florian Willomitzer, Prasanna V Rangarajan, Fengqiang Li, Muralidhar M Balaji, Marc P Christensen, and Oliver Cossaint. Fast non-line-of-sight imaging with high-resolution and wide field of view using synthetic wavelength holography. Nature communications, 12(1):6647, 2021. 2",
1566
+ "[42] Shumian Xin, Sotiris Nousias, Kiriakos N Kutulakos, Aswin C Sankaranarayanan, Srinivasa G Narasimhan, and Ioannis Gkioulekas. A theory of Fermat paths for nonline-of-sight shape reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6800-6809, 2019. 2, 3",
1567
+ "[43] Shichao Yue, Hao He, Peng Cao, Kaiwen Zha, Masayuki Koizumi, and Dina Katabi. Cornerradar: Rf-based indoor localization around corners. Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies, 6(1):1-24, 2022. 2",
1568
+ "[44] Chao Zhang, Ning Zhang, Zhijie Ma, Letian Wang, Yu Qin, Jieyang Jia, and Kai Zang. A $240 \\times 160$ 3d-stacked spad dtof image sensor with rolling shutter and in-pixel histogram for mobile devices. IEEE Open Journal of the Solid-State Circuits Society, 2:3-11, 2021. 2"
1569
+ ],
1570
+ "bbox": [
1571
+ 501,
1572
+ 92,
1573
+ 890,
1574
+ 428
1575
+ ],
1576
+ "page_idx": 9
1577
+ },
1578
+ {
1579
+ "type": "page_number",
1580
+ "text": "9201",
1581
+ "bbox": [
1582
+ 482,
1583
+ 945,
1584
+ 513,
1585
+ 955
1586
+ ],
1587
+ "page_idx": 9
1588
+ }
1589
+ ]
2023/Role of Transients in Two-Bounce Non-Line-of-Sight Imaging/dd8b3d32-1f20-46bb-9a36-d731b9d46ed2_model.json ADDED
@@ -0,0 +1,2178 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ [
2
+ [
3
+ {
4
+ "type": "header",
5
+ "bbox": [
6
+ 0.107,
7
+ 0.003,
8
+ 0.182,
9
+ 0.043
10
+ ],
11
+ "angle": 0,
12
+ "content": "CVF"
13
+ },
14
+ {
15
+ "type": "header",
16
+ "bbox": [
17
+ 0.237,
18
+ 0.0,
19
+ 0.812,
20
+ 0.047
21
+ ],
22
+ "angle": 0,
23
+ "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore."
24
+ },
25
+ {
26
+ "type": "title",
27
+ "bbox": [
28
+ 0.176,
29
+ 0.131,
30
+ 0.797,
31
+ 0.155
32
+ ],
33
+ "angle": 0,
34
+ "content": "Role of Transients in Two-Bounce Non-Line-of-Sight Imaging"
35
+ },
36
+ {
37
+ "type": "text",
38
+ "bbox": [
39
+ 0.241,
40
+ 0.18,
41
+ 0.721,
42
+ 0.216
43
+ ],
44
+ "angle": 0,
45
+ "content": "Siddharth Somasundaram<sup>1</sup> Akshit Dave<sup>2</sup> Connor Henley<sup>1</sup> Ashok Veeraraghavan<sup>2</sup> Ramesh Raskar<sup>1</sup>"
46
+ },
47
+ {
48
+ "type": "text",
49
+ "bbox": [
50
+ 0.249,
51
+ 0.217,
52
+ 0.723,
53
+ 0.236
54
+ ],
55
+ "angle": 0,
56
+ "content": "<sup>1</sup>Massachusetts Institute of Technology <sup>2</sup>Rice University"
57
+ },
58
+ {
59
+ "type": "text",
60
+ "bbox": [
61
+ 0.229,
62
+ 0.237,
63
+ 0.737,
64
+ 0.253
65
+ ],
66
+ "angle": 0,
67
+ "content": "{sidsoma,co24401,raskar}@mit.edu,{akshat,vashok}@rice.edu"
68
+ },
69
+ {
70
+ "type": "image",
71
+ "bbox": [
72
+ 0.088,
73
+ 0.287,
74
+ 0.277,
75
+ 0.473
76
+ ],
77
+ "angle": 0,
78
+ "content": null
79
+ },
80
+ {
81
+ "type": "image_caption",
82
+ "bbox": [
83
+ 0.115,
84
+ 0.477,
85
+ 0.25,
86
+ 0.489
87
+ ],
88
+ "angle": 0,
89
+ "content": "(a) Two-Bounce NLOS"
90
+ },
91
+ {
92
+ "type": "image",
93
+ "bbox": [
94
+ 0.281,
95
+ 0.287,
96
+ 0.449,
97
+ 0.473
98
+ ],
99
+ "angle": 0,
100
+ "content": null
101
+ },
102
+ {
103
+ "type": "image_caption",
104
+ "bbox": [
105
+ 0.294,
106
+ 0.477,
107
+ 0.433,
108
+ 0.491
109
+ ],
110
+ "angle": 0,
111
+ "content": "(b) Application Settings"
112
+ },
113
+ {
114
+ "type": "image",
115
+ "bbox": [
116
+ 0.453,
117
+ 0.287,
118
+ 0.603,
119
+ 0.473
120
+ ],
121
+ "angle": 0,
122
+ "content": null
123
+ },
124
+ {
125
+ "type": "image_caption",
126
+ "bbox": [
127
+ 0.458,
128
+ 0.477,
129
+ 0.595,
130
+ 0.491
131
+ ],
132
+ "angle": 0,
133
+ "content": "(c) Experimental Setup"
134
+ },
135
+ {
136
+ "type": "image",
137
+ "bbox": [
138
+ 0.605,
139
+ 0.287,
140
+ 0.882,
141
+ 0.473
142
+ ],
143
+ "angle": 0,
144
+ "content": null
145
+ },
146
+ {
147
+ "type": "image_caption",
148
+ "bbox": [
149
+ 0.624,
150
+ 0.477,
151
+ 0.866,
152
+ 0.49
153
+ ],
154
+ "angle": 0,
155
+ "content": "(d) Reconstruction with 10 Measurements"
156
+ },
157
+ {
158
+ "type": "image_caption",
159
+ "bbox": [
160
+ 0.075,
161
+ 0.501,
162
+ 0.894,
163
+ 0.572
164
+ ],
165
+ "angle": 0,
166
+ "content": "Figure 1. Two-bounce Transient NLOS Imaging. (a) We propose using two-bounce transients for NLOS imaging. Two-bounce signals capture information about the shadows in a scene, and use of transient information can reduce the number of measurements needed for reconstruction by using multiplexed illumination. (b) Our work can be applicable in many scenarios, such as tunnels, corridors, streets, and forests. (c)-(d) We experimentally validate the use of two-bounce transients on real objects. We show that time-of-flight information enables more robust reconstructions of occluded objects while using fewer measurements."
167
+ },
168
+ {
169
+ "type": "title",
170
+ "bbox": [
171
+ 0.235,
172
+ 0.584,
173
+ 0.314,
174
+ 0.6
175
+ ],
176
+ "angle": 0,
177
+ "content": "Abstract"
178
+ },
179
+ {
180
+ "type": "text",
181
+ "bbox": [
182
+ 0.076,
183
+ 0.612,
184
+ 0.473,
185
+ 0.884
186
+ ],
187
+ "angle": 0,
188
+ "content": "The goal of non-line-of-sight (NLOS) imaging is to image objects occluded from the camera's field of view using multiply scattered light. Recent works have demonstrated the feasibility of two-bounce (2B) NLOS imaging by scanning a laser and measuring cast shadows of occluded objects in scenes with two relay surfaces. In this work, we study the role of time-of-flight (ToF) measurements, i.e. transients, in 2B-NLOS under multiplexed illumination. Specifically, we study how ToF information can reduce the number of measurements and spatial resolution needed for shape reconstruction. We present our findings with respect to trade-offs in (1) temporal resolution, (2) spatial resolution, and (3) number of image captures by studying SNR and recoverability as functions of system parameters. This leads to a formal definition of the mathematical constraints for 2B lidar. We believe that our work lays an analytical groundwork for design of future NLOS imaging systems, especially as ToF sensors become increasingly ubiquitous."
189
+ },
190
+ {
191
+ "type": "title",
192
+ "bbox": [
193
+ 0.502,
194
+ 0.584,
195
+ 0.631,
196
+ 0.6
197
+ ],
198
+ "angle": 0,
199
+ "content": "1. Introduction"
200
+ },
201
+ {
202
+ "type": "text",
203
+ "bbox": [
204
+ 0.498,
205
+ 0.609,
206
+ 0.892,
207
+ 0.836
208
+ ],
209
+ "angle": 0,
210
+ "content": "Non-line-of-sight (NLOS) imaging aims to reconstruct objects occluded from direct line of sight and has the potential to be transformative in numerous applications across autonomous driving, search and rescue, and non-invasive medical imaging [24]. The key approach is to measure light that has undergone multiple surface scattering events and computationally invert these measurements to estimate hidden geometries. Recent work used two-bounce light, as shown in Fig. 2a, to reconstruct high quality shapes behind occluders [15]. The key idea is that two-bounce light captures information about the shadows of the occluded object. By scanning the laser source at different points 1 on a relay surface and measuring multiple shadow images, it is possible to reconstruct the hidden object by computing the visual hull [20] of the measured shadows."
211
+ },
212
+ {
213
+ "type": "text",
214
+ "bbox": [
215
+ 0.498,
216
+ 0.84,
217
+ 0.894,
218
+ 0.901
219
+ ],
220
+ "angle": 0,
221
+ "content": "Benefits of Two-Bounce Two-bounce (2B) light can be captured using two relay surfaces on opposite sides of the hidden scene (Fig. 1a). This can occur in a variety of real-world settings such as tunnels, hallways, streets, and clit"
222
+ },
223
+ {
224
+ "type": "page_number",
225
+ "bbox": [
226
+ 0.482,
227
+ 0.945,
228
+ 0.516,
229
+ 0.957
230
+ ],
231
+ "angle": 0,
232
+ "content": "9192"
233
+ }
234
+ ],
235
+ [
236
+ {
237
+ "type": "image",
238
+ "bbox": [
239
+ 0.081,
240
+ 0.089,
241
+ 0.268,
242
+ 0.201
243
+ ],
244
+ "angle": 0,
245
+ "content": null
246
+ },
247
+ {
248
+ "type": "image_caption",
249
+ "bbox": [
250
+ 0.097,
251
+ 0.204,
252
+ 0.253,
253
+ 0.216
254
+ ],
255
+ "angle": 0,
256
+ "content": "(a) Two-Bounce NLOS Imaging"
257
+ },
258
+ {
259
+ "type": "image",
260
+ "bbox": [
261
+ 0.271,
262
+ 0.089,
263
+ 0.468,
264
+ 0.201
265
+ ],
266
+ "angle": 0,
267
+ "content": null
268
+ },
269
+ {
270
+ "type": "image_caption",
271
+ "bbox": [
272
+ 0.297,
273
+ 0.204,
274
+ 0.459,
275
+ 0.215
276
+ ],
277
+ "angle": 0,
278
+ "content": "(b) Two-Bounce Transient NLOS"
279
+ },
280
+ {
281
+ "type": "image_caption",
282
+ "bbox": [
283
+ 0.076,
284
+ 0.227,
285
+ 0.47,
286
+ 0.31
287
+ ],
288
+ "angle": 0,
289
+ "content": "Figure 2. Two-Bounce NLOS Imaging. (a) Two-bounce NLOS imaging performs 3D reconstruction by using information contained in the shadows of the hidden object when illuminated by one virtual source 1 at a time. (b) In this work, we study the benefits of using transient information for two-bounce NLOS imaging under the presence of multiplexed illumination."
290
+ },
291
+ {
292
+ "type": "text",
293
+ "bbox": [
294
+ 0.076,
295
+ 0.34,
296
+ 0.468,
297
+ 0.505
298
+ ],
299
+ "angle": 0,
300
+ "content": "tered environments like forests (Fig. 1b). For example, imagine an autonomous vehicle in a tunnel being able to see ahead of the car in front of it by using two-bounce signals from the sides of the tunnel. Two-bounce light also helps alleviate the signal-to-noise ratio (SNR) issue in 3B-NLOS imaging, since 3B signals experience significant signal attenuation at each surface scatter event. Furthermore, 2B light captures information about the cast shadows of the hidden object instead of reflectance, as in the 3B case. As a result, the SNR of 2B signals is independent of the object's albedo and can therefore robustly image darker objects."
301
+ },
302
+ {
303
+ "type": "text",
304
+ "bbox": [
305
+ 0.076,
306
+ 0.512,
307
+ 0.468,
308
+ 0.707
309
+ ],
310
+ "angle": 0,
311
+ "content": "Contributions In this work, we propose using two-bounce transients for NLOS imaging to enable the use of multiplexed illumination. 3B-NLOS transient imaging systems have been analyzed previously, but we believe we are the first to analyze the use of multiplexed illumination and transient measurements for 2B-NLOS (Fig. 2b). Our key insight is that multiplexed illumination enables measurement of occluded objects with fewer image captures, and 2B transients enable demultiplexing shadows, as shown in Fig. 3. The robust SNR and promise of few-shot capture afforded by 2B transients makes the idea an important direction for few-shot NLOS imaging in general environments. We summarize our contributions below."
312
+ },
313
+ {
314
+ "type": "text",
315
+ "bbox": [
316
+ 0.097,
317
+ 0.722,
318
+ 0.468,
319
+ 0.752
320
+ ],
321
+ "angle": 0,
322
+ "content": "- Model: We present new insights into the algebraic formulations of two-bounce lidar for NLOS imaging."
323
+ },
324
+ {
325
+ "type": "text",
326
+ "bbox": [
327
+ 0.097,
328
+ 0.767,
329
+ 0.468,
330
+ 0.84
331
+ ],
332
+ "angle": 0,
333
+ "content": "- Analysis: We analyze tradeoffs with spatial resolution, temporal resolution, and number of image captures by using quantitative metrics such as SNR and recoverability, as well as qualitative validation from real and simulated results."
334
+ },
335
+ {
336
+ "type": "list",
337
+ "bbox": [
338
+ 0.097,
339
+ 0.722,
340
+ 0.468,
341
+ 0.84
342
+ ],
343
+ "angle": 0,
344
+ "content": null
345
+ },
346
+ {
347
+ "type": "text",
348
+ "bbox": [
349
+ 0.077,
350
+ 0.856,
351
+ 0.468,
352
+ 0.901
353
+ ],
354
+ "angle": 0,
355
+ "content": "Although we envision the ideas introduced here to inspire future works in single-shot NLOS imaging, we do not claim this as a contribution in this paper."
356
+ },
357
+ {
358
+ "type": "text",
359
+ "bbox": [
360
+ 0.498,
361
+ 0.092,
362
+ 0.892,
363
+ 0.304
364
+ ],
365
+ "angle": 0,
366
+ "content": "Scope of This Work All data used for this work is from simulation and a single-pixel SPAD sensor. We do not physically realize few-shot results in this paper due to limited availability of high-resolution SPAD arrays presently. Instead, we emulate SPAD array measurements by scanning a single-pixel SPAD across the field of view and and emulate multiplexed illumination by scanning a laser spot, acquiring individual transient images, and summing the images in post-processing. These measurements, however, are more conducive to our goal of performing analysis that explores the landscape of possibilities in NLOS imaging with advances in ToF sensors. We believe that the ideas and analysis introduced in this work will be increasingly relevant as SPAD array technology matures [19,26,44]."
367
+ },
368
+ {
369
+ "type": "title",
370
+ "bbox": [
371
+ 0.5,
372
+ 0.317,
373
+ 0.642,
374
+ 0.332
375
+ ],
376
+ "angle": 0,
377
+ "content": "2. Related Work"
378
+ },
379
+ {
380
+ "type": "title",
381
+ "bbox": [
382
+ 0.5,
383
+ 0.341,
384
+ 0.715,
385
+ 0.358
386
+ ],
387
+ "angle": 0,
388
+ "content": "Non-Line-of-Sight Imaging"
389
+ },
390
+ {
391
+ "type": "text",
392
+ "bbox": [
393
+ 0.498,
394
+ 0.364,
395
+ 0.892,
396
+ 0.576
397
+ ],
398
+ "angle": 0,
399
+ "content": "ToF Methods: The non-line-of-sight imaging (NLOS) problem aims to see around a corner by measuring light reflected off intermediary surfaces. A large body of work models the response of a hidden scene to a temporal pulse of light and generally perform a variant of ellipsoidal tomography or solve a linear inverse problem [3, 5, 11, 12, 18, 39]. The temporal profiles can probe features of the hidden space, such as Fermat paths [42], scene impulse response [23], and individual voxels [28]. Other works derive efficient algorithms using a confocal setup [2, 22, 27] and data-driven techniques [8]. NLOS has also been demonstrated with amplitude modulated continuous wave (AMCW) ToF by exploiting sparsity and fitting measurements to a Lambertian reflectance model [13, 16]."
400
+ },
401
+ {
402
+ "type": "text",
403
+ "bbox": [
404
+ 0.498,
405
+ 0.58,
406
+ 0.892,
407
+ 0.761
408
+ ],
409
+ "angle": 0,
410
+ "content": "Other Active Methods: Chen et al. use steady-state (i.e. time-integrated) intensity information by leveraging partial directionality of diffuse reflectors [7]. Radio wavelengths can also be used to image through walls and/or around them [1, 32, 43]. There has also been significant progress in exploiting speckle correlations [9, 17, 25, 34] and synthetic wave holography [40, 41] for NLOS imaging. Henley et al. demonstrate the use of two-bounce light (i.e. shadows) for NLOS using intensity sensors [15]. An extension of this work uses ToF constraints on two-bounce pathlengths to isolate shadows from ambient background and higher-order inter-reflections [14]. However multiplexing was not used."
411
+ },
412
+ {
413
+ "type": "text",
414
+ "bbox": [
415
+ 0.498,
416
+ 0.765,
417
+ 0.892,
418
+ 0.901
419
+ ],
420
+ "angle": 0,
421
+ "content": "Passive Methods: Passive methods make use of optical information already present in the scene without an active source. Many passive approaches exploit polarization cues [10,36], ambient thermal light [21,36], shadows [35], multiview reflections [38], and neural networks [33,37] to reconstruct or make inferences about the scene outside the camera's field of view. Occlusion-assisted imaging makes use of information present at sharp discontinuities in LOS and NLOS scenes by turning them into virtual pinholes [4,30] or"
422
+ },
423
+ {
424
+ "type": "page_number",
425
+ "bbox": [
426
+ 0.482,
427
+ 0.945,
428
+ 0.516,
429
+ 0.957
430
+ ],
431
+ "angle": 0,
432
+ "content": "9193"
433
+ }
434
+ ],
435
+ [
436
+ {
437
+ "type": "table",
438
+ "bbox": [
439
+ 0.105,
440
+ 0.089,
441
+ 0.868,
442
+ 0.273
443
+ ],
444
+ "angle": 0,
445
+ "content": "<table><tr><td rowspan=\"2\">NLOS Imaging Domain</td><td colspan=\"3\">Key Components</td><td colspan=\"3\">Key Features</td></tr><tr><td>Assisting Surfaces</td><td>Sensor Type</td><td>Light Source</td><td>Number of Captures</td><td>Recovered Dimensions</td><td>Signal Quality</td></tr><tr><td rowspan=\"2\">Passive Occlusion-Aided [4,31,33]</td><td>Relay Wall/Floor</td><td rowspan=\"2\">Intensity</td><td rowspan=\"2\">Passive</td><td rowspan=\"2\">Low</td><td rowspan=\"2\">2D</td><td rowspan=\"2\">Low (strong background)</td></tr><tr><td>Edge Occluders</td></tr><tr><td rowspan=\"2\">Three-Bounce Transient [18,27,30,42]</td><td>Relay Wall/Floor</td><td rowspan=\"2\">ToF</td><td rowspan=\"2\">Active (Scan)</td><td rowspan=\"2\">High</td><td rowspan=\"2\">3D</td><td rowspan=\"2\">Medium (3-bounce)</td></tr><tr><td>Edge Occluders</td></tr><tr><td>Two-Bounce Transient (Ours)</td><td>Two LOS Walls</td><td>ToF</td><td>Active (Flash)</td><td>Low</td><td>3D</td><td>High (2-bounce)</td></tr></table>"
446
+ },
447
+ {
448
+ "type": "table_caption",
449
+ "bbox": [
450
+ 0.076,
451
+ 0.281,
452
+ 0.895,
453
+ 0.325
454
+ ],
455
+ "angle": 0,
456
+ "content": "Table 1. Comparison of NLOS Methods. Passive approaches require fewer measurements but can either recover up to two dimensions or are strongly affected by ambient noise. Works exploiting three-bounce light paths demonstrate 3D imaging around the corner but require a bounce on the hidden surface, reducing the signal quality. We analyze exploiting two-bounce transients for NLOS imaging."
457
+ },
458
+ {
459
+ "type": "text",
460
+ "bbox": [
461
+ 0.077,
462
+ 0.35,
463
+ 0.471,
464
+ 0.381
465
+ ],
466
+ "angle": 0,
467
+ "content": "virtual pinspecks [31]. Intensity-only methods are difficult to scale to scenarios with arbitrary lighting conditions."
468
+ },
469
+ {
470
+ "type": "title",
471
+ "bbox": [
472
+ 0.077,
473
+ 0.405,
474
+ 0.373,
475
+ 0.423
476
+ ],
477
+ "angle": 0,
478
+ "content": "3. Forward and Adjoint Operators"
479
+ },
480
+ {
481
+ "type": "title",
482
+ "bbox": [
483
+ 0.078,
484
+ 0.438,
485
+ 0.229,
486
+ 0.455
487
+ ],
488
+ "angle": 0,
489
+ "content": "3.1. Problem Setup"
490
+ },
491
+ {
492
+ "type": "text",
493
+ "bbox": [
494
+ 0.076,
495
+ 0.47,
496
+ 0.473,
497
+ 0.787
498
+ ],
499
+ "angle": 0,
500
+ "content": "We first introduce terminology and notation for our problem formulation. The object of interest lies behind an occluder and is flanked on two sides by planar Lambertian relay surfaces. Without loss of generality, the right wall is the illumination wall \\(\\mathcal{L}\\) and the left wall is the observation wall \\(\\mathcal{O}\\). Light incident on the illumination wall at a point \\(l\\) is a virtual source. Each virtual source diffusely radiates light towards the observation wall. Light will arrive at a virtual detector \\(s\\) from \\(l\\) if the object doesn't lie along the ray connecting \\(s\\) and \\(l\\), as shown in Fig. 2a. If the object obstructs this ray, \\(s\\) is said to lie in the shadow of the object. We assume that every virtual source and virtual detector is visible to the laser and camera field of view, respectively. The hidden scene is discretized into a voxel grid described by a binary occupancy (or transparency) function. This implicitly assumes that we are reconstructing opaque surfaces. We neglect the effect of higher order bounces (e.g. interreflections, 3B light) because SPADs can temporally gate out the longer ToF of 3B light. Furthermore, the intensity of higher order bounces will be much lower than that of two-bounce light because all surfaces are assumed to be diffuse."
501
+ },
502
+ {
503
+ "type": "text",
504
+ "bbox": [
505
+ 0.076,
506
+ 0.795,
507
+ 0.47,
508
+ 0.903
509
+ ],
510
+ "angle": 0,
511
+ "content": "We denote light transients as ToF measurements for light that arrives at the virtual detector and shadow transients as ToF measurements for light that didn't arrive at the virtual detector because of the object's shadow. Both of these are two-bounce transients. We show that modeling the problem in terms of shadow transients enables reconstruction of the hidden objects shape by solving a linear inverse problem."
512
+ },
513
+ {
514
+ "type": "image",
515
+ "bbox": [
516
+ 0.546,
517
+ 0.348,
518
+ 0.845,
519
+ 0.537
520
+ ],
521
+ "angle": 0,
522
+ "content": null
523
+ },
524
+ {
525
+ "type": "image_caption",
526
+ "bbox": [
527
+ 0.499,
528
+ 0.545,
529
+ 0.892,
530
+ 0.629
531
+ ],
532
+ "angle": 0,
533
+ "content": "Figure 3. Time-of-Flight Information for Demultiplexing Shadows. Assume the point \\( s \\) lies in shadow. When using multiplexed illumination, there exists an ambiguity as to whether \\( s \\) lies in the shadow of \\( l_1 \\), \\( l_2 \\), or both. However, the paths \\( L - l_1 - s - C \\) and \\( L - l_2 - s - C \\) have different pathlengths, suggesting the usefulness of ToF information for demultiplexing shadows."
534
+ },
535
+ {
536
+ "type": "title",
537
+ "bbox": [
538
+ 0.5,
539
+ 0.651,
540
+ 0.715,
541
+ 0.667
542
+ ],
543
+ "angle": 0,
544
+ "content": "3.2. Linear Forward Model"
545
+ },
546
+ {
547
+ "type": "text",
548
+ "bbox": [
549
+ 0.5,
550
+ 0.674,
551
+ 0.892,
552
+ 0.706
553
+ ],
554
+ "angle": 0,
555
+ "content": "Light Transients Consider the case where the occluded scene is empty. The measured transient is then given by"
556
+ },
557
+ {
558
+ "type": "equation",
559
+ "bbox": [
560
+ 0.538,
561
+ 0.71,
562
+ 0.892,
563
+ 0.744
564
+ ],
565
+ "angle": 0,
566
+ "content": "\\[\ni _ {0} (t; \\mathbf {s}) = \\alpha \\iint_ {\\mathcal {L}} \\frac {\\delta (\\| \\mathbf {l} - \\mathbf {g} \\| + \\| \\mathbf {l} - \\mathbf {s} \\| - c t)}{\\| \\mathbf {l} - \\mathbf {s} \\| ^ {2}} d \\mathbf {l}, \\tag {1}\n\\]"
567
+ },
568
+ {
569
+ "type": "text",
570
+ "bbox": [
571
+ 0.498,
572
+ 0.749,
573
+ 0.894,
574
+ 0.84
575
+ ],
576
+ "angle": 0,
577
+ "content": "where \\(\\mathbf{g}\\) is the location of the laser, \\(c\\) is the speed of light, and \\(\\alpha\\) is the intensity of the laser. We refer to \\(i_0\\) as the empty transient. For simplicity, we ignore \\(\\alpha\\) for the rest of the derivations in this section and assume that the length between the virtual detector and camera is calibrated or known. The numerator of the integrand"
578
+ },
579
+ {
580
+ "type": "equation",
581
+ "bbox": [
582
+ 0.57,
583
+ 0.847,
584
+ 0.892,
585
+ 0.865
586
+ ],
587
+ "angle": 0,
588
+ "content": "\\[\ni (t; \\mathbf {s}, \\mathbf {l}) = \\delta (\\| \\mathbf {l} - \\mathbf {g} \\| + \\| \\mathbf {l} - \\mathbf {s} \\| - c t) \\tag {2}\n\\]"
589
+ },
590
+ {
591
+ "type": "text",
592
+ "bbox": [
593
+ 0.499,
594
+ 0.871,
595
+ 0.892,
596
+ 0.902
597
+ ],
598
+ "angle": 0,
599
+ "content": "describes the equation of an ellipsoid, with foci at \\(\\mathbf{g}\\) and s and major axis length \\(ct\\). When an object is present in the"
600
+ },
601
+ {
602
+ "type": "page_number",
603
+ "bbox": [
604
+ 0.482,
605
+ 0.945,
606
+ 0.516,
607
+ 0.957
608
+ ],
609
+ "angle": 0,
610
+ "content": "9194"
611
+ }
612
+ ],
613
+ [
614
+ {
615
+ "type": "image",
616
+ "bbox": [
617
+ 0.091,
618
+ 0.09,
619
+ 0.46,
620
+ 0.155
621
+ ],
622
+ "angle": 0,
623
+ "content": null
624
+ },
625
+ {
626
+ "type": "image_caption",
627
+ "bbox": [
628
+ 0.077,
629
+ 0.164,
630
+ 0.47,
631
+ 0.233
632
+ ],
633
+ "angle": 0,
634
+ "content": "Figure 4. Light and Shadow Transients. Three spacetime curves are depicted. (left) Empty transient \\( i_0 \\) with no hidden object, (middle) light transient \\( i_l \\) with an object, and (right) shadow transient \\( i_s = i_0 - i_l \\). Switching from light transients \\( i_l \\) to shadow transients \\( i_s \\) enables derivation of a linear forward model."
635
+ },
636
+ {
637
+ "type": "text",
638
+ "bbox": [
639
+ 0.077,
640
+ 0.262,
641
+ 0.445,
642
+ 0.277
643
+ ],
644
+ "angle": 0,
645
+ "content": "occluded scene, the measured light transient is given by"
646
+ },
647
+ {
648
+ "type": "equation",
649
+ "bbox": [
650
+ 0.164,
651
+ 0.289,
652
+ 0.469,
653
+ 0.322
654
+ ],
655
+ "angle": 0,
656
+ "content": "\\[\ni _ {l} (t; \\mathbf {s}) = \\iint_ {\\mathcal {L}} V _ {\\mathbf {s}, \\mathbf {l}} \\cdot \\frac {i (t ; \\mathbf {s} , \\mathbf {l})}{\\| \\mathbf {l} - \\mathbf {s} \\| ^ {2}} d \\mathbf {l}, \\tag {3}\n\\]"
657
+ },
658
+ {
659
+ "type": "text",
660
+ "bbox": [
661
+ 0.078,
662
+ 0.333,
663
+ 0.361,
664
+ 0.349
665
+ ],
666
+ "angle": 0,
667
+ "content": "where \\(V_{\\mathbf{s},1}\\) is the visibility between s and l."
668
+ },
669
+ {
670
+ "type": "equation",
671
+ "bbox": [
672
+ 0.168,
673
+ 0.354,
674
+ 0.469,
675
+ 0.388
676
+ ],
677
+ "angle": 0,
678
+ "content": "\\[\nV _ {\\mathbf {s}, \\mathbf {l}} = \\exp \\left(\\oint_ {\\mathbf {l}} ^ {\\mathbf {s}} \\ln \\left(f _ {t} (\\mathbf {x})\\right) d \\mathbf {x}\\right) \\tag {4}\n\\]"
679
+ },
680
+ {
681
+ "type": "text",
682
+ "bbox": [
683
+ 0.077,
684
+ 0.399,
685
+ 0.469,
686
+ 0.445
687
+ ],
688
+ "angle": 0,
689
+ "content": "where Eq. (4) computes the product integral along the line connecting \\(\\mathbf{l}\\) and \\(\\mathbf{s}\\) and \\(f_{t}(\\mathbf{x})\\in \\{0,1\\}\\) is the transparency function of voxel \\(\\mathbf{x}\\) in the occluded region."
690
+ },
691
+ {
692
+ "type": "text",
693
+ "bbox": [
694
+ 0.077,
695
+ 0.452,
696
+ 0.469,
697
+ 0.528
698
+ ],
699
+ "angle": 0,
700
+ "content": "Shadow Transients The non-linearity of Eq. (4) makes model inversion challenging with light transients. Casting the formulation from light space-time to shadow space-time enables a linear formulation of the problem. A shadow transient at \\(s\\) is given by"
701
+ },
702
+ {
703
+ "type": "equation",
704
+ "bbox": [
705
+ 0.18,
706
+ 0.541,
707
+ 0.469,
708
+ 0.557
709
+ ],
710
+ "angle": 0,
711
+ "content": "\\[\ni _ {s} (t; \\mathbf {s}) = i _ {0} (t; \\mathbf {s}) - i _ {l} (t; \\mathbf {s}). \\tag {5}\n\\]"
712
+ },
713
+ {
714
+ "type": "text",
715
+ "bbox": [
716
+ 0.077,
717
+ 0.57,
718
+ 0.469,
719
+ 0.646
720
+ ],
721
+ "angle": 0,
722
+ "content": "Fig. 4 illustrates the difference between light transients and shadow transients. Shadow transients directly probe voxel occupancy, whereas light transients probe for voxel transparency. For example, if the occluded scene is empty except for voxel \\(\\mathbf{x}\\), the shadow transient is given by"
723
+ },
724
+ {
725
+ "type": "equation",
726
+ "bbox": [
727
+ 0.09,
728
+ 0.658,
729
+ 0.469,
730
+ 0.691
731
+ ],
732
+ "angle": 0,
733
+ "content": "\\[\ni _ {s} (t; \\mathbf {s}, \\mathbf {x}) = \\frac {\\delta (\\| \\mathbf {l} (\\mathbf {x} , \\mathbf {s}) - \\mathbf {g} \\| + \\| \\mathbf {l} (\\mathbf {x} , \\mathbf {s}) - \\mathbf {s} \\| - c t)}{\\| \\mathbf {l} (\\mathbf {x} , \\mathbf {s}) - \\mathbf {s} \\| ^ {2}}, \\tag {6}\n\\]"
734
+ },
735
+ {
736
+ "type": "text",
737
+ "bbox": [
738
+ 0.077,
739
+ 0.703,
740
+ 0.468,
741
+ 0.749
742
+ ],
743
+ "angle": 0,
744
+ "content": "where \\(\\mathbf{l}(\\mathbf{x},\\mathbf{s})\\) is a point on the illumination wall that lies along the ray connecting \\(\\mathbf{x}\\) and \\(\\mathbf{s}\\). If the wall doesn't intersect this ray, \\(i_s(t;\\mathbf{x},\\mathbf{s}) = 0\\)."
745
+ },
746
+ {
747
+ "type": "text",
748
+ "bbox": [
749
+ 0.077,
750
+ 0.756,
751
+ 0.469,
752
+ 0.8
753
+ ],
754
+ "angle": 0,
755
+ "content": "Linear Relaxation Now, consider the general case where a single opaque object lies in the hidden scene. We can describe the shadow transient as"
756
+ },
757
+ {
758
+ "type": "equation",
759
+ "bbox": [
760
+ 0.161,
761
+ 0.81,
762
+ 0.469,
763
+ 0.844
764
+ ],
765
+ "angle": 0,
766
+ "content": "\\[\ni _ {s} (t; \\mathbf {s}) = \\iint_ {\\mathcal {L}} O _ {\\mathbf {s}, \\mathbf {l}} \\cdot \\frac {i (t ; \\mathbf {s} , \\mathbf {l})}{\\| \\mathbf {l} - \\mathbf {s} \\| ^ {2}} d \\mathbf {l}, \\tag {7}\n\\]"
767
+ },
768
+ {
769
+ "type": "text",
770
+ "bbox": [
771
+ 0.077,
772
+ 0.856,
773
+ 0.469,
774
+ 0.9
775
+ ],
776
+ "angle": 0,
777
+ "content": "where \\( O_{\\mathbf{s},1} \\in \\{0,1\\} \\) is the occlusion function (denotes if the ray between 1 and s is occluded). It can be expressed as the indicator function"
778
+ },
779
+ {
780
+ "type": "equation",
781
+ "bbox": [
782
+ 0.585,
783
+ 0.085,
784
+ 0.891,
785
+ 0.16
786
+ ],
787
+ "angle": 0,
788
+ "content": "\\[\n\\begin{array}{l} O _ {\\mathbf {s}, \\mathrm {l}} = \\left\\{ \\begin{array}{l l} 1, & \\text {i f} \\oint_ {\\mathrm {l}} ^ {\\mathbf {s}} f _ {o} (\\mathbf {x}) d \\mathbf {x} > 0 \\\\ 0, & \\text {o t h e r w i s e} \\end{array} \\right. \\tag {8} \\\\ \\approx \\oint_ {1} ^ {\\mathbf {s}} f _ {o} (\\mathbf {x}) d \\mathbf {x}, \\\\ \\end{array}\n\\]"
789
+ },
790
+ {
791
+ "type": "text",
792
+ "bbox": [
793
+ 0.499,
794
+ 0.171,
795
+ 0.892,
796
+ 0.276
797
+ ],
798
+ "angle": 0,
799
+ "content": "where \\( f_{o}(\\mathbf{x}) \\) is the voxel occupancy function of the hidden scene. The linear relaxation in Eq. (8) is possible because NLOS reconstructions typically only consider ToF information, but not radiometric information. The above formulation will preserve the structure of the streak images in spacetime, even if the intensities are overestimated. Combining Eq. (5), (7), and (8), we obtain"
800
+ },
801
+ {
802
+ "type": "equation",
803
+ "bbox": [
804
+ 0.508,
805
+ 0.287,
806
+ 0.891,
807
+ 0.321
808
+ ],
809
+ "angle": 0,
810
+ "content": "\\[\ni _ {0} (t; \\mathbf {s}) - i _ {l} (t; \\mathbf {s}) = \\iint_ {\\mathcal {L}} \\frac {i (t ; \\mathbf {s} , \\mathbf {l})}{\\| \\mathbf {l} - \\mathbf {s} \\| ^ {2}} \\cdot \\left(\\oint_ {\\mathbf {l}} ^ {\\mathbf {s}} f _ {o} (\\mathbf {x}) d \\mathbf {x}\\right) d \\mathbf {l}, \\tag {9}\n\\]"
811
+ },
812
+ {
813
+ "type": "text",
814
+ "bbox": [
815
+ 0.499,
816
+ 0.332,
817
+ 0.891,
818
+ 0.362
819
+ ],
820
+ "angle": 0,
821
+ "content": "where we solve for \\( f_{o}(\\mathbf{x}) \\). We can estimate \\( i_0 \\) from the scene geometry and \\( i_l \\) is the measured transient."
822
+ },
823
+ {
824
+ "type": "text",
825
+ "bbox": [
826
+ 0.499,
827
+ 0.366,
828
+ 0.892,
829
+ 0.412
830
+ ],
831
+ "angle": 0,
832
+ "content": "Discrete Beam Illumination Discretizing \\(\\mathcal{L}\\) as \\(K\\) discrete laser spots and the hidden space into a voxel grid, we can write Eq. (9) as"
833
+ },
834
+ {
835
+ "type": "equation",
836
+ "bbox": [
837
+ 0.5,
838
+ 0.423,
839
+ 0.901,
840
+ 0.48
841
+ ],
842
+ "angle": 0,
843
+ "content": "\\[\ni _ {s} [ t; \\mathbf {s} ] = \\sum_ {k = 1} ^ {K} \\frac {\\delta [ \\| \\mathbf {l} _ {k} - \\mathbf {g} \\| + \\| \\mathbf {l} _ {k} - \\mathbf {s} \\| - c t ]}{\\| \\mathbf {l} _ {k} - \\mathbf {s} \\| ^ {2}} \\cdot \\left(\\sum_ {\\mathbf {x} \\in \\mathcal {P} _ {\\mathbf {l} _ {k}, \\mathbf {s}}} f _ {o} [ \\mathbf {x} ]\\right), \\tag {10}\n\\]"
844
+ },
845
+ {
846
+ "type": "text",
847
+ "bbox": [
848
+ 0.5,
849
+ 0.481,
850
+ 0.89,
851
+ 0.511
852
+ ],
853
+ "angle": 0,
854
+ "content": "where \\(\\mathcal{P}_{1,\\mathrm{s}}\\) denotes the set of voxels lying along the path between 1 and s. Eq. (10) can be expressed in matrix form"
855
+ },
856
+ {
857
+ "type": "equation",
858
+ "bbox": [
859
+ 0.634,
860
+ 0.524,
861
+ 0.891,
862
+ 0.54
863
+ ],
864
+ "angle": 0,
865
+ "content": "\\[\n\\mathbf {I} _ {s} = \\mathbf {P B f} = \\mathbf {A f}, \\tag {11}\n\\]"
866
+ },
867
+ {
868
+ "type": "text",
869
+ "bbox": [
870
+ 0.499,
871
+ 0.552,
872
+ 0.892,
873
+ 0.656
874
+ ],
875
+ "angle": 0,
876
+ "content": "where \\(\\mathbf{I}_s\\in \\mathbb{R}^{n_u n_v n_t}\\) is the vectorized space-time measurements, \\(\\mathbf{f}\\in \\mathbb{R}^{n_x n_y n_z}\\) is the voxel occupancies, and \\(\\mathbf{A}\\in \\mathbb{R}^{n_u n_v n_t\\times n_x n_y n_z}\\) is the measurement operator mapping voxel occupancies to space-time measurements. \\(n_u\\) and \\(n_v\\) are the number of pixels along the \\(u\\) and \\(v\\) direction, and \\(n_t\\) is the number of timing bins. \\(n_x,n_y,\\) and \\(n_z\\) are the number of voxels in the \\(x,y,\\) and \\(z\\) directions."
877
+ },
878
+ {
879
+ "type": "text",
880
+ "bbox": [
881
+ 0.499,
882
+ 0.657,
883
+ 0.892,
884
+ 0.901
885
+ ],
886
+ "angle": 0,
887
+ "content": "\\(\\mathbf{P}\\) and \\(\\mathbf{B}\\) model the temporal and spatial relationships, respectively, between virtual sources, virtual detectors, and voxels. \\(\\mathbf{P} \\in \\mathbb{R}^{n_u n_v n_t \\times n_u n_v K}\\) is a block diagonal matrix, with the column space of the \\(i\\)th block \\(\\mathbf{P}^i \\in \\mathbb{R}^{n_t \\times K}\\) containing the corresponding \\(\\delta(\\cdot)\\) functions from Eq. (2) for virtual detector \\(\\mathbf{s}_i\\) and all \\(K\\) virtual sources. \\(\\mathbf{B} \\in \\mathbb{R}^{n_u n_v K \\times n_x n_y n_z}\\) is composed of \\(n_u n_v\\) vertical sub-blocks. Each entry \\(\\mathbf{B}_{kj}^i \\in \\{0,1\\}\\) of the \\(i\\)th sub-block \\(\\mathbf{B}^i \\in \\mathbb{R}^{K \\times n_x n_y n_z}\\) of sub-block \\(i\\) indicates whether voxel \\(j\\) lies along the line connecting virtual detector \\(\\mathbf{s}_i\\) and virtual source \\(\\mathbf{l}_k\\). The multiplication \\(\\mathbf{Bf}\\) performs the line integral in Eq. (8). The \\(k\\)th column vector of \\(\\mathbf{P}\\) is the space-time impulse response (STIR) to the \\(k\\)th illumination source when the hidden scene is completely occupied (note that we are dealing with shadow transients, not light transients). The \\(i\\)th column of \\(\\mathbf{A}\\) forms the expected space-time measurements"
888
+ },
889
+ {
890
+ "type": "page_number",
891
+ "bbox": [
892
+ 0.482,
893
+ 0.945,
894
+ 0.516,
895
+ 0.957
896
+ ],
897
+ "angle": 0,
898
+ "content": "9195"
899
+ }
900
+ ],
901
+ [
902
+ {
903
+ "type": "text",
904
+ "bbox": [
905
+ 0.076,
906
+ 0.092,
907
+ 0.473,
908
+ 0.258
909
+ ],
910
+ "angle": 0,
911
+ "content": "if only the \\(i\\)th voxel is occupied (i.e. the STIR to a single voxel). We neglect the \\(r^2\\) falloff term in our reconstructions. Additional explanation of the A matrix is provided in the supplementary. The linear model in Eq. (11) resembles that of 3B-bounce NLOS, where \\(\\mathbf{A}\\) is a measurement operator that maps voxels \\(\\mathbf{f}\\) to measurements \\(\\mathbf{I}\\). However, in 3B-NLOS, the information is encoded into the reflectance of the object, whereas in 2B-NLOS, the information is encoded into the shadows. Therefore, the former is linear with respects to light transients, and the latter is linear with respect to shadow transients."
912
+ },
913
+ {
914
+ "type": "title",
915
+ "bbox": [
916
+ 0.077,
917
+ 0.27,
918
+ 0.317,
919
+ 0.287
920
+ ],
921
+ "angle": 0,
922
+ "content": "3.3. Reconstruction Algorithm"
923
+ },
924
+ {
925
+ "type": "text",
926
+ "bbox": [
927
+ 0.097,
928
+ 0.295,
929
+ 0.47,
930
+ 0.31
931
+ ],
932
+ "angle": 0,
933
+ "content": "Reconstruction of \\(\\mathbf{f}\\) can be performed by backprojection"
934
+ },
935
+ {
936
+ "type": "equation",
937
+ "bbox": [
938
+ 0.228,
939
+ 0.318,
940
+ 0.47,
941
+ 0.337
942
+ ],
943
+ "angle": 0,
944
+ "content": "\\[\n\\mathbf {f} _ {b p} = \\mathbf {A} ^ {\\top} \\mathbf {I} _ {s}. \\tag {12}\n\\]"
945
+ },
946
+ {
947
+ "type": "text",
948
+ "bbox": [
949
+ 0.077,
950
+ 0.347,
951
+ 0.471,
952
+ 0.379
953
+ ],
954
+ "angle": 0,
955
+ "content": "Backprojection produces a low-frequency reconstruction of \\(\\mathbf{f}\\). In practice, filtering \\(\\mathbf{f}_{bp}\\) with a Laplacian filter"
956
+ },
957
+ {
958
+ "type": "equation",
959
+ "bbox": [
960
+ 0.159,
961
+ 0.391,
962
+ 0.47,
963
+ 0.433
964
+ ],
965
+ "angle": 0,
966
+ "content": "\\[\n\\mathbf {f} _ {f b p} = \\left(\\frac {\\partial^ {2}}{\\partial x ^ {2}} + \\frac {\\partial^ {2}}{\\partial y ^ {2}} + \\frac {\\partial^ {2}}{\\partial z ^ {2}}\\right) \\mathbf {f} _ {b p} \\tag {13}\n\\]"
967
+ },
968
+ {
969
+ "type": "text",
970
+ "bbox": [
971
+ 0.076,
972
+ 0.444,
973
+ 0.47,
974
+ 0.595
975
+ ],
976
+ "angle": 0,
977
+ "content": "has been shown to yield reasonably good results in ellipsoidal tomography in 3-bounce NLOS and line-based tomography in medical imaging. Our linear formulation also enables reconstruction with priors. We recognize that these optimization-based methods would improve reconstruction over unfiltered backprojection, which is equivalent to a low-frequency version of voxel carving in our setup. However, we restrict our focus to unfiltered backprojection because of its convenient relationship with the Gram matrix, which enables further analysis of our method in Sec. 4.2."
978
+ },
979
+ {
980
+ "type": "title",
981
+ "bbox": [
982
+ 0.077,
983
+ 0.612,
984
+ 0.273,
985
+ 0.629
986
+ ],
987
+ "angle": 0,
988
+ "content": "4. Design and Analysis"
989
+ },
990
+ {
991
+ "type": "text",
992
+ "bbox": [
993
+ 0.076,
994
+ 0.638,
995
+ 0.47,
996
+ 0.79
997
+ ],
998
+ "angle": 0,
999
+ "content": "In this section, we analyze how the performance of the proposed two-bounce transient system varies with system parameters and the geometry of LOS relay surfaces. We also analyze the impact of temporal resolution, spatial resolution, and number of measurements on recoverability. The amount of multiplexing in the scene is inversely related to the number of measurements. For example, if 60 laser spots are used and 15 measurements are captured, then each image contains 4 multiplexed shadows. We use SNR, coherence, and FWHM of the PSF to quantify performance."
1000
+ },
1001
+ {
1002
+ "type": "title",
1003
+ "bbox": [
1004
+ 0.077,
1005
+ 0.801,
1006
+ 0.267,
1007
+ 0.817
1008
+ ],
1009
+ "angle": 0,
1010
+ "content": "4.1. Voxel SNR Analysis"
1011
+ },
1012
+ {
1013
+ "type": "text",
1014
+ "bbox": [
1015
+ 0.076,
1016
+ 0.825,
1017
+ 0.471,
1018
+ 0.903
1019
+ ],
1020
+ "angle": 0,
1021
+ "content": "We quantify the noise characteristics of the measured shadow two-bounce transients using per-voxel SNR. Consider the case where only voxel \\(\\mathbf{x}\\) is occupied in the hidden region. We derive the SNR of the number of \"shadow\" photons received in the measured shadow transient \\(\\mathbf{I}_s^{\\mathbf{x}}\\). In vector"
1022
+ },
1023
+ {
1024
+ "type": "image",
1025
+ "bbox": [
1026
+ 0.513,
1027
+ 0.09,
1028
+ 0.878,
1029
+ 0.222
1030
+ ],
1031
+ "angle": 0,
1032
+ "content": null
1033
+ },
1034
+ {
1035
+ "type": "image_caption",
1036
+ "bbox": [
1037
+ 0.498,
1038
+ 0.231,
1039
+ 0.894,
1040
+ 0.342
1041
+ ],
1042
+ "angle": 0,
1043
+ "content": "Figure 5. SNR Analysis. The goal of this analysis is to determine how scene geometry affects the SNR of the two-bounce measurement at each voxel in the hidden scene. The first two cases consider the orientations of the relay surfaces, and the third considers the field of view of the camera. SNR is highest in the middle of the hidden scene for parallel surfaces, orienting the walls closer together increases SNR due to reduced \\( r^2 \\) falloff, and SNR is highest closer to the virtual detectors for small FOV cameras."
1044
+ },
1045
+ {
1046
+ "type": "text",
1047
+ "bbox": [
1048
+ 0.499,
1049
+ 0.366,
1050
+ 0.893,
1051
+ 0.397
1052
+ ],
1053
+ "angle": 0,
1054
+ "content": "form, the occupancy function can be described as \\(\\delta^{\\mathbf{x}}\\). From Eq. (11), the shadow transient is given by"
1055
+ },
1056
+ {
1057
+ "type": "equation",
1058
+ "bbox": [
1059
+ 0.658,
1060
+ 0.405,
1061
+ 0.891,
1062
+ 0.422
1063
+ ],
1064
+ "angle": 0,
1065
+ "content": "\\[\n\\mathbf {I} _ {s} ^ {\\mathbf {x}} = \\mathbf {A} \\delta^ {\\mathbf {x}} \\tag {14}\n\\]"
1066
+ },
1067
+ {
1068
+ "type": "text",
1069
+ "bbox": [
1070
+ 0.499,
1071
+ 0.43,
1072
+ 0.893,
1073
+ 0.492
1074
+ ],
1075
+ "angle": 0,
1076
+ "content": "and the light transients measured by the SPAD sensor are given by \\(\\mathbf{I}_l^{\\mathbf{x}}\\). For this analysis, we account for \\(r^2\\) falloff because it is a physically accurate phenomenon that impacts SNR. From Eq. (5),"
1077
+ },
1078
+ {
1079
+ "type": "equation",
1080
+ "bbox": [
1081
+ 0.641,
1082
+ 0.499,
1083
+ 0.891,
1084
+ 0.516
1085
+ ],
1086
+ "angle": 0,
1087
+ "content": "\\[\n\\mathbf {I} _ {l} ^ {\\mathbf {x}} = \\mathbf {I} _ {0} - \\mathbf {A} \\delta^ {\\mathbf {x}} \\tag {15}\n\\]"
1088
+ },
1089
+ {
1090
+ "type": "text",
1091
+ "bbox": [
1092
+ 0.498,
1093
+ 0.524,
1094
+ 0.893,
1095
+ 0.631
1096
+ ],
1097
+ "angle": 0,
1098
+ "content": "where \\(\\mathbf{I}_0\\) is the empty transient and assumed to be calibrated. Assuming no SPAD pile-up distortion, the number of photons arriving at the sensor for each time bin of \\(\\mathbf{I}_l^{\\mathbf{x}}\\) follows an independent Poisson distribution [29]. Thus, the total number of photons \\(P^{\\mathbf{x}}\\) measured follows a Poisson distribution with the rate parameter being the sum of all the photon counts for all time bins for all pixels."
1099
+ },
1100
+ {
1101
+ "type": "equation",
1102
+ "bbox": [
1103
+ 0.591,
1104
+ 0.638,
1105
+ 0.891,
1106
+ 0.656
1107
+ ],
1108
+ "angle": 0,
1109
+ "content": "\\[\nP ^ {\\mathbf {x}} \\sim \\operatorname {P o i s s o n} \\left(\\mathbf {1} ^ {T} \\left(\\mathbf {I} _ {0} - \\mathbf {A} \\delta^ {\\mathbf {x}}\\right)\\right) \\tag {16}\n\\]"
1110
+ },
1111
+ {
1112
+ "type": "text",
1113
+ "bbox": [
1114
+ 0.498,
1115
+ 0.664,
1116
+ 0.893,
1117
+ 0.71
1118
+ ],
1119
+ "angle": 0,
1120
+ "content": "where \\(\\mathbf{1} \\in \\mathbb{R}^{n_u n_v n_t}\\) is the one vector. Using the fact that SNR of a Poisson distribution is \\(\\mu / \\sigma = \\sqrt{\\lambda}\\) with rate parameter \\(\\lambda\\), the per-voxel SNR \\(S^{\\mathbf{x}}\\) is given by"
1121
+ },
1122
+ {
1123
+ "type": "equation",
1124
+ "bbox": [
1125
+ 0.615,
1126
+ 0.719,
1127
+ 0.891,
1128
+ 0.743
1129
+ ],
1130
+ "angle": 0,
1131
+ "content": "\\[\nS ^ {\\mathbf {x}} = \\sqrt {\\mathbf {1} ^ {T} \\left(\\mathbf {I} _ {0} - \\mathbf {A} \\boldsymbol {\\delta} ^ {\\mathbf {x}}\\right)} \\tag {17}\n\\]"
1132
+ },
1133
+ {
1134
+ "type": "text",
1135
+ "bbox": [
1136
+ 0.498,
1137
+ 0.75,
1138
+ 0.895,
1139
+ 0.902
1140
+ ],
1141
+ "angle": 0,
1142
+ "content": "The per-voxel SNR is plotted in Fig. 5 for three different scenarios. The first two scenarios deal with the orientation of the walls, and the third considers the field of view of the camera. In Fig. 5a, we can see that the SNR is higher for center voxels, as more rays connecting the illumination wall and observation wall pass through the center voxels. When the relay walls are oriented towards each other (Fig. 5b), the SNR increases due to reduced \\( r^2 \\) falloff. When the camera's field of view is small compared to the laser (Fig. 5c), regions closer to the observation wall have the highest SNR."
1143
+ },
1144
+ {
1145
+ "type": "page_number",
1146
+ "bbox": [
1147
+ 0.482,
1148
+ 0.945,
1149
+ 0.516,
1150
+ 0.957
1151
+ ],
1152
+ "angle": 0,
1153
+ "content": "9196"
1154
+ }
1155
+ ],
1156
+ [
1157
+ {
1158
+ "type": "image",
1159
+ "bbox": [
1160
+ 0.104,
1161
+ 0.1,
1162
+ 0.437,
1163
+ 0.252
1164
+ ],
1165
+ "angle": 0,
1166
+ "content": null
1167
+ },
1168
+ {
1169
+ "type": "image_caption",
1170
+ "bbox": [
1171
+ 0.076,
1172
+ 0.27,
1173
+ 0.473,
1174
+ 0.327
1175
+ ],
1176
+ "angle": 0,
1177
+ "content": "Figure 6. Spatially Varying PSF. In two-bounce lidar, the PSF has a double cone shape. The orientation of the double cone varies strongly as a function of the orientation and length of the relay surfaces (i.e. the virtual apertures) and the voxel position."
1178
+ },
1179
+ {
1180
+ "type": "title",
1181
+ "bbox": [
1182
+ 0.077,
1183
+ 0.352,
1184
+ 0.295,
1185
+ 0.369
1186
+ ],
1187
+ "angle": 0,
1188
+ "content": "4.2. Recoverability Analysis"
1189
+ },
1190
+ {
1191
+ "type": "text",
1192
+ "bbox": [
1193
+ 0.076,
1194
+ 0.376,
1195
+ 0.47,
1196
+ 0.497
1197
+ ],
1198
+ "angle": 0,
1199
+ "content": "Resolution We now study the resolution limits of our method. In particular, we consider two key metrics in our measurement operator: mutual coherence and full width at half maximum (FWHM) of the PSF. We are interested in how (1) temporal resolution, (2) spatial resolution \\( N \\), and (3) number of measurements \\( M \\) impact each of our metrics. From Eq. (12), we know that the backprojected occupancy can be obtained from the occupancy function by"
1200
+ },
1201
+ {
1202
+ "type": "equation",
1203
+ "bbox": [
1204
+ 0.218,
1205
+ 0.508,
1206
+ 0.47,
1207
+ 0.527
1208
+ ],
1209
+ "angle": 0,
1210
+ "content": "\\[\n\\mathbf {f} _ {b p} = \\left(\\mathbf {A} ^ {\\top} \\mathbf {A}\\right) \\mathbf {f}, \\tag {18}\n\\]"
1211
+ },
1212
+ {
1213
+ "type": "text",
1214
+ "bbox": [
1215
+ 0.076,
1216
+ 0.538,
1217
+ 0.47,
1218
+ 0.614
1219
+ ],
1220
+ "angle": 0,
1221
+ "content": "where \\(\\mathbf{G} = \\mathbf{A}^{\\top}\\mathbf{A}\\) is the Gram matrix. We can interpret the columns (or rows) of the Gram matrix to be the spatially varying point spread function (PSF) of the imaging system, as shown in Fig. 6. Each column \\(\\mathbf{G}_i\\) describes how a voxel \\(\\mathbf{x}_i\\) is blurred during backprojection reconstruction."
1222
+ },
1223
+ {
1224
+ "type": "text",
1225
+ "bbox": [
1226
+ 0.076,
1227
+ 0.615,
1228
+ 0.473,
1229
+ 0.902
1230
+ ],
1231
+ "angle": 0,
1232
+ "content": "We can see the effect of different imaging parameters on the PSF in Fig. 7, where we sweep different values of (1) \\(M\\) and (2) \\(N\\). In particular, we note that the lobes of the PSF are suppressed when using temporal information, even with low spatial resolution and few number of image measurements. When using no temporal information, the PSF is still roughly centered at the correct location, but has large FWHM. The sharpness of the PSF when using temporal information indicates that it is better suited to applications that require finer voxel resolution with few-shot capture. In case (1) without transients, we see that the FWHM of the PSF is large along both the \\(x\\) and \\(y\\) direction for small values of \\(M\\). As \\(M\\) approaches the number of illumination spots, the FWHM approaches the theoretical lower bound. In case (2), increased spatial resolution helps better localize the center of the PSF, but the FWHM does not improve. The transient cases are quite robust to both low number of measurements and low spatial resolution. Intuitively, higher values of \\(N\\) in the transient case slightly help accentuate the center voxel"
1233
+ },
1234
+ {
1235
+ "type": "text",
1236
+ "bbox": [
1237
+ 0.499,
1238
+ 0.092,
1239
+ 0.878,
1240
+ 0.108
1241
+ ],
1242
+ "angle": 0,
1243
+ "content": "by projecting more diverse rays through the center voxel."
1244
+ },
1245
+ {
1246
+ "type": "text",
1247
+ "bbox": [
1248
+ 0.499,
1249
+ 0.111,
1250
+ 0.892,
1251
+ 0.202
1252
+ ],
1253
+ "angle": 0,
1254
+ "content": "Mutual Coherence We wish to further quantify the impact of spatio-temporal resolution and number of measurements on the recoverability of the hidden scene. We know from compressive sensing theory that sparse recovery of the hidden scene is possible when the mutual coherence of \\(\\mathbf{A}\\) is sufficiently small. The mutual coherence is defined as"
1255
+ },
1256
+ {
1257
+ "type": "equation",
1258
+ "bbox": [
1259
+ 0.614,
1260
+ 0.213,
1261
+ 0.892,
1262
+ 0.248
1263
+ ],
1264
+ "angle": 0,
1265
+ "content": "\\[\n\\mu = \\max _ {1 \\leq i \\neq j \\leq n} \\frac {\\left| \\mathbf {A} _ {i} ^ {\\top} \\mathbf {A} _ {j} \\right|}{\\left| \\mathbf {A} _ {i} \\right| \\left| \\mathbf {A} _ {j} \\right|}. \\tag {19}\n\\]"
1266
+ },
1267
+ {
1268
+ "type": "text",
1269
+ "bbox": [
1270
+ 0.498,
1271
+ 0.257,
1272
+ 0.893,
1273
+ 0.469
1274
+ ],
1275
+ "angle": 0,
1276
+ "content": "Intuitively, the coherence measures the largest \"similarity\" or correlation between any two column vectors in measurement matrix \\(\\mathbf{A}\\) [6]. Physically, this quantifies how similar the STIRs of two different voxels are. If \\(\\mu = 1\\), then two voxels have the exact same STIR, which means that there exists an ambiguity as to which voxel resulted in a certain spatio-temporal response. Alternatively, \\(\\mu\\) can be defined as the maximum off-diagonal entry of the Gram matrix. Fig. 8 illustrates how the Gram matrix (and subsequently the mutual coherence) is affected by coarser timing resolutions. From compressive sensing theory, we know that a \\(K\\)-sparse scene can be recovered if \\(K \\leq 0.5(1 + 1 / \\mu)\\), which motivates our empirical analysis of the effect of spatial resolution and temporal resolution on \\(\\mu\\)."
1277
+ },
1278
+ {
1279
+ "type": "text",
1280
+ "bbox": [
1281
+ 0.498,
1282
+ 0.469,
1283
+ 0.893,
1284
+ 0.742
1285
+ ],
1286
+ "angle": 0,
1287
+ "content": "From Fig. 9, we can see the impact that spatial and temporal resolution have on mutual coherence in a single-shot setup with multiplexed illumination. Consider a scene similar to that of Fig. 6, where the walls are \\(2\\mathrm{m}\\) apart, the walls are \\(3\\mathrm{m}\\) long, the axial distance to the walls is \\(2\\mathrm{m}\\), and the voxel resolution is discretized to \\(2\\mathrm{cm}\\). Increasing the spatial resolution at a fixed temporal resolution appears to have little to no effect on the mutual coherence. As Kadambi et al. also point out, improvements of coherence through higher spatial resolution are limited by the size of the virtual aperture (i.e. the wall) [16]. On the other hand, the difference between \\(100~\\mathrm{ps}\\) resolution and \\(10~\\mathrm{ps}\\) resolution reduces the coherence from 0.9 to 0.3 for a fixed spatial resolution. Any timing resolution coarser than \\(\\sim 120\\) ps results in a coherence of 1 for multiplexed measurements because pathlength differences for different shadows are no longer resolvable. This analysis motivates the importance of temporal resolution when using multiplexed illumination."
1288
+ },
1289
+ {
1290
+ "type": "title",
1291
+ "bbox": [
1292
+ 0.5,
1293
+ 0.755,
1294
+ 0.734,
1295
+ 0.772
1296
+ ],
1297
+ "angle": 0,
1298
+ "content": "5. Experiments and Results"
1299
+ },
1300
+ {
1301
+ "type": "text",
1302
+ "bbox": [
1303
+ 0.498,
1304
+ 0.78,
1305
+ 0.892,
1306
+ 0.901
1307
+ ],
1308
+ "angle": 0,
1309
+ "content": "We scan a single-pixel SPAD across the field of view on the left wall to mimic the output of a SPAD array. We scan a single laser spot across the right wall, and sum up the corresponding transient images in post-processing to mimic the effect of multiplexing. Further details are provided in the supplementary. We first validate our model using simulated data in Fig. 10, where we demonstrate that we can localize four letters and reconstruct their 2D shape with just a single"
1310
+ },
1311
+ {
1312
+ "type": "page_number",
1313
+ "bbox": [
1314
+ 0.482,
1315
+ 0.945,
1316
+ 0.516,
1317
+ 0.957
1318
+ ],
1319
+ "angle": 0,
1320
+ "content": "9197"
1321
+ }
1322
+ ],
1323
+ [
1324
+ {
1325
+ "type": "image",
1326
+ "bbox": [
1327
+ 0.084,
1328
+ 0.09,
1329
+ 0.498,
1330
+ 0.341
1331
+ ],
1332
+ "angle": 0,
1333
+ "content": null
1334
+ },
1335
+ {
1336
+ "type": "image",
1337
+ "bbox": [
1338
+ 0.501,
1339
+ 0.091,
1340
+ 0.892,
1341
+ 0.341
1342
+ ],
1343
+ "angle": 0,
1344
+ "content": null
1345
+ },
1346
+ {
1347
+ "type": "image_caption",
1348
+ "bbox": [
1349
+ 0.076,
1350
+ 0.353,
1351
+ 0.894,
1352
+ 0.409
1353
+ ],
1354
+ "angle": 0,
1355
+ "content": "Figure 7. PSF as a Function of Number of Measurements and Spatial Resolution. The PSF of the center voxel is shown in the transient and intensity-only case when multiplexed with 60 laser spots. (1) the number of pixels \\( N = 40 \\) are kept fixed with the number of measurements \\( M \\) being swept (left) and (2) \\( M = 10 \\) is kept fixed and \\( N \\) is swept (right). Transient measurements result in PSFs with narrower FWHMs, even with low spatial resolution and low number of measurements."
1356
+ },
1357
+ {
1358
+ "type": "image",
1359
+ "bbox": [
1360
+ 0.09,
1361
+ 0.434,
1362
+ 0.456,
1363
+ 0.515
1364
+ ],
1365
+ "angle": 0,
1366
+ "content": null
1367
+ },
1368
+ {
1369
+ "type": "image_caption",
1370
+ "bbox": [
1371
+ 0.076,
1372
+ 0.53,
1373
+ 0.47,
1374
+ 0.628
1375
+ ],
1376
+ "angle": 0,
1377
+ "content": "Figure 8. Gram Matrix as a Function of Temporal Resolution. For a defined voxel resolution and fixed spatial resolution. The columns (or rows) of the Gram matrix form the spatially varying PSFs shown in Fig. 6. The off-diagonal streaks describe the coherence between different voxels. As expected, higher temporal resolution decreases the coherence of the Gram matrix, which increases the likelihood of guaranteed sparse recovery."
1378
+ },
1379
+ {
1380
+ "type": "text",
1381
+ "bbox": [
1382
+ 0.076,
1383
+ 0.66,
1384
+ 0.473,
1385
+ 0.902
1386
+ ],
1387
+ "angle": 0,
1388
+ "content": "capture. We render \\( i_{o} \\) and \\( i_{l} \\) to obtain a noiseless version of the shadow transient for this experiment. Then, we reconstruct the shapes of real mannequins and must estimate \\( i_{0} \\). Fig. 11 shows that ToF measurements provide a noticeable improvement in image reconstruction. When there's no multiplexing (i.e., the number of laser spots is the same as number of measurements), the intensity-based method is similar to the ToF-based method. However, in the intensity-only case, the object becomes increasingly blurry until it is no longer visible as the amount of multiplexing increases. In the ToF case for single-shot, we can see a blurry image roughly localized where the center of the mannequin is as well as some high-frequency detail. Furthermore, the top view of the ToF reconstructions show that the high-frequency shape of the arms can be reconstructed even with multiplexed illumination, indicating the promise of single"
1389
+ },
1390
+ {
1391
+ "type": "image",
1392
+ "bbox": [
1393
+ 0.569,
1394
+ 0.435,
1395
+ 0.825,
1396
+ 0.602
1397
+ ],
1398
+ "angle": 0,
1399
+ "content": null
1400
+ },
1401
+ {
1402
+ "type": "image_caption",
1403
+ "bbox": [
1404
+ 0.498,
1405
+ 0.616,
1406
+ 0.892,
1407
+ 0.768
1408
+ ],
1409
+ "angle": 0,
1410
+ "content": "Figure 9. Tradeoffs Between Spatial Resolution and Temporal Resolution. We quantify the tradeoff using mutual coherence. We use the same setup as Fig. 6, but the walls have length \\(3\\mathrm{m}\\), the walls are space \\(2\\mathrm{m}\\) apart, and the distance to the walls is \\(2\\mathrm{m}\\). We discretize the voxel resolution to \\(2\\mathrm{cm}(\\sim 67\\mathrm{ps})\\). The bottom right of the figure, where spatial and temporal resolution are highest is where mutual coherence is the lowest. For a fixed temporal resolution, improvements in coherence quickly plateau. Improvements in temporal resolution dominate improvements in spatial resolution when reducing coherence, suggesting the importance of ToF for demultiplexing illumination in two-bounce NLOS imaging."
1411
+ },
1412
+ {
1413
+ "type": "text",
1414
+ "bbox": [
1415
+ 0.498,
1416
+ 0.795,
1417
+ 0.892,
1418
+ 0.901
1419
+ ],
1420
+ "angle": 0,
1421
+ "content": "shot capture with two-bounce ToF information. Fig. 12 shows that poorer temporal resolution results in aliasing artifacts (repeated arms and legs), whereas the rough shape of the object can still be seen with lower spatial resolution. Finally, we compare our method to the probabilistic voxel carving method used in [15] in Fig. 13. Voxel carving results in sharper reconstructions than backprojection."
1422
+ },
1423
+ {
1424
+ "type": "page_number",
1425
+ "bbox": [
1426
+ 0.483,
1427
+ 0.945,
1428
+ 0.516,
1429
+ 0.956
1430
+ ],
1431
+ "angle": 0,
1432
+ "content": "9198"
1433
+ }
1434
+ ],
1435
+ [
1436
+ {
1437
+ "type": "image",
1438
+ "bbox": [
1439
+ 0.092,
1440
+ 0.089,
1441
+ 0.459,
1442
+ 0.26
1443
+ ],
1444
+ "angle": 0,
1445
+ "content": null
1446
+ },
1447
+ {
1448
+ "type": "image_caption",
1449
+ "bbox": [
1450
+ 0.076,
1451
+ 0.271,
1452
+ 0.473,
1453
+ 0.355
1454
+ ],
1455
+ "angle": 0,
1456
+ "content": "Figure 10. Single-Shot Localization. We obtain single-shot reconstructions of the \"N\", \"L\", \"O\", and \"S\" letters in the hidden scene with and without transient information. We see that using transient information, we are able to effectively localize the hidden object and even get a sharp outline of the letters, demonstrating the promise of single-shot capture with two-bounce transients."
1457
+ },
1458
+ {
1459
+ "type": "image",
1460
+ "bbox": [
1461
+ 0.093,
1462
+ 0.392,
1463
+ 0.46,
1464
+ 0.652
1465
+ ],
1466
+ "angle": 0,
1467
+ "content": null
1468
+ },
1469
+ {
1470
+ "type": "image_caption",
1471
+ "bbox": [
1472
+ 0.076,
1473
+ 0.667,
1474
+ 0.473,
1475
+ 0.807
1476
+ ],
1477
+ "angle": 0,
1478
+ "content": "Figure 11. Transients Enable Scene Recovery from Multiplexed Illumination. We perform backprojection of measurements with varying levels of multiplexing. Plotted in the top two rows are \\( x \\) slices of the reconstructed volume. Under no multiplexing (right), intensity is sufficient to carve the shadow along detected pixel. But reconstruction quality sharply deteriorates as we move towards the multiplexing regime (left). When measuring with transients, time-of-flight information allows us to disambiguate between different shadows and recover a rough silhouette of the object even with just a single measurement."
1479
+ },
1480
+ {
1481
+ "type": "text",
1482
+ "bbox": [
1483
+ 0.076,
1484
+ 0.856,
1485
+ 0.471,
1486
+ 0.903
1487
+ ],
1488
+ "angle": 0,
1489
+ "content": "However, our method can be improved by filtering or incorporating priors, and is also generalizable to multiplexed illumination by utilizing ToF information."
1490
+ },
1491
+ {
1492
+ "type": "image",
1493
+ "bbox": [
1494
+ 0.518,
1495
+ 0.093,
1496
+ 0.852,
1497
+ 0.281
1498
+ ],
1499
+ "angle": 0,
1500
+ "content": null
1501
+ },
1502
+ {
1503
+ "type": "image_caption",
1504
+ "bbox": [
1505
+ 0.499,
1506
+ 0.289,
1507
+ 0.895,
1508
+ 0.348
1509
+ ],
1510
+ "angle": 0,
1511
+ "content": "Figure 12. Spatial Resolution vs. Timing Resolution. We plot reconstructions with different spatio-temporal resolutions. Lower timing resolution results in aliasing artifacts (i.e. repeated arms and legs). Lower spatial resolution introduces hole artifacts."
1512
+ },
1513
+ {
1514
+ "type": "image",
1515
+ "bbox": [
1516
+ 0.513,
1517
+ 0.359,
1518
+ 0.88,
1519
+ 0.502
1520
+ ],
1521
+ "angle": 0,
1522
+ "content": null
1523
+ },
1524
+ {
1525
+ "type": "image_caption",
1526
+ "bbox": [
1527
+ 0.499,
1528
+ 0.51,
1529
+ 0.895,
1530
+ 0.568
1531
+ ],
1532
+ "angle": 0,
1533
+ "content": "Figure 13. Scanning Baseline Comparison. We compare to the voxel carving method used in [15] (a) before and (b) after thresholding when scanning 1 laser spot. Our method is more general, but voxel carving yields sharper results with laser scanning."
1534
+ },
1535
+ {
1536
+ "type": "title",
1537
+ "bbox": [
1538
+ 0.501,
1539
+ 0.592,
1540
+ 0.614,
1541
+ 0.608
1542
+ ],
1543
+ "angle": 0,
1544
+ "content": "6. Discussion"
1545
+ },
1546
+ {
1547
+ "type": "text",
1548
+ "bbox": [
1549
+ 0.498,
1550
+ 0.618,
1551
+ 0.895,
1552
+ 0.816
1553
+ ],
1554
+ "angle": 0,
1555
+ "content": "In this paper, we introduced the use of two-bounce transients for NLOS imaging. Compared to 3B-NLOS, 2B-NLOS enables more robust imaging of occluded objects in the presence of scenes with arbitrary lighting conditions and low albedo objects. Transient 2B-NLOS is promising for few-shot capture of occluded objects by using multiplexed illumination, especially as SPAD array technology matures. When comparing tradeoffs in temporal resolution, spatial resolution, and number of measurements, we qualitatively and quantitatively verify with real and simulated results that temporal resolution is most important for demultiplexing shadows. We anticipate future work to extend the analysis in this paper to SPAD arrays for real-time NLOS imaging."
1556
+ },
1557
+ {
1558
+ "type": "title",
1559
+ "bbox": [
1560
+ 0.5,
1561
+ 0.83,
1562
+ 0.669,
1563
+ 0.848
1564
+ ],
1565
+ "angle": 0,
1566
+ "content": "Acknowledgements"
1567
+ },
1568
+ {
1569
+ "type": "text",
1570
+ "bbox": [
1571
+ 0.499,
1572
+ 0.856,
1573
+ 0.895,
1574
+ 0.903
1575
+ ],
1576
+ "angle": 0,
1577
+ "content": "AD and AV are supported by NSF Expeditions Award #IIS-1730574 and NSF CAREER Award #IIS-1652633. CH was supported by a Draper Scholarship."
1578
+ },
1579
+ {
1580
+ "type": "page_number",
1581
+ "bbox": [
1582
+ 0.483,
1583
+ 0.945,
1584
+ 0.516,
1585
+ 0.957
1586
+ ],
1587
+ "angle": 0,
1588
+ "content": "9199"
1589
+ }
1590
+ ],
1591
+ [
1592
+ {
1593
+ "type": "title",
1594
+ "bbox": [
1595
+ 0.08,
1596
+ 0.09,
1597
+ 0.174,
1598
+ 0.106
1599
+ ],
1600
+ "angle": 0,
1601
+ "content": "References"
1602
+ },
1603
+ {
1604
+ "type": "ref_text",
1605
+ "bbox": [
1606
+ 0.087,
1607
+ 0.116,
1608
+ 0.469,
1609
+ 0.158
1610
+ ],
1611
+ "angle": 0,
1612
+ "content": "[1] Fadel Adib and Dina Katabi. See through walls with wifi! In Proceedings of the ACM SIGCOMM 2013 conference on SIGCOMM, pages 75-86, 2013. 2"
1613
+ },
1614
+ {
1615
+ "type": "ref_text",
1616
+ "bbox": [
1617
+ 0.087,
1618
+ 0.16,
1619
+ 0.469,
1620
+ 0.229
1621
+ ],
1622
+ "angle": 0,
1623
+ "content": "[2] Byeongjoo Ahn, Akshat Dave, Ashok Veeraraghavan, Ioannis Gkioulekas, and Aswin C Sankaranarayanan. Convolutional approximations to the general non-line-of-sight imaging operator. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7889-7899, 2019. 2"
1624
+ },
1625
+ {
1626
+ "type": "ref_text",
1627
+ "bbox": [
1628
+ 0.087,
1629
+ 0.231,
1630
+ 0.469,
1631
+ 0.273
1632
+ ],
1633
+ "angle": 0,
1634
+ "content": "[3] Victor Arellano, Diego Gutierrez, and Adrian Jarabo. Fast back-projection for non-line of sight reconstruction. In ACM SIGGRAPH 2017 Posters, pages 1-2. 2017. 2"
1635
+ },
1636
+ {
1637
+ "type": "ref_text",
1638
+ "bbox": [
1639
+ 0.087,
1640
+ 0.274,
1641
+ 0.469,
1642
+ 0.355
1643
+ ],
1644
+ "angle": 0,
1645
+ "content": "[4] Katherine L Bouman, Vickie Ye, Adam B Yedidia, Frédo Durand, Gregory W Wornell, Antonio Torralba, and William T Freeman. Turning corners into cameras: Principles and methods. In Proceedings of the IEEE International Conference on Computer Vision, pages 2270-2278, 2017. 2, 3"
1646
+ },
1647
+ {
1648
+ "type": "ref_text",
1649
+ "bbox": [
1650
+ 0.087,
1651
+ 0.359,
1652
+ 0.469,
1653
+ 0.413
1654
+ ],
1655
+ "angle": 0,
1656
+ "content": "[5] Mauro Buttafava, Jessica Zeman, Alberto Tosi, Kevin Eliceiri, and Andreas Velten. Non-line-of-sight imaging using a time-gated single photon avalanche diode. Optics express, 23(16):20997-21011, 2015. 2"
1657
+ },
1658
+ {
1659
+ "type": "ref_text",
1660
+ "bbox": [
1661
+ 0.087,
1662
+ 0.417,
1663
+ 0.469,
1664
+ 0.457
1665
+ ],
1666
+ "angle": 0,
1667
+ "content": "[6] Emmanuel J Candès and Michael B Wakin. An introduction to compressive sampling. IEEE signal processing magazine, 25(2):21-30, 2008. 6"
1668
+ },
1669
+ {
1670
+ "type": "ref_text",
1671
+ "bbox": [
1672
+ 0.087,
1673
+ 0.46,
1674
+ 0.469,
1675
+ 0.515
1676
+ ],
1677
+ "angle": 0,
1678
+ "content": "[7] Wenzheng Chen, Simon Daneau, Fahim Mannan, and Felix Heide. Steady-state non-line-of-sight imaging. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6790-6799, 2019. 2"
1679
+ },
1680
+ {
1681
+ "type": "ref_text",
1682
+ "bbox": [
1683
+ 0.087,
1684
+ 0.518,
1685
+ 0.469,
1686
+ 0.572
1687
+ ],
1688
+ "angle": 0,
1689
+ "content": "[8] Wenzheng Chen, Fangyin Wei, Kiriakos N Kutulakos, Szymon Rusinkiewicz, and Felix Heide. Learned feature embeddings for non-line-of-sight imaging and recognition. ACM Transactions on Graphics (ToG), 39(6):1-18, 2020. 2"
1690
+ },
1691
+ {
1692
+ "type": "ref_text",
1693
+ "bbox": [
1694
+ 0.087,
1695
+ 0.575,
1696
+ 0.469,
1697
+ 0.644
1698
+ ],
1699
+ "angle": 0,
1700
+ "content": "[9] Akshit Dave, Muralidhar Madabhushi Balaji, Prasanna Rangarajan, Ashok Veeraraghavan, and Marc P Christensen. Foveated non-line-of-sight imaging. In Computational Optical Sensing and Imaging, pages CTh5C-6. Optica Publishing Group, 2020. 2"
1701
+ },
1702
+ {
1703
+ "type": "ref_text",
1704
+ "bbox": [
1705
+ 0.08,
1706
+ 0.647,
1707
+ 0.469,
1708
+ 0.715
1709
+ ],
1710
+ "angle": 0,
1711
+ "content": "[10] Akshit Dave, Yongyi Zhao, and Ashok Veeraraghavan. Pandora: Polarization-aided neural decomposition of radiance. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23–27, 2022, Proceedings, Part VII, pages 538–556. Springer, 2022. 2"
1712
+ },
1713
+ {
1714
+ "type": "ref_text",
1715
+ "bbox": [
1716
+ 0.08,
1717
+ 0.718,
1718
+ 0.469,
1719
+ 0.771
1720
+ ],
1721
+ "angle": 0,
1722
+ "content": "[11] Genevieve Gariepy, Francesco Tonolini, Robert Henderson, Jonathan Leach, and Daniele Faccio. Detection and tracking of moving objects hidden from view. Nature Photonics, 10(1):23-26, 2016. 2"
1723
+ },
1724
+ {
1725
+ "type": "ref_text",
1726
+ "bbox": [
1727
+ 0.08,
1728
+ 0.774,
1729
+ 0.469,
1730
+ 0.829
1731
+ ],
1732
+ "angle": 0,
1733
+ "content": "[12] Felix Heide, Matthew O'Toole, Kai Zang, David B Lindell, Steven Diamond, and Gordon Wetzstein. Non-line-of-sight imaging with partial occluders and surface normals. ACM Transactions on Graphics (ToG), 38(3):1-10, 2019. 2"
1734
+ },
1735
+ {
1736
+ "type": "ref_text",
1737
+ "bbox": [
1738
+ 0.08,
1739
+ 0.832,
1740
+ 0.469,
1741
+ 0.901
1742
+ ],
1743
+ "angle": 0,
1744
+ "content": "[13] Felix Heide, Lei Xiao, Wolfgang Heidrich, and Matthias B Hullin. Diffuse mirrors: 3d reconstruction from diffuse indirect illumination using inexpensive time-of-flight sensors. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3222-3229, 2014. 2"
1745
+ },
1746
+ {
1747
+ "type": "list",
1748
+ "bbox": [
1749
+ 0.08,
1750
+ 0.116,
1751
+ 0.469,
1752
+ 0.901
1753
+ ],
1754
+ "angle": 0,
1755
+ "content": null
1756
+ },
1757
+ {
1758
+ "type": "ref_text",
1759
+ "bbox": [
1760
+ 0.503,
1761
+ 0.093,
1762
+ 0.892,
1763
+ 0.134
1764
+ ],
1765
+ "angle": 0,
1766
+ "content": "[14] Connor Henley, Joseph Hollmann, and Ramesh Raskar. Bounce-flash lidar. IEEE Transactions on Computational Imaging, 8:411-424, 2022. 2"
1767
+ },
1768
+ {
1769
+ "type": "ref_text",
1770
+ "bbox": [
1771
+ 0.503,
1772
+ 0.136,
1773
+ 0.892,
1774
+ 0.19
1775
+ ],
1776
+ "angle": 0,
1777
+ "content": "[15] Connor Henley, Tomohiro Maeda, Tristan Swedish, and Ramesh Raskar. Imaging behind occluders using two-bounce light. In ECCV, pages 573-588. Springer, 2020. 1, 2, 7, 8"
1778
+ },
1779
+ {
1780
+ "type": "ref_text",
1781
+ "bbox": [
1782
+ 0.503,
1783
+ 0.193,
1784
+ 0.892,
1785
+ 0.234
1786
+ ],
1787
+ "angle": 0,
1788
+ "content": "[16] Achuta Kadambi, Hang Zhao, Boxin Shi, and Ramesh Raskar. Occluded imaging with time-of-flight sensors. ACM Transactions on Graphics (ToG), 35(2):1-12, 2016. 2, 6"
1789
+ },
1790
+ {
1791
+ "type": "ref_text",
1792
+ "bbox": [
1793
+ 0.503,
1794
+ 0.236,
1795
+ 0.892,
1796
+ 0.291
1797
+ ],
1798
+ "angle": 0,
1799
+ "content": "[17] Ori Katz, Pierre Heidmann, Mathias Fink, and Sylvain Gigan. Non-invasive single-shot imaging through scattering layers and around corners via speckle correlations. Nature photonics, 8(10):784-790, 2014. 2"
1800
+ },
1801
+ {
1802
+ "type": "ref_text",
1803
+ "bbox": [
1804
+ 0.504,
1805
+ 0.293,
1806
+ 0.892,
1807
+ 0.347
1808
+ ],
1809
+ "angle": 0,
1810
+ "content": "[18] Ahmed Kirmani, Tyler Hutchison, James Davis, and Ramesh Raskar. Looking around the corner using transient imaging. In 2009 IEEE 12th International Conference on Computer Vision, pages 159-166. IEEE, 2009. 2, 3"
1811
+ },
1812
+ {
1813
+ "type": "ref_text",
1814
+ "bbox": [
1815
+ 0.504,
1816
+ 0.349,
1817
+ 0.892,
1818
+ 0.445
1819
+ ],
1820
+ "angle": 0,
1821
+ "content": "[19] Oichi Kumagai, Junichi Ohmachi, Masao Matsumura, Shinichiro Yagi, Kenichi Tayu, Keitaro Amagawa, Tomohiro Matsukawa, Osamu Ozawa, Daisuke Hirono, Yasuhiro Shinozuka, et al. 7.3 a \\(189 \\times 600\\) back-illuminated stacked spad direct time-of-flight depth sensor for automotive lidar systems. In 2021 IEEE International Solid-State Circuits Conference (ISSCC), volume 64, pages 110-112. IEEE, 2021. 2"
1822
+ },
1823
+ {
1824
+ "type": "ref_text",
1825
+ "bbox": [
1826
+ 0.504,
1827
+ 0.447,
1828
+ 0.892,
1829
+ 0.474
1830
+ ],
1831
+ "angle": 0,
1832
+ "content": "[20] Aldo Laurentini. The visual hull concept for silhouette-based image understanding. PAMI, 16(2):150-162, 1994. 1"
1833
+ },
1834
+ {
1835
+ "type": "ref_text",
1836
+ "bbox": [
1837
+ 0.504,
1838
+ 0.476,
1839
+ 0.892,
1840
+ 0.517
1841
+ ],
1842
+ "angle": 0,
1843
+ "content": "[21] Di Lin, Connor Hashemi, and James R Leger. Passive non-line-of-sight imaging using plenoptic information. JOSA A, 37(4):540-551, 2020. 2"
1844
+ },
1845
+ {
1846
+ "type": "ref_text",
1847
+ "bbox": [
1848
+ 0.504,
1849
+ 0.519,
1850
+ 0.892,
1851
+ 0.572
1852
+ ],
1853
+ "angle": 0,
1854
+ "content": "[22] David B Lindell, Gordon Wetzstein, and Matthew O'Toole. Wave-based non-line-of-sight imaging using fast fkr migration. ACM Transactions on Graphics (ToG), 38(4):1-13, 2019. 2"
1855
+ },
1856
+ {
1857
+ "type": "ref_text",
1858
+ "bbox": [
1859
+ 0.504,
1860
+ 0.576,
1861
+ 0.892,
1862
+ 0.644
1863
+ ],
1864
+ "angle": 0,
1865
+ "content": "[23] Xiaochun Liu, Ibón Guillén, Marco La Manna, Ji Hyun Nam, Syed Azer Reza, Toan Huu Le, Adrian Jarabo, Diego Gutierrez, and Andreas Velten. Non-line-of-sight imaging using phasor-field virtual wave optics. Nature, 572(7771):620-623, 2019. 2"
1866
+ },
1867
+ {
1868
+ "type": "ref_text",
1869
+ "bbox": [
1870
+ 0.504,
1871
+ 0.647,
1872
+ 0.892,
1873
+ 0.687
1874
+ ],
1875
+ "angle": 0,
1876
+ "content": "[24] Tomohiro Maeda, Guy Satat, Tristan Swedish, Lagnojita Sinha, and Ramesh Raskar. Recent advances in imaging around corners. arXiv preprint arXiv:1910.05613, 2019. 1"
1877
+ },
1878
+ {
1879
+ "type": "ref_text",
1880
+ "bbox": [
1881
+ 0.504,
1882
+ 0.689,
1883
+ 0.892,
1884
+ 0.758
1885
+ ],
1886
+ "angle": 0,
1887
+ "content": "[25] Christopher A Metzler, Felix Heide, Prasana Rangarajan, Muralidhar Madabhushi Balaji, Aparna Viswanath, Ashok Veeraraghavan, and Richard G Baraniuk. Deep-inverse correlography: towards real-time high-resolution non-line-of-sight imaging. Optica, 7(1):63-71, 2020. 2"
1888
+ },
1889
+ {
1890
+ "type": "ref_text",
1891
+ "bbox": [
1892
+ 0.504,
1893
+ 0.76,
1894
+ 0.892,
1895
+ 0.828
1896
+ ],
1897
+ "angle": 0,
1898
+ "content": "[26] Kazuhiro Morimoto, Andrei Ardelean, Ming-Lo Wu, Arin Can Ulku, Ivan Michel Antolovic, Claudio Bruschini, and Edoardo Charbon. Megapixel time-gated spad image sensor for 2d and 3d imaging applications. Optica, 7(4):346–354, 2020. 2"
1899
+ },
1900
+ {
1901
+ "type": "ref_text",
1902
+ "bbox": [
1903
+ 0.504,
1904
+ 0.831,
1905
+ 0.892,
1906
+ 0.871
1907
+ ],
1908
+ "angle": 0,
1909
+ "content": "[27] Matthew O'Toole, David B Lindell, and Gordon Wetzstein. Confocal non-line-of-sight imaging based on the light-cone transform. Nature, 555(7696):338-341, 2018. 2, 3"
1910
+ },
1911
+ {
1912
+ "type": "ref_text",
1913
+ "bbox": [
1914
+ 0.504,
1915
+ 0.874,
1916
+ 0.892,
1917
+ 0.901
1918
+ ],
1919
+ "angle": 0,
1920
+ "content": "[28] Adithya Pediredla, Akshit Dave, and Ashok Veeraraghavan. Snlos: Non-line-of-sight scanning through temporal focus"
1921
+ },
1922
+ {
1923
+ "type": "list",
1924
+ "bbox": [
1925
+ 0.503,
1926
+ 0.093,
1927
+ 0.892,
1928
+ 0.901
1929
+ ],
1930
+ "angle": 0,
1931
+ "content": null
1932
+ },
1933
+ {
1934
+ "type": "page_number",
1935
+ "bbox": [
1936
+ 0.483,
1937
+ 0.946,
1938
+ 0.516,
1939
+ 0.956
1940
+ ],
1941
+ "angle": 0,
1942
+ "content": "9200"
1943
+ }
1944
+ ],
1945
+ [
1946
+ {
1947
+ "type": "ref_text",
1948
+ "bbox": [
1949
+ 0.11,
1950
+ 0.093,
1951
+ 0.468,
1952
+ 0.12
1953
+ ],
1954
+ "angle": 0,
1955
+ "content": "ing. In 2019 IEEE International Conference on Computational Photography (ICCP), pages 1-13. IEEE, 2019. 2"
1956
+ },
1957
+ {
1958
+ "type": "ref_text",
1959
+ "bbox": [
1960
+ 0.08,
1961
+ 0.122,
1962
+ 0.469,
1963
+ 0.19
1964
+ ],
1965
+ "angle": 0,
1966
+ "content": "[29] Adithya K Pediredla, Aswin C Sankaranarayanan, Mauro Buttafava, Alberto Tosi, and Ashok Veeraraghavan. Signal processing based pile-up compensation for gated single-photon avalanche diodes. arXiv preprint arXiv:1806.07437, 2018.5"
1967
+ },
1968
+ {
1969
+ "type": "ref_text",
1970
+ "bbox": [
1971
+ 0.08,
1972
+ 0.193,
1973
+ 0.469,
1974
+ 0.274
1975
+ ],
1976
+ "angle": 0,
1977
+ "content": "[30] Joshua Rapp, Charles Saunders, Julián Tachella, John Murray-Bruce, Yoann Altmann, Jean-Yves Tourneret, Stephen McLaughlin, Robin Dawson, Franco NC Wong, and Vivek K Goyal. Seeing around corners with edge-resolved transient imaging. Nature communications, 11(1):1-10, 2020. 2, 3"
1978
+ },
1979
+ {
1980
+ "type": "ref_text",
1981
+ "bbox": [
1982
+ 0.08,
1983
+ 0.277,
1984
+ 0.468,
1985
+ 0.318
1986
+ ],
1987
+ "angle": 0,
1988
+ "content": "[31] Charles Saunders, John Murray-Bruce, and Vivek K Goyal. Computational periscopy with an ordinary digital camera. Nature, 565(7740):472-475, 2019. 3"
1989
+ },
1990
+ {
1991
+ "type": "ref_text",
1992
+ "bbox": [
1993
+ 0.08,
1994
+ 0.321,
1995
+ 0.469,
1996
+ 0.417
1997
+ ],
1998
+ "angle": 0,
1999
+ "content": "[32] Nicolas Scheiner, Florian Kraus, Fangyin Wei, Buu Phan, Fahim Mannan, Nils Appenrodt, Werner Ritter, Jurgen Dickmann, Klaus Dietmayer, Bernhard Sick, et al. Seeing around street corners: Non-line-of-sight detection and tracking in-the-wild using doppler radar. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2068-2077, 2020. 2"
2000
+ },
2001
+ {
2002
+ "type": "ref_text",
2003
+ "bbox": [
2004
+ 0.08,
2005
+ 0.42,
2006
+ 0.469,
2007
+ 0.488
2008
+ ],
2009
+ "angle": 0,
2010
+ "content": "[33] Prafull Sharma, Miika Aittala, Yoav Y Schechner, Antonio Torralba, Gregory W Wornell, William T Freeman, and Frédo Durand. What you can learn by staring at a blank wall. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2330-2339, 2021. 2, 3"
2011
+ },
2012
+ {
2013
+ "type": "ref_text",
2014
+ "bbox": [
2015
+ 0.08,
2016
+ 0.49,
2017
+ 0.469,
2018
+ 0.558
2019
+ ],
2020
+ "angle": 0,
2021
+ "content": "[34] Brandon M Smith, Matthew O'Toole, and Mohit Gupta. Tracking multiple objects outside the line of sight using speckle imaging. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6258-6266, 2018. 2"
2022
+ },
2023
+ {
2024
+ "type": "ref_text",
2025
+ "bbox": [
2026
+ 0.08,
2027
+ 0.562,
2028
+ 0.469,
2029
+ 0.628
2030
+ ],
2031
+ "angle": 0,
2032
+ "content": "[35] Tristan Swedish, Connor Henley, and Ramesh Raskar. Objects as cameras: Estimating high-frequency illumination from shadows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2593-2602, 2021. 2"
2033
+ },
2034
+ {
2035
+ "type": "ref_text",
2036
+ "bbox": [
2037
+ 0.08,
2038
+ 0.632,
2039
+ 0.469,
2040
+ 0.687
2041
+ ],
2042
+ "angle": 0,
2043
+ "content": "[36] Kenichiro Tanaka, Yasuhiro Mukaigawa, and Achuta Kadambi. Polarized non-line-of-sight imaging. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2136-2145, 2020. 2"
2044
+ },
2045
+ {
2046
+ "type": "ref_text",
2047
+ "bbox": [
2048
+ 0.08,
2049
+ 0.689,
2050
+ 0.469,
2051
+ 0.73
2052
+ ],
2053
+ "angle": 0,
2054
+ "content": "[37] Matthew Tancik, Guy Satat, and Ramesh Raskar. Flash photography for data-driven hidden scene recovery. arXiv preprint arXiv:1810.11710, 2018. 2"
2055
+ },
2056
+ {
2057
+ "type": "ref_text",
2058
+ "bbox": [
2059
+ 0.08,
2060
+ 0.732,
2061
+ 0.469,
2062
+ 0.786
2063
+ ],
2064
+ "angle": 0,
2065
+ "content": "[38] Kushagra Tiwary, Askhat Dave, Nikhil Behari, Tzofi Klinghoffer, Ashok Veeraraghavan, and Ramesh Raskar. Orca: Glossy objects as radiance field cameras. arXiv preprint arXiv:2212.04531, 2022. 2"
2066
+ },
2067
+ {
2068
+ "type": "ref_text",
2069
+ "bbox": [
2070
+ 0.08,
2071
+ 0.789,
2072
+ 0.469,
2073
+ 0.856
2074
+ ],
2075
+ "angle": 0,
2076
+ "content": "[39] Andreas Velten, Thomas Willwacher, Otkrist Gupta, Ashok Veeraraghavan, Moungi G Bawendi, and Ramesh Raskar. Recovering three-dimensional shape around a corner using ultrafast time-of-flight imaging. Nature communications, 3(1):1-8, 2012. 2"
2077
+ },
2078
+ {
2079
+ "type": "ref_text",
2080
+ "bbox": [
2081
+ 0.08,
2082
+ 0.859,
2083
+ 0.469,
2084
+ 0.901
2085
+ ],
2086
+ "angle": 0,
2087
+ "content": "[40] Florian Willomitzer, Fengqiang Li, Muralidhar Madabhushi Balaji, Prasanna Ranganajan, and Oliver Cossairt. High resolution non-line-of-sight imaging with superheterodyne re"
2088
+ },
2089
+ {
2090
+ "type": "list",
2091
+ "bbox": [
2092
+ 0.08,
2093
+ 0.093,
2094
+ 0.469,
2095
+ 0.901
2096
+ ],
2097
+ "angle": 0,
2098
+ "content": null
2099
+ },
2100
+ {
2101
+ "type": "ref_text",
2102
+ "bbox": [
2103
+ 0.535,
2104
+ 0.093,
2105
+ 0.892,
2106
+ 0.133
2107
+ ],
2108
+ "angle": 0,
2109
+ "content": "mote digital holography. In Computational Optical Sensing and Imaging, pages CM2A-2. Optica Publishing Group, 2019. 2"
2110
+ },
2111
+ {
2112
+ "type": "ref_text",
2113
+ "bbox": [
2114
+ 0.503,
2115
+ 0.136,
2116
+ 0.892,
2117
+ 0.205
2118
+ ],
2119
+ "angle": 0,
2120
+ "content": "[41] Florian Willomitzer, Prasanna V Rangarajan, Fengqiang Li, Muralidhar M Balaji, Marc P Christensen, and Oliver Cossaint. Fast non-line-of-sight imaging with high-resolution and wide field of view using synthetic wavelength holography. Nature communications, 12(1):6647, 2021. 2"
2121
+ },
2122
+ {
2123
+ "type": "ref_text",
2124
+ "bbox": [
2125
+ 0.503,
2126
+ 0.207,
2127
+ 0.892,
2128
+ 0.289
2129
+ ],
2130
+ "angle": 0,
2131
+ "content": "[42] Shumian Xin, Sotiris Nousias, Kiriakos N Kutulakos, Aswin C Sankaranarayanan, Srinivasa G Narasimhan, and Ioannis Gkioulekas. A theory of Fermat paths for nonline-of-sight shape reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6800-6809, 2019. 2, 3"
2132
+ },
2133
+ {
2134
+ "type": "ref_text",
2135
+ "bbox": [
2136
+ 0.503,
2137
+ 0.291,
2138
+ 0.892,
2139
+ 0.358
2140
+ ],
2141
+ "angle": 0,
2142
+ "content": "[43] Shichao Yue, Hao He, Peng Cao, Kaiwen Zha, Masayuki Koizumi, and Dina Katabi. Cornerradar: Rf-based indoor localization around corners. Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies, 6(1):1-24, 2022. 2"
2143
+ },
2144
+ {
2145
+ "type": "ref_text",
2146
+ "bbox": [
2147
+ 0.504,
2148
+ 0.361,
2149
+ 0.892,
2150
+ 0.429
2151
+ ],
2152
+ "angle": 0,
2153
+ "content": "[44] Chao Zhang, Ning Zhang, Zhijie Ma, Letian Wang, Yu Qin, Jieyang Jia, and Kai Zang. A \\(240 \\times 160\\) 3d-stacked spad dtof image sensor with rolling shutter and in-pixel histogram for mobile devices. IEEE Open Journal of the Solid-State Circuits Society, 2:3-11, 2021. 2"
2154
+ },
2155
+ {
2156
+ "type": "list",
2157
+ "bbox": [
2158
+ 0.503,
2159
+ 0.093,
2160
+ 0.892,
2161
+ 0.429
2162
+ ],
2163
+ "angle": 0,
2164
+ "content": null
2165
+ },
2166
+ {
2167
+ "type": "page_number",
2168
+ "bbox": [
2169
+ 0.483,
2170
+ 0.946,
2171
+ 0.514,
2172
+ 0.956
2173
+ ],
2174
+ "angle": 0,
2175
+ "content": "9201"
2176
+ }
2177
+ ]
2178
+ ]
2023/Role of Transients in Two-Bounce Non-Line-of-Sight Imaging/dd8b3d32-1f20-46bb-9a36-d731b9d46ed2_origin.pdf ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:7d685831a1ac2b9a357c07e2025cac0753c54c4661e70cef2ce46207cf71f579
3
+ size 7207324
2023/Role of Transients in Two-Bounce Non-Line-of-Sight Imaging/full.md ADDED
@@ -0,0 +1,333 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Role of Transients in Two-Bounce Non-Line-of-Sight Imaging
2
+
3
+ Siddharth Somasundaram<sup>1</sup> Akshit Dave<sup>2</sup> Connor Henley<sup>1</sup> Ashok Veeraraghavan<sup>2</sup> Ramesh Raskar<sup>1</sup>
4
+
5
+ <sup>1</sup>Massachusetts Institute of Technology <sup>2</sup>Rice University
6
+
7
+ {sidsoma,co24401,raskar}@mit.edu,{akshat,vashok}@rice.edu
8
+
9
+ ![](images/7aeab3f60d5fffa2bd2ad3467548e8845c54e83e6be1cc168a60191a04250c76.jpg)
10
+ (a) Two-Bounce NLOS
11
+
12
+ ![](images/118345b65ad62ac1f6b1a164a33c258b22fb20ad950318959ea43658f57ef8d0.jpg)
13
+ (b) Application Settings
14
+
15
+ ![](images/52c0223be557335de100b98aeb25d634c86691168774718254326334a72d2710.jpg)
16
+ (c) Experimental Setup
17
+ Figure 1. Two-bounce Transient NLOS Imaging. (a) We propose using two-bounce transients for NLOS imaging. Two-bounce signals capture information about the shadows in a scene, and use of transient information can reduce the number of measurements needed for reconstruction by using multiplexed illumination. (b) Our work can be applicable in many scenarios, such as tunnels, corridors, streets, and forests. (c)-(d) We experimentally validate the use of two-bounce transients on real objects. We show that time-of-flight information enables more robust reconstructions of occluded objects while using fewer measurements.
18
+
19
+ ![](images/c2882cdcc836f088655b91a89b156cfbbe64c7a1a2d68e2e585d6f6782bc3ac9.jpg)
20
+ (d) Reconstruction with 10 Measurements
21
+
22
+ # Abstract
23
+
24
+ The goal of non-line-of-sight (NLOS) imaging is to image objects occluded from the camera's field of view using multiply scattered light. Recent works have demonstrated the feasibility of two-bounce (2B) NLOS imaging by scanning a laser and measuring cast shadows of occluded objects in scenes with two relay surfaces. In this work, we study the role of time-of-flight (ToF) measurements, i.e. transients, in 2B-NLOS under multiplexed illumination. Specifically, we study how ToF information can reduce the number of measurements and spatial resolution needed for shape reconstruction. We present our findings with respect to trade-offs in (1) temporal resolution, (2) spatial resolution, and (3) number of image captures by studying SNR and recoverability as functions of system parameters. This leads to a formal definition of the mathematical constraints for 2B lidar. We believe that our work lays an analytical groundwork for design of future NLOS imaging systems, especially as ToF sensors become increasingly ubiquitous.
25
+
26
+ # 1. Introduction
27
+
28
+ Non-line-of-sight (NLOS) imaging aims to reconstruct objects occluded from direct line of sight and has the potential to be transformative in numerous applications across autonomous driving, search and rescue, and non-invasive medical imaging [24]. The key approach is to measure light that has undergone multiple surface scattering events and computationally invert these measurements to estimate hidden geometries. Recent work used two-bounce light, as shown in Fig. 2a, to reconstruct high quality shapes behind occluders [15]. The key idea is that two-bounce light captures information about the shadows of the occluded object. By scanning the laser source at different points 1 on a relay surface and measuring multiple shadow images, it is possible to reconstruct the hidden object by computing the visual hull [20] of the measured shadows.
29
+
30
+ Benefits of Two-Bounce Two-bounce (2B) light can be captured using two relay surfaces on opposite sides of the hidden scene (Fig. 1a). This can occur in a variety of real-world settings such as tunnels, hallways, streets, and clit
31
+
32
+ ![](images/dac7c60efc57b3d2780b5a921c14a32785da739cb4a0e855d406b94c62a09000.jpg)
33
+ (a) Two-Bounce NLOS Imaging
34
+
35
+ ![](images/afdd89e2ce005569ede0fd9f3f24ea328c4a24daef53b9947c7063d6f5e8f52d.jpg)
36
+ (b) Two-Bounce Transient NLOS
37
+ Figure 2. Two-Bounce NLOS Imaging. (a) Two-bounce NLOS imaging performs 3D reconstruction by using information contained in the shadows of the hidden object when illuminated by one virtual source 1 at a time. (b) In this work, we study the benefits of using transient information for two-bounce NLOS imaging under the presence of multiplexed illumination.
38
+
39
+ tered environments like forests (Fig. 1b). For example, imagine an autonomous vehicle in a tunnel being able to see ahead of the car in front of it by using two-bounce signals from the sides of the tunnel. Two-bounce light also helps alleviate the signal-to-noise ratio (SNR) issue in 3B-NLOS imaging, since 3B signals experience significant signal attenuation at each surface scatter event. Furthermore, 2B light captures information about the cast shadows of the hidden object instead of reflectance, as in the 3B case. As a result, the SNR of 2B signals is independent of the object's albedo and can therefore robustly image darker objects.
40
+
41
+ Contributions In this work, we propose using two-bounce transients for NLOS imaging to enable the use of multiplexed illumination. 3B-NLOS transient imaging systems have been analyzed previously, but we believe we are the first to analyze the use of multiplexed illumination and transient measurements for 2B-NLOS (Fig. 2b). Our key insight is that multiplexed illumination enables measurement of occluded objects with fewer image captures, and 2B transients enable demultiplexing shadows, as shown in Fig. 3. The robust SNR and promise of few-shot capture afforded by 2B transients makes the idea an important direction for few-shot NLOS imaging in general environments. We summarize our contributions below.
42
+
43
+ - Model: We present new insights into the algebraic formulations of two-bounce lidar for NLOS imaging.
44
+ - Analysis: We analyze tradeoffs with spatial resolution, temporal resolution, and number of image captures by using quantitative metrics such as SNR and recoverability, as well as qualitative validation from real and simulated results.
45
+
46
+ Although we envision the ideas introduced here to inspire future works in single-shot NLOS imaging, we do not claim this as a contribution in this paper.
47
+
48
+ Scope of This Work All data used for this work is from simulation and a single-pixel SPAD sensor. We do not physically realize few-shot results in this paper due to limited availability of high-resolution SPAD arrays presently. Instead, we emulate SPAD array measurements by scanning a single-pixel SPAD across the field of view and and emulate multiplexed illumination by scanning a laser spot, acquiring individual transient images, and summing the images in post-processing. These measurements, however, are more conducive to our goal of performing analysis that explores the landscape of possibilities in NLOS imaging with advances in ToF sensors. We believe that the ideas and analysis introduced in this work will be increasingly relevant as SPAD array technology matures [19,26,44].
49
+
50
+ # 2. Related Work
51
+
52
+ # Non-Line-of-Sight Imaging
53
+
54
+ ToF Methods: The non-line-of-sight imaging (NLOS) problem aims to see around a corner by measuring light reflected off intermediary surfaces. A large body of work models the response of a hidden scene to a temporal pulse of light and generally perform a variant of ellipsoidal tomography or solve a linear inverse problem [3, 5, 11, 12, 18, 39]. The temporal profiles can probe features of the hidden space, such as Fermat paths [42], scene impulse response [23], and individual voxels [28]. Other works derive efficient algorithms using a confocal setup [2, 22, 27] and data-driven techniques [8]. NLOS has also been demonstrated with amplitude modulated continuous wave (AMCW) ToF by exploiting sparsity and fitting measurements to a Lambertian reflectance model [13, 16].
55
+
56
+ Other Active Methods: Chen et al. use steady-state (i.e. time-integrated) intensity information by leveraging partial directionality of diffuse reflectors [7]. Radio wavelengths can also be used to image through walls and/or around them [1, 32, 43]. There has also been significant progress in exploiting speckle correlations [9, 17, 25, 34] and synthetic wave holography [40, 41] for NLOS imaging. Henley et al. demonstrate the use of two-bounce light (i.e. shadows) for NLOS using intensity sensors [15]. An extension of this work uses ToF constraints on two-bounce pathlengths to isolate shadows from ambient background and higher-order inter-reflections [14]. However multiplexing was not used.
57
+
58
+ Passive Methods: Passive methods make use of optical information already present in the scene without an active source. Many passive approaches exploit polarization cues [10,36], ambient thermal light [21,36], shadows [35], multiview reflections [38], and neural networks [33,37] to reconstruct or make inferences about the scene outside the camera's field of view. Occlusion-assisted imaging makes use of information present at sharp discontinuities in LOS and NLOS scenes by turning them into virtual pinholes [4,30] or
59
+
60
+ <table><tr><td rowspan="2">NLOS Imaging Domain</td><td colspan="3">Key Components</td><td colspan="3">Key Features</td></tr><tr><td>Assisting Surfaces</td><td>Sensor Type</td><td>Light Source</td><td>Number of Captures</td><td>Recovered Dimensions</td><td>Signal Quality</td></tr><tr><td rowspan="2">Passive Occlusion-Aided [4,31,33]</td><td>Relay Wall/Floor</td><td rowspan="2">Intensity</td><td rowspan="2">Passive</td><td rowspan="2">Low</td><td rowspan="2">2D</td><td rowspan="2">Low (strong background)</td></tr><tr><td>Edge Occluders</td></tr><tr><td rowspan="2">Three-Bounce Transient [18,27,30,42]</td><td>Relay Wall/Floor</td><td rowspan="2">ToF</td><td rowspan="2">Active (Scan)</td><td rowspan="2">High</td><td rowspan="2">3D</td><td rowspan="2">Medium (3-bounce)</td></tr><tr><td>Edge Occluders</td></tr><tr><td>Two-Bounce Transient (Ours)</td><td>Two LOS Walls</td><td>ToF</td><td>Active (Flash)</td><td>Low</td><td>3D</td><td>High (2-bounce)</td></tr></table>
61
+
62
+ Table 1. Comparison of NLOS Methods. Passive approaches require fewer measurements but can either recover up to two dimensions or are strongly affected by ambient noise. Works exploiting three-bounce light paths demonstrate 3D imaging around the corner but require a bounce on the hidden surface, reducing the signal quality. We analyze exploiting two-bounce transients for NLOS imaging.
63
+
64
+ virtual pinspecks [31]. Intensity-only methods are difficult to scale to scenarios with arbitrary lighting conditions.
65
+
66
+ # 3. Forward and Adjoint Operators
67
+
68
+ # 3.1. Problem Setup
69
+
70
+ We first introduce terminology and notation for our problem formulation. The object of interest lies behind an occluder and is flanked on two sides by planar Lambertian relay surfaces. Without loss of generality, the right wall is the illumination wall $\mathcal{L}$ and the left wall is the observation wall $\mathcal{O}$ . Light incident on the illumination wall at a point $l$ is a virtual source. Each virtual source diffusely radiates light towards the observation wall. Light will arrive at a virtual detector $s$ from $l$ if the object doesn't lie along the ray connecting $s$ and $l$ , as shown in Fig. 2a. If the object obstructs this ray, $s$ is said to lie in the shadow of the object. We assume that every virtual source and virtual detector is visible to the laser and camera field of view, respectively. The hidden scene is discretized into a voxel grid described by a binary occupancy (or transparency) function. This implicitly assumes that we are reconstructing opaque surfaces. We neglect the effect of higher order bounces (e.g. interreflections, 3B light) because SPADs can temporally gate out the longer ToF of 3B light. Furthermore, the intensity of higher order bounces will be much lower than that of two-bounce light because all surfaces are assumed to be diffuse.
71
+
72
+ We denote light transients as ToF measurements for light that arrives at the virtual detector and shadow transients as ToF measurements for light that didn't arrive at the virtual detector because of the object's shadow. Both of these are two-bounce transients. We show that modeling the problem in terms of shadow transients enables reconstruction of the hidden objects shape by solving a linear inverse problem.
73
+
74
+ ![](images/6f41010c71b9b4e84377d5a0bb51f56fd62c5842531823a5568b6a7bf5682fd8.jpg)
75
+ Figure 3. Time-of-Flight Information for Demultiplexing Shadows. Assume the point $s$ lies in shadow. When using multiplexed illumination, there exists an ambiguity as to whether $s$ lies in the shadow of $l_1$ , $l_2$ , or both. However, the paths $L - l_1 - s - C$ and $L - l_2 - s - C$ have different pathlengths, suggesting the usefulness of ToF information for demultiplexing shadows.
76
+
77
+ # 3.2. Linear Forward Model
78
+
79
+ Light Transients Consider the case where the occluded scene is empty. The measured transient is then given by
80
+
81
+ $$
82
+ i _ {0} (t; \mathbf {s}) = \alpha \iint_ {\mathcal {L}} \frac {\delta (\| \mathbf {l} - \mathbf {g} \| + \| \mathbf {l} - \mathbf {s} \| - c t)}{\| \mathbf {l} - \mathbf {s} \| ^ {2}} d \mathbf {l}, \tag {1}
83
+ $$
84
+
85
+ where $\mathbf{g}$ is the location of the laser, $c$ is the speed of light, and $\alpha$ is the intensity of the laser. We refer to $i_0$ as the empty transient. For simplicity, we ignore $\alpha$ for the rest of the derivations in this section and assume that the length between the virtual detector and camera is calibrated or known. The numerator of the integrand
86
+
87
+ $$
88
+ i (t; \mathbf {s}, \mathbf {l}) = \delta (\| \mathbf {l} - \mathbf {g} \| + \| \mathbf {l} - \mathbf {s} \| - c t) \tag {2}
89
+ $$
90
+
91
+ describes the equation of an ellipsoid, with foci at $\mathbf{g}$ and s and major axis length $ct$ . When an object is present in the
92
+
93
+ ![](images/dc5705168e811c254e22b1d68ae3ba330bd0192dfe1a920cfa8f9180350463ca.jpg)
94
+ Figure 4. Light and Shadow Transients. Three spacetime curves are depicted. (left) Empty transient $i_0$ with no hidden object, (middle) light transient $i_l$ with an object, and (right) shadow transient $i_s = i_0 - i_l$ . Switching from light transients $i_l$ to shadow transients $i_s$ enables derivation of a linear forward model.
95
+
96
+ occluded scene, the measured light transient is given by
97
+
98
+ $$
99
+ i _ {l} (t; \mathbf {s}) = \iint_ {\mathcal {L}} V _ {\mathbf {s}, \mathbf {l}} \cdot \frac {i (t ; \mathbf {s} , \mathbf {l})}{\| \mathbf {l} - \mathbf {s} \| ^ {2}} d \mathbf {l}, \tag {3}
100
+ $$
101
+
102
+ where $V_{\mathbf{s},1}$ is the visibility between s and l.
103
+
104
+ $$
105
+ V _ {\mathbf {s}, \mathbf {l}} = \exp \left(\oint_ {\mathbf {l}} ^ {\mathbf {s}} \ln \left(f _ {t} (\mathbf {x})\right) d \mathbf {x}\right) \tag {4}
106
+ $$
107
+
108
+ where Eq. (4) computes the product integral along the line connecting $\mathbf{l}$ and $\mathbf{s}$ and $f_{t}(\mathbf{x})\in \{0,1\}$ is the transparency function of voxel $\mathbf{x}$ in the occluded region.
109
+
110
+ Shadow Transients The non-linearity of Eq. (4) makes model inversion challenging with light transients. Casting the formulation from light space-time to shadow space-time enables a linear formulation of the problem. A shadow transient at $s$ is given by
111
+
112
+ $$
113
+ i _ {s} (t; \mathbf {s}) = i _ {0} (t; \mathbf {s}) - i _ {l} (t; \mathbf {s}). \tag {5}
114
+ $$
115
+
116
+ Fig. 4 illustrates the difference between light transients and shadow transients. Shadow transients directly probe voxel occupancy, whereas light transients probe for voxel transparency. For example, if the occluded scene is empty except for voxel $\mathbf{x}$ , the shadow transient is given by
117
+
118
+ $$
119
+ i _ {s} (t; \mathbf {s}, \mathbf {x}) = \frac {\delta (\| \mathbf {l} (\mathbf {x} , \mathbf {s}) - \mathbf {g} \| + \| \mathbf {l} (\mathbf {x} , \mathbf {s}) - \mathbf {s} \| - c t)}{\| \mathbf {l} (\mathbf {x} , \mathbf {s}) - \mathbf {s} \| ^ {2}}, \tag {6}
120
+ $$
121
+
122
+ where $\mathbf{l}(\mathbf{x},\mathbf{s})$ is a point on the illumination wall that lies along the ray connecting $\mathbf{x}$ and $\mathbf{s}$ . If the wall doesn't intersect this ray, $i_s(t;\mathbf{x},\mathbf{s}) = 0$ .
123
+
124
+ Linear Relaxation Now, consider the general case where a single opaque object lies in the hidden scene. We can describe the shadow transient as
125
+
126
+ $$
127
+ i _ {s} (t; \mathbf {s}) = \iint_ {\mathcal {L}} O _ {\mathbf {s}, \mathbf {l}} \cdot \frac {i (t ; \mathbf {s} , \mathbf {l})}{\| \mathbf {l} - \mathbf {s} \| ^ {2}} d \mathbf {l}, \tag {7}
128
+ $$
129
+
130
+ where $O_{\mathbf{s},1} \in \{0,1\}$ is the occlusion function (denotes if the ray between 1 and s is occluded). It can be expressed as the indicator function
131
+
132
+ $$
133
+ \begin{array}{l} O _ {\mathbf {s}, \mathrm {l}} = \left\{ \begin{array}{l l} 1, & \text {i f} \oint_ {\mathrm {l}} ^ {\mathbf {s}} f _ {o} (\mathbf {x}) d \mathbf {x} > 0 \\ 0, & \text {o t h e r w i s e} \end{array} \right. \tag {8} \\ \approx \oint_ {1} ^ {\mathbf {s}} f _ {o} (\mathbf {x}) d \mathbf {x}, \\ \end{array}
134
+ $$
135
+
136
+ where $f_{o}(\mathbf{x})$ is the voxel occupancy function of the hidden scene. The linear relaxation in Eq. (8) is possible because NLOS reconstructions typically only consider ToF information, but not radiometric information. The above formulation will preserve the structure of the streak images in spacetime, even if the intensities are overestimated. Combining Eq. (5), (7), and (8), we obtain
137
+
138
+ $$
139
+ i _ {0} (t; \mathbf {s}) - i _ {l} (t; \mathbf {s}) = \iint_ {\mathcal {L}} \frac {i (t ; \mathbf {s} , \mathbf {l})}{\| \mathbf {l} - \mathbf {s} \| ^ {2}} \cdot \left(\oint_ {\mathbf {l}} ^ {\mathbf {s}} f _ {o} (\mathbf {x}) d \mathbf {x}\right) d \mathbf {l}, \tag {9}
140
+ $$
141
+
142
+ where we solve for $f_{o}(\mathbf{x})$ . We can estimate $i_0$ from the scene geometry and $i_l$ is the measured transient.
143
+
144
+ Discrete Beam Illumination Discretizing $\mathcal{L}$ as $K$ discrete laser spots and the hidden space into a voxel grid, we can write Eq. (9) as
145
+
146
+ $$
147
+ i _ {s} [ t; \mathbf {s} ] = \sum_ {k = 1} ^ {K} \frac {\delta [ \| \mathbf {l} _ {k} - \mathbf {g} \| + \| \mathbf {l} _ {k} - \mathbf {s} \| - c t ]}{\| \mathbf {l} _ {k} - \mathbf {s} \| ^ {2}} \cdot \left(\sum_ {\mathbf {x} \in \mathcal {P} _ {\mathbf {l} _ {k}, \mathbf {s}}} f _ {o} [ \mathbf {x} ]\right), \tag {10}
148
+ $$
149
+
150
+ where $\mathcal{P}_{1,\mathrm{s}}$ denotes the set of voxels lying along the path between 1 and s. Eq. (10) can be expressed in matrix form
151
+
152
+ $$
153
+ \mathbf {I} _ {s} = \mathbf {P B f} = \mathbf {A f}, \tag {11}
154
+ $$
155
+
156
+ where $\mathbf{I}_s\in \mathbb{R}^{n_u n_v n_t}$ is the vectorized space-time measurements, $\mathbf{f}\in \mathbb{R}^{n_x n_y n_z}$ is the voxel occupancies, and $\mathbf{A}\in \mathbb{R}^{n_u n_v n_t\times n_x n_y n_z}$ is the measurement operator mapping voxel occupancies to space-time measurements. $n_u$ and $n_v$ are the number of pixels along the $u$ and $v$ direction, and $n_t$ is the number of timing bins. $n_x,n_y,$ and $n_z$ are the number of voxels in the $x,y,$ and $z$ directions.
157
+
158
+ $\mathbf{P}$ and $\mathbf{B}$ model the temporal and spatial relationships, respectively, between virtual sources, virtual detectors, and voxels. $\mathbf{P} \in \mathbb{R}^{n_u n_v n_t \times n_u n_v K}$ is a block diagonal matrix, with the column space of the $i$ th block $\mathbf{P}^i \in \mathbb{R}^{n_t \times K}$ containing the corresponding $\delta(\cdot)$ functions from Eq. (2) for virtual detector $\mathbf{s}_i$ and all $K$ virtual sources. $\mathbf{B} \in \mathbb{R}^{n_u n_v K \times n_x n_y n_z}$ is composed of $n_u n_v$ vertical sub-blocks. Each entry $\mathbf{B}_{kj}^i \in \{0,1\}$ of the $i$ th sub-block $\mathbf{B}^i \in \mathbb{R}^{K \times n_x n_y n_z}$ of sub-block $i$ indicates whether voxel $j$ lies along the line connecting virtual detector $\mathbf{s}_i$ and virtual source $\mathbf{l}_k$ . The multiplication $\mathbf{Bf}$ performs the line integral in Eq. (8). The $k$ th column vector of $\mathbf{P}$ is the space-time impulse response (STIR) to the $k$ th illumination source when the hidden scene is completely occupied (note that we are dealing with shadow transients, not light transients). The $i$ th column of $\mathbf{A}$ forms the expected space-time measurements
159
+
160
+ if only the $i$ th voxel is occupied (i.e. the STIR to a single voxel). We neglect the $r^2$ falloff term in our reconstructions. Additional explanation of the A matrix is provided in the supplementary. The linear model in Eq. (11) resembles that of 3B-bounce NLOS, where $\mathbf{A}$ is a measurement operator that maps voxels $\mathbf{f}$ to measurements $\mathbf{I}$ . However, in 3B-NLOS, the information is encoded into the reflectance of the object, whereas in 2B-NLOS, the information is encoded into the shadows. Therefore, the former is linear with respects to light transients, and the latter is linear with respect to shadow transients.
161
+
162
+ # 3.3. Reconstruction Algorithm
163
+
164
+ Reconstruction of $\mathbf{f}$ can be performed by backprojection
165
+
166
+ $$
167
+ \mathbf {f} _ {b p} = \mathbf {A} ^ {\top} \mathbf {I} _ {s}. \tag {12}
168
+ $$
169
+
170
+ Backprojection produces a low-frequency reconstruction of $\mathbf{f}$ . In practice, filtering $\mathbf{f}_{bp}$ with a Laplacian filter
171
+
172
+ $$
173
+ \mathbf {f} _ {f b p} = \left(\frac {\partial^ {2}}{\partial x ^ {2}} + \frac {\partial^ {2}}{\partial y ^ {2}} + \frac {\partial^ {2}}{\partial z ^ {2}}\right) \mathbf {f} _ {b p} \tag {13}
174
+ $$
175
+
176
+ has been shown to yield reasonably good results in ellipsoidal tomography in 3-bounce NLOS and line-based tomography in medical imaging. Our linear formulation also enables reconstruction with priors. We recognize that these optimization-based methods would improve reconstruction over unfiltered backprojection, which is equivalent to a low-frequency version of voxel carving in our setup. However, we restrict our focus to unfiltered backprojection because of its convenient relationship with the Gram matrix, which enables further analysis of our method in Sec. 4.2.
177
+
178
+ # 4. Design and Analysis
179
+
180
+ In this section, we analyze how the performance of the proposed two-bounce transient system varies with system parameters and the geometry of LOS relay surfaces. We also analyze the impact of temporal resolution, spatial resolution, and number of measurements on recoverability. The amount of multiplexing in the scene is inversely related to the number of measurements. For example, if 60 laser spots are used and 15 measurements are captured, then each image contains 4 multiplexed shadows. We use SNR, coherence, and FWHM of the PSF to quantify performance.
181
+
182
+ # 4.1. Voxel SNR Analysis
183
+
184
+ We quantify the noise characteristics of the measured shadow two-bounce transients using per-voxel SNR. Consider the case where only voxel $\mathbf{x}$ is occupied in the hidden region. We derive the SNR of the number of "shadow" photons received in the measured shadow transient $\mathbf{I}_s^{\mathbf{x}}$ . In vector
185
+
186
+ ![](images/4f3db07091b483a62612cdaf18a59324896d0484893c3c68aaf647b579b2ea8a.jpg)
187
+ Figure 5. SNR Analysis. The goal of this analysis is to determine how scene geometry affects the SNR of the two-bounce measurement at each voxel in the hidden scene. The first two cases consider the orientations of the relay surfaces, and the third considers the field of view of the camera. SNR is highest in the middle of the hidden scene for parallel surfaces, orienting the walls closer together increases SNR due to reduced $r^2$ falloff, and SNR is highest closer to the virtual detectors for small FOV cameras.
188
+
189
+ form, the occupancy function can be described as $\delta^{\mathbf{x}}$ . From Eq. (11), the shadow transient is given by
190
+
191
+ $$
192
+ \mathbf {I} _ {s} ^ {\mathbf {x}} = \mathbf {A} \delta^ {\mathbf {x}} \tag {14}
193
+ $$
194
+
195
+ and the light transients measured by the SPAD sensor are given by $\mathbf{I}_l^{\mathbf{x}}$ . For this analysis, we account for $r^2$ falloff because it is a physically accurate phenomenon that impacts SNR. From Eq. (5),
196
+
197
+ $$
198
+ \mathbf {I} _ {l} ^ {\mathbf {x}} = \mathbf {I} _ {0} - \mathbf {A} \delta^ {\mathbf {x}} \tag {15}
199
+ $$
200
+
201
+ where $\mathbf{I}_0$ is the empty transient and assumed to be calibrated. Assuming no SPAD pile-up distortion, the number of photons arriving at the sensor for each time bin of $\mathbf{I}_l^{\mathbf{x}}$ follows an independent Poisson distribution [29]. Thus, the total number of photons $P^{\mathbf{x}}$ measured follows a Poisson distribution with the rate parameter being the sum of all the photon counts for all time bins for all pixels.
202
+
203
+ $$
204
+ P ^ {\mathbf {x}} \sim \operatorname {P o i s s o n} \left(\mathbf {1} ^ {T} \left(\mathbf {I} _ {0} - \mathbf {A} \delta^ {\mathbf {x}}\right)\right) \tag {16}
205
+ $$
206
+
207
+ where $\mathbf{1} \in \mathbb{R}^{n_u n_v n_t}$ is the one vector. Using the fact that SNR of a Poisson distribution is $\mu / \sigma = \sqrt{\lambda}$ with rate parameter $\lambda$ , the per-voxel SNR $S^{\mathbf{x}}$ is given by
208
+
209
+ $$
210
+ S ^ {\mathbf {x}} = \sqrt {\mathbf {1} ^ {T} \left(\mathbf {I} _ {0} - \mathbf {A} \boldsymbol {\delta} ^ {\mathbf {x}}\right)} \tag {17}
211
+ $$
212
+
213
+ The per-voxel SNR is plotted in Fig. 5 for three different scenarios. The first two scenarios deal with the orientation of the walls, and the third considers the field of view of the camera. In Fig. 5a, we can see that the SNR is higher for center voxels, as more rays connecting the illumination wall and observation wall pass through the center voxels. When the relay walls are oriented towards each other (Fig. 5b), the SNR increases due to reduced $r^2$ falloff. When the camera's field of view is small compared to the laser (Fig. 5c), regions closer to the observation wall have the highest SNR.
214
+
215
+ ![](images/5e8c362db0744c8c9755964b4fda3be0b9a90703ab8cbc9ccf84fdd8847089e7.jpg)
216
+ Figure 6. Spatially Varying PSF. In two-bounce lidar, the PSF has a double cone shape. The orientation of the double cone varies strongly as a function of the orientation and length of the relay surfaces (i.e. the virtual apertures) and the voxel position.
217
+
218
+ # 4.2. Recoverability Analysis
219
+
220
+ Resolution We now study the resolution limits of our method. In particular, we consider two key metrics in our measurement operator: mutual coherence and full width at half maximum (FWHM) of the PSF. We are interested in how (1) temporal resolution, (2) spatial resolution $N$ , and (3) number of measurements $M$ impact each of our metrics. From Eq. (12), we know that the backprojected occupancy can be obtained from the occupancy function by
221
+
222
+ $$
223
+ \mathbf {f} _ {b p} = \left(\mathbf {A} ^ {\top} \mathbf {A}\right) \mathbf {f}, \tag {18}
224
+ $$
225
+
226
+ where $\mathbf{G} = \mathbf{A}^{\top}\mathbf{A}$ is the Gram matrix. We can interpret the columns (or rows) of the Gram matrix to be the spatially varying point spread function (PSF) of the imaging system, as shown in Fig. 6. Each column $\mathbf{G}_i$ describes how a voxel $\mathbf{x}_i$ is blurred during backprojection reconstruction.
227
+
228
+ We can see the effect of different imaging parameters on the PSF in Fig. 7, where we sweep different values of (1) $M$ and (2) $N$ . In particular, we note that the lobes of the PSF are suppressed when using temporal information, even with low spatial resolution and few number of image measurements. When using no temporal information, the PSF is still roughly centered at the correct location, but has large FWHM. The sharpness of the PSF when using temporal information indicates that it is better suited to applications that require finer voxel resolution with few-shot capture. In case (1) without transients, we see that the FWHM of the PSF is large along both the $x$ and $y$ direction for small values of $M$ . As $M$ approaches the number of illumination spots, the FWHM approaches the theoretical lower bound. In case (2), increased spatial resolution helps better localize the center of the PSF, but the FWHM does not improve. The transient cases are quite robust to both low number of measurements and low spatial resolution. Intuitively, higher values of $N$ in the transient case slightly help accentuate the center voxel
229
+
230
+ by projecting more diverse rays through the center voxel.
231
+
232
+ Mutual Coherence We wish to further quantify the impact of spatio-temporal resolution and number of measurements on the recoverability of the hidden scene. We know from compressive sensing theory that sparse recovery of the hidden scene is possible when the mutual coherence of $\mathbf{A}$ is sufficiently small. The mutual coherence is defined as
233
+
234
+ $$
235
+ \mu = \max _ {1 \leq i \neq j \leq n} \frac {\left| \mathbf {A} _ {i} ^ {\top} \mathbf {A} _ {j} \right|}{\left| \mathbf {A} _ {i} \right| \left| \mathbf {A} _ {j} \right|}. \tag {19}
236
+ $$
237
+
238
+ Intuitively, the coherence measures the largest "similarity" or correlation between any two column vectors in measurement matrix $\mathbf{A}$ [6]. Physically, this quantifies how similar the STIRs of two different voxels are. If $\mu = 1$ , then two voxels have the exact same STIR, which means that there exists an ambiguity as to which voxel resulted in a certain spatio-temporal response. Alternatively, $\mu$ can be defined as the maximum off-diagonal entry of the Gram matrix. Fig. 8 illustrates how the Gram matrix (and subsequently the mutual coherence) is affected by coarser timing resolutions. From compressive sensing theory, we know that a $K$ -sparse scene can be recovered if $K \leq 0.5(1 + 1 / \mu)$ , which motivates our empirical analysis of the effect of spatial resolution and temporal resolution on $\mu$ .
239
+
240
+ From Fig. 9, we can see the impact that spatial and temporal resolution have on mutual coherence in a single-shot setup with multiplexed illumination. Consider a scene similar to that of Fig. 6, where the walls are $2\mathrm{m}$ apart, the walls are $3\mathrm{m}$ long, the axial distance to the walls is $2\mathrm{m}$ , and the voxel resolution is discretized to $2\mathrm{cm}$ . Increasing the spatial resolution at a fixed temporal resolution appears to have little to no effect on the mutual coherence. As Kadambi et al. also point out, improvements of coherence through higher spatial resolution are limited by the size of the virtual aperture (i.e. the wall) [16]. On the other hand, the difference between $100~\mathrm{ps}$ resolution and $10~\mathrm{ps}$ resolution reduces the coherence from 0.9 to 0.3 for a fixed spatial resolution. Any timing resolution coarser than $\sim 120$ ps results in a coherence of 1 for multiplexed measurements because pathlength differences for different shadows are no longer resolvable. This analysis motivates the importance of temporal resolution when using multiplexed illumination.
241
+
242
+ # 5. Experiments and Results
243
+
244
+ We scan a single-pixel SPAD across the field of view on the left wall to mimic the output of a SPAD array. We scan a single laser spot across the right wall, and sum up the corresponding transient images in post-processing to mimic the effect of multiplexing. Further details are provided in the supplementary. We first validate our model using simulated data in Fig. 10, where we demonstrate that we can localize four letters and reconstruct their 2D shape with just a single
245
+
246
+ ![](images/8917c0c3376b274e4c1aec23ce917105bd1496bf57c63b2ca47e202dafefab2f.jpg)
247
+ Figure 7. PSF as a Function of Number of Measurements and Spatial Resolution. The PSF of the center voxel is shown in the transient and intensity-only case when multiplexed with 60 laser spots. (1) the number of pixels $N = 40$ are kept fixed with the number of measurements $M$ being swept (left) and (2) $M = 10$ is kept fixed and $N$ is swept (right). Transient measurements result in PSFs with narrower FWHMs, even with low spatial resolution and low number of measurements.
248
+
249
+ ![](images/1edc73563e372c28f865e5f33e9f5ed0763c3157ca12399ed9b489b4d511ba4e.jpg)
250
+
251
+ ![](images/16f0d107255a743cc7c7b3edbfebd1230b98d899f7b505bddff4798abb3c252b.jpg)
252
+ Figure 8. Gram Matrix as a Function of Temporal Resolution. For a defined voxel resolution and fixed spatial resolution. The columns (or rows) of the Gram matrix form the spatially varying PSFs shown in Fig. 6. The off-diagonal streaks describe the coherence between different voxels. As expected, higher temporal resolution decreases the coherence of the Gram matrix, which increases the likelihood of guaranteed sparse recovery.
253
+
254
+ capture. We render $i_{o}$ and $i_{l}$ to obtain a noiseless version of the shadow transient for this experiment. Then, we reconstruct the shapes of real mannequins and must estimate $i_{0}$ . Fig. 11 shows that ToF measurements provide a noticeable improvement in image reconstruction. When there's no multiplexing (i.e., the number of laser spots is the same as number of measurements), the intensity-based method is similar to the ToF-based method. However, in the intensity-only case, the object becomes increasingly blurry until it is no longer visible as the amount of multiplexing increases. In the ToF case for single-shot, we can see a blurry image roughly localized where the center of the mannequin is as well as some high-frequency detail. Furthermore, the top view of the ToF reconstructions show that the high-frequency shape of the arms can be reconstructed even with multiplexed illumination, indicating the promise of single
255
+
256
+ ![](images/a939948ccd7e2441a4107bbde6b02c4135b2b8f32bbefa8083f6280b2129e310.jpg)
257
+ Figure 9. Tradeoffs Between Spatial Resolution and Temporal Resolution. We quantify the tradeoff using mutual coherence. We use the same setup as Fig. 6, but the walls have length $3\mathrm{m}$ , the walls are space $2\mathrm{m}$ apart, and the distance to the walls is $2\mathrm{m}$ . We discretize the voxel resolution to $2\mathrm{cm}(\sim 67\mathrm{ps})$ . The bottom right of the figure, where spatial and temporal resolution are highest is where mutual coherence is the lowest. For a fixed temporal resolution, improvements in coherence quickly plateau. Improvements in temporal resolution dominate improvements in spatial resolution when reducing coherence, suggesting the importance of ToF for demultiplexing illumination in two-bounce NLOS imaging.
258
+
259
+ shot capture with two-bounce ToF information. Fig. 12 shows that poorer temporal resolution results in aliasing artifacts (repeated arms and legs), whereas the rough shape of the object can still be seen with lower spatial resolution. Finally, we compare our method to the probabilistic voxel carving method used in [15] in Fig. 13. Voxel carving results in sharper reconstructions than backprojection.
260
+
261
+ ![](images/73d56e3e1d2c94eac3c110d2cccf8f917bb7ef872fe2dbc3535b42a6c8e81187.jpg)
262
+ Figure 10. Single-Shot Localization. We obtain single-shot reconstructions of the "N", "L", "O", and "S" letters in the hidden scene with and without transient information. We see that using transient information, we are able to effectively localize the hidden object and even get a sharp outline of the letters, demonstrating the promise of single-shot capture with two-bounce transients.
263
+
264
+ ![](images/052e98ac960dfbd0dda5d08d4c729a58a63172b1586062dfa28ba96a9705a46b.jpg)
265
+ Figure 11. Transients Enable Scene Recovery from Multiplexed Illumination. We perform backprojection of measurements with varying levels of multiplexing. Plotted in the top two rows are $x$ slices of the reconstructed volume. Under no multiplexing (right), intensity is sufficient to carve the shadow along detected pixel. But reconstruction quality sharply deteriorates as we move towards the multiplexing regime (left). When measuring with transients, time-of-flight information allows us to disambiguate between different shadows and recover a rough silhouette of the object even with just a single measurement.
266
+
267
+ However, our method can be improved by filtering or incorporating priors, and is also generalizable to multiplexed illumination by utilizing ToF information.
268
+
269
+ ![](images/ad76127c3b9f4afd934c1107a982ed0d2ca3477e6dc88bdfb5454de7fd317d7e.jpg)
270
+ Figure 12. Spatial Resolution vs. Timing Resolution. We plot reconstructions with different spatio-temporal resolutions. Lower timing resolution results in aliasing artifacts (i.e. repeated arms and legs). Lower spatial resolution introduces hole artifacts.
271
+
272
+ ![](images/fd216f0dc7a5bd6663190c42867f44743f539f7e6a813ffe1cf67a6a302b7c89.jpg)
273
+ Figure 13. Scanning Baseline Comparison. We compare to the voxel carving method used in [15] (a) before and (b) after thresholding when scanning 1 laser spot. Our method is more general, but voxel carving yields sharper results with laser scanning.
274
+
275
+ # 6. Discussion
276
+
277
+ In this paper, we introduced the use of two-bounce transients for NLOS imaging. Compared to 3B-NLOS, 2B-NLOS enables more robust imaging of occluded objects in the presence of scenes with arbitrary lighting conditions and low albedo objects. Transient 2B-NLOS is promising for few-shot capture of occluded objects by using multiplexed illumination, especially as SPAD array technology matures. When comparing tradeoffs in temporal resolution, spatial resolution, and number of measurements, we qualitatively and quantitatively verify with real and simulated results that temporal resolution is most important for demultiplexing shadows. We anticipate future work to extend the analysis in this paper to SPAD arrays for real-time NLOS imaging.
278
+
279
+ # Acknowledgements
280
+
281
+ AD and AV are supported by NSF Expeditions Award #IIS-1730574 and NSF CAREER Award #IIS-1652633. CH was supported by a Draper Scholarship.
282
+
283
+ # References
284
+
285
+ [1] Fadel Adib and Dina Katabi. See through walls with wifi! In Proceedings of the ACM SIGCOMM 2013 conference on SIGCOMM, pages 75-86, 2013. 2
286
+ [2] Byeongjoo Ahn, Akshat Dave, Ashok Veeraraghavan, Ioannis Gkioulekas, and Aswin C Sankaranarayanan. Convolutional approximations to the general non-line-of-sight imaging operator. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7889-7899, 2019. 2
287
+ [3] Victor Arellano, Diego Gutierrez, and Adrian Jarabo. Fast back-projection for non-line of sight reconstruction. In ACM SIGGRAPH 2017 Posters, pages 1-2. 2017. 2
288
+ [4] Katherine L Bouman, Vickie Ye, Adam B Yedidia, Frédo Durand, Gregory W Wornell, Antonio Torralba, and William T Freeman. Turning corners into cameras: Principles and methods. In Proceedings of the IEEE International Conference on Computer Vision, pages 2270-2278, 2017. 2, 3
289
+ [5] Mauro Buttafava, Jessica Zeman, Alberto Tosi, Kevin Eliceiri, and Andreas Velten. Non-line-of-sight imaging using a time-gated single photon avalanche diode. Optics express, 23(16):20997-21011, 2015. 2
290
+ [6] Emmanuel J Candès and Michael B Wakin. An introduction to compressive sampling. IEEE signal processing magazine, 25(2):21-30, 2008. 6
291
+ [7] Wenzheng Chen, Simon Daneau, Fahim Mannan, and Felix Heide. Steady-state non-line-of-sight imaging. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6790-6799, 2019. 2
292
+ [8] Wenzheng Chen, Fangyin Wei, Kiriakos N Kutulakos, Szymon Rusinkiewicz, and Felix Heide. Learned feature embeddings for non-line-of-sight imaging and recognition. ACM Transactions on Graphics (ToG), 39(6):1-18, 2020. 2
293
+ [9] Akshit Dave, Muralidhar Madabhushi Balaji, Prasanna Rangarajan, Ashok Veeraraghavan, and Marc P Christensen. Foveated non-line-of-sight imaging. In Computational Optical Sensing and Imaging, pages CTh5C-6. Optica Publishing Group, 2020. 2
294
+ [10] Akshit Dave, Yongyi Zhao, and Ashok Veeraraghavan. Pandora: Polarization-aided neural decomposition of radiance. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23–27, 2022, Proceedings, Part VII, pages 538–556. Springer, 2022. 2
295
+ [11] Genevieve Gariepy, Francesco Tonolini, Robert Henderson, Jonathan Leach, and Daniele Faccio. Detection and tracking of moving objects hidden from view. Nature Photonics, 10(1):23-26, 2016. 2
296
+ [12] Felix Heide, Matthew O'Toole, Kai Zang, David B Lindell, Steven Diamond, and Gordon Wetzstein. Non-line-of-sight imaging with partial occluders and surface normals. ACM Transactions on Graphics (ToG), 38(3):1-10, 2019. 2
297
+ [13] Felix Heide, Lei Xiao, Wolfgang Heidrich, and Matthias B Hullin. Diffuse mirrors: 3d reconstruction from diffuse indirect illumination using inexpensive time-of-flight sensors. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3222-3229, 2014. 2
298
+
299
+ [14] Connor Henley, Joseph Hollmann, and Ramesh Raskar. Bounce-flash lidar. IEEE Transactions on Computational Imaging, 8:411-424, 2022. 2
300
+ [15] Connor Henley, Tomohiro Maeda, Tristan Swedish, and Ramesh Raskar. Imaging behind occluders using two-bounce light. In ECCV, pages 573-588. Springer, 2020. 1, 2, 7, 8
301
+ [16] Achuta Kadambi, Hang Zhao, Boxin Shi, and Ramesh Raskar. Occluded imaging with time-of-flight sensors. ACM Transactions on Graphics (ToG), 35(2):1-12, 2016. 2, 6
302
+ [17] Ori Katz, Pierre Heidmann, Mathias Fink, and Sylvain Gigan. Non-invasive single-shot imaging through scattering layers and around corners via speckle correlations. Nature photonics, 8(10):784-790, 2014. 2
303
+ [18] Ahmed Kirmani, Tyler Hutchison, James Davis, and Ramesh Raskar. Looking around the corner using transient imaging. In 2009 IEEE 12th International Conference on Computer Vision, pages 159-166. IEEE, 2009. 2, 3
304
+ [19] Oichi Kumagai, Junichi Ohmachi, Masao Matsumura, Shinichiro Yagi, Kenichi Tayu, Keitaro Amagawa, Tomohiro Matsukawa, Osamu Ozawa, Daisuke Hirono, Yasuhiro Shinozuka, et al. 7.3 a $189 \times 600$ back-illuminated stacked spad direct time-of-flight depth sensor for automotive lidar systems. In 2021 IEEE International Solid-State Circuits Conference (ISSCC), volume 64, pages 110-112. IEEE, 2021. 2
305
+ [20] Aldo Laurentini. The visual hull concept for silhouette-based image understanding. PAMI, 16(2):150-162, 1994. 1
306
+ [21] Di Lin, Connor Hashemi, and James R Leger. Passive non-line-of-sight imaging using plenoptic information. JOSA A, 37(4):540-551, 2020. 2
307
+ [22] David B Lindell, Gordon Wetzstein, and Matthew O'Toole. Wave-based non-line-of-sight imaging using fast fkr migration. ACM Transactions on Graphics (ToG), 38(4):1-13, 2019. 2
308
+ [23] Xiaochun Liu, Ibón Guillén, Marco La Manna, Ji Hyun Nam, Syed Azer Reza, Toan Huu Le, Adrian Jarabo, Diego Gutierrez, and Andreas Velten. Non-line-of-sight imaging using phasor-field virtual wave optics. Nature, 572(7771):620-623, 2019. 2
309
+ [24] Tomohiro Maeda, Guy Satat, Tristan Swedish, Lagnojita Sinha, and Ramesh Raskar. Recent advances in imaging around corners. arXiv preprint arXiv:1910.05613, 2019. 1
310
+ [25] Christopher A Metzler, Felix Heide, Prasana Rangarajan, Muralidhar Madabhushi Balaji, Aparna Viswanath, Ashok Veeraraghavan, and Richard G Baraniuk. Deep-inverse correlography: towards real-time high-resolution non-line-of-sight imaging. Optica, 7(1):63-71, 2020. 2
311
+ [26] Kazuhiro Morimoto, Andrei Ardelean, Ming-Lo Wu, Arin Can Ulku, Ivan Michel Antolovic, Claudio Bruschini, and Edoardo Charbon. Megapixel time-gated spad image sensor for 2d and 3d imaging applications. Optica, 7(4):346–354, 2020. 2
312
+ [27] Matthew O'Toole, David B Lindell, and Gordon Wetzstein. Confocal non-line-of-sight imaging based on the light-cone transform. Nature, 555(7696):338-341, 2018. 2, 3
313
+ [28] Adithya Pediredla, Akshit Dave, and Ashok Veeraraghavan. Snlos: Non-line-of-sight scanning through temporal focus
314
+
315
+ ing. In 2019 IEEE International Conference on Computational Photography (ICCP), pages 1-13. IEEE, 2019. 2
316
+ [29] Adithya K Pediredla, Aswin C Sankaranarayanan, Mauro Buttafava, Alberto Tosi, and Ashok Veeraraghavan. Signal processing based pile-up compensation for gated single-photon avalanche diodes. arXiv preprint arXiv:1806.07437, 2018.5
317
+ [30] Joshua Rapp, Charles Saunders, Julián Tachella, John Murray-Bruce, Yoann Altmann, Jean-Yves Tourneret, Stephen McLaughlin, Robin Dawson, Franco NC Wong, and Vivek K Goyal. Seeing around corners with edge-resolved transient imaging. Nature communications, 11(1):1-10, 2020. 2, 3
318
+ [31] Charles Saunders, John Murray-Bruce, and Vivek K Goyal. Computational periscopy with an ordinary digital camera. Nature, 565(7740):472-475, 2019. 3
319
+ [32] Nicolas Scheiner, Florian Kraus, Fangyin Wei, Buu Phan, Fahim Mannan, Nils Appenrodt, Werner Ritter, Jurgen Dickmann, Klaus Dietmayer, Bernhard Sick, et al. Seeing around street corners: Non-line-of-sight detection and tracking in-the-wild using doppler radar. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2068-2077, 2020. 2
320
+ [33] Prafull Sharma, Miika Aittala, Yoav Y Schechner, Antonio Torralba, Gregory W Wornell, William T Freeman, and Frédo Durand. What you can learn by staring at a blank wall. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2330-2339, 2021. 2, 3
321
+ [34] Brandon M Smith, Matthew O'Toole, and Mohit Gupta. Tracking multiple objects outside the line of sight using speckle imaging. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6258-6266, 2018. 2
322
+ [35] Tristan Swedish, Connor Henley, and Ramesh Raskar. Objects as cameras: Estimating high-frequency illumination from shadows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2593-2602, 2021. 2
323
+ [36] Kenichiro Tanaka, Yasuhiro Mukaigawa, and Achuta Kadambi. Polarized non-line-of-sight imaging. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2136-2145, 2020. 2
324
+ [37] Matthew Tancik, Guy Satat, and Ramesh Raskar. Flash photography for data-driven hidden scene recovery. arXiv preprint arXiv:1810.11710, 2018. 2
325
+ [38] Kushagra Tiwary, Askhat Dave, Nikhil Behari, Tzofi Klinghoffer, Ashok Veeraraghavan, and Ramesh Raskar. Orca: Glossy objects as radiance field cameras. arXiv preprint arXiv:2212.04531, 2022. 2
326
+ [39] Andreas Velten, Thomas Willwacher, Otkrist Gupta, Ashok Veeraraghavan, Moungi G Bawendi, and Ramesh Raskar. Recovering three-dimensional shape around a corner using ultrafast time-of-flight imaging. Nature communications, 3(1):1-8, 2012. 2
327
+ [40] Florian Willomitzer, Fengqiang Li, Muralidhar Madabhushi Balaji, Prasanna Ranganajan, and Oliver Cossairt. High resolution non-line-of-sight imaging with superheterodyne re
328
+
329
+ mote digital holography. In Computational Optical Sensing and Imaging, pages CM2A-2. Optica Publishing Group, 2019. 2
330
+ [41] Florian Willomitzer, Prasanna V Rangarajan, Fengqiang Li, Muralidhar M Balaji, Marc P Christensen, and Oliver Cossaint. Fast non-line-of-sight imaging with high-resolution and wide field of view using synthetic wavelength holography. Nature communications, 12(1):6647, 2021. 2
331
+ [42] Shumian Xin, Sotiris Nousias, Kiriakos N Kutulakos, Aswin C Sankaranarayanan, Srinivasa G Narasimhan, and Ioannis Gkioulekas. A theory of Fermat paths for nonline-of-sight shape reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6800-6809, 2019. 2, 3
332
+ [43] Shichao Yue, Hao He, Peng Cao, Kaiwen Zha, Masayuki Koizumi, and Dina Katabi. Cornerradar: Rf-based indoor localization around corners. Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies, 6(1):1-24, 2022. 2
333
+ [44] Chao Zhang, Ning Zhang, Zhijie Ma, Letian Wang, Yu Qin, Jieyang Jia, and Kai Zang. A $240 \times 160$ 3d-stacked spad dtof image sensor with rolling shutter and in-pixel histogram for mobile devices. IEEE Open Journal of the Solid-State Circuits Society, 2:3-11, 2021. 2
2023/Role of Transients in Two-Bounce Non-Line-of-Sight Imaging/images.zip ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:575dfc5ec860fea9e31e97dcce08e3593a0c5b9b594a3c3b94ea9d8691610702
3
+ size 610077
2023/Role of Transients in Two-Bounce Non-Line-of-Sight Imaging/layout.json ADDED
The diff for this file is too large to render. See raw diff
 
2023/Rotation-Invariant Transformer for Point Cloud Matching/424bbd8d-fe04-4142-91bc-ea8eb4c92ccb_content_list.json ADDED
@@ -0,0 +1,1538 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ [
2
+ {
3
+ "type": "text",
4
+ "text": "Rotation-Invariant Transformer for Point Cloud Matching",
5
+ "text_level": 1,
6
+ "bbox": [
7
+ 189,
8
+ 130,
9
+ 782,
10
+ 152
11
+ ],
12
+ "page_idx": 0
13
+ },
14
+ {
15
+ "type": "text",
16
+ "text": "Hao Yu $^{1}$ Zheng Qin $^{2}$ Ji Hou $^{3}$ Mahdi Saleh $^{1,4}$ Dongsheng Li $^{2}$ Benjamin Busam $^{1,4}$ Slobodan Ilic $^{1,5}$ TU Munich $^{2}$ NUDT $^{3}$ Meta Reality Labs $^{4}$ 3Dwe.ai $^{5}$ Siemens AG, Munich",
17
+ "bbox": [
18
+ 86,
19
+ 179,
20
+ 915,
21
+ 217
22
+ ],
23
+ "page_idx": 0
24
+ },
25
+ {
26
+ "type": "text",
27
+ "text": "Abstract",
28
+ "text_level": 1,
29
+ "bbox": [
30
+ 233,
31
+ 251,
32
+ 313,
33
+ 267
34
+ ],
35
+ "page_idx": 0
36
+ },
37
+ {
38
+ "type": "text",
39
+ "text": "The intrinsic rotation invariance lies at the core of matching point clouds with handcrafted descriptors. However, it is widely despised by recent deep matchers that obtain the rotation invariance extrinsically via data augmentation. As the finite number of augmented rotations can never span the continuous $SO(3)$ space, these methods usually show instability when facing rotations that are rarely seen. To this end, we introduce RoITr, a Rotation-Invariant Transformer to cope with the pose variations in the point cloud matching task. We contribute both on the local and global levels. Starting from the local level, we introduce an attention mechanism embedded with Point Pair Feature (PPF)-based coordinates to describe the pose-invariant geometry, upon which a novel attention-based encoder-decoder architecture is constructed. We further propose a global transformer with rotation-invariant cross-frame spatial awareness learned by the self-attention mechanism, which significantly improves the feature distinctiveness and makes the model robust with respect to the low overlap. Experiments are conducted on both the rigid and non-rigid public benchmarks, where RoITr outperforms all the state-of-the-art models by a considerable margin in the low-overlapping scenarios. Especially when the rotations are enlarged on the challenging 3DLoMatch benchmark, RoITr surpasses the existing methods by at least 13 and 5 percentage points in terms of Inlier Ratio and Registration Recall, respectively. Code is publicly available<sup>1</sup>.",
40
+ "bbox": [
41
+ 76,
42
+ 282,
43
+ 473,
44
+ 691
45
+ ],
46
+ "page_idx": 0
47
+ },
48
+ {
49
+ "type": "text",
50
+ "text": "1. Introduction",
51
+ "text_level": 1,
52
+ "bbox": [
53
+ 76,
54
+ 700,
55
+ 209,
56
+ 715
57
+ ],
58
+ "page_idx": 0
59
+ },
60
+ {
61
+ "type": "text",
62
+ "text": "The correspondence estimation between a pair of partially-overlapping point clouds is a long-standing task that lies at the core of many computer vision applications, such as tracking [17, 18], reconstruction [22, 23, 39], pose estimation [19, 27, 50] and 3D representation learning [15, 16, 46], etc. In a typical solution, geometry is first encoded into descriptors, and correspondences are then established between two frames by matching the most similar descriptors. As the two frames are observed from different views, depicting the same geometry under different trans",
63
+ "bbox": [
64
+ 76,
65
+ 726,
66
+ 468,
67
+ 878
68
+ ],
69
+ "page_idx": 0
70
+ },
71
+ {
72
+ "type": "image",
73
+ "img_path": "images/81825f8e94531760f124a36a4048f3485bea7faa274c6d8e55532d5c6f8d23a7.jpg",
74
+ "image_caption": [
75
+ "Figure 1. Feature Matching Recall (FMR) on 3DLoMatch [19] and Rotated 3DLoMatch. Distance to the diagonal represents the robustness against rotations. Among all the state-of-the-art approaches, RoITr not only ranks first on both benchmarks but also shows the best robustness against the enlarged rotations."
76
+ ],
77
+ "image_footnote": [],
78
+ "bbox": [
79
+ 501,
80
+ 250,
81
+ 893,
82
+ 422
83
+ ],
84
+ "page_idx": 0
85
+ },
86
+ {
87
+ "type": "text",
88
+ "text": "formations identically, i.e., the pose-invariance, becomes the key to success in the point cloud matching task.",
89
+ "bbox": [
90
+ 498,
91
+ 503,
92
+ 890,
93
+ 534
94
+ ],
95
+ "page_idx": 0
96
+ },
97
+ {
98
+ "type": "text",
99
+ "text": "Since the side effects caused by a global translation can always be easily eliminated, e.g., by aligning the barycenter with the origin, the attention naturally shifts to coping with the rotations. In the past, handcrafted local descriptors [11, 30, 31, 41] were designed to be rotation-invariant so that the same geometry observed from different views can be correctly matched. With the emergence of deep neural models for 3D point analysis, e.g., multilayer perceptrons (MLPs)-based like PointNet [25, 26], convolutions-based like KPConv [6, 40], and the attention-based like PointTransformer [33, 53], recent approaches [1, 7, 9, 10, 13, 19, 20, 27, 32, 48-51] propose to learn descriptors from raw points as an alternative to handcrafted features that are less robust to occlusion and noise. The majority of deep point matchers [7, 10, 19, 20, 27, 34, 48, 50-52] is sensitive to rotations. Consequently, their invariance to rotations must be obtained extrinsically via augmented training to ensure that the same geometry under different poses can be depicted similarly. However, as the training cases can never span the continuous $SO(3)$ space, they always suffer from instability when facing rotations that are rarely seen during training. This can be observed by a significant performance drop under enlarged rotations at inference time. (See Fig. 1.)",
100
+ "bbox": [
101
+ 496,
102
+ 536,
103
+ 892,
104
+ 883
105
+ ],
106
+ "page_idx": 0
107
+ },
108
+ {
109
+ "type": "text",
110
+ "text": "There are other works [1, 9, 13, 32, 44] that only lever",
111
+ "bbox": [
112
+ 519,
113
+ 885,
114
+ 890,
115
+ 901
116
+ ],
117
+ "page_idx": 0
118
+ },
119
+ {
120
+ "type": "header",
121
+ "text": "CVF",
122
+ "bbox": [
123
+ 106,
124
+ 2,
125
+ 181,
126
+ 42
127
+ ],
128
+ "page_idx": 0
129
+ },
130
+ {
131
+ "type": "header",
132
+ "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.",
133
+ "bbox": [
134
+ 236,
135
+ 1,
136
+ 807,
137
+ 46
138
+ ],
139
+ "page_idx": 0
140
+ },
141
+ {
142
+ "type": "page_footnote",
143
+ "text": "1https://github.com/haoyu94/RoITr",
144
+ "bbox": [
145
+ 94,
146
+ 886,
147
+ 284,
148
+ 898
149
+ ],
150
+ "page_idx": 0
151
+ },
152
+ {
153
+ "type": "page_number",
154
+ "text": "5384",
155
+ "bbox": [
156
+ 482,
157
+ 944,
158
+ 514,
159
+ 955
160
+ ],
161
+ "page_idx": 0
162
+ },
163
+ {
164
+ "type": "text",
165
+ "text": "age deep neural networks to encode the pure geometry with the intrinsically-designed rotation invariance. However, the intrinsic rotation invariance comes at the cost of losing global context. For example, a human's left and right halves are almost identically described, which naturally degrades the distinctiveness of features. Most recently, RIGA [49] is proposed to enhance the distinctiveness of the rotation-invariant descriptors by incorporating a global context, e.g., the left and right halves of a human become distinguishable by knowing there is a chair on the left while a table on the right. However, it lacks a highly-representative geometry encoder since it relies on PointNet [25], which accounts for an ineffective local geometry description. Moreover, as depicting the cross-frame spatial relationships is non-trivial, previous works [19, 27, 34, 50] merely leverage the contextual features in the cross-frame context aggregation, which neglects the positional information. Although RIGA proposes to learn a rotation-invariant position representation by leveraging an additional PointNet, this simple design is hard to model the complex cross-frame positional relationships and leads to less distinctive descriptors.",
166
+ "bbox": [
167
+ 76,
168
+ 90,
169
+ 472,
170
+ 407
171
+ ],
172
+ "page_idx": 1
173
+ },
174
+ {
175
+ "type": "text",
176
+ "text": "In this paper, we present Rotation-Invariant Transformer (RoITr) to tackle the problem of point cloud matching under arbitrary pose variations. By using Point Pair Features (PPFs) as the local coordinates, we propose an attention mechanism to learn the pure geometry regardless of the varying poses. Upon it, attention-based layers are further proposed to compose the encoder-decoder architecture for highly-discriminative and rotation-invariant geometry encoding. We demonstrate its superiority over PointTransformer [53], a state-of-the-art attention-based backbone network, in terms of both efficiency and efficacy in Fig. 8 and Tab. 4 (a), respectively. On the global level, the cross-frame position awareness is introduced in a rotation-invariant fashion to facilitate feature distinctiveness. We illustrate its significance over the state-of-the-art design [27] in Tab. 4 (d). Our main contributions are summarized as:",
177
+ "bbox": [
178
+ 76,
179
+ 407,
180
+ 472,
181
+ 650
182
+ ],
183
+ "page_idx": 1
184
+ },
185
+ {
186
+ "type": "list",
187
+ "sub_type": "text",
188
+ "list_items": [
189
+ "- An attention mechanism designed to disentangle the geometry and poses, which enables the pose-agnostic geometry description.",
190
+ "- An attention-based encoder-decoder architecture that learns highly-representative local geometry in a rotation-invariant fashion.",
191
+ "- A global transformer with rotation-invariant cross-frame position awareness that significantly enhances the feature distinctiveness."
192
+ ],
193
+ "bbox": [
194
+ 94,
195
+ 659,
196
+ 468,
197
+ 801
198
+ ],
199
+ "page_idx": 1
200
+ },
201
+ {
202
+ "type": "text",
203
+ "text": "2. Related Work",
204
+ "text_level": 1,
205
+ "bbox": [
206
+ 76,
207
+ 814,
208
+ 218,
209
+ 830
210
+ ],
211
+ "page_idx": 1
212
+ },
213
+ {
214
+ "type": "text",
215
+ "text": "Models with Extrinsic Rotation Invariance. The mainstream of deep learning-based point cloud matching approaches is intrinsically rotation-sensitive. Pioneers [10,51] learn to describe local patches from a rotation-variant input.",
216
+ "bbox": [
217
+ 76,
218
+ 839,
219
+ 468,
220
+ 901
221
+ ],
222
+ "page_idx": 1
223
+ },
224
+ {
225
+ "type": "text",
226
+ "text": "FCGF [7] leverages fully-convolutional networks to accelerate the geometry description. D3Feat [3] jointly detects and describes sparse keypoints for matching. Predator [19] incorporates the global context to enhance the local descriptors and predicts the overlap regions for keypoint sampling. CoFiNet [50] extracts coarse-to-fine correspondences to alleviate the repeatability issue of keypoints. GeoTrans [27] considers the geometric information in fusing the intra-frame context globally. However, the awareness of spatial positions is missing in the cross-frame aggregation. Lepard [20] extends the non-rigid shape matching [34, 38, 42] to point clouds [28] and proposes a re-positioning module to alleviate the pose variations. REGTR [48] directly regresses the corresponding coordinates and registers point clouds in an end-to-end fashion. Nonetheless, all of these methods suffer from instability with additional rotations.",
227
+ "bbox": [
228
+ 496,
229
+ 90,
230
+ 890,
231
+ 330
232
+ ],
233
+ "page_idx": 1
234
+ },
235
+ {
236
+ "type": "text",
237
+ "text": "Methods with Intrinsic Rotation Invariance. A branch of handcrafted descriptors [8, 14, 41] aligns the input to a canonical representation according to an estimated local reference frame (LRF), while the others [11, 30, 31] mine the rotation-invariant components and encode them as the representation of the local geometry. Inspired by that, some deep learning-based methods [1, 4, 9, 13, 32, 34, 49] are designed to be intrinsically rotation-invariant to make the neural models focus on the pose-agnostic pure geometry. As a pioneer, PPF-FoldNet [9] consumes PPF-based patches and learns the descriptors using a FoldingNet [47]-based architecture without supervision. LRF-based works [1, 13, 32, 34] achieve rotation invariance by aligning their input to the defined canonical representation. YOHO [44] adopts a group of rotations to learn a rotation-equivariant feature group and further obtain the invariance via group pooling. A common problem of the rotation-invariant methods is the less distinctive features. Although RIGA [49] incorporates the global context into local descriptors to enhance the feature distinctiveness, its ineffective local geometry encoding and global position description learned by PointNet [25] still constraint the representation ability of its descriptors.",
238
+ "bbox": [
239
+ 496,
240
+ 333,
241
+ 892,
242
+ 665
243
+ ],
244
+ "page_idx": 1
245
+ },
246
+ {
247
+ "type": "text",
248
+ "text": "3. Method",
249
+ "text_level": 1,
250
+ "bbox": [
251
+ 500,
252
+ 672,
253
+ 589,
254
+ 688
255
+ ],
256
+ "page_idx": 1
257
+ },
258
+ {
259
+ "type": "text",
260
+ "text": "Problem Statement. We tackle the problem of matching a pair of partially-overlapping point clouds $\\mathbf{P} \\in \\mathbb{R}^{n \\times 3}$ and $\\mathbf{Q} \\in \\mathbb{R}^{m \\times 3}$ , and extract a correspondence set $\\hat{\\mathcal{C}} = \\{(\\hat{\\mathbf{p}}_i, \\hat{\\mathbf{q}}_j) | \\hat{\\mathbf{p}}_i \\in \\hat{\\mathbf{P}} \\subseteq \\mathbf{P}, \\hat{\\mathbf{q}}_j \\in \\hat{\\mathbf{Q}} \\subseteq \\mathbf{Q}\\}$ that minimizes:",
261
+ "bbox": [
262
+ 498,
263
+ 699,
264
+ 890,
265
+ 761
266
+ ],
267
+ "page_idx": 1
268
+ },
269
+ {
270
+ "type": "equation",
271
+ "text": "\n$$\n\\frac {1}{| \\hat {\\mathcal {C}} |} \\sum_ {(\\hat {\\mathbf {p}} _ {i}, \\hat {\\mathbf {q}} _ {j}) \\in \\hat {\\mathcal {C}}} \\| \\mathcal {M} ^ {*} (\\hat {\\mathbf {p}} _ {i}) - \\hat {\\mathbf {q}} _ {j} \\| _ {2}, \\tag {1}\n$$\n",
272
+ "text_format": "latex",
273
+ "bbox": [
274
+ 591,
275
+ 767,
276
+ 890,
277
+ 808
278
+ ],
279
+ "page_idx": 1
280
+ },
281
+ {
282
+ "type": "text",
283
+ "text": "where $\\| \\cdot \\| _2$ denotes the Euclidean norm and $|\\cdot |$ is the set cardinality. $\\mathcal{M}^{*}(\\cdot)$ stands for the ground-truth mapping function that maps $\\hat{\\mathbf{p}}_i$ to its corresponding position in $\\hat{\\mathbf{Q}}$ . In rigid scenarios, it is defined by a transformation $\\mathbf{T}^{*}\\in SE(3)$ . For the non-rigid cases it can be denoted as a per-point flow $\\mathbf{f}_i^*\\in \\mathbb{R}^3$ known as the deformation field.",
284
+ "bbox": [
285
+ 496,
286
+ 808,
287
+ 890,
288
+ 901
289
+ ],
290
+ "page_idx": 1
291
+ },
292
+ {
293
+ "type": "page_number",
294
+ "text": "5385",
295
+ "bbox": [
296
+ 482,
297
+ 944,
298
+ 514,
299
+ 955
300
+ ],
301
+ "page_idx": 1
302
+ },
303
+ {
304
+ "type": "image",
305
+ "img_path": "images/352f993e80b8039c3fff593ff3b3710cc2efb699d5437ea5534831fd617b8f8e.jpg",
306
+ "image_caption": [
307
+ "(0). Input"
308
+ ],
309
+ "image_footnote": [],
310
+ "bbox": [
311
+ 86,
312
+ 89,
313
+ 305,
314
+ 164
315
+ ],
316
+ "page_idx": 2
317
+ },
318
+ {
319
+ "type": "image",
320
+ "img_path": "images/db87a1d376332b7b03618dd5794cdaa08a7516290f18cc9528e01d25c924b84b.jpg",
321
+ "image_caption": [
322
+ "(2).Aggregation"
323
+ ],
324
+ "image_footnote": [],
325
+ "bbox": [
326
+ 308,
327
+ 89,
328
+ 656,
329
+ 178
330
+ ],
331
+ "page_idx": 2
332
+ },
333
+ {
334
+ "type": "image",
335
+ "img_path": "images/974288bc32060553067e5aa14d8a1b36e4504339339ba487679a2a0f3eddfa37.jpg",
336
+ "image_caption": [
337
+ "(5). Output"
338
+ ],
339
+ "image_footnote": [],
340
+ "bbox": [
341
+ 648,
342
+ 97,
343
+ 883,
344
+ 169
345
+ ],
346
+ "page_idx": 2
347
+ },
348
+ {
349
+ "type": "image",
350
+ "img_path": "images/4e5ef51993cd3abfeb8ba841a2aba0b59cbc3d108e6b058fc3f2d67369ad52e6.jpg",
351
+ "image_caption": [
352
+ "(1).Encoder"
353
+ ],
354
+ "image_footnote": [],
355
+ "bbox": [
356
+ 91,
357
+ 205,
358
+ 303,
359
+ 354
360
+ ],
361
+ "page_idx": 2
362
+ },
363
+ {
364
+ "type": "image",
365
+ "img_path": "images/09130879a56551a0ae9902d954d95aa14320b2226c7bf4ba302d260c1e117218.jpg",
366
+ "image_caption": [
367
+ "(3). Decoder"
368
+ ],
369
+ "image_footnote": [],
370
+ "bbox": [
371
+ 310,
372
+ 205,
373
+ 658,
374
+ 354
375
+ ],
376
+ "page_idx": 2
377
+ },
378
+ {
379
+ "type": "image",
380
+ "img_path": "images/ebbdcf4c197d78aed4fe93f07f359b890d055be2d6966b8ff2e9b4e25a40194c.jpg",
381
+ "image_caption": [
382
+ "(4). Matching"
383
+ ],
384
+ "image_footnote": [],
385
+ "bbox": [
386
+ 669,
387
+ 207,
388
+ 883,
389
+ 354
390
+ ],
391
+ "page_idx": 2
392
+ },
393
+ {
394
+ "type": "text",
395
+ "text": "Method Overview. An overview of RoITr is shown in Fig. 2. RoITr consists of an encoder-decoder architecture named Point Pair Feature Transformer (PPFTrans) for local geometry encoding and a stack of $g \\times$ global transformers for global context aggregation. Correspondence set $\\hat{C}$ is extracted by the coarse-to-fine matching [50].",
396
+ "bbox": [
397
+ 75,
398
+ 542,
399
+ 468,
400
+ 632
401
+ ],
402
+ "page_idx": 2
403
+ },
404
+ {
405
+ "type": "text",
406
+ "text": "3.1. PPF Attention Mechanism",
407
+ "text_level": 1,
408
+ "bbox": [
409
+ 76,
410
+ 648,
411
+ 316,
412
+ 662
413
+ ],
414
+ "page_idx": 2
415
+ },
416
+ {
417
+ "type": "text",
418
+ "text": "Overview. Fig. 3 compares three different self-attention mechanisms. The standard attention [43] only leverages the input context to obtain the Query $\\mathbf{Q}$ and Key $\\mathbf{K}$ to compute the contextual attention $\\mathbf{A}_C$ , as well as the Value $\\mathbf{V}$ that encodes information for the contextual message $\\mathbf{M}_C$ . GeoTrans [27] proposes to learn the positional encoding $\\mathbf{E}$ from the geometry and calculates a second attention $\\mathbf{A}_G$ to reweigh $\\mathbf{A}_C$ . However, the cues contained in the raw geometry are totally neglected. To this end, we propose to learn the pose-agnostic geometric cues $\\mathbf{G}$ and further generate the geometric message $\\mathbf{M}_G$ in the PPF Attention Mechanism (PAM). On the local level, $\\mathbf{M}_G$ is combined with $\\mathbf{M}_C$ for feature enhancement, while on the global level, it is used to learn the rotation-invariant position representation for the cross-frame context aggregation. More specifically, we de",
419
+ "bbox": [
420
+ 75,
421
+ 674,
422
+ 468,
423
+ 900
424
+ ],
425
+ "page_idx": 2
426
+ },
427
+ {
428
+ "type": "image",
429
+ "img_path": "images/c0b712b6186087050f178483da8e9448683cbb6f8c24b4fae9914085547debf4.jpg",
430
+ "image_caption": [
431
+ "(a). Standard Attention"
432
+ ],
433
+ "image_footnote": [],
434
+ "bbox": [
435
+ 514,
436
+ 544,
437
+ 614,
438
+ 628
439
+ ],
440
+ "page_idx": 2
441
+ },
442
+ {
443
+ "type": "image",
444
+ "img_path": "images/050ce13ccd9012d289ebabdf10f1e5aa082cc00677dad1952586dbd8194f9d92.jpg",
445
+ "image_caption": [
446
+ "Figure 2. An Overview of RoITr. From left to right: $(\\pmb{0})$ . RoITr takes as input a pair of triplets $\\mathcal{P} = (\\mathbf{P}, \\mathbf{N}, \\mathbf{X})$ and $\\mathcal{Q} = (\\mathbf{Q}, \\mathbf{M}, \\mathbf{Y})$ , each with three dimensions referring to the point cloud, the estimated normals, and the initial features. (1).[§ 3.2] A stack of encoder blocks hierarchically downsamples the points to coarser superpoints and encodes the local geometry, yielding superpoint triplets $\\mathcal{P}'$ and $\\mathcal{Q}'$ . Each encoder block consists of an Attentional Abstraction Layer (AAL) for downsampling and abstraction, followed by $e \\times$ PPF Attention Layers (PALs) for local geometry encoding and context aggregation. Both of them are based on our proposed PPF Attention Mechanism (PAM), which enables the pose-agnostic encoding of pure geometry. (See Fig. 3 and Fig. 4). (2).[§ 3.3] Global information is fused to enhance the superpoint features of $\\mathcal{P}'$ and $\\mathcal{Q}'$ . The geometric cues are globally aggregated as a rotation-invariant position representation, which introduces spatial awareness in the consecutive cross-frame context aggregation. After a stack of $g \\times$ global transformers, the globally-enhanced triplets $\\widetilde{\\mathcal{P}}'$ and $\\widetilde{\\mathcal{Q}}'$ are produced. (3).[§ 3.2] Superpoint triplets $\\mathcal{P}'$ and $\\mathcal{Q}'$ are decoded to point triplets $\\hat{\\mathcal{P}}$ and $\\hat{\\mathcal{Q}}$ by a stack of decoder blocks. Each block consists of a Transition Up Layer (TUL) for upsampling and context aggregation, followed by $d \\times$ PALs. (4).[§ 3.4] By adopting the coarse-to-fine matching [50], $\\widetilde{\\mathcal{P}}'$ and $\\widetilde{\\mathcal{Q}}'$ are matched to generate superpoint correspondences, which are consecutively refined to point correspondences between $\\hat{\\mathcal{P}}$ and $\\hat{\\mathcal{Q}}$ . (5). $\\hat{\\mathcal{C}}$ is established between $\\hat{\\mathcal{P}}$ and $\\hat{\\mathcal{Q}}$ .",
447
+ "(b). GeoTrans"
448
+ ],
449
+ "image_footnote": [],
450
+ "bbox": [
451
+ 620,
452
+ 542,
453
+ 741,
454
+ 628
455
+ ],
456
+ "page_idx": 2
457
+ },
458
+ {
459
+ "type": "image",
460
+ "img_path": "images/1ae7e04b44a0ede10cb1f694b3023f8ff8b9887eed7c1e32a5a65db1f3724d8b.jpg",
461
+ "image_caption": [
462
+ "(c).PAM (Ours)",
463
+ "Figure 3. Illustration of different self-attention computation in the standard attention [43], GeoTrans [27], and PAM."
464
+ ],
465
+ "image_footnote": [],
466
+ "bbox": [
467
+ 746,
468
+ 542,
469
+ 885,
470
+ 628
471
+ ],
472
+ "page_idx": 2
473
+ },
474
+ {
475
+ "type": "text",
476
+ "text": "fine PAM on an Anchor triplet $\\mathcal{P}^A = (\\mathbf{P}^A, \\mathbf{N}^A, \\mathbf{X}^A)$ and a Support triplet $\\mathcal{P}^S = (\\mathbf{P}^S, \\mathbf{N}^S, \\mathbf{X}^S)$ , both with three dimensions referring to the point cloud, the estimated normals, and the associated features, respectively. PAM aggregates the learned context and geometric cues from $\\mathcal{P}^S$ and flows the messages to $\\mathcal{P}^A$ .",
477
+ "bbox": [
478
+ 496,
479
+ 702,
480
+ 890,
481
+ 792
482
+ ],
483
+ "page_idx": 2
484
+ },
485
+ {
486
+ "type": "text",
487
+ "text": "Pose-Agnostic Coordinate Representation. The basis of PAM is the pose-agnostic local coordinate representation that we construct based on PPFs [11]. Let $\\mathcal{P}_i^A \\coloneqq (\\mathbf{p}_i^A \\in \\mathbf{P}^A, \\mathbf{n}_i^A \\in \\mathbf{N}^A, \\mathbf{x}_i^A \\in \\mathbf{X}^A) \\in \\mathcal{P}^A$ denote the triplet constructed by picking the $i^{th}$ item on each dimension. For each $\\mathbf{p}_i^A$ , a subset of $\\mathcal{P}^S$ is first retrieved according to the Euclidean distance w.r.t. $\\mathbf{P}^S$ , denoted as",
488
+ "bbox": [
489
+ 496,
490
+ 795,
491
+ 890,
492
+ 898
493
+ ],
494
+ "page_idx": 2
495
+ },
496
+ {
497
+ "type": "page_number",
498
+ "text": "5386",
499
+ "bbox": [
500
+ 482,
501
+ 945,
502
+ 514,
503
+ 955
504
+ ],
505
+ "page_idx": 2
506
+ },
507
+ {
508
+ "type": "image",
509
+ "img_path": "images/5081a262fa30968f29b7c9345f2e9eb18f9480785ff18366f83ba9c4b08632b3.jpg",
510
+ "image_caption": [
511
+ "Figure 4. Left: The workflow of the PPF Attention Mechanism (PAM). Right: Detailed calculation of the attention."
512
+ ],
513
+ "image_footnote": [],
514
+ "bbox": [
515
+ 81,
516
+ 88,
517
+ 472,
518
+ 233
519
+ ],
520
+ "page_idx": 3
521
+ },
522
+ {
523
+ "type": "text",
524
+ "text": "$\\mathcal{P}_{\\mathcal{N}(i)}^{S} \\coloneqq (\\mathbf{P}_{\\mathcal{N}(i)}^{S}, \\mathbf{N}_{\\mathcal{N}(i)}^{S}, \\mathbf{X}_{\\mathcal{N}(i)}^{S}) \\subseteq \\mathcal{P}^{S}$ , with $\\mathcal{N}(i)$ the indices of $k$ -nearest neighbors. We then adopt PPFs [11] to construct a local coordinate system around each $\\mathbf{p}_i^A$ to represent the pose-agnostic position of $\\mathbf{P}_{\\mathcal{N}(i)}^{S}$ w.r.t. it. The coordinate of point $\\mathbf{p}_j^S \\in \\mathbf{P}_{\\mathcal{N}(i)}^{S}$ is transferred to:",
525
+ "bbox": [
526
+ 75,
527
+ 280,
528
+ 470,
529
+ 364
530
+ ],
531
+ "page_idx": 3
532
+ },
533
+ {
534
+ "type": "equation",
535
+ "text": "\n$$\n\\mathbf {e} _ {j} ^ {S} = \\left(\\| \\mathbf {d} \\| _ {2}, \\angle \\left(\\mathbf {n} _ {i} ^ {A}, \\mathbf {d}\\right), \\angle \\left(\\mathbf {n} _ {j} ^ {S}, \\mathbf {d}\\right), \\angle \\left(\\mathbf {n} _ {j} ^ {S}, \\mathbf {n} _ {i} ^ {A}\\right)\\right), \\quad (2)\n$$\n",
536
+ "text_format": "latex",
537
+ "bbox": [
538
+ 114,
539
+ 375,
540
+ 468,
541
+ 395
542
+ ],
543
+ "page_idx": 3
544
+ },
545
+ {
546
+ "type": "text",
547
+ "text": "with $\\mathbf{d} = \\mathbf{p}_j^S -\\mathbf{p}_i^A$ , and $\\mathbf{n}_i^A$ and $\\mathbf{n}_j^S$ the estimated normals of $\\mathbf{p}_i^A$ and $\\mathbf{p}_j^S$ , respectively. $\\angle (\\mathbf{v}_1,\\mathbf{v}_2)$ computes the angles between the two vectors [5, 9]. The transferred coordinates of $\\mathbf{P}_{\\mathcal{N}(i)}^{S}$ are denoted as $\\mathbf{E}_{\\mathcal{N}(i)}^{S}$",
548
+ "bbox": [
549
+ 75,
550
+ 402,
551
+ 468,
552
+ 469
553
+ ],
554
+ "page_idx": 3
555
+ },
556
+ {
557
+ "type": "text",
558
+ "text": "PPF Attention Mechanism. PPF Attention Mechanism (PAM) takes as input the Support triplet $\\mathcal{P}^S$ and the Anchor point cloud $\\mathbf{P}^A$ with estimated normals $\\mathbf{N}^A$ . PAM generates the Anchor features $\\mathbf{X}^A$ by aggregating the pose-agnostic local geometry and highly-representative learned context from $\\mathcal{P}^S$ , which is defined as:",
559
+ "bbox": [
560
+ 75,
561
+ 468,
562
+ 468,
563
+ 556
564
+ ],
565
+ "page_idx": 3
566
+ },
567
+ {
568
+ "type": "equation",
569
+ "text": "\n$$\n\\mathcal {P} ^ {A} = \\delta \\left(\\mathbf {P} ^ {A}, \\mathbf {N} ^ {A} \\mid \\mathcal {P} ^ {S}\\right), \\tag {3}\n$$\n",
570
+ "text_format": "latex",
571
+ "bbox": [
572
+ 192,
573
+ 570,
574
+ 468,
575
+ 590
576
+ ],
577
+ "page_idx": 3
578
+ },
579
+ {
580
+ "type": "text",
581
+ "text": "with $\\delta(\\cdot)$ representing PAM. As shown in Fig. 4 (a), for each $\\mathbf{p}_i^A \\in \\mathbf{P}^A$ with normal $\\mathbf{n}_i^A$ , we find its nearest point $\\mathbf{p}_j^S \\in \\mathbf{P}^S$ whose associated feature $\\mathbf{x}_j^S$ is assigned to $\\mathbf{p}_i^A$ as the initial description. Then, $k$ -nearest neighbors from $\\mathbf{P}^S$ are retrieved according to the Euclidean distance in 3D space, yielding $\\mathbf{P}_{\\mathcal{N}(i)}^S \\subseteq \\mathbf{P}^S$ and $\\mathbf{X}_{\\mathcal{N}(i)}^S \\subseteq \\mathbf{X}^S$ . Following Eq. 2, $\\mathbf{P}_{\\mathcal{N}(i)}^S$ is transferred to the pose-agnostic position representation $\\mathbf{E}_{\\mathcal{N}(i)}^S$ , which is consecutively projected to the coordinate embedding $\\mathbf{E}_S$ via a linear layer. $\\mathbf{x}_j^S$ and $\\mathbf{X}_{\\mathcal{N}(i)}^S$ are projected to the contextual features $\\mathbf{x}_S$ and $\\mathbf{X}_S$ by a second shared linear layer, respectively. In Fig. 4 (b), the attention mechanism uses five learnable matrices $\\mathbf{W}_G$ , $\\mathbf{W}_E$ , $\\mathbf{W}_Q$ , $\\mathbf{W}_K$ , and $\\mathbf{W}_V$ to project the input. Specifically, $\\mathbf{W}_G$ and $\\mathbf{W}_E$ project the input coordinate representation to the geometric cues and positional encoding by:",
582
+ "bbox": [
583
+ 75,
584
+ 595,
585
+ 470,
586
+ 834
587
+ ],
588
+ "page_idx": 3
589
+ },
590
+ {
591
+ "type": "equation",
592
+ "text": "\n$$\n\\mathbf {G} = \\mathbf {E} _ {S} \\mathbf {W} _ {G} \\quad \\text {a n d} \\quad \\mathbf {E} = \\mathbf {E} _ {S} \\mathbf {W} _ {E}, \\tag {4}\n$$\n",
593
+ "text_format": "latex",
594
+ "bbox": [
595
+ 151,
596
+ 847,
597
+ 468,
598
+ 864
599
+ ],
600
+ "page_idx": 3
601
+ },
602
+ {
603
+ "type": "text",
604
+ "text": "respectively. Similarly, $\\mathbf{W}_Q$ , $\\mathbf{W}_K$ , and $\\mathbf{W}_V$ project the learned context to Query, Key, and Value as:",
605
+ "bbox": [
606
+ 75,
607
+ 869,
608
+ 468,
609
+ 901
610
+ ],
611
+ "page_idx": 3
612
+ },
613
+ {
614
+ "type": "equation",
615
+ "text": "\n$$\n\\mathbf {q} = \\mathbf {x} _ {S} \\mathbf {W} _ {Q}, \\mathbf {K} = \\mathbf {X} _ {S} \\mathbf {W} _ {K}, \\text {a n d} \\mathbf {V} = \\mathbf {X} _ {S} \\mathbf {W} _ {V}, \\tag {5}\n$$\n",
616
+ "text_format": "latex",
617
+ "bbox": [
618
+ 522,
619
+ 119,
620
+ 890,
621
+ 136
622
+ ],
623
+ "page_idx": 3
624
+ },
625
+ {
626
+ "type": "text",
627
+ "text": "respectively. The attention a that measures the feature similarity, and the message m that encodes both the pose-agnostic geometry and the representative context read as:",
628
+ "bbox": [
629
+ 498,
630
+ 142,
631
+ 890,
632
+ 186
633
+ ],
634
+ "page_idx": 3
635
+ },
636
+ {
637
+ "type": "equation",
638
+ "text": "\n$$\n\\mathbf {a} = \\operatorname {S o f t m a x} \\left(\\frac {\\mathbf {q} \\mathbf {E} ^ {T} + \\mathbf {q} \\mathbf {K} ^ {T}}{\\sqrt {c _ {0}}}\\right) \\quad \\text {a n d} \\quad \\mathbf {m} = \\mathbf {a} \\mathbf {G} + \\mathbf {a} \\mathbf {V}, \\tag {6}\n$$\n",
639
+ "text_format": "latex",
640
+ "bbox": [
641
+ 506,
642
+ 198,
643
+ 890,
644
+ 234
645
+ ],
646
+ "page_idx": 3
647
+ },
648
+ {
649
+ "type": "text",
650
+ "text": "respectively. The message $\\mathbf{m}$ is projected and aggregated to $\\mathbf{x}_j^S$ via an element-wise addition followed by a normalization through LayerNorm [2]. The final linear layer projects the obtained feature to $\\mathbf{x}_i^A$ , from which $\\mathbf{X}^A$ is obtained to formulate the output $\\mathcal{P}^A$ with the known $\\mathbf{P}^A$ and $\\mathbf{N}^A$ .",
651
+ "bbox": [
652
+ 498,
653
+ 244,
654
+ 890,
655
+ 321
656
+ ],
657
+ "page_idx": 3
658
+ },
659
+ {
660
+ "type": "text",
661
+ "text": "3.2. PPFTrans for Local Geometry Description",
662
+ "text_level": 1,
663
+ "bbox": [
664
+ 498,
665
+ 332,
666
+ 862,
667
+ 349
668
+ ],
669
+ "page_idx": 3
670
+ },
671
+ {
672
+ "type": "text",
673
+ "text": "Overview. As illustrated in Fig. 2, PPFTrans consumes triplets $\\mathcal{P}$ and $\\mathcal{Q}$ . Taking $\\mathcal{P} = (\\mathbf{P},\\mathbf{N},\\mathbf{X})$ as an example, it consists of $\\mathbf{P} \\in \\mathbb{R}^{n \\times 3}$ the points cloud, $\\mathbf{N} \\in \\mathbb{R}^{n \\times 3}$ the normals estimated from $\\mathbf{P}$ , and $\\mathbf{X} = \\vec{\\mathbf{1}} \\in \\mathbb{R}^{n \\times 1}$ the initial point features. The encoder produces the superpoint triplet $\\mathcal{P}' = (\\mathbf{P}',\\mathbf{N}',\\mathbf{X}')$ with $\\mathbf{P}' \\in \\mathbb{R}^{n' \\times 3}$ and $\\mathbf{X}' \\in \\mathbb{R}^{n' \\times c'}$ . With the consecutive decoder, $\\mathcal{P}'$ is decoded to a triplet $\\hat{\\mathcal{P}} = (\\hat{\\mathbf{P}},\\hat{\\mathbf{N}},\\hat{\\mathbf{X}})$ including $\\hat{n}$ points with features $\\hat{\\mathbf{X}} \\in \\mathbb{R}^{\\hat{n} \\times \\hat{c}}$ . Notably, as we adopt a Farthest Point Sampling (FPS) strategy [26], it always satisfies that $\\mathbf{P}' \\subseteq \\hat{\\mathbf{P}} \\subseteq \\mathbf{P}$ . The same goes for a second point cloud $\\mathbf{Q}$ with an input triplet $\\mathcal{Q} = (\\mathbf{Q} \\in \\mathbb{R}^{m \\times 3},\\mathbf{M} \\in \\mathbb{R}^{m \\times 3},\\mathbf{Y} = \\vec{\\mathbf{1}} \\in \\mathbb{R}^{m \\times 1})$ by the shared architecture. In the rest of this paper, we only demonstrate for $\\mathcal{P}$ unless the model processes $\\mathcal{Q}$ differently. Encoder. The encoder is constructed by stacking several encoder blocks, each including an Attentional Abstraction Layer (AAL) followed by $e \\times$ PPF Attention Layers (PALs). Each block consumes the output of the previous block as the Support triplet $\\mathcal{P}^S$ ( $\\mathcal{P}^S = \\mathcal{P}$ for the first block). $\\mathcal{P}^S$ first flows to AAL, where Anchor points $\\mathbf{P}^A$ with associated normals $\\mathbf{N}^A$ are obtained via FPS [26]. The Anchor triplet $\\mathcal{P}^A$ is then generated in AAL via a PAM following Eq. 3. A sequence of PALs is applied for enhancing the Anchor features $\\mathbf{X}^A$ , each updating the features as:",
674
+ "bbox": [
675
+ 498,
676
+ 357,
677
+ 890,
678
+ 720
679
+ ],
680
+ "page_idx": 3
681
+ },
682
+ {
683
+ "type": "equation",
684
+ "text": "\n$$\n\\mathcal {P} ^ {A} \\leftarrow \\theta (\\mathcal {P} ^ {A}) = \\operatorname {R e L U} \\left(\\mathbf {X} ^ {A} + \\phi \\left(\\delta \\left(\\mathbf {P} ^ {A}, \\mathbf {N} ^ {A} \\mid \\mathcal {P} ^ {A}\\right)\\right)\\right), \\tag {7}\n$$\n",
685
+ "text_format": "latex",
686
+ "bbox": [
687
+ 509,
688
+ 731,
689
+ 890,
690
+ 751
691
+ ],
692
+ "page_idx": 3
693
+ },
694
+ {
695
+ "type": "text",
696
+ "text": "with $\\phi$ the LayerNorm [2], $\\delta$ the PAM, and $\\theta$ the PAL. $\\leftarrow$ depicts feature updating. The encoder block outputs the updated $\\mathcal{P}^A$ , and the output of the whole encoder is defined as $\\mathcal{P}'$ , which is the output of the final encoder block.",
697
+ "bbox": [
698
+ 498,
699
+ 763,
700
+ 890,
701
+ 823
702
+ ],
703
+ "page_idx": 3
704
+ },
705
+ {
706
+ "type": "text",
707
+ "text": "Decoder. We build the decoder by stacking a series of decoder blocks, each consisting of a Transition Up Layer (TUL) followed by $d \\times \\mathrm{PAL}$ . Each block takes the output of the previous block as the Anchor triplet $\\mathcal{P}^A$ ( $\\mathcal{P}^A = \\mathcal{P}'$ for the first block), and takes the Support",
708
+ "bbox": [
709
+ 498,
710
+ 824,
711
+ 890,
712
+ 900
713
+ ],
714
+ "page_idx": 3
715
+ },
716
+ {
717
+ "type": "page_number",
718
+ "text": "5387",
719
+ "bbox": [
720
+ 482,
721
+ 944,
722
+ 514,
723
+ 955
724
+ ],
725
+ "page_idx": 3
726
+ },
727
+ {
728
+ "type": "image",
729
+ "img_path": "images/f4b3f87f9e83497fcc7b10f0844a7cef1f6e7b788a4cf5fc18e8d0079f4a86bd.jpg",
730
+ "image_caption": [
731
+ "(a). Geometry-Aware Self-Attention"
732
+ ],
733
+ "image_footnote": [],
734
+ "bbox": [
735
+ 83,
736
+ 90,
737
+ 316,
738
+ 239
739
+ ],
740
+ "page_idx": 4
741
+ },
742
+ {
743
+ "type": "image",
744
+ "img_path": "images/d764f44f9d306301505fa374dfc9c5743da0b1b9c69a915b710abec5d66a715b.jpg",
745
+ "image_caption": [
746
+ "(b). Position-Aware Cross-Attention",
747
+ "Figure 5. The computation graph of our global transformer consisting of the Geometry-Aware Self-Attention Module (GSM) and Position-Aware Cross-Attention Module (PCM)."
748
+ ],
749
+ "image_footnote": [],
750
+ "bbox": [
751
+ 320,
752
+ 90,
753
+ 468,
754
+ 241
755
+ ],
756
+ "page_idx": 4
757
+ },
758
+ {
759
+ "type": "text",
760
+ "text": "triplet $\\mathcal{P}^S$ from the encoder via skip connections. The input flows to TUL, where each feature $\\widetilde{\\mathbf{x}}_j^S\\in \\widetilde{\\mathbf{X}}^S$ assigned to $\\mathbf{p}_j^S\\in \\mathbf{P}^S$ is interpolated by:",
761
+ "bbox": [
762
+ 75,
763
+ 306,
764
+ 468,
765
+ 357
766
+ ],
767
+ "page_idx": 4
768
+ },
769
+ {
770
+ "type": "equation",
771
+ "text": "\n$$\n\\widetilde {\\mathbf {x}} _ {j} ^ {S} = \\frac {\\sum_ {i \\in \\mathcal {N} (j)} w _ {i} ^ {j} \\mathbf {x} _ {i} ^ {A}}{\\sum_ {i \\in \\mathcal {N} (j)} w _ {i} ^ {j}}, \\text {w i t h} w _ {i} ^ {j} = \\frac {1}{\\left\\| \\mathbf {p} _ {j} ^ {S} - \\mathbf {p} _ {i} ^ {A} \\right\\| _ {2}}, \\tag {8}\n$$\n",
772
+ "text_format": "latex",
773
+ "bbox": [
774
+ 96,
775
+ 359,
776
+ 468,
777
+ 400
778
+ ],
779
+ "page_idx": 4
780
+ },
781
+ {
782
+ "type": "text",
783
+ "text": "with $\\mathcal{N}(j)$ the $k$ -nearest neighbors of $\\mathbf{p}_j^S$ in $\\mathbf{P}^A$ . Features are updated by two linear layers as $\\mathcal{P}^S \\gets \\zeta_1(\\mathbf{X}^S) + \\zeta_2(\\widetilde{\\mathbf{X}}^S)$ . A sequence of PALs is adopted after TUL, each enhancing the features as $\\mathcal{P}^S \\gets \\theta(\\mathcal{P}^S)$ according to Eq. 7. The decoder block outputs the updated $\\mathcal{P}^S$ , and the output of the whole decoder is denoted as $\\hat{\\mathcal{P}}$ , which is the output of the final decoder block.",
784
+ "bbox": [
785
+ 75,
786
+ 410,
787
+ 468,
788
+ 521
789
+ ],
790
+ "page_idx": 4
791
+ },
792
+ {
793
+ "type": "text",
794
+ "text": "3.3. Global Transformer for Context Aggregation",
795
+ "text_level": 1,
796
+ "bbox": [
797
+ 76,
798
+ 530,
799
+ 459,
800
+ 547
801
+ ],
802
+ "page_idx": 4
803
+ },
804
+ {
805
+ "type": "text",
806
+ "text": "Overview. Our designed global transformer takes as input a pair of triplets $\\mathcal{P}'$ and $\\mathcal{Q}'$ , and enhances the features with the global context, yielding $\\widetilde{\\mathbf{X}}' \\in \\mathbb{R}^{n' \\times c'}$ and $\\widetilde{\\mathbf{Y}}' \\in \\mathbb{R}^{m' \\times c'}$ , respectively. We stack $g \\times$ global transformers, with each including a Geometry-Aware Self-Attention Module (GSM) and a Position-Aware Cross-Attention Module (PCM) (See Fig. 2 and Fig. 5). Different from previous works [19, 27, 48-50] that totally neglect the cross-frame spatial relationships, we propose to learn a rotation-invariant position representation for each superpoint to enable the position-aware cross-frame context aggregation.",
807
+ "bbox": [
808
+ 75,
809
+ 551,
810
+ 468,
811
+ 718
812
+ ],
813
+ "page_idx": 4
814
+ },
815
+ {
816
+ "type": "text",
817
+ "text": "Geometry-Aware Self-Attention Module. On the global level, we modify PAM to learn the rotation-invariant position representation and to aggregate the learned context across the whole frame simultaneously. The design of GSM is detailed in Fig. 5 (a). GSM has two branches, where the geometry branch mines the geometric cues from the pairwise rotation-invariant geometry representation proposed in [27], and the context branch aggregates the global context across the frame. We refer the readers to the Appendix for the detailed construction of $\\mathbf{R}' \\in \\mathbb{R}^{n' \\times n' \\times c'}$ and the ablation study on it. Similar to Eq. 4, the geometric cues $\\mathbf{G}'$ and the positional encoding $\\mathbf{E}'$ are linearly projected from",
818
+ "bbox": [
819
+ 75,
820
+ 719,
821
+ 468,
822
+ 901
823
+ ],
824
+ "page_idx": 4
825
+ },
826
+ {
827
+ "type": "text",
828
+ "text": "$\\mathbf{R}^{\\prime}$ . $\\mathbf{E}^{\\prime}$ is further processed in the geometry branch and finally leveraged as the rotation-invariant position representation. In the context branch, $\\mathbf{Q}^{\\prime}$ , $\\mathbf{K}^{\\prime}$ , and $\\mathbf{V}^{\\prime}$ are obtained by linearly mapping the input features $\\mathbf{X}^{\\prime}$ similar to Eq. 5. The hybrid score matrix $\\mathbf{S}^{\\prime} \\in \\mathbb{R}^{n^{\\prime} \\times n^{\\prime}}$ is computed as:",
829
+ "bbox": [
830
+ 496,
831
+ 90,
832
+ 890,
833
+ 167
834
+ ],
835
+ "page_idx": 4
836
+ },
837
+ {
838
+ "type": "equation",
839
+ "text": "\n$$\n\\mathbf {S} ^ {\\prime} (i, j) = \\frac {\\left(\\mathbf {q} _ {i} ^ {\\prime}\\right) \\left(\\mathbf {e} _ {i , j} ^ {\\prime} + \\mathbf {k} _ {j} ^ {\\prime}\\right) ^ {T}}{\\sqrt {c ^ {\\prime}}}, \\tag {9}\n$$\n",
840
+ "text_format": "latex",
841
+ "bbox": [
842
+ 596,
843
+ 176,
844
+ 890,
845
+ 212
846
+ ],
847
+ "page_idx": 4
848
+ },
849
+ {
850
+ "type": "text",
851
+ "text": "with $\\mathbf{e}_{i,j}^{\\prime} := \\mathbf{E}^{\\prime}(i,j,:)$ , $\\mathbf{q}_i^{\\prime} := \\mathbf{Q}^{\\prime}(i,:)$ , and $\\mathbf{k}_j^{\\prime} := \\mathbf{K}^{\\prime}(j,:)$ the $c^{\\prime}$ -dimension vectors. The hybrid attention $\\mathbf{A}^{\\prime}$ is obtained via a Softmax function over each row of $\\mathbf{S}^{\\prime}$ , and the geometric messages $\\mathbf{M}_G^{\\prime} \\in \\mathbb{R}^{n^{\\prime} \\times c^{\\prime}}$ are computed as:",
852
+ "bbox": [
853
+ 496,
854
+ 217,
855
+ 890,
856
+ 279
857
+ ],
858
+ "page_idx": 4
859
+ },
860
+ {
861
+ "type": "equation",
862
+ "text": "\n$$\n\\mathbf {M} _ {G} ^ {\\prime} (i,:) = \\sum_ {1 \\leq j \\leq n ^ {\\prime}} a _ {i, j} ^ {\\prime} \\mathbf {g} _ {i, j} ^ {\\prime}, \\tag {10}\n$$\n",
863
+ "text_format": "latex",
864
+ "bbox": [
865
+ 599,
866
+ 290,
867
+ 890,
868
+ 324
869
+ ],
870
+ "page_idx": 4
871
+ },
872
+ {
873
+ "type": "text",
874
+ "text": "with $a_{i,j}^{\\prime} \\coloneqq \\mathbf{A}^{\\prime}(i,j)$ and $\\mathbf{g}_{i,j}^{\\prime} \\coloneqq \\mathbf{G}^{\\prime}(i,j,:)$ . The contextual messages $\\mathbf{M}_V^{\\prime} \\in \\mathbb{R}^{n^{\\prime} \\times c^{\\prime}}$ are computed by $\\mathbf{A}^{\\prime}\\mathbf{V}^{\\prime}$ . After a feed-forward network [43], the position representation $\\mathbf{E}_P^{\\prime}$ and globally-enhanced context $\\mathbf{C}_P^{\\prime}$ are generated.",
875
+ "bbox": [
876
+ 496,
877
+ 329,
878
+ 890,
879
+ 393
880
+ ],
881
+ "page_idx": 4
882
+ },
883
+ {
884
+ "type": "text",
885
+ "text": "Position-Aware Cross-Attention Module. PCM consumes a pair of doublets $(\\mathbf{E}_P',\\mathbf{C}_P')$ and $(\\mathbf{E}_Q',\\mathbf{C}_Q')$ that are generated from $\\mathcal{P}'$ and $\\mathcal{Q}'$ by a shared GSM, respectively. As the cross-attention is directional, we apply the same PCM twice, with the first aggregation from $\\mathcal{Q}'$ to $\\mathcal{P}'$ (See Fig. 5 (b)), and the second reversed. As the first step, the rotation-invariant position representation is incorporated to make the consecutive cross-attention position-aware, yielding position-aware features $\\mathbf{F}_P' = \\mathbf{E}_P' + \\mathbf{C}_P'$ and $\\mathbf{F}_Q' = \\mathbf{E}_Q' + \\mathbf{C}_Q'$ . Similar to Eq. 5, $\\widetilde{\\mathbf{Q}}',\\widetilde{\\mathbf{K}}'$ , and $\\widetilde{\\mathbf{V}}'$ are computed as the linear projection of $\\mathbf{F}_P',\\mathbf{F}_Q'$ , and $\\mathbf{F}_Q'$ , respectively. The attention matrix $\\widetilde{\\mathbf{A}}\\in \\mathbb{R}^{n'\\times m'}$ is computed via a row-wise softmax function applied on $\\widetilde{\\mathbf{Q}}'\\widetilde{\\mathbf{K}}'^T$ . The fused messages are presented as $\\widetilde{\\mathbf{A}}\\widetilde{\\mathbf{V}}'$ , which are finally mapped to the output features $\\widetilde{\\mathbf{X}}'$ through a feedforward network. As we introduce spatial awareness at the beginning of PCM, both the attention computation and message fusion are aware of the cross-frame positions. After the twice application of PCM, the input features are enhanced as $\\mathcal{P}'\\gets \\widetilde{\\mathbf{X}}'$ and $\\mathcal{Q}'\\gets \\widetilde{\\mathbf{Y}}'$ , respectively. The global aggregation stage finally generates a pair of triplets $\\widetilde{\\mathcal{P}}':=(\\mathbf{P}',\\mathbf{N}',\\widetilde{\\mathbf{X}}')$ and $\\widetilde{\\mathcal{Q}}':=(\\mathbf{Q}',\\mathbf{M}',\\widetilde{\\mathbf{Y}}')$ , with the enhanced features from the last global transformer.",
886
+ "bbox": [
887
+ 496,
888
+ 393,
889
+ 892,
890
+ 746
891
+ ],
892
+ "page_idx": 4
893
+ },
894
+ {
895
+ "type": "text",
896
+ "text": "3.4. Point Matching and Loss Function",
897
+ "text_level": 1,
898
+ "bbox": [
899
+ 500,
900
+ 756,
901
+ 795,
902
+ 772
903
+ ],
904
+ "page_idx": 4
905
+ },
906
+ {
907
+ "type": "text",
908
+ "text": "Superpoint Matching. As shown in Fig. 2, the point matching stage consumes a pair of superpoint triplets $\\hat{\\mathcal{P}}'$ and $\\hat{\\mathcal{Q}}'$ obtained from the global transformer, as well as a pair of point triplets $\\hat{\\mathcal{P}}$ and $\\hat{\\mathcal{Q}}$ produced by the decoder. We adopt the coarse-to-fine matching proposed in [50]. Following [27], we first normalize the superpoint features $\\widetilde{\\mathbf{X}}'$ and $\\widetilde{\\mathbf{Y}}'$ onto a unit hypersphere, and measure the pairwise similarity using a Gaussian correlation matrix $\\widetilde{\\mathbf{S}}$ with $\\widetilde{\\mathbf{S}}(i,j) =$",
909
+ "bbox": [
910
+ 496,
911
+ 779,
912
+ 890,
913
+ 900
914
+ ],
915
+ "page_idx": 4
916
+ },
917
+ {
918
+ "type": "page_number",
919
+ "text": "5388",
920
+ "bbox": [
921
+ 482,
922
+ 944,
923
+ 514,
924
+ 955
925
+ ],
926
+ "page_idx": 4
927
+ },
928
+ {
929
+ "type": "text",
930
+ "text": "$-\\exp \\left(-\\|\\widetilde{\\mathbf{x}}_i'\\widetilde{\\mathbf{y}}_j'\\|_2^2\\right)$ . After a dual-normalization [27,29,37] on $\\widetilde{\\mathbf{S}}$ for global feature correlation, superpoints associated to the top- $k$ entries are selected as the coarse correspondence set $\\mathcal{C}' = \\{(\\mathbf{p}_i',\\mathbf{q}_i')|\\mathbf{p}_i' \\in \\mathbf{P}',\\mathbf{q}_i' \\in \\mathbf{Q}'\\}$ .",
931
+ "bbox": [
932
+ 76,
933
+ 90,
934
+ 467,
935
+ 155
936
+ ],
937
+ "page_idx": 5
938
+ },
939
+ {
940
+ "type": "text",
941
+ "text": "Point Matching. For extracting point correspondences, denser points $\\hat{\\mathbf{P}}$ and $\\hat{\\mathbf{Q}}$ are first assigned to superpoints. To this end, the point-to-node strategy [50] is leveraged, where each point is assigned to its closest superpoint in 3D space. Given a superpoint $\\mathbf{p}_i' \\in \\mathbf{P}'$ , the group of points assigned to it is denoted as $\\hat{\\mathbf{G}}_i^P \\subseteq \\hat{\\mathbf{P}}$ . The group of features associated to $\\hat{\\mathbf{G}}_i^P$ is further defined as $\\hat{\\mathbf{G}}_i^X$ with $\\hat{\\mathbf{G}}_i^X \\subseteq \\hat{X}$ . For each superpoint correspondence $\\mathcal{C}_l' = (\\mathbf{p}_i', \\mathbf{q}_j')$ , the similarity between the corresponding feature groups $\\hat{\\mathbf{G}}_i^X$ and $\\hat{\\mathbf{G}}_j^Y$ is calculated as $\\hat{\\mathbf{S}}_l = \\hat{\\mathbf{G}}_i^X (\\hat{\\mathbf{G}}_j^Y)^T / \\sqrt{\\hat{c}}$ , with $\\hat{c}$ the feature dimension. We then follow [35] to append a stack row and column to $\\hat{\\mathbf{S}}_l$ filled with a learnable parameter $\\alpha$ , and iteratively run the Sinkhorn Algorithm [36]. After removing the slack row and column of $\\hat{\\mathbf{S}}_l$ , the mutual top-k entries, i.e., entries with top- $k$ confidence on both the row and the column, are selected to formulate a point correspondence set $\\hat{\\mathcal{C}}_l$ . The final correspondence set $\\hat{\\mathcal{C}}$ is collected by $\\hat{\\mathcal{C}} = \\bigcup_{l=1}^{|\\mathcal{C}'|} \\hat{\\mathcal{C}}_l$ .",
942
+ "bbox": [
943
+ 75,
944
+ 156,
945
+ 467,
946
+ 425
947
+ ],
948
+ "page_idx": 5
949
+ },
950
+ {
951
+ "type": "text",
952
+ "text": "Loss Function. Our loss function reads as $\\mathcal{L} = \\mathcal{L}_s + \\lambda \\mathcal{L}_p$ , with a superpoint matching loss $\\mathcal{L}_s$ and a point matching loss $\\mathcal{L}_p$ balanced by a hyper-parameter $\\lambda$ ( $\\lambda = 1$ by default). The detailed definition is introduced in the Appendix.",
953
+ "bbox": [
954
+ 76,
955
+ 425,
956
+ 467,
957
+ 486
958
+ ],
959
+ "page_idx": 5
960
+ },
961
+ {
962
+ "type": "text",
963
+ "text": "4. Experiment",
964
+ "text_level": 1,
965
+ "bbox": [
966
+ 76,
967
+ 503,
968
+ 200,
969
+ 518
970
+ ],
971
+ "page_idx": 5
972
+ },
973
+ {
974
+ "type": "text",
975
+ "text": "We evaluate RoITr on both rigid (3DMatch [51] & 3DLoMatch [19]) and non-rigid (4DMatch [20] & 4DLoMatch [20]) benchmarks. For the rigid matching, we further evaluate our correspondences on the registration task, where RANSAC [12] is used. Details of implementation, metrics, and runtime analysis are introduced in the Appendix.",
976
+ "bbox": [
977
+ 75,
978
+ 529,
979
+ 467,
980
+ 619
981
+ ],
982
+ "page_idx": 5
983
+ },
984
+ {
985
+ "type": "text",
986
+ "text": "4.1. Rigid Indoor Scenes: 3DMatch & 3DLoMatch",
987
+ "text_level": 1,
988
+ "bbox": [
989
+ 76,
990
+ 633,
991
+ 467,
992
+ 648
993
+ ],
994
+ "page_idx": 5
995
+ },
996
+ {
997
+ "type": "text",
998
+ "text": "Dataset. 3DMatch [51] collects 62 indoor scenes, among which 46 are used for training, 8 for validation, and 8 for testing. We use the data processed by [19] where the 3DMatch data is spilt as 3DMatch ( $>30\\%$ overlap) and 3DLoMatch ( $10\\% \\sim 30\\%$ overlap). To evaluate robustness to arbitrary rotations, we follow [49] for creating the rotated benchmarks, where full-range rotations are individually added to the two frames of each point cloud pair.",
999
+ "bbox": [
1000
+ 75,
1001
+ 657,
1002
+ 467,
1003
+ 777
1004
+ ],
1005
+ "page_idx": 5
1006
+ },
1007
+ {
1008
+ "type": "text",
1009
+ "text": "Metrics. We follow [19] to use three metrics for evaluation: (1). Inlier Ratio (IR) that computes the ratio of putative correspondences whose residual distance is smaller than a threshold (i.e., $0.1\\mathrm{m}$ ) under the ground-truth transformation; (2). Feature Matching Recall (FMR) that calculates the fraction of point cloud pairs whose IR is larger than a threshold (i.e., $5\\%$ ); (3). Registration Recall (RR) that counts the fraction of point cloud pairs that are cor",
1010
+ "bbox": [
1011
+ 75,
1012
+ 779,
1013
+ 467,
1014
+ 900
1015
+ ],
1016
+ "page_idx": 5
1017
+ },
1018
+ {
1019
+ "type": "table",
1020
+ "img_path": "images/0cb444ff44e69bfba329c78511ad9a59e2054d53c905eacf0c3487fa87501608.jpg",
1021
+ "table_caption": [],
1022
+ "table_footnote": [],
1023
+ "table_body": "<table><tr><td rowspan=\"2\"># Samples=5,000</td><td colspan=\"2\">3DMatch</td><td colspan=\"2\">3DLoMatch</td></tr><tr><td>Origin</td><td>Rotated</td><td>Origin</td><td>Rotated</td></tr><tr><td colspan=\"5\">Feature Matching Recall (%) ↑</td></tr><tr><td>SpinNet [1]</td><td>97.4</td><td>97.4</td><td>75.5</td><td>75.2</td></tr><tr><td>Predator [19]</td><td>96.6</td><td>96.2</td><td>78.6</td><td>73.7</td></tr><tr><td>CoFiNet [50]</td><td>98.1</td><td>97.4</td><td>83.1</td><td>78.6</td></tr><tr><td>YOHO [44]</td><td>98.2</td><td>97.8</td><td>79.4</td><td>77.8</td></tr><tr><td>RIGA [49]</td><td>97.9</td><td>98.2</td><td>85.1</td><td>84.5</td></tr><tr><td>Lepard [20]</td><td>98.0</td><td>97.4</td><td>83.1</td><td>79.5</td></tr><tr><td>GeoTrans [27]</td><td>97.9</td><td>97.8</td><td>88.3</td><td>85.8</td></tr><tr><td>RoITr (Ours)</td><td>98.0</td><td>98.2</td><td>89.6</td><td>89.4</td></tr><tr><td colspan=\"5\">Inlier Ratio (%) ↑</td></tr><tr><td>SpinNet [1]</td><td>48.5</td><td>48.7</td><td>25.7</td><td>25.7</td></tr><tr><td>Predator [19]</td><td>58.0</td><td>52.8</td><td>26.7</td><td>22.4</td></tr><tr><td>CoFiNet [50]</td><td>49.8</td><td>46.8</td><td>24.4</td><td>21.5</td></tr><tr><td>YOHO [44]</td><td>64.4</td><td>64.1</td><td>25.9</td><td>23.2</td></tr><tr><td>RIGA [49]</td><td>68.4</td><td>68.5</td><td>32.1</td><td>32.1</td></tr><tr><td>Lepard [20]</td><td>58.6</td><td>53.7</td><td>28.4</td><td>24.4</td></tr><tr><td>GeoTrans [27]</td><td>71.9</td><td>68.2</td><td>43.5</td><td>40.0</td></tr><tr><td>RoITr (Ours)</td><td>82.6</td><td>82.3</td><td>54.3</td><td>53.2</td></tr><tr><td colspan=\"5\">Registration Recall (%) ↑</td></tr><tr><td>SpinNet [1]</td><td>88.8</td><td>93.2</td><td>58.2</td><td>61.8</td></tr><tr><td>Predator [19]</td><td>89.0</td><td>92.0</td><td>59.8</td><td>58.6</td></tr><tr><td>CoFiNet [50]</td><td>89.3</td><td>92.0</td><td>67.5</td><td>62.5</td></tr><tr><td>YOHO [44]</td><td>90.8</td><td>92.5</td><td>65.2</td><td>66.8</td></tr><tr><td>RIGA [49]</td><td>89.3</td><td>93.0</td><td>65.1</td><td>66.9</td></tr><tr><td>Lepard [20]</td><td>92.7</td><td>84.9</td><td>65.4</td><td>49.0</td></tr><tr><td>GeoTrans [27]</td><td>92.0</td><td>92.0</td><td>75.0</td><td>71.8</td></tr><tr><td>RoITr (Ours)</td><td>91.9</td><td>94.7</td><td>74.8</td><td>77.2</td></tr></table>",
1024
+ "bbox": [
1025
+ 504,
1026
+ 89,
1027
+ 893,
1028
+ 470
1029
+ ],
1030
+ "page_idx": 5
1031
+ },
1032
+ {
1033
+ "type": "text",
1034
+ "text": "Table 1. Quantitative results on (Rotated) 3DMatch & 3DLoMatch. 5,000 points/correspondences are used for the evaluation.",
1035
+ "bbox": [
1036
+ 500,
1037
+ 474,
1038
+ 890,
1039
+ 502
1040
+ ],
1041
+ "page_idx": 5
1042
+ },
1043
+ {
1044
+ "type": "text",
1045
+ "text": "rectly registered (i.e., with $\\mathrm{RMSE} < 0.2\\mathrm{m}$ ).",
1046
+ "bbox": [
1047
+ 500,
1048
+ 508,
1049
+ 799,
1050
+ 523
1051
+ ],
1052
+ "page_idx": 5
1053
+ },
1054
+ {
1055
+ "type": "text",
1056
+ "text": "Comparison with the State-of-the-Art. We compare RoITr with 7 state-of-the-art methods, among which Predator [19], CoFiNet [50], Lepard [20], and GeoTrans [27] are rotation-sensitive models, while SpinNet [1], YOHO [44], and RIGA [49] guarantee the rotation invariance by design. In Tab. 1 we demonstrate the matching and registration results on 3DMatch and 3DLoMatch, as well as on their rotated versions, with 5,000 sampled points/correspondences. Regarding IR, RoITr outperforms all the others by a large margin on both datasets, which indicates our method matches points more correctly. For FMR, we significantly surpass all the others on 3DLoMatch, while staying on par with CoFiNet and YOHO on 3DMatch, which indicates that our model is good at coping with hard cases, i.e., we find at least $5\\%$ inliers on more test data. For the registration evaluation in terms of RR, RoITr achieves comparable performance with GeoTrans and Lepard on 3DMatch, but leads the board together with GeoTrans on 3DLoMatch with an overwhelming advantage over the others. Our stability against additional rotations is further demonstrated on the rotated data, where we outperform all the others with a substantial margin. Qualitative results can be found in Fig. 6.",
1057
+ "bbox": [
1058
+ 496,
1059
+ 526,
1060
+ 890,
1061
+ 859
1062
+ ],
1063
+ "page_idx": 5
1064
+ },
1065
+ {
1066
+ "type": "page_footnote",
1067
+ "text": "2We follow [49] to calculate the RR strictly with RMSE $< 0.2\\mathrm{m}$ on the rotated data, which is slightly different from the RR on the original data.",
1068
+ "bbox": [
1069
+ 500,
1070
+ 875,
1071
+ 890,
1072
+ 900
1073
+ ],
1074
+ "page_idx": 5
1075
+ },
1076
+ {
1077
+ "type": "page_number",
1078
+ "text": "5389",
1079
+ "bbox": [
1080
+ 482,
1081
+ 944,
1082
+ 514,
1083
+ 955
1084
+ ],
1085
+ "page_idx": 5
1086
+ },
1087
+ {
1088
+ "type": "image",
1089
+ "img_path": "images/ac23312a02beb4a475dda47e424d95133ebb2541ce2ad59f1a91cd2ede8584a7.jpg",
1090
+ "image_caption": [
1091
+ "Figure 6. Qualitative results on 3DLoMatch. GeoTrans [27] is used as the baseline. Columns (b) and (c) show the correspondences, while columns (d) and (e) demonstrate the registration results. Green/red lines indicate inliers/outliers. More cases are shown in the Appendix."
1092
+ ],
1093
+ "image_footnote": [],
1094
+ "bbox": [
1095
+ 83,
1096
+ 90,
1097
+ 877,
1098
+ 296
1099
+ ],
1100
+ "page_idx": 6
1101
+ },
1102
+ {
1103
+ "type": "table",
1104
+ "img_path": "images/1e8eb32eb063570480f6d3285b0a77b8f012cf5924f07754a7dfde9c894e5a77.jpg",
1105
+ "table_caption": [],
1106
+ "table_footnote": [],
1107
+ "table_body": "<table><tr><td rowspan=\"2\"># Samples</td><td colspan=\"5\">3DMatch</td><td colspan=\"5\">3DLoMatch</td></tr><tr><td>5000</td><td>2500</td><td>1000</td><td>500</td><td>250</td><td>5000</td><td>2500</td><td>1000</td><td>500</td><td>250</td></tr><tr><td colspan=\"11\">Feature Matching Recall (%)↑</td></tr><tr><td>SpinNet [1]</td><td>97.4</td><td>97.0</td><td>96.4</td><td>96.7</td><td>94.8</td><td>75.5</td><td>75.1</td><td>74.2</td><td>69.0</td><td>62.7</td></tr><tr><td>Predator [19]</td><td>96.6</td><td>96.6</td><td>96.5</td><td>96.3</td><td>96.5</td><td>78.6</td><td>77.4</td><td>76.3</td><td>75.7</td><td>75.3</td></tr><tr><td>CoFiNet [50]</td><td>98.1</td><td>98.3</td><td>98.1</td><td>98.2</td><td>98.3</td><td>83.1</td><td>83.5</td><td>83.3</td><td>83.1</td><td>82.6</td></tr><tr><td>YOHO [44]</td><td>98.2</td><td>97.6</td><td>97.5</td><td>97.7</td><td>96.0</td><td>79.4</td><td>78.1</td><td>76.3</td><td>73.8</td><td>69.1</td></tr><tr><td>RIGA [49]</td><td>97.9</td><td>97.8</td><td>97.7</td><td>97.7</td><td>97.6</td><td>85.1</td><td>85.0</td><td>85.1</td><td>84.3</td><td>85.1</td></tr><tr><td>GeoTrans [27]</td><td>97.9</td><td>97.9</td><td>97.9</td><td>97.9</td><td>97.6</td><td>88.3</td><td>88.6</td><td>88.8</td><td>88.6</td><td>88.3</td></tr><tr><td>RoITr (Ours)</td><td>98.0</td><td>98.0</td><td>97.9</td><td>98.0</td><td>97.9</td><td>89.6</td><td>89.6</td><td>89.5</td><td>89.4</td><td>89.3</td></tr><tr><td colspan=\"11\">Inlier Ratio (%)↑</td></tr><tr><td>SpinNet [1]</td><td>48.5</td><td>46.2</td><td>40.8</td><td>35.1</td><td>29.0</td><td>25.7</td><td>23.7</td><td>20.6</td><td>18.2</td><td>13.1</td></tr><tr><td>Predator [19]</td><td>58.0</td><td>58.4</td><td>57.1</td><td>54.1</td><td>49.3</td><td>26.7</td><td>28.1</td><td>28.3</td><td>27.5</td><td>25.8</td></tr><tr><td>CoFiNet [50]</td><td>49.8</td><td>51.2</td><td>51.9</td><td>52.2</td><td>52.2</td><td>24.4</td><td>25.9</td><td>26.7</td><td>26.8</td><td>26.9</td></tr><tr><td>YOHO [44]</td><td>64.4</td><td>60.7</td><td>55.7</td><td>46.4</td><td>41.2</td><td>25.9</td><td>23.3</td><td>22.6</td><td>18.2</td><td>15.0</td></tr><tr><td>RIGA [49]</td><td>68.4</td><td>69.7</td><td>70.6</td><td>70.9</td><td>71.0</td><td>32.1</td><td>33.4</td><td>34.3</td><td>34.5</td><td>34.6</td></tr><tr><td>GeoTrans [27]</td><td>71.9</td><td>75.2</td><td>76.0</td><td>82.2</td><td>85.1</td><td>43.5</td><td>45.3</td><td>46.2</td><td>52.9</td><td>57.7</td></tr><tr><td>RoITr (Ours)</td><td>82.6</td><td>82.8</td><td>83.0</td><td>83.0</td><td>83.0</td><td>54.3</td><td>54.6</td><td>55.1</td><td>55.2</td><td>55.3</td></tr><tr><td colspan=\"11\">Registration Recall (%)↑</td></tr><tr><td>SpinNet [1]</td><td>88.8</td><td>88.0</td><td>84.5</td><td>79.0</td><td>69.2</td><td>58.2</td><td>56.7</td><td>49.8</td><td>41.0</td><td>26.7</td></tr><tr><td>Predator [19]</td><td>89.0</td><td>89.9</td><td>90.6</td><td>88.5</td><td>86.6</td><td>59.8</td><td>61.2</td><td>62.4</td><td>60.8</td><td>58.1</td></tr><tr><td>CoFiNet [50]</td><td>89.3</td><td>88.9</td><td>88.4</td><td>87.4</td><td>87.0</td><td>67.5</td><td>66.2</td><td>64.2</td><td>63.1</td><td>61.0</td></tr><tr><td>YOHO [44]</td><td>90.8</td><td>90.3</td><td>89.1</td><td>88.6</td><td>84.5</td><td>65.2</td><td>65.5</td><td>63.2</td><td>56.5</td><td>48.0</td></tr><tr><td>RIGA [49]</td><td>89.3</td><td>88.4</td><td>89.1</td><td>89.0</td><td>87.7</td><td>65.1</td><td>64.7</td><td>64.5</td><td>64.1</td><td>61.8</td></tr><tr><td>GeoTrans [27]</td><td>92.0</td><td>91.8</td><td>91.8</td><td>91.4</td><td>91.2</td><td>75.0</td><td>74.8</td><td>74.2</td><td>74.1</td><td>73.5</td></tr><tr><td>RoITr (Ours)</td><td>91.9</td><td>91.7</td><td>91.8</td><td>91.4</td><td>91.0</td><td>74.7</td><td>74.8</td><td>74.8</td><td>74.2</td><td>73.6</td></tr></table>",
1108
+ "bbox": [
1109
+ 81,
1110
+ 333,
1111
+ 472,
1112
+ 599
1113
+ ],
1114
+ "page_idx": 6
1115
+ },
1116
+ {
1117
+ "type": "text",
1118
+ "text": "Analysis on the Number of Correspondences. We further analyze the influence of a varying number of correspondences. As illustrated in Tab. 2, RoITr shows outstanding performance on both datasets with various correspondences, proving its stability when only a few correspondences are accessible. The same test on the rotated benchmarks is given in the Appendix.",
1119
+ "bbox": [
1120
+ 75,
1121
+ 654,
1122
+ 468,
1123
+ 760
1124
+ ],
1125
+ "page_idx": 6
1126
+ },
1127
+ {
1128
+ "type": "text",
1129
+ "text": "4.2. Deformable Objects: 4DMatch & 4DLoMatch",
1130
+ "text_level": 1,
1131
+ "bbox": [
1132
+ 76,
1133
+ 770,
1134
+ 467,
1135
+ 787
1136
+ ],
1137
+ "page_idx": 6
1138
+ },
1139
+ {
1140
+ "type": "text",
1141
+ "text": "Dataset. 4DMatch [20] contains 1,761 animations randomly selected from DeformingThings4D [21]. The 1,761 sequences are divided into 1,232/176/353 as train/val/test, where the test set is further split into 4DMatch and 4DLoMatch based on an overlap ratio threshold of $45\\%$ . Metrics. We follow [20] to use two different metrics: (1). Inlier Ratio (IR) which is defined as same as the IR on 3DMatch,",
1142
+ "bbox": [
1143
+ 75,
1144
+ 794,
1145
+ 470,
1146
+ 900
1147
+ ],
1148
+ "page_idx": 6
1149
+ },
1150
+ {
1151
+ "type": "table",
1152
+ "img_path": "images/9236b432a4eb3b18d0f396703602995a38897a205f43521d78cd042deb473363.jpg",
1153
+ "table_caption": [
1154
+ "Table 2. Quantitative results on 3DMatch & 3DLoMatch with a varying number of points/correspondences. See the results on rotated data in the Appendix."
1155
+ ],
1156
+ "table_footnote": [],
1157
+ "table_body": "<table><tr><td rowspan=\"2\">Category</td><td rowspan=\"2\">Method</td><td colspan=\"2\">4DMatch</td><td colspan=\"2\">4DLoMatch</td></tr><tr><td>NFMR(%)↑</td><td>IR(%)↑</td><td>NFMR(%)↑</td><td>IR(%)↑</td></tr><tr><td rowspan=\"2\">Scene Flow</td><td>PWC [45]</td><td>21.6</td><td>20.0</td><td>10.0</td><td>7.2</td></tr><tr><td>FLOT [24]</td><td>27.1</td><td>24.9</td><td>15.2</td><td>10.7</td></tr><tr><td rowspan=\"4\">Feature Matching</td><td>Predator [19]</td><td>56.4</td><td>60.4</td><td>32.1</td><td>27.5</td></tr><tr><td>GeoTrans [27]</td><td>83.2</td><td>82.2</td><td>65.4</td><td>63.6</td></tr><tr><td>Lepard [20]</td><td>83.7</td><td>82.7</td><td>66.9</td><td>55.7</td></tr><tr><td>RoITr (Ours)</td><td>83.0</td><td>84.4</td><td>69.4</td><td>67.6</td></tr></table>",
1158
+ "bbox": [
1159
+ 503,
1160
+ 333,
1161
+ 893,
1162
+ 421
1163
+ ],
1164
+ "page_idx": 6
1165
+ },
1166
+ {
1167
+ "type": "text",
1168
+ "text": "Table 3. Quantitative results on 4DMatch & 4DLoMatch.",
1169
+ "bbox": [
1170
+ 524,
1171
+ 426,
1172
+ 864,
1173
+ 439
1174
+ ],
1175
+ "page_idx": 6
1176
+ },
1177
+ {
1178
+ "type": "text",
1179
+ "text": "but with a different threshold (i.e., $0.04\\mathrm{m}$ ); (2). Non-rigid Feature Matching Recall (NFRM) that measures the fraction of ground-truth matches that can be successfully recovered by the putative correspondences.",
1180
+ "bbox": [
1181
+ 496,
1182
+ 460,
1183
+ 892,
1184
+ 520
1185
+ ],
1186
+ "page_idx": 6
1187
+ },
1188
+ {
1189
+ "type": "text",
1190
+ "text": "Comparison with the State-of-the-Art. We compare RoITr with 5 baselines, among which PwC [45] and FLOT [24] are scene flow-based methods, while Predator [19], Lepard [20], GeoTrans [27] are based on feature matching. The results shown in Tab. 3 indicate that although our rotation-invariance is mainly designed for rigid scenarios, RoITr could also achieve outstanding performance in the non-rigid matching task, which further confirms the superiority of our model design. Qualitative results are demonstrated in Fig. 7.",
1191
+ "bbox": [
1192
+ 496,
1193
+ 521,
1194
+ 890,
1195
+ 672
1196
+ ],
1197
+ "page_idx": 6
1198
+ },
1199
+ {
1200
+ "type": "text",
1201
+ "text": "4.3. Ablation Study",
1202
+ "text_level": 1,
1203
+ "bbox": [
1204
+ 500,
1205
+ 681,
1206
+ 653,
1207
+ 696
1208
+ ],
1209
+ "page_idx": 6
1210
+ },
1211
+ {
1212
+ "type": "text",
1213
+ "text": "Local Attention. We first replace our PPFTrans with PointTransformer (PT) [53] in Tab. 4 (a.1), which leads to a sharp performance drop. We then ablate by embedding our PPF-based local coordinates into PT (Tab. 4 (a.2)) and by adopting the relative coordinates, i.e., $\\mathbf{p}_j - \\mathbf{p}_j$ , used by PT in our PAM (Tab. 4 (a.3)). Our local coordinate representation significantly boosts the performance of PT in the task of point cloud matching and meanwhile makes it rotation-invariant, although its performance is still far behind ours. However, the relative coordinates fail to work in our PAM, as we adopt a more efficient attention mechanism [43] that learns a scalar attention value for each feature $\\mathbf{x} \\in \\mathbb{R}^c$ and is consequently hard to work under varying poses with a",
1214
+ "bbox": [
1215
+ 496,
1216
+ 703,
1217
+ 892,
1218
+ 900
1219
+ ],
1220
+ "page_idx": 6
1221
+ },
1222
+ {
1223
+ "type": "page_number",
1224
+ "text": "5390",
1225
+ "bbox": [
1226
+ 482,
1227
+ 944,
1228
+ 514,
1229
+ 955
1230
+ ],
1231
+ "page_idx": 6
1232
+ },
1233
+ {
1234
+ "type": "image",
1235
+ "img_path": "images/05f21fec257031ee45502721928229e415a594471edac0c8ac5a61e3a223fa87.jpg",
1236
+ "image_caption": [
1237
+ "Figure 7. Qualitative results of non-rigid matching on 4DLoMatch with Lepard [20] as the baseline. Green/red lines indicate inliers/outliers. See the Appendix for more examples."
1238
+ ],
1239
+ "image_footnote": [],
1240
+ "bbox": [
1241
+ 83,
1242
+ 88,
1243
+ 468,
1244
+ 281
1245
+ ],
1246
+ "page_idx": 7
1247
+ },
1248
+ {
1249
+ "type": "text",
1250
+ "text": "rotation-sensitive design. As a comparison, PT learns a per-channel vector attention $\\mathbf{a} \\in \\mathbb{R}^c$ for the same feature $\\mathbf{x}$ and could deal with the pose variations, but at the cost of the efficiency as shown in Fig. 8. When the number of channels is increased, our advantage in terms of efficiency is enlarged. As we achieve that with more parameters, the gap becomes more significant when runtime is normalized with the number of parameters in the right figure. With our PPF-based local coordinate, the scalar attention could focus on the pose-agnostic pure geometry and therefore achieves the best performance shown in Tab. 4 (a.4).",
1251
+ "bbox": [
1252
+ 75,
1253
+ 335,
1254
+ 468,
1255
+ 500
1256
+ ],
1257
+ "page_idx": 7
1258
+ },
1259
+ {
1260
+ "type": "text",
1261
+ "text": "Abstraction Layer. We ablate our Attentional Abstraction Layer (AAL) by replacing it with the pooling-based abstraction design used in [25, 26, 53]. We test the max pooling in Tab. 4 (b.1) and the average pooling in Tab. 4 (b.2), both showing a degrading performance compared with our AAL, which demonstrates our superiority.",
1262
+ "bbox": [
1263
+ 75,
1264
+ 503,
1265
+ 468,
1266
+ 594
1267
+ ],
1268
+ "page_idx": 7
1269
+ },
1270
+ {
1271
+ "type": "text",
1272
+ "text": "Backbone. In Tab. 4 (a.1) we have shown our superiority compared with PT [53]. We further replace our PPFTrans with the KPCnv-based backbone network which is widely used in previous deep matchers [19, 27, 50]. The fact that KPCnv falls behind our design demonstrates the advantage of PPFTrans in geometry encoding.",
1273
+ "bbox": [
1274
+ 75,
1275
+ 595,
1276
+ 468,
1277
+ 686
1278
+ ],
1279
+ "page_idx": 7
1280
+ },
1281
+ {
1282
+ "type": "text",
1283
+ "text": "Global Transformer. We replace our design with the global transformer of GeoTrans [27] which performs state-of-the-art but without the cross-frame spatial awareness. The dropping results in Tab. 4 (d.1) proves the excellence of our design with the cross-frame position awareness.",
1284
+ "bbox": [
1285
+ 75,
1286
+ 686,
1287
+ 468,
1288
+ 762
1289
+ ],
1290
+ "page_idx": 7
1291
+ },
1292
+ {
1293
+ "type": "text",
1294
+ "text": "The Number of Global Transformers. To demonstrate the importance of being globally aware, we first remove the global transformer. The substantial performance drop confirms the significance of global awareness. Then we add one global transformer and observe an increased performance. In our default setting with 3 global transformers, the model performs the best. However, when the number is increased to 5, the model shows a slight performance drop, which we owe to overfitting. As the data augmentation of rotations",
1295
+ "bbox": [
1296
+ 75,
1297
+ 763,
1298
+ 470,
1299
+ 901
1300
+ ],
1301
+ "page_idx": 7
1302
+ },
1303
+ {
1304
+ "type": "image",
1305
+ "img_path": "images/c79f5972814873dd0680a193dcb9720a19fc8606439b7e3d540e215126e359a0.jpg",
1306
+ "image_caption": [
1307
+ "Figure 8. Left: Runtime comparison between our PPF attention Mechanism (PAM) and the local attention in PointTransformer [53]. Right: Runtime normalized by aligning the number of parameters."
1308
+ ],
1309
+ "image_footnote": [],
1310
+ "bbox": [
1311
+ 501,
1312
+ 88,
1313
+ 877,
1314
+ 247
1315
+ ],
1316
+ "page_idx": 7
1317
+ },
1318
+ {
1319
+ "type": "table",
1320
+ "img_path": "images/24fe69e841b131b2c4d375ae270ff0ab9af5012c9189f0f4fcbbbc40a0140dca.jpg",
1321
+ "table_caption": [],
1322
+ "table_footnote": [],
1323
+ "table_body": "<table><tr><td rowspan=\"2\">Category</td><td rowspan=\"2\">Model</td><td colspan=\"3\">Origin</td><td colspan=\"3\">Rotated</td></tr><tr><td>FMR</td><td>IR</td><td>RR</td><td>FMR</td><td>IR</td><td>RR</td></tr><tr><td rowspan=\"4\">a. Local</td><td>1. PT [53]</td><td>79.0</td><td>36.5</td><td>61.6</td><td>76.5</td><td>34.7</td><td>60.0</td></tr><tr><td>*2. PPF+PT [53]</td><td>87.0</td><td>49.9</td><td>69.9</td><td>86.8</td><td>49.4</td><td>71.2</td></tr><tr><td>3. Δxyz+Ours</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td></tr><tr><td>*4. Ours</td><td>89.6</td><td>54.3</td><td>74.7</td><td>89.4</td><td>53.2</td><td>77.2</td></tr><tr><td rowspan=\"3\">b. Aggregation</td><td>*1. max pooling</td><td>85.2</td><td>50.1</td><td>70.5</td><td>85.4</td><td>50.2</td><td>71.9</td></tr><tr><td>*2. avg pooling</td><td>87.8</td><td>52.6</td><td>73.8</td><td>87.2</td><td>52.5</td><td>74.7</td></tr><tr><td>*3. Ours</td><td>89.6</td><td>54.3</td><td>74.7</td><td>89.4</td><td>53.2</td><td>77.2</td></tr><tr><td rowspan=\"2\">c. Backbone</td><td>1. KPCov [40]</td><td>85.2</td><td>44.4</td><td>70.6</td><td>83.0</td><td>42.3</td><td>71.5</td></tr><tr><td>*2. Ours</td><td>89.6</td><td>54.3</td><td>74.7</td><td>89.4</td><td>53.2</td><td>77.2</td></tr><tr><td rowspan=\"2\">d. Global</td><td>*1. GeoTrans [27]</td><td>87.7</td><td>53.6</td><td>73.0</td><td>87.5</td><td>53.2</td><td>75.1</td></tr><tr><td>*2. Ours</td><td>89.6</td><td>54.3</td><td>74.7</td><td>89.4</td><td>53.2</td><td>77.2</td></tr><tr><td rowspan=\"4\">e. #Global</td><td>*1. g = 0</td><td>87.2</td><td>37.6</td><td>70.7</td><td>87.5</td><td>37.6</td><td>72.7</td></tr><tr><td>*2. g = 1</td><td>87.1</td><td>42.1</td><td>70.8</td><td>86.8</td><td>42.1</td><td>73.0</td></tr><tr><td>*3. g = 3 (Ours)</td><td>89.6</td><td>54.3</td><td>74.7</td><td>89.4</td><td>53.2</td><td>77.2</td></tr><tr><td>*4. g = 5</td><td>87.1</td><td>52.5</td><td>72.1</td><td>87.0</td><td>52.4</td><td>73.3</td></tr></table>",
1324
+ "bbox": [
1325
+ 503,
1326
+ 309,
1327
+ 893,
1328
+ 484
1329
+ ],
1330
+ "page_idx": 7
1331
+ },
1332
+ {
1333
+ "type": "text",
1334
+ "text": "Table 4. Ablation study on (rotated) 3DLoMatch. 5,000 points/correspondences are leveraged. * indicates the methods with intrinsic rotation invariance. See the Appendix for the results on (rotated) 3DMatch.",
1335
+ "bbox": [
1336
+ 496,
1337
+ 489,
1338
+ 893,
1339
+ 546
1340
+ ],
1341
+ "page_idx": 7
1342
+ },
1343
+ {
1344
+ "type": "text",
1345
+ "text": "has less effect on an intrinsically rotation-invariant method, more data is required for training a larger model.",
1346
+ "bbox": [
1347
+ 496,
1348
+ 551,
1349
+ 890,
1350
+ 583
1351
+ ],
1352
+ "page_idx": 7
1353
+ },
1354
+ {
1355
+ "type": "text",
1356
+ "text": "5. Conclusion",
1357
+ "text_level": 1,
1358
+ "bbox": [
1359
+ 500,
1360
+ 588,
1361
+ 619,
1362
+ 604
1363
+ ],
1364
+ "page_idx": 7
1365
+ },
1366
+ {
1367
+ "type": "text",
1368
+ "text": "We introduced RoITr - an intrinsically rotation-invariant model for point cloud matching. We proposed PAM (PPF Attention Mechanism) that embeds PPF-based local coordinates to encode rotation-invariant geometry. This design lies at the core of AAL (Attention Abstraction Layer), PAL (PPF Attention Layer), and TUL (Transition Up Layer) which are consecutively stacked to compose PPFTrans (PPF Transformer) for representative and pose-agnostic geometry description. We further enhanced features by introducing a novel global transformer architecture, which ensures the rotation-invariant cross-frame spatial awareness. Extensive experiments are conducted on both rigid and non-rigid benchmarks to demonstrate the superiority of our approach, especially the remarkable robustness against arbitrary rotations. Limitations are discussed in the Appendix.",
1369
+ "bbox": [
1370
+ 496,
1371
+ 613,
1372
+ 890,
1373
+ 839
1374
+ ],
1375
+ "page_idx": 7
1376
+ },
1377
+ {
1378
+ "type": "text",
1379
+ "text": "Acknowledgment. This paper is supported by the National Natural Science Foundation of China under Grant No. 62025208. We appreciate the help from Lennart Bastian, Mert Karaoglu, Ning Liu, and Zhiying Leng.",
1380
+ "bbox": [
1381
+ 496,
1382
+ 840,
1383
+ 890,
1384
+ 901
1385
+ ],
1386
+ "page_idx": 7
1387
+ },
1388
+ {
1389
+ "type": "page_number",
1390
+ "text": "5391",
1391
+ "bbox": [
1392
+ 482,
1393
+ 944,
1394
+ 513,
1395
+ 957
1396
+ ],
1397
+ "page_idx": 7
1398
+ },
1399
+ {
1400
+ "type": "text",
1401
+ "text": "References",
1402
+ "text_level": 1,
1403
+ "bbox": [
1404
+ 78,
1405
+ 89,
1406
+ 174,
1407
+ 104
1408
+ ],
1409
+ "page_idx": 8
1410
+ },
1411
+ {
1412
+ "type": "list",
1413
+ "sub_type": "ref_text",
1414
+ "list_items": [
1415
+ "[1] Sheng Ao, Qingyong Hu, Bo Yang, Andrew Markham, and Yulan Guo. Spinnet: Learning a general surface descriptor for 3d point cloud registration. In CVPR, 2021. 1, 2, 6, 7",
1416
+ "[2] Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hinton. Layer normalization. arXiv preprint arXiv:1607.06450, 2016. 4",
1417
+ "[3] Xuyang Bai, Zixin Luo, Lei Zhou, Hongbo Fu, Long Quan, and Chiew-Lan Tai. D3feat: Joint learning of dense detection and description of 3d local features. In CVPR, 2020. 2",
1418
+ "[4] Axel Barroso-Laguna, Yannick Verdie, Benjamin Busam, and Krystian Mikolajczyk. Hdd-net: Hybrid detector descriptor with mutual interactive learning. In Proceedings of the Asian Conference on Computer Vision, 2020. 2",
1419
+ "[5] Tolga Birdal and Slobodan Ilic. Point pair features based object detection and pose estimation revisited. In 3DV, 2015. 4",
1420
+ "[6] Christopher Choy, JunYoung Gwak, and Silvio Savarese. 4d spatio-temporal convnets: Minkowski convolutional neural networks. In CVPR, 2019. 1",
1421
+ "[7] Christopher Choy, Jaesik Park, and Vladlen Koltun. Fully convolutional geometric features. In ICCV, 2019. 1, 2",
1422
+ "[8] Chin Seng Chua and Ray Jarvis. Point signatures: A new representation for 3d object recognition. IJCV, 1997. 2",
1423
+ "[9] Haowen Deng, Tolga Birdal, and Slobodan Ilic. Ppf-foldnet: Unsupervised learning of rotation invariant 3d local descriptors. In ECCV, 2018. 1, 2, 4",
1424
+ "[10] Haowen Deng, Tolga Birdal, and Slobodan Ilic. Ppfnet: Global context aware local features for robust 3d point matching. In CVPR, 2018. 1, 2",
1425
+ "[11] Bertram Drost, Markus Ulrich, Nassir Navab, and Slobodan Ilic. Model globally, match locally: Efficient and robust 3d object recognition. In CVPR, 2010. 1, 2, 3, 4",
1426
+ "[12] Martin A Fischler and Robert C Bolles. Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. Communications of the ACM, 1981. 6",
1427
+ "[13] Zan Gojcic, Caifa Zhou, Jan D Wegner, and Andreas Wieser. The perfect match: 3d point cloud matching with smoothed densities. In CVPR, 2019. 1, 2",
1428
+ "[14] Yulan Guo, Ferdous Sohel, Mohammed Bennamoun, Min Lu, and Jianwei Wan. Rotational projection statistics for 3d local surface description and object recognition. IJCV, 2013. 2",
1429
+ "[15] Ji Hou, Benjamin Graham, Matthias Nießner, and Saining Xie. Exploring data-efficient 3d scene understanding with contrastive scene contexts. In CVPR, 2021. 1",
1430
+ "[16] Ji Hou, Saining Xie, Benjamin Graham, Angela Dai, and Matthias Nießner. Pri3d: Can 3d priors help 2d representation learning? In ICCV, 2021. 1",
1431
+ "[17] Chun-Hao Huang, Benjamin Allain, Jean-Sébastien Franco, Nassir Navab, Slobodan Ilic, and Edmond Boyer. Volumetric 3d tracking by detection. In CVPR, 2016. 1",
1432
+ "[18] Chun-Hao Paul Huang, Benjamin Allain, Edmond Boyer, Jean-Sébastien Franco, Federico Tombari, Nassir Navab, and Slobodan Ilic. Tracking-by-detection of 3d human shapes: from surfaces to volumes. TPAMI, 2017. 1"
1433
+ ],
1434
+ "bbox": [
1435
+ 78,
1436
+ 114,
1437
+ 470,
1438
+ 898
1439
+ ],
1440
+ "page_idx": 8
1441
+ },
1442
+ {
1443
+ "type": "list",
1444
+ "sub_type": "ref_text",
1445
+ "list_items": [
1446
+ "[19] Shengyu Huang, Zan Gojcic, Mikhail Usvyatsov, Andreas Wieser, and Konrad Schindler. Predator: Registration of 3d point clouds with low overlap. In CVPR, 2021. 1, 2, 5, 6, 7, 8",
1447
+ "[20] Yang Li and Tatsuya Harada. Lepard: Learning partial point cloud matching in rigid and deformable scenes. In CVPR, 2022. 1, 2, 6, 7, 8",
1448
+ "[21] Yang Li, Hikari Takehara, Takafumi Taketomi, Bo Zheng, and Matthias Nießner. 4dcomplete: Non-rigid motion estimation beyond the observable surface. In ICCV, 2021. 7",
1449
+ "[22] Richard A Newcombe, Dieter Fox, and Steven M Seitz. Dynamicfusion: Reconstruction and tracking of non-rigid scenes in real-time. In CVPR, 2015. 1",
1450
+ "[23] Richard A Newcombe, Shahram Izadi, Otmar Hilliges, David Molyneaux, David Kim, Andrew J Davison, Pushmeet Kohi, Jamie Shotton, Steve Hodges, and Andrew Fitzgibbon. Kinectfusion: Real-time dense surface mapping and tracking. In IEEE International Symposium on Mixed and Augmented Reality, 2011. 1",
1451
+ "[24] Gilles Puy, Alexandre Boulch, and Renaud Marlet. Flot: Scene flow on point clouds guided by optimal transport. In ECCV, 2020. 7",
1452
+ "[25] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In CVPR, 2017. 1, 2, 8",
1453
+ "[26] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. NeurIPS, 2017. 1, 4, 8",
1454
+ "[27] Zheng Qin, Hao Yu, Changjian Wang, Yulan Guo, Yuxing Peng, and Kai Xu. Geometric transformer for fast and robust point cloud registration. In CVPR, 2022. 1, 2, 3, 5, 6, 7, 8",
1455
+ "[28] Zheng Qin, Hao Yu, Changjian Wang, Yuxing Peng, and Kai Xu. Deep graph-based spatial consistency for robust non-rigid point cloud registration. arXiv preprint arXiv:2303.09950, 2023. 2",
1456
+ "[29] Ignacio Rocco, Mircea Cimpoi, Relja Arandjelović, Akihiko Torii, Tomas Pajdla, and Josef Sivic. Neighbourhood consensus networks. *NeurIPS*, 2018. 6",
1457
+ "[30] Radu Bogdan Rusu, Nico Blodow, and Michael Beetz. Fast point feature histograms (fpfh) for 3d registration. In ICRA, 2009. 1, 2",
1458
+ "[31] Radu Bogdan Rusu, Nico Blodow, Zoltan Csaba Marton, and Michael Beetz. Aligning point cloud views using persistent feature histograms. In IROS, 2008. 1, 2",
1459
+ "[32] Mahdi Saleh, Shervin Dehghani, Benjamin Busam, Nassir Navab, and Federico Tombari. Graphite: Graph-induced feature extraction for point cloud registration. In 3DV, 2020. 1, 2",
1460
+ "[33] Mahdi Saleh, Yige Wang, Nassir Navab, Benjamin Busam, and Federico Tombari. Cloudattention: Efficient multi-scale attention scheme for 3d point cloud learning. arXiv preprint arXiv:2208.00524, 2022. 1",
1461
+ "[34] Mahdi Saleh, Shun-Cheng Wu, Luca Cosmo, Nassir Navab, Benjamin Busam, and Federico Tombari. Bending graphs: Hierarchical shape matching using gated optimal transport. In CVPR, 2022. 1, 2"
1462
+ ],
1463
+ "bbox": [
1464
+ 501,
1465
+ 92,
1466
+ 892,
1467
+ 898
1468
+ ],
1469
+ "page_idx": 8
1470
+ },
1471
+ {
1472
+ "type": "page_number",
1473
+ "text": "5392",
1474
+ "bbox": [
1475
+ 482,
1476
+ 944,
1477
+ 516,
1478
+ 955
1479
+ ],
1480
+ "page_idx": 8
1481
+ },
1482
+ {
1483
+ "type": "list",
1484
+ "sub_type": "ref_text",
1485
+ "list_items": [
1486
+ "[35] Paul-Edouard Sarlin, Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superglue: Learning feature matching with graph neural networks. In CVPR, 2020. 6",
1487
+ "[36] Richard Sinkhorn and Paul Knopp. Concerning nonnegative matrices and doubly stochastic matrices. Pacific Journal of Mathematics, 1967. 6",
1488
+ "[37] Jiaming Sun, Zehong Shen, Yuang Wang, Hujun Bao, and Xiaowei Zhou. Loftr: Detector-free local feature matching with transformers. In CVPR, 2021. 6",
1489
+ "[38] Jiapeng Tang, Lev Markhasin, Bi Wang, Justus Thies, and Matthias Nießner. Neural shape deformation priors. arXiv preprint arXiv:2210.05616, 2022. 2",
1490
+ "[39] Jiapeng Tang, Dan Xu, Kui Jia, and Lei Zhang. Learning parallel dense correspondence from spatio-temporal descriptors for efficient and robust 4d reconstruction. In CVPR, 2021. 1",
1491
+ "[40] Hugues Thomas, Charles R Qi, Jean-Emmanuel Deschaud, Beatrix Marcotegui, François Goulette, and Leonidas J Guibas. Kpconv: Flexible and deformable convolution for point clouds. In ICCV, 2019. 1, 8",
1492
+ "[41] Federico Tombari, Samuele Salti, and Luigi Di Stefano. Unique signatures of histograms for local surface description. In ECCV, 2010. 1, 2",
1493
+ "[42] Giovanni Trappolini, Luca Cosmo, Luca Moschella, Ricardo Marin, Simone Melzi, and Emanuele Rodola. Shape registration in the time of transformers. NeurIPS, 2021. 2",
1494
+ "[43] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. NeurIPS, 2017. 3, 5, 7",
1495
+ "[44] Haiping Wang, Yuan Liu, Zhen Dong, and Wenping Wang. You only hypothesize once: Point cloud registration with rotation-equivariant descriptors. In ACM MM, 2022. 1, 2, 6, 7",
1496
+ "[45] Wenxuan Wu, Zhiyuan Wang, Zhuwen Li, Wei Liu, and Li Fuxin. Pointpwc-net: A coarse-to-fine network for supervised and self-supervised scene flow estimation on 3d point clouds. arXiv preprint arXiv:1911.12408, 2019. 7",
1497
+ "[46] Saining Xie, Jiatao Gu, Demi Guo, Charles R Qi, Leonidas Guibas, and Or Litany. Pointcontrast: Unsupervised pretraining for 3d point cloud understanding. In ECCV, 2020. 1",
1498
+ "[47] Yaoqing Yang, Chen Feng, Yiru Shen, and Dong Tian. Foldingnet: Point cloud auto-encoder via deep grid deformation. In CVPR, 2018. 2",
1499
+ "[48] Zi Jian Yew and Gim Hee Lee. Regtr: End-to-end point cloud correspondences with transformers. In CVPR, 2022. 1, 2, 5",
1500
+ "[49] Hao Yu, Ji Hou, Zheng Qin, Mahdi Saleh, Ivan Shugurov, Kai Wang, Benjamin Busam, and Slobodan Ilic. Riga: Rotation-invariant and globally-aware descriptors for point cloud registration. arXiv preprint arXiv:2209.13252, 2022. 1, 2, 5, 6, 7",
1501
+ "[50] Hao Yu, Fu Li, Mahdi Saleh, Benjamin Busam, and Slobodan Ilic. Cofinet: Reliable coarse-to-fine correspondences for robust pointcloud registration. NeurIPS, 2021. 1, 2, 3, 5, 6, 7, 8"
1502
+ ],
1503
+ "bbox": [
1504
+ 78,
1505
+ 90,
1506
+ 470,
1507
+ 898
1508
+ ],
1509
+ "page_idx": 9
1510
+ },
1511
+ {
1512
+ "type": "list",
1513
+ "sub_type": "ref_text",
1514
+ "list_items": [
1515
+ "[51] Andy Zeng, Shuran Song, Matthias Nießner, Matthew Fisher, Jianxiong Xiao, and Thomas Funkhouser. 3dmatch: Learning local geometric descriptors from rgb-d reconstructions. In CVPR, 2017. 1, 2, 6",
1516
+ "[52] Yu Zhang, Junle Yu, Xiaolin Huang, Wenhui Zhou, and Ji Hou. Pcr-cg: Point cloud registration via deep explicit color and geometry. In ECCV, 2022. 1",
1517
+ "[53] Hengshuang Zhao, Li Jiang, Jiaya Jia, Philip HS Torr, and Vladlen Koltun. Point transformer. In ICCV, 2021. 1, 2, 7, 8"
1518
+ ],
1519
+ "bbox": [
1520
+ 501,
1521
+ 92,
1522
+ 890,
1523
+ 219
1524
+ ],
1525
+ "page_idx": 9
1526
+ },
1527
+ {
1528
+ "type": "page_number",
1529
+ "text": "5393",
1530
+ "bbox": [
1531
+ 482,
1532
+ 945,
1533
+ 514,
1534
+ 955
1535
+ ],
1536
+ "page_idx": 9
1537
+ }
1538
+ ]
2023/Rotation-Invariant Transformer for Point Cloud Matching/424bbd8d-fe04-4142-91bc-ea8eb4c92ccb_model.json ADDED
The diff for this file is too large to render. See raw diff
 
2023/Rotation-Invariant Transformer for Point Cloud Matching/424bbd8d-fe04-4142-91bc-ea8eb4c92ccb_origin.pdf ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:54b2fc1c3b8e7234e8108c3ea833f77ab38af5ef54492f4b2756246747ba1d30
3
+ size 1245502
2023/Rotation-Invariant Transformer for Point Cloud Matching/full.md ADDED
@@ -0,0 +1,311 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Rotation-Invariant Transformer for Point Cloud Matching
2
+
3
+ Hao Yu $^{1}$ Zheng Qin $^{2}$ Ji Hou $^{3}$ Mahdi Saleh $^{1,4}$ Dongsheng Li $^{2}$ Benjamin Busam $^{1,4}$ Slobodan Ilic $^{1,5}$ TU Munich $^{2}$ NUDT $^{3}$ Meta Reality Labs $^{4}$ 3Dwe.ai $^{5}$ Siemens AG, Munich
4
+
5
+ # Abstract
6
+
7
+ The intrinsic rotation invariance lies at the core of matching point clouds with handcrafted descriptors. However, it is widely despised by recent deep matchers that obtain the rotation invariance extrinsically via data augmentation. As the finite number of augmented rotations can never span the continuous $SO(3)$ space, these methods usually show instability when facing rotations that are rarely seen. To this end, we introduce RoITr, a Rotation-Invariant Transformer to cope with the pose variations in the point cloud matching task. We contribute both on the local and global levels. Starting from the local level, we introduce an attention mechanism embedded with Point Pair Feature (PPF)-based coordinates to describe the pose-invariant geometry, upon which a novel attention-based encoder-decoder architecture is constructed. We further propose a global transformer with rotation-invariant cross-frame spatial awareness learned by the self-attention mechanism, which significantly improves the feature distinctiveness and makes the model robust with respect to the low overlap. Experiments are conducted on both the rigid and non-rigid public benchmarks, where RoITr outperforms all the state-of-the-art models by a considerable margin in the low-overlapping scenarios. Especially when the rotations are enlarged on the challenging 3DLoMatch benchmark, RoITr surpasses the existing methods by at least 13 and 5 percentage points in terms of Inlier Ratio and Registration Recall, respectively. Code is publicly available<sup>1</sup>.
8
+
9
+ # 1. Introduction
10
+
11
+ The correspondence estimation between a pair of partially-overlapping point clouds is a long-standing task that lies at the core of many computer vision applications, such as tracking [17, 18], reconstruction [22, 23, 39], pose estimation [19, 27, 50] and 3D representation learning [15, 16, 46], etc. In a typical solution, geometry is first encoded into descriptors, and correspondences are then established between two frames by matching the most similar descriptors. As the two frames are observed from different views, depicting the same geometry under different trans
12
+
13
+ ![](images/81825f8e94531760f124a36a4048f3485bea7faa274c6d8e55532d5c6f8d23a7.jpg)
14
+ Figure 1. Feature Matching Recall (FMR) on 3DLoMatch [19] and Rotated 3DLoMatch. Distance to the diagonal represents the robustness against rotations. Among all the state-of-the-art approaches, RoITr not only ranks first on both benchmarks but also shows the best robustness against the enlarged rotations.
15
+
16
+ formations identically, i.e., the pose-invariance, becomes the key to success in the point cloud matching task.
17
+
18
+ Since the side effects caused by a global translation can always be easily eliminated, e.g., by aligning the barycenter with the origin, the attention naturally shifts to coping with the rotations. In the past, handcrafted local descriptors [11, 30, 31, 41] were designed to be rotation-invariant so that the same geometry observed from different views can be correctly matched. With the emergence of deep neural models for 3D point analysis, e.g., multilayer perceptrons (MLPs)-based like PointNet [25, 26], convolutions-based like KPConv [6, 40], and the attention-based like PointTransformer [33, 53], recent approaches [1, 7, 9, 10, 13, 19, 20, 27, 32, 48-51] propose to learn descriptors from raw points as an alternative to handcrafted features that are less robust to occlusion and noise. The majority of deep point matchers [7, 10, 19, 20, 27, 34, 48, 50-52] is sensitive to rotations. Consequently, their invariance to rotations must be obtained extrinsically via augmented training to ensure that the same geometry under different poses can be depicted similarly. However, as the training cases can never span the continuous $SO(3)$ space, they always suffer from instability when facing rotations that are rarely seen during training. This can be observed by a significant performance drop under enlarged rotations at inference time. (See Fig. 1.)
19
+
20
+ There are other works [1, 9, 13, 32, 44] that only lever
21
+
22
+ age deep neural networks to encode the pure geometry with the intrinsically-designed rotation invariance. However, the intrinsic rotation invariance comes at the cost of losing global context. For example, a human's left and right halves are almost identically described, which naturally degrades the distinctiveness of features. Most recently, RIGA [49] is proposed to enhance the distinctiveness of the rotation-invariant descriptors by incorporating a global context, e.g., the left and right halves of a human become distinguishable by knowing there is a chair on the left while a table on the right. However, it lacks a highly-representative geometry encoder since it relies on PointNet [25], which accounts for an ineffective local geometry description. Moreover, as depicting the cross-frame spatial relationships is non-trivial, previous works [19, 27, 34, 50] merely leverage the contextual features in the cross-frame context aggregation, which neglects the positional information. Although RIGA proposes to learn a rotation-invariant position representation by leveraging an additional PointNet, this simple design is hard to model the complex cross-frame positional relationships and leads to less distinctive descriptors.
23
+
24
+ In this paper, we present Rotation-Invariant Transformer (RoITr) to tackle the problem of point cloud matching under arbitrary pose variations. By using Point Pair Features (PPFs) as the local coordinates, we propose an attention mechanism to learn the pure geometry regardless of the varying poses. Upon it, attention-based layers are further proposed to compose the encoder-decoder architecture for highly-discriminative and rotation-invariant geometry encoding. We demonstrate its superiority over PointTransformer [53], a state-of-the-art attention-based backbone network, in terms of both efficiency and efficacy in Fig. 8 and Tab. 4 (a), respectively. On the global level, the cross-frame position awareness is introduced in a rotation-invariant fashion to facilitate feature distinctiveness. We illustrate its significance over the state-of-the-art design [27] in Tab. 4 (d). Our main contributions are summarized as:
25
+
26
+ - An attention mechanism designed to disentangle the geometry and poses, which enables the pose-agnostic geometry description.
27
+ - An attention-based encoder-decoder architecture that learns highly-representative local geometry in a rotation-invariant fashion.
28
+ - A global transformer with rotation-invariant cross-frame position awareness that significantly enhances the feature distinctiveness.
29
+
30
+ # 2. Related Work
31
+
32
+ Models with Extrinsic Rotation Invariance. The mainstream of deep learning-based point cloud matching approaches is intrinsically rotation-sensitive. Pioneers [10,51] learn to describe local patches from a rotation-variant input.
33
+
34
+ FCGF [7] leverages fully-convolutional networks to accelerate the geometry description. D3Feat [3] jointly detects and describes sparse keypoints for matching. Predator [19] incorporates the global context to enhance the local descriptors and predicts the overlap regions for keypoint sampling. CoFiNet [50] extracts coarse-to-fine correspondences to alleviate the repeatability issue of keypoints. GeoTrans [27] considers the geometric information in fusing the intra-frame context globally. However, the awareness of spatial positions is missing in the cross-frame aggregation. Lepard [20] extends the non-rigid shape matching [34, 38, 42] to point clouds [28] and proposes a re-positioning module to alleviate the pose variations. REGTR [48] directly regresses the corresponding coordinates and registers point clouds in an end-to-end fashion. Nonetheless, all of these methods suffer from instability with additional rotations.
35
+
36
+ Methods with Intrinsic Rotation Invariance. A branch of handcrafted descriptors [8, 14, 41] aligns the input to a canonical representation according to an estimated local reference frame (LRF), while the others [11, 30, 31] mine the rotation-invariant components and encode them as the representation of the local geometry. Inspired by that, some deep learning-based methods [1, 4, 9, 13, 32, 34, 49] are designed to be intrinsically rotation-invariant to make the neural models focus on the pose-agnostic pure geometry. As a pioneer, PPF-FoldNet [9] consumes PPF-based patches and learns the descriptors using a FoldingNet [47]-based architecture without supervision. LRF-based works [1, 13, 32, 34] achieve rotation invariance by aligning their input to the defined canonical representation. YOHO [44] adopts a group of rotations to learn a rotation-equivariant feature group and further obtain the invariance via group pooling. A common problem of the rotation-invariant methods is the less distinctive features. Although RIGA [49] incorporates the global context into local descriptors to enhance the feature distinctiveness, its ineffective local geometry encoding and global position description learned by PointNet [25] still constraint the representation ability of its descriptors.
37
+
38
+ # 3. Method
39
+
40
+ Problem Statement. We tackle the problem of matching a pair of partially-overlapping point clouds $\mathbf{P} \in \mathbb{R}^{n \times 3}$ and $\mathbf{Q} \in \mathbb{R}^{m \times 3}$ , and extract a correspondence set $\hat{\mathcal{C}} = \{(\hat{\mathbf{p}}_i, \hat{\mathbf{q}}_j) | \hat{\mathbf{p}}_i \in \hat{\mathbf{P}} \subseteq \mathbf{P}, \hat{\mathbf{q}}_j \in \hat{\mathbf{Q}} \subseteq \mathbf{Q}\}$ that minimizes:
41
+
42
+ $$
43
+ \frac {1}{| \hat {\mathcal {C}} |} \sum_ {(\hat {\mathbf {p}} _ {i}, \hat {\mathbf {q}} _ {j}) \in \hat {\mathcal {C}}} \| \mathcal {M} ^ {*} (\hat {\mathbf {p}} _ {i}) - \hat {\mathbf {q}} _ {j} \| _ {2}, \tag {1}
44
+ $$
45
+
46
+ where $\| \cdot \| _2$ denotes the Euclidean norm and $|\cdot |$ is the set cardinality. $\mathcal{M}^{*}(\cdot)$ stands for the ground-truth mapping function that maps $\hat{\mathbf{p}}_i$ to its corresponding position in $\hat{\mathbf{Q}}$ . In rigid scenarios, it is defined by a transformation $\mathbf{T}^{*}\in SE(3)$ . For the non-rigid cases it can be denoted as a per-point flow $\mathbf{f}_i^*\in \mathbb{R}^3$ known as the deformation field.
47
+
48
+ ![](images/352f993e80b8039c3fff593ff3b3710cc2efb699d5437ea5534831fd617b8f8e.jpg)
49
+ (0). Input
50
+
51
+ ![](images/db87a1d376332b7b03618dd5794cdaa08a7516290f18cc9528e01d25c924b84b.jpg)
52
+ (2).Aggregation
53
+
54
+ ![](images/974288bc32060553067e5aa14d8a1b36e4504339339ba487679a2a0f3eddfa37.jpg)
55
+ (5). Output
56
+
57
+ ![](images/4e5ef51993cd3abfeb8ba841a2aba0b59cbc3d108e6b058fc3f2d67369ad52e6.jpg)
58
+ (1).Encoder
59
+
60
+ ![](images/09130879a56551a0ae9902d954d95aa14320b2226c7bf4ba302d260c1e117218.jpg)
61
+ (3). Decoder
62
+
63
+ ![](images/ebbdcf4c197d78aed4fe93f07f359b890d055be2d6966b8ff2e9b4e25a40194c.jpg)
64
+ (4). Matching
65
+
66
+ Method Overview. An overview of RoITr is shown in Fig. 2. RoITr consists of an encoder-decoder architecture named Point Pair Feature Transformer (PPFTrans) for local geometry encoding and a stack of $g \times$ global transformers for global context aggregation. Correspondence set $\hat{C}$ is extracted by the coarse-to-fine matching [50].
67
+
68
+ # 3.1. PPF Attention Mechanism
69
+
70
+ Overview. Fig. 3 compares three different self-attention mechanisms. The standard attention [43] only leverages the input context to obtain the Query $\mathbf{Q}$ and Key $\mathbf{K}$ to compute the contextual attention $\mathbf{A}_C$ , as well as the Value $\mathbf{V}$ that encodes information for the contextual message $\mathbf{M}_C$ . GeoTrans [27] proposes to learn the positional encoding $\mathbf{E}$ from the geometry and calculates a second attention $\mathbf{A}_G$ to reweigh $\mathbf{A}_C$ . However, the cues contained in the raw geometry are totally neglected. To this end, we propose to learn the pose-agnostic geometric cues $\mathbf{G}$ and further generate the geometric message $\mathbf{M}_G$ in the PPF Attention Mechanism (PAM). On the local level, $\mathbf{M}_G$ is combined with $\mathbf{M}_C$ for feature enhancement, while on the global level, it is used to learn the rotation-invariant position representation for the cross-frame context aggregation. More specifically, we de
71
+
72
+ ![](images/c0b712b6186087050f178483da8e9448683cbb6f8c24b4fae9914085547debf4.jpg)
73
+ (a). Standard Attention
74
+
75
+ ![](images/050ce13ccd9012d289ebabdf10f1e5aa082cc00677dad1952586dbd8194f9d92.jpg)
76
+ Figure 2. An Overview of RoITr. From left to right: $(\pmb{0})$ . RoITr takes as input a pair of triplets $\mathcal{P} = (\mathbf{P}, \mathbf{N}, \mathbf{X})$ and $\mathcal{Q} = (\mathbf{Q}, \mathbf{M}, \mathbf{Y})$ , each with three dimensions referring to the point cloud, the estimated normals, and the initial features. (1).[§ 3.2] A stack of encoder blocks hierarchically downsamples the points to coarser superpoints and encodes the local geometry, yielding superpoint triplets $\mathcal{P}'$ and $\mathcal{Q}'$ . Each encoder block consists of an Attentional Abstraction Layer (AAL) for downsampling and abstraction, followed by $e \times$ PPF Attention Layers (PALs) for local geometry encoding and context aggregation. Both of them are based on our proposed PPF Attention Mechanism (PAM), which enables the pose-agnostic encoding of pure geometry. (See Fig. 3 and Fig. 4). (2).[§ 3.3] Global information is fused to enhance the superpoint features of $\mathcal{P}'$ and $\mathcal{Q}'$ . The geometric cues are globally aggregated as a rotation-invariant position representation, which introduces spatial awareness in the consecutive cross-frame context aggregation. After a stack of $g \times$ global transformers, the globally-enhanced triplets $\widetilde{\mathcal{P}}'$ and $\widetilde{\mathcal{Q}}'$ are produced. (3).[§ 3.2] Superpoint triplets $\mathcal{P}'$ and $\mathcal{Q}'$ are decoded to point triplets $\hat{\mathcal{P}}$ and $\hat{\mathcal{Q}}$ by a stack of decoder blocks. Each block consists of a Transition Up Layer (TUL) for upsampling and context aggregation, followed by $d \times$ PALs. (4).[§ 3.4] By adopting the coarse-to-fine matching [50], $\widetilde{\mathcal{P}}'$ and $\widetilde{\mathcal{Q}}'$ are matched to generate superpoint correspondences, which are consecutively refined to point correspondences between $\hat{\mathcal{P}}$ and $\hat{\mathcal{Q}}$ . (5). $\hat{\mathcal{C}}$ is established between $\hat{\mathcal{P}}$ and $\hat{\mathcal{Q}}$ .
77
+ (b). GeoTrans
78
+
79
+ ![](images/1ae7e04b44a0ede10cb1f694b3023f8ff8b9887eed7c1e32a5a65db1f3724d8b.jpg)
80
+ (c).PAM (Ours)
81
+ Figure 3. Illustration of different self-attention computation in the standard attention [43], GeoTrans [27], and PAM.
82
+
83
+ fine PAM on an Anchor triplet $\mathcal{P}^A = (\mathbf{P}^A, \mathbf{N}^A, \mathbf{X}^A)$ and a Support triplet $\mathcal{P}^S = (\mathbf{P}^S, \mathbf{N}^S, \mathbf{X}^S)$ , both with three dimensions referring to the point cloud, the estimated normals, and the associated features, respectively. PAM aggregates the learned context and geometric cues from $\mathcal{P}^S$ and flows the messages to $\mathcal{P}^A$ .
84
+
85
+ Pose-Agnostic Coordinate Representation. The basis of PAM is the pose-agnostic local coordinate representation that we construct based on PPFs [11]. Let $\mathcal{P}_i^A \coloneqq (\mathbf{p}_i^A \in \mathbf{P}^A, \mathbf{n}_i^A \in \mathbf{N}^A, \mathbf{x}_i^A \in \mathbf{X}^A) \in \mathcal{P}^A$ denote the triplet constructed by picking the $i^{th}$ item on each dimension. For each $\mathbf{p}_i^A$ , a subset of $\mathcal{P}^S$ is first retrieved according to the Euclidean distance w.r.t. $\mathbf{P}^S$ , denoted as
86
+
87
+ ![](images/5081a262fa30968f29b7c9345f2e9eb18f9480785ff18366f83ba9c4b08632b3.jpg)
88
+ Figure 4. Left: The workflow of the PPF Attention Mechanism (PAM). Right: Detailed calculation of the attention.
89
+
90
+ $\mathcal{P}_{\mathcal{N}(i)}^{S} \coloneqq (\mathbf{P}_{\mathcal{N}(i)}^{S}, \mathbf{N}_{\mathcal{N}(i)}^{S}, \mathbf{X}_{\mathcal{N}(i)}^{S}) \subseteq \mathcal{P}^{S}$ , with $\mathcal{N}(i)$ the indices of $k$ -nearest neighbors. We then adopt PPFs [11] to construct a local coordinate system around each $\mathbf{p}_i^A$ to represent the pose-agnostic position of $\mathbf{P}_{\mathcal{N}(i)}^{S}$ w.r.t. it. The coordinate of point $\mathbf{p}_j^S \in \mathbf{P}_{\mathcal{N}(i)}^{S}$ is transferred to:
91
+
92
+ $$
93
+ \mathbf {e} _ {j} ^ {S} = \left(\| \mathbf {d} \| _ {2}, \angle \left(\mathbf {n} _ {i} ^ {A}, \mathbf {d}\right), \angle \left(\mathbf {n} _ {j} ^ {S}, \mathbf {d}\right), \angle \left(\mathbf {n} _ {j} ^ {S}, \mathbf {n} _ {i} ^ {A}\right)\right), \quad (2)
94
+ $$
95
+
96
+ with $\mathbf{d} = \mathbf{p}_j^S -\mathbf{p}_i^A$ , and $\mathbf{n}_i^A$ and $\mathbf{n}_j^S$ the estimated normals of $\mathbf{p}_i^A$ and $\mathbf{p}_j^S$ , respectively. $\angle (\mathbf{v}_1,\mathbf{v}_2)$ computes the angles between the two vectors [5, 9]. The transferred coordinates of $\mathbf{P}_{\mathcal{N}(i)}^{S}$ are denoted as $\mathbf{E}_{\mathcal{N}(i)}^{S}$
97
+
98
+ PPF Attention Mechanism. PPF Attention Mechanism (PAM) takes as input the Support triplet $\mathcal{P}^S$ and the Anchor point cloud $\mathbf{P}^A$ with estimated normals $\mathbf{N}^A$ . PAM generates the Anchor features $\mathbf{X}^A$ by aggregating the pose-agnostic local geometry and highly-representative learned context from $\mathcal{P}^S$ , which is defined as:
99
+
100
+ $$
101
+ \mathcal {P} ^ {A} = \delta \left(\mathbf {P} ^ {A}, \mathbf {N} ^ {A} \mid \mathcal {P} ^ {S}\right), \tag {3}
102
+ $$
103
+
104
+ with $\delta(\cdot)$ representing PAM. As shown in Fig. 4 (a), for each $\mathbf{p}_i^A \in \mathbf{P}^A$ with normal $\mathbf{n}_i^A$ , we find its nearest point $\mathbf{p}_j^S \in \mathbf{P}^S$ whose associated feature $\mathbf{x}_j^S$ is assigned to $\mathbf{p}_i^A$ as the initial description. Then, $k$ -nearest neighbors from $\mathbf{P}^S$ are retrieved according to the Euclidean distance in 3D space, yielding $\mathbf{P}_{\mathcal{N}(i)}^S \subseteq \mathbf{P}^S$ and $\mathbf{X}_{\mathcal{N}(i)}^S \subseteq \mathbf{X}^S$ . Following Eq. 2, $\mathbf{P}_{\mathcal{N}(i)}^S$ is transferred to the pose-agnostic position representation $\mathbf{E}_{\mathcal{N}(i)}^S$ , which is consecutively projected to the coordinate embedding $\mathbf{E}_S$ via a linear layer. $\mathbf{x}_j^S$ and $\mathbf{X}_{\mathcal{N}(i)}^S$ are projected to the contextual features $\mathbf{x}_S$ and $\mathbf{X}_S$ by a second shared linear layer, respectively. In Fig. 4 (b), the attention mechanism uses five learnable matrices $\mathbf{W}_G$ , $\mathbf{W}_E$ , $\mathbf{W}_Q$ , $\mathbf{W}_K$ , and $\mathbf{W}_V$ to project the input. Specifically, $\mathbf{W}_G$ and $\mathbf{W}_E$ project the input coordinate representation to the geometric cues and positional encoding by:
105
+
106
+ $$
107
+ \mathbf {G} = \mathbf {E} _ {S} \mathbf {W} _ {G} \quad \text {a n d} \quad \mathbf {E} = \mathbf {E} _ {S} \mathbf {W} _ {E}, \tag {4}
108
+ $$
109
+
110
+ respectively. Similarly, $\mathbf{W}_Q$ , $\mathbf{W}_K$ , and $\mathbf{W}_V$ project the learned context to Query, Key, and Value as:
111
+
112
+ $$
113
+ \mathbf {q} = \mathbf {x} _ {S} \mathbf {W} _ {Q}, \mathbf {K} = \mathbf {X} _ {S} \mathbf {W} _ {K}, \text {a n d} \mathbf {V} = \mathbf {X} _ {S} \mathbf {W} _ {V}, \tag {5}
114
+ $$
115
+
116
+ respectively. The attention a that measures the feature similarity, and the message m that encodes both the pose-agnostic geometry and the representative context read as:
117
+
118
+ $$
119
+ \mathbf {a} = \operatorname {S o f t m a x} \left(\frac {\mathbf {q} \mathbf {E} ^ {T} + \mathbf {q} \mathbf {K} ^ {T}}{\sqrt {c _ {0}}}\right) \quad \text {a n d} \quad \mathbf {m} = \mathbf {a} \mathbf {G} + \mathbf {a} \mathbf {V}, \tag {6}
120
+ $$
121
+
122
+ respectively. The message $\mathbf{m}$ is projected and aggregated to $\mathbf{x}_j^S$ via an element-wise addition followed by a normalization through LayerNorm [2]. The final linear layer projects the obtained feature to $\mathbf{x}_i^A$ , from which $\mathbf{X}^A$ is obtained to formulate the output $\mathcal{P}^A$ with the known $\mathbf{P}^A$ and $\mathbf{N}^A$ .
123
+
124
+ # 3.2. PPFTrans for Local Geometry Description
125
+
126
+ Overview. As illustrated in Fig. 2, PPFTrans consumes triplets $\mathcal{P}$ and $\mathcal{Q}$ . Taking $\mathcal{P} = (\mathbf{P},\mathbf{N},\mathbf{X})$ as an example, it consists of $\mathbf{P} \in \mathbb{R}^{n \times 3}$ the points cloud, $\mathbf{N} \in \mathbb{R}^{n \times 3}$ the normals estimated from $\mathbf{P}$ , and $\mathbf{X} = \vec{\mathbf{1}} \in \mathbb{R}^{n \times 1}$ the initial point features. The encoder produces the superpoint triplet $\mathcal{P}' = (\mathbf{P}',\mathbf{N}',\mathbf{X}')$ with $\mathbf{P}' \in \mathbb{R}^{n' \times 3}$ and $\mathbf{X}' \in \mathbb{R}^{n' \times c'}$ . With the consecutive decoder, $\mathcal{P}'$ is decoded to a triplet $\hat{\mathcal{P}} = (\hat{\mathbf{P}},\hat{\mathbf{N}},\hat{\mathbf{X}})$ including $\hat{n}$ points with features $\hat{\mathbf{X}} \in \mathbb{R}^{\hat{n} \times \hat{c}}$ . Notably, as we adopt a Farthest Point Sampling (FPS) strategy [26], it always satisfies that $\mathbf{P}' \subseteq \hat{\mathbf{P}} \subseteq \mathbf{P}$ . The same goes for a second point cloud $\mathbf{Q}$ with an input triplet $\mathcal{Q} = (\mathbf{Q} \in \mathbb{R}^{m \times 3},\mathbf{M} \in \mathbb{R}^{m \times 3},\mathbf{Y} = \vec{\mathbf{1}} \in \mathbb{R}^{m \times 1})$ by the shared architecture. In the rest of this paper, we only demonstrate for $\mathcal{P}$ unless the model processes $\mathcal{Q}$ differently. Encoder. The encoder is constructed by stacking several encoder blocks, each including an Attentional Abstraction Layer (AAL) followed by $e \times$ PPF Attention Layers (PALs). Each block consumes the output of the previous block as the Support triplet $\mathcal{P}^S$ ( $\mathcal{P}^S = \mathcal{P}$ for the first block). $\mathcal{P}^S$ first flows to AAL, where Anchor points $\mathbf{P}^A$ with associated normals $\mathbf{N}^A$ are obtained via FPS [26]. The Anchor triplet $\mathcal{P}^A$ is then generated in AAL via a PAM following Eq. 3. A sequence of PALs is applied for enhancing the Anchor features $\mathbf{X}^A$ , each updating the features as:
127
+
128
+ $$
129
+ \mathcal {P} ^ {A} \leftarrow \theta (\mathcal {P} ^ {A}) = \operatorname {R e L U} \left(\mathbf {X} ^ {A} + \phi \left(\delta \left(\mathbf {P} ^ {A}, \mathbf {N} ^ {A} \mid \mathcal {P} ^ {A}\right)\right)\right), \tag {7}
130
+ $$
131
+
132
+ with $\phi$ the LayerNorm [2], $\delta$ the PAM, and $\theta$ the PAL. $\leftarrow$ depicts feature updating. The encoder block outputs the updated $\mathcal{P}^A$ , and the output of the whole encoder is defined as $\mathcal{P}'$ , which is the output of the final encoder block.
133
+
134
+ Decoder. We build the decoder by stacking a series of decoder blocks, each consisting of a Transition Up Layer (TUL) followed by $d \times \mathrm{PAL}$ . Each block takes the output of the previous block as the Anchor triplet $\mathcal{P}^A$ ( $\mathcal{P}^A = \mathcal{P}'$ for the first block), and takes the Support
135
+
136
+ ![](images/f4b3f87f9e83497fcc7b10f0844a7cef1f6e7b788a4cf5fc18e8d0079f4a86bd.jpg)
137
+ (a). Geometry-Aware Self-Attention
138
+
139
+ ![](images/d764f44f9d306301505fa374dfc9c5743da0b1b9c69a915b710abec5d66a715b.jpg)
140
+ (b). Position-Aware Cross-Attention
141
+ Figure 5. The computation graph of our global transformer consisting of the Geometry-Aware Self-Attention Module (GSM) and Position-Aware Cross-Attention Module (PCM).
142
+
143
+ triplet $\mathcal{P}^S$ from the encoder via skip connections. The input flows to TUL, where each feature $\widetilde{\mathbf{x}}_j^S\in \widetilde{\mathbf{X}}^S$ assigned to $\mathbf{p}_j^S\in \mathbf{P}^S$ is interpolated by:
144
+
145
+ $$
146
+ \widetilde {\mathbf {x}} _ {j} ^ {S} = \frac {\sum_ {i \in \mathcal {N} (j)} w _ {i} ^ {j} \mathbf {x} _ {i} ^ {A}}{\sum_ {i \in \mathcal {N} (j)} w _ {i} ^ {j}}, \text {w i t h} w _ {i} ^ {j} = \frac {1}{\left\| \mathbf {p} _ {j} ^ {S} - \mathbf {p} _ {i} ^ {A} \right\| _ {2}}, \tag {8}
147
+ $$
148
+
149
+ with $\mathcal{N}(j)$ the $k$ -nearest neighbors of $\mathbf{p}_j^S$ in $\mathbf{P}^A$ . Features are updated by two linear layers as $\mathcal{P}^S \gets \zeta_1(\mathbf{X}^S) + \zeta_2(\widetilde{\mathbf{X}}^S)$ . A sequence of PALs is adopted after TUL, each enhancing the features as $\mathcal{P}^S \gets \theta(\mathcal{P}^S)$ according to Eq. 7. The decoder block outputs the updated $\mathcal{P}^S$ , and the output of the whole decoder is denoted as $\hat{\mathcal{P}}$ , which is the output of the final decoder block.
150
+
151
+ # 3.3. Global Transformer for Context Aggregation
152
+
153
+ Overview. Our designed global transformer takes as input a pair of triplets $\mathcal{P}'$ and $\mathcal{Q}'$ , and enhances the features with the global context, yielding $\widetilde{\mathbf{X}}' \in \mathbb{R}^{n' \times c'}$ and $\widetilde{\mathbf{Y}}' \in \mathbb{R}^{m' \times c'}$ , respectively. We stack $g \times$ global transformers, with each including a Geometry-Aware Self-Attention Module (GSM) and a Position-Aware Cross-Attention Module (PCM) (See Fig. 2 and Fig. 5). Different from previous works [19, 27, 48-50] that totally neglect the cross-frame spatial relationships, we propose to learn a rotation-invariant position representation for each superpoint to enable the position-aware cross-frame context aggregation.
154
+
155
+ Geometry-Aware Self-Attention Module. On the global level, we modify PAM to learn the rotation-invariant position representation and to aggregate the learned context across the whole frame simultaneously. The design of GSM is detailed in Fig. 5 (a). GSM has two branches, where the geometry branch mines the geometric cues from the pairwise rotation-invariant geometry representation proposed in [27], and the context branch aggregates the global context across the frame. We refer the readers to the Appendix for the detailed construction of $\mathbf{R}' \in \mathbb{R}^{n' \times n' \times c'}$ and the ablation study on it. Similar to Eq. 4, the geometric cues $\mathbf{G}'$ and the positional encoding $\mathbf{E}'$ are linearly projected from
156
+
157
+ $\mathbf{R}^{\prime}$ . $\mathbf{E}^{\prime}$ is further processed in the geometry branch and finally leveraged as the rotation-invariant position representation. In the context branch, $\mathbf{Q}^{\prime}$ , $\mathbf{K}^{\prime}$ , and $\mathbf{V}^{\prime}$ are obtained by linearly mapping the input features $\mathbf{X}^{\prime}$ similar to Eq. 5. The hybrid score matrix $\mathbf{S}^{\prime} \in \mathbb{R}^{n^{\prime} \times n^{\prime}}$ is computed as:
158
+
159
+ $$
160
+ \mathbf {S} ^ {\prime} (i, j) = \frac {\left(\mathbf {q} _ {i} ^ {\prime}\right) \left(\mathbf {e} _ {i , j} ^ {\prime} + \mathbf {k} _ {j} ^ {\prime}\right) ^ {T}}{\sqrt {c ^ {\prime}}}, \tag {9}
161
+ $$
162
+
163
+ with $\mathbf{e}_{i,j}^{\prime} := \mathbf{E}^{\prime}(i,j,:)$ , $\mathbf{q}_i^{\prime} := \mathbf{Q}^{\prime}(i,:)$ , and $\mathbf{k}_j^{\prime} := \mathbf{K}^{\prime}(j,:)$ the $c^{\prime}$ -dimension vectors. The hybrid attention $\mathbf{A}^{\prime}$ is obtained via a Softmax function over each row of $\mathbf{S}^{\prime}$ , and the geometric messages $\mathbf{M}_G^{\prime} \in \mathbb{R}^{n^{\prime} \times c^{\prime}}$ are computed as:
164
+
165
+ $$
166
+ \mathbf {M} _ {G} ^ {\prime} (i,:) = \sum_ {1 \leq j \leq n ^ {\prime}} a _ {i, j} ^ {\prime} \mathbf {g} _ {i, j} ^ {\prime}, \tag {10}
167
+ $$
168
+
169
+ with $a_{i,j}^{\prime} \coloneqq \mathbf{A}^{\prime}(i,j)$ and $\mathbf{g}_{i,j}^{\prime} \coloneqq \mathbf{G}^{\prime}(i,j,:)$ . The contextual messages $\mathbf{M}_V^{\prime} \in \mathbb{R}^{n^{\prime} \times c^{\prime}}$ are computed by $\mathbf{A}^{\prime}\mathbf{V}^{\prime}$ . After a feed-forward network [43], the position representation $\mathbf{E}_P^{\prime}$ and globally-enhanced context $\mathbf{C}_P^{\prime}$ are generated.
170
+
171
+ Position-Aware Cross-Attention Module. PCM consumes a pair of doublets $(\mathbf{E}_P',\mathbf{C}_P')$ and $(\mathbf{E}_Q',\mathbf{C}_Q')$ that are generated from $\mathcal{P}'$ and $\mathcal{Q}'$ by a shared GSM, respectively. As the cross-attention is directional, we apply the same PCM twice, with the first aggregation from $\mathcal{Q}'$ to $\mathcal{P}'$ (See Fig. 5 (b)), and the second reversed. As the first step, the rotation-invariant position representation is incorporated to make the consecutive cross-attention position-aware, yielding position-aware features $\mathbf{F}_P' = \mathbf{E}_P' + \mathbf{C}_P'$ and $\mathbf{F}_Q' = \mathbf{E}_Q' + \mathbf{C}_Q'$ . Similar to Eq. 5, $\widetilde{\mathbf{Q}}',\widetilde{\mathbf{K}}'$ , and $\widetilde{\mathbf{V}}'$ are computed as the linear projection of $\mathbf{F}_P',\mathbf{F}_Q'$ , and $\mathbf{F}_Q'$ , respectively. The attention matrix $\widetilde{\mathbf{A}}\in \mathbb{R}^{n'\times m'}$ is computed via a row-wise softmax function applied on $\widetilde{\mathbf{Q}}'\widetilde{\mathbf{K}}'^T$ . The fused messages are presented as $\widetilde{\mathbf{A}}\widetilde{\mathbf{V}}'$ , which are finally mapped to the output features $\widetilde{\mathbf{X}}'$ through a feedforward network. As we introduce spatial awareness at the beginning of PCM, both the attention computation and message fusion are aware of the cross-frame positions. After the twice application of PCM, the input features are enhanced as $\mathcal{P}'\gets \widetilde{\mathbf{X}}'$ and $\mathcal{Q}'\gets \widetilde{\mathbf{Y}}'$ , respectively. The global aggregation stage finally generates a pair of triplets $\widetilde{\mathcal{P}}':=(\mathbf{P}',\mathbf{N}',\widetilde{\mathbf{X}}')$ and $\widetilde{\mathcal{Q}}':=(\mathbf{Q}',\mathbf{M}',\widetilde{\mathbf{Y}}')$ , with the enhanced features from the last global transformer.
172
+
173
+ # 3.4. Point Matching and Loss Function
174
+
175
+ Superpoint Matching. As shown in Fig. 2, the point matching stage consumes a pair of superpoint triplets $\hat{\mathcal{P}}'$ and $\hat{\mathcal{Q}}'$ obtained from the global transformer, as well as a pair of point triplets $\hat{\mathcal{P}}$ and $\hat{\mathcal{Q}}$ produced by the decoder. We adopt the coarse-to-fine matching proposed in [50]. Following [27], we first normalize the superpoint features $\widetilde{\mathbf{X}}'$ and $\widetilde{\mathbf{Y}}'$ onto a unit hypersphere, and measure the pairwise similarity using a Gaussian correlation matrix $\widetilde{\mathbf{S}}$ with $\widetilde{\mathbf{S}}(i,j) =$
176
+
177
+ $-\exp \left(-\|\widetilde{\mathbf{x}}_i'\widetilde{\mathbf{y}}_j'\|_2^2\right)$ . After a dual-normalization [27,29,37] on $\widetilde{\mathbf{S}}$ for global feature correlation, superpoints associated to the top- $k$ entries are selected as the coarse correspondence set $\mathcal{C}' = \{(\mathbf{p}_i',\mathbf{q}_i')|\mathbf{p}_i' \in \mathbf{P}',\mathbf{q}_i' \in \mathbf{Q}'\}$ .
178
+
179
+ Point Matching. For extracting point correspondences, denser points $\hat{\mathbf{P}}$ and $\hat{\mathbf{Q}}$ are first assigned to superpoints. To this end, the point-to-node strategy [50] is leveraged, where each point is assigned to its closest superpoint in 3D space. Given a superpoint $\mathbf{p}_i' \in \mathbf{P}'$ , the group of points assigned to it is denoted as $\hat{\mathbf{G}}_i^P \subseteq \hat{\mathbf{P}}$ . The group of features associated to $\hat{\mathbf{G}}_i^P$ is further defined as $\hat{\mathbf{G}}_i^X$ with $\hat{\mathbf{G}}_i^X \subseteq \hat{X}$ . For each superpoint correspondence $\mathcal{C}_l' = (\mathbf{p}_i', \mathbf{q}_j')$ , the similarity between the corresponding feature groups $\hat{\mathbf{G}}_i^X$ and $\hat{\mathbf{G}}_j^Y$ is calculated as $\hat{\mathbf{S}}_l = \hat{\mathbf{G}}_i^X (\hat{\mathbf{G}}_j^Y)^T / \sqrt{\hat{c}}$ , with $\hat{c}$ the feature dimension. We then follow [35] to append a stack row and column to $\hat{\mathbf{S}}_l$ filled with a learnable parameter $\alpha$ , and iteratively run the Sinkhorn Algorithm [36]. After removing the slack row and column of $\hat{\mathbf{S}}_l$ , the mutual top-k entries, i.e., entries with top- $k$ confidence on both the row and the column, are selected to formulate a point correspondence set $\hat{\mathcal{C}}_l$ . The final correspondence set $\hat{\mathcal{C}}$ is collected by $\hat{\mathcal{C}} = \bigcup_{l=1}^{|\mathcal{C}'|} \hat{\mathcal{C}}_l$ .
180
+
181
+ Loss Function. Our loss function reads as $\mathcal{L} = \mathcal{L}_s + \lambda \mathcal{L}_p$ , with a superpoint matching loss $\mathcal{L}_s$ and a point matching loss $\mathcal{L}_p$ balanced by a hyper-parameter $\lambda$ ( $\lambda = 1$ by default). The detailed definition is introduced in the Appendix.
182
+
183
+ # 4. Experiment
184
+
185
+ We evaluate RoITr on both rigid (3DMatch [51] & 3DLoMatch [19]) and non-rigid (4DMatch [20] & 4DLoMatch [20]) benchmarks. For the rigid matching, we further evaluate our correspondences on the registration task, where RANSAC [12] is used. Details of implementation, metrics, and runtime analysis are introduced in the Appendix.
186
+
187
+ # 4.1. Rigid Indoor Scenes: 3DMatch & 3DLoMatch
188
+
189
+ Dataset. 3DMatch [51] collects 62 indoor scenes, among which 46 are used for training, 8 for validation, and 8 for testing. We use the data processed by [19] where the 3DMatch data is spilt as 3DMatch ( $>30\%$ overlap) and 3DLoMatch ( $10\% \sim 30\%$ overlap). To evaluate robustness to arbitrary rotations, we follow [49] for creating the rotated benchmarks, where full-range rotations are individually added to the two frames of each point cloud pair.
190
+
191
+ Metrics. We follow [19] to use three metrics for evaluation: (1). Inlier Ratio (IR) that computes the ratio of putative correspondences whose residual distance is smaller than a threshold (i.e., $0.1\mathrm{m}$ ) under the ground-truth transformation; (2). Feature Matching Recall (FMR) that calculates the fraction of point cloud pairs whose IR is larger than a threshold (i.e., $5\%$ ); (3). Registration Recall (RR) that counts the fraction of point cloud pairs that are cor
192
+
193
+ <table><tr><td rowspan="2"># Samples=5,000</td><td colspan="2">3DMatch</td><td colspan="2">3DLoMatch</td></tr><tr><td>Origin</td><td>Rotated</td><td>Origin</td><td>Rotated</td></tr><tr><td colspan="5">Feature Matching Recall (%) ↑</td></tr><tr><td>SpinNet [1]</td><td>97.4</td><td>97.4</td><td>75.5</td><td>75.2</td></tr><tr><td>Predator [19]</td><td>96.6</td><td>96.2</td><td>78.6</td><td>73.7</td></tr><tr><td>CoFiNet [50]</td><td>98.1</td><td>97.4</td><td>83.1</td><td>78.6</td></tr><tr><td>YOHO [44]</td><td>98.2</td><td>97.8</td><td>79.4</td><td>77.8</td></tr><tr><td>RIGA [49]</td><td>97.9</td><td>98.2</td><td>85.1</td><td>84.5</td></tr><tr><td>Lepard [20]</td><td>98.0</td><td>97.4</td><td>83.1</td><td>79.5</td></tr><tr><td>GeoTrans [27]</td><td>97.9</td><td>97.8</td><td>88.3</td><td>85.8</td></tr><tr><td>RoITr (Ours)</td><td>98.0</td><td>98.2</td><td>89.6</td><td>89.4</td></tr><tr><td colspan="5">Inlier Ratio (%) ↑</td></tr><tr><td>SpinNet [1]</td><td>48.5</td><td>48.7</td><td>25.7</td><td>25.7</td></tr><tr><td>Predator [19]</td><td>58.0</td><td>52.8</td><td>26.7</td><td>22.4</td></tr><tr><td>CoFiNet [50]</td><td>49.8</td><td>46.8</td><td>24.4</td><td>21.5</td></tr><tr><td>YOHO [44]</td><td>64.4</td><td>64.1</td><td>25.9</td><td>23.2</td></tr><tr><td>RIGA [49]</td><td>68.4</td><td>68.5</td><td>32.1</td><td>32.1</td></tr><tr><td>Lepard [20]</td><td>58.6</td><td>53.7</td><td>28.4</td><td>24.4</td></tr><tr><td>GeoTrans [27]</td><td>71.9</td><td>68.2</td><td>43.5</td><td>40.0</td></tr><tr><td>RoITr (Ours)</td><td>82.6</td><td>82.3</td><td>54.3</td><td>53.2</td></tr><tr><td colspan="5">Registration Recall (%) ↑</td></tr><tr><td>SpinNet [1]</td><td>88.8</td><td>93.2</td><td>58.2</td><td>61.8</td></tr><tr><td>Predator [19]</td><td>89.0</td><td>92.0</td><td>59.8</td><td>58.6</td></tr><tr><td>CoFiNet [50]</td><td>89.3</td><td>92.0</td><td>67.5</td><td>62.5</td></tr><tr><td>YOHO [44]</td><td>90.8</td><td>92.5</td><td>65.2</td><td>66.8</td></tr><tr><td>RIGA [49]</td><td>89.3</td><td>93.0</td><td>65.1</td><td>66.9</td></tr><tr><td>Lepard [20]</td><td>92.7</td><td>84.9</td><td>65.4</td><td>49.0</td></tr><tr><td>GeoTrans [27]</td><td>92.0</td><td>92.0</td><td>75.0</td><td>71.8</td></tr><tr><td>RoITr (Ours)</td><td>91.9</td><td>94.7</td><td>74.8</td><td>77.2</td></tr></table>
194
+
195
+ Table 1. Quantitative results on (Rotated) 3DMatch & 3DLoMatch. 5,000 points/correspondences are used for the evaluation.
196
+
197
+ rectly registered (i.e., with $\mathrm{RMSE} < 0.2\mathrm{m}$ ).
198
+
199
+ Comparison with the State-of-the-Art. We compare RoITr with 7 state-of-the-art methods, among which Predator [19], CoFiNet [50], Lepard [20], and GeoTrans [27] are rotation-sensitive models, while SpinNet [1], YOHO [44], and RIGA [49] guarantee the rotation invariance by design. In Tab. 1 we demonstrate the matching and registration results on 3DMatch and 3DLoMatch, as well as on their rotated versions, with 5,000 sampled points/correspondences. Regarding IR, RoITr outperforms all the others by a large margin on both datasets, which indicates our method matches points more correctly. For FMR, we significantly surpass all the others on 3DLoMatch, while staying on par with CoFiNet and YOHO on 3DMatch, which indicates that our model is good at coping with hard cases, i.e., we find at least $5\%$ inliers on more test data. For the registration evaluation in terms of RR, RoITr achieves comparable performance with GeoTrans and Lepard on 3DMatch, but leads the board together with GeoTrans on 3DLoMatch with an overwhelming advantage over the others. Our stability against additional rotations is further demonstrated on the rotated data, where we outperform all the others with a substantial margin. Qualitative results can be found in Fig. 6.
200
+
201
+ ![](images/ac23312a02beb4a475dda47e424d95133ebb2541ce2ad59f1a91cd2ede8584a7.jpg)
202
+ Figure 6. Qualitative results on 3DLoMatch. GeoTrans [27] is used as the baseline. Columns (b) and (c) show the correspondences, while columns (d) and (e) demonstrate the registration results. Green/red lines indicate inliers/outliers. More cases are shown in the Appendix.
203
+
204
+ <table><tr><td rowspan="2"># Samples</td><td colspan="5">3DMatch</td><td colspan="5">3DLoMatch</td></tr><tr><td>5000</td><td>2500</td><td>1000</td><td>500</td><td>250</td><td>5000</td><td>2500</td><td>1000</td><td>500</td><td>250</td></tr><tr><td colspan="11">Feature Matching Recall (%)↑</td></tr><tr><td>SpinNet [1]</td><td>97.4</td><td>97.0</td><td>96.4</td><td>96.7</td><td>94.8</td><td>75.5</td><td>75.1</td><td>74.2</td><td>69.0</td><td>62.7</td></tr><tr><td>Predator [19]</td><td>96.6</td><td>96.6</td><td>96.5</td><td>96.3</td><td>96.5</td><td>78.6</td><td>77.4</td><td>76.3</td><td>75.7</td><td>75.3</td></tr><tr><td>CoFiNet [50]</td><td>98.1</td><td>98.3</td><td>98.1</td><td>98.2</td><td>98.3</td><td>83.1</td><td>83.5</td><td>83.3</td><td>83.1</td><td>82.6</td></tr><tr><td>YOHO [44]</td><td>98.2</td><td>97.6</td><td>97.5</td><td>97.7</td><td>96.0</td><td>79.4</td><td>78.1</td><td>76.3</td><td>73.8</td><td>69.1</td></tr><tr><td>RIGA [49]</td><td>97.9</td><td>97.8</td><td>97.7</td><td>97.7</td><td>97.6</td><td>85.1</td><td>85.0</td><td>85.1</td><td>84.3</td><td>85.1</td></tr><tr><td>GeoTrans [27]</td><td>97.9</td><td>97.9</td><td>97.9</td><td>97.9</td><td>97.6</td><td>88.3</td><td>88.6</td><td>88.8</td><td>88.6</td><td>88.3</td></tr><tr><td>RoITr (Ours)</td><td>98.0</td><td>98.0</td><td>97.9</td><td>98.0</td><td>97.9</td><td>89.6</td><td>89.6</td><td>89.5</td><td>89.4</td><td>89.3</td></tr><tr><td colspan="11">Inlier Ratio (%)↑</td></tr><tr><td>SpinNet [1]</td><td>48.5</td><td>46.2</td><td>40.8</td><td>35.1</td><td>29.0</td><td>25.7</td><td>23.7</td><td>20.6</td><td>18.2</td><td>13.1</td></tr><tr><td>Predator [19]</td><td>58.0</td><td>58.4</td><td>57.1</td><td>54.1</td><td>49.3</td><td>26.7</td><td>28.1</td><td>28.3</td><td>27.5</td><td>25.8</td></tr><tr><td>CoFiNet [50]</td><td>49.8</td><td>51.2</td><td>51.9</td><td>52.2</td><td>52.2</td><td>24.4</td><td>25.9</td><td>26.7</td><td>26.8</td><td>26.9</td></tr><tr><td>YOHO [44]</td><td>64.4</td><td>60.7</td><td>55.7</td><td>46.4</td><td>41.2</td><td>25.9</td><td>23.3</td><td>22.6</td><td>18.2</td><td>15.0</td></tr><tr><td>RIGA [49]</td><td>68.4</td><td>69.7</td><td>70.6</td><td>70.9</td><td>71.0</td><td>32.1</td><td>33.4</td><td>34.3</td><td>34.5</td><td>34.6</td></tr><tr><td>GeoTrans [27]</td><td>71.9</td><td>75.2</td><td>76.0</td><td>82.2</td><td>85.1</td><td>43.5</td><td>45.3</td><td>46.2</td><td>52.9</td><td>57.7</td></tr><tr><td>RoITr (Ours)</td><td>82.6</td><td>82.8</td><td>83.0</td><td>83.0</td><td>83.0</td><td>54.3</td><td>54.6</td><td>55.1</td><td>55.2</td><td>55.3</td></tr><tr><td colspan="11">Registration Recall (%)↑</td></tr><tr><td>SpinNet [1]</td><td>88.8</td><td>88.0</td><td>84.5</td><td>79.0</td><td>69.2</td><td>58.2</td><td>56.7</td><td>49.8</td><td>41.0</td><td>26.7</td></tr><tr><td>Predator [19]</td><td>89.0</td><td>89.9</td><td>90.6</td><td>88.5</td><td>86.6</td><td>59.8</td><td>61.2</td><td>62.4</td><td>60.8</td><td>58.1</td></tr><tr><td>CoFiNet [50]</td><td>89.3</td><td>88.9</td><td>88.4</td><td>87.4</td><td>87.0</td><td>67.5</td><td>66.2</td><td>64.2</td><td>63.1</td><td>61.0</td></tr><tr><td>YOHO [44]</td><td>90.8</td><td>90.3</td><td>89.1</td><td>88.6</td><td>84.5</td><td>65.2</td><td>65.5</td><td>63.2</td><td>56.5</td><td>48.0</td></tr><tr><td>RIGA [49]</td><td>89.3</td><td>88.4</td><td>89.1</td><td>89.0</td><td>87.7</td><td>65.1</td><td>64.7</td><td>64.5</td><td>64.1</td><td>61.8</td></tr><tr><td>GeoTrans [27]</td><td>92.0</td><td>91.8</td><td>91.8</td><td>91.4</td><td>91.2</td><td>75.0</td><td>74.8</td><td>74.2</td><td>74.1</td><td>73.5</td></tr><tr><td>RoITr (Ours)</td><td>91.9</td><td>91.7</td><td>91.8</td><td>91.4</td><td>91.0</td><td>74.7</td><td>74.8</td><td>74.8</td><td>74.2</td><td>73.6</td></tr></table>
205
+
206
+ Analysis on the Number of Correspondences. We further analyze the influence of a varying number of correspondences. As illustrated in Tab. 2, RoITr shows outstanding performance on both datasets with various correspondences, proving its stability when only a few correspondences are accessible. The same test on the rotated benchmarks is given in the Appendix.
207
+
208
+ # 4.2. Deformable Objects: 4DMatch & 4DLoMatch
209
+
210
+ Dataset. 4DMatch [20] contains 1,761 animations randomly selected from DeformingThings4D [21]. The 1,761 sequences are divided into 1,232/176/353 as train/val/test, where the test set is further split into 4DMatch and 4DLoMatch based on an overlap ratio threshold of $45\%$ . Metrics. We follow [20] to use two different metrics: (1). Inlier Ratio (IR) which is defined as same as the IR on 3DMatch,
211
+
212
+ Table 2. Quantitative results on 3DMatch & 3DLoMatch with a varying number of points/correspondences. See the results on rotated data in the Appendix.
213
+
214
+ <table><tr><td rowspan="2">Category</td><td rowspan="2">Method</td><td colspan="2">4DMatch</td><td colspan="2">4DLoMatch</td></tr><tr><td>NFMR(%)↑</td><td>IR(%)↑</td><td>NFMR(%)↑</td><td>IR(%)↑</td></tr><tr><td rowspan="2">Scene Flow</td><td>PWC [45]</td><td>21.6</td><td>20.0</td><td>10.0</td><td>7.2</td></tr><tr><td>FLOT [24]</td><td>27.1</td><td>24.9</td><td>15.2</td><td>10.7</td></tr><tr><td rowspan="4">Feature Matching</td><td>Predator [19]</td><td>56.4</td><td>60.4</td><td>32.1</td><td>27.5</td></tr><tr><td>GeoTrans [27]</td><td>83.2</td><td>82.2</td><td>65.4</td><td>63.6</td></tr><tr><td>Lepard [20]</td><td>83.7</td><td>82.7</td><td>66.9</td><td>55.7</td></tr><tr><td>RoITr (Ours)</td><td>83.0</td><td>84.4</td><td>69.4</td><td>67.6</td></tr></table>
215
+
216
+ Table 3. Quantitative results on 4DMatch & 4DLoMatch.
217
+
218
+ but with a different threshold (i.e., $0.04\mathrm{m}$ ); (2). Non-rigid Feature Matching Recall (NFRM) that measures the fraction of ground-truth matches that can be successfully recovered by the putative correspondences.
219
+
220
+ Comparison with the State-of-the-Art. We compare RoITr with 5 baselines, among which PwC [45] and FLOT [24] are scene flow-based methods, while Predator [19], Lepard [20], GeoTrans [27] are based on feature matching. The results shown in Tab. 3 indicate that although our rotation-invariance is mainly designed for rigid scenarios, RoITr could also achieve outstanding performance in the non-rigid matching task, which further confirms the superiority of our model design. Qualitative results are demonstrated in Fig. 7.
221
+
222
+ # 4.3. Ablation Study
223
+
224
+ Local Attention. We first replace our PPFTrans with PointTransformer (PT) [53] in Tab. 4 (a.1), which leads to a sharp performance drop. We then ablate by embedding our PPF-based local coordinates into PT (Tab. 4 (a.2)) and by adopting the relative coordinates, i.e., $\mathbf{p}_j - \mathbf{p}_j$ , used by PT in our PAM (Tab. 4 (a.3)). Our local coordinate representation significantly boosts the performance of PT in the task of point cloud matching and meanwhile makes it rotation-invariant, although its performance is still far behind ours. However, the relative coordinates fail to work in our PAM, as we adopt a more efficient attention mechanism [43] that learns a scalar attention value for each feature $\mathbf{x} \in \mathbb{R}^c$ and is consequently hard to work under varying poses with a
225
+
226
+ ![](images/05f21fec257031ee45502721928229e415a594471edac0c8ac5a61e3a223fa87.jpg)
227
+ Figure 7. Qualitative results of non-rigid matching on 4DLoMatch with Lepard [20] as the baseline. Green/red lines indicate inliers/outliers. See the Appendix for more examples.
228
+
229
+ rotation-sensitive design. As a comparison, PT learns a per-channel vector attention $\mathbf{a} \in \mathbb{R}^c$ for the same feature $\mathbf{x}$ and could deal with the pose variations, but at the cost of the efficiency as shown in Fig. 8. When the number of channels is increased, our advantage in terms of efficiency is enlarged. As we achieve that with more parameters, the gap becomes more significant when runtime is normalized with the number of parameters in the right figure. With our PPF-based local coordinate, the scalar attention could focus on the pose-agnostic pure geometry and therefore achieves the best performance shown in Tab. 4 (a.4).
230
+
231
+ Abstraction Layer. We ablate our Attentional Abstraction Layer (AAL) by replacing it with the pooling-based abstraction design used in [25, 26, 53]. We test the max pooling in Tab. 4 (b.1) and the average pooling in Tab. 4 (b.2), both showing a degrading performance compared with our AAL, which demonstrates our superiority.
232
+
233
+ Backbone. In Tab. 4 (a.1) we have shown our superiority compared with PT [53]. We further replace our PPFTrans with the KPCnv-based backbone network which is widely used in previous deep matchers [19, 27, 50]. The fact that KPCnv falls behind our design demonstrates the advantage of PPFTrans in geometry encoding.
234
+
235
+ Global Transformer. We replace our design with the global transformer of GeoTrans [27] which performs state-of-the-art but without the cross-frame spatial awareness. The dropping results in Tab. 4 (d.1) proves the excellence of our design with the cross-frame position awareness.
236
+
237
+ The Number of Global Transformers. To demonstrate the importance of being globally aware, we first remove the global transformer. The substantial performance drop confirms the significance of global awareness. Then we add one global transformer and observe an increased performance. In our default setting with 3 global transformers, the model performs the best. However, when the number is increased to 5, the model shows a slight performance drop, which we owe to overfitting. As the data augmentation of rotations
238
+
239
+ ![](images/c79f5972814873dd0680a193dcb9720a19fc8606439b7e3d540e215126e359a0.jpg)
240
+ Figure 8. Left: Runtime comparison between our PPF attention Mechanism (PAM) and the local attention in PointTransformer [53]. Right: Runtime normalized by aligning the number of parameters.
241
+
242
+ <table><tr><td rowspan="2">Category</td><td rowspan="2">Model</td><td colspan="3">Origin</td><td colspan="3">Rotated</td></tr><tr><td>FMR</td><td>IR</td><td>RR</td><td>FMR</td><td>IR</td><td>RR</td></tr><tr><td rowspan="4">a. Local</td><td>1. PT [53]</td><td>79.0</td><td>36.5</td><td>61.6</td><td>76.5</td><td>34.7</td><td>60.0</td></tr><tr><td>*2. PPF+PT [53]</td><td>87.0</td><td>49.9</td><td>69.9</td><td>86.8</td><td>49.4</td><td>71.2</td></tr><tr><td>3. Δxyz+Ours</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td></tr><tr><td>*4. Ours</td><td>89.6</td><td>54.3</td><td>74.7</td><td>89.4</td><td>53.2</td><td>77.2</td></tr><tr><td rowspan="3">b. Aggregation</td><td>*1. max pooling</td><td>85.2</td><td>50.1</td><td>70.5</td><td>85.4</td><td>50.2</td><td>71.9</td></tr><tr><td>*2. avg pooling</td><td>87.8</td><td>52.6</td><td>73.8</td><td>87.2</td><td>52.5</td><td>74.7</td></tr><tr><td>*3. Ours</td><td>89.6</td><td>54.3</td><td>74.7</td><td>89.4</td><td>53.2</td><td>77.2</td></tr><tr><td rowspan="2">c. Backbone</td><td>1. KPCov [40]</td><td>85.2</td><td>44.4</td><td>70.6</td><td>83.0</td><td>42.3</td><td>71.5</td></tr><tr><td>*2. Ours</td><td>89.6</td><td>54.3</td><td>74.7</td><td>89.4</td><td>53.2</td><td>77.2</td></tr><tr><td rowspan="2">d. Global</td><td>*1. GeoTrans [27]</td><td>87.7</td><td>53.6</td><td>73.0</td><td>87.5</td><td>53.2</td><td>75.1</td></tr><tr><td>*2. Ours</td><td>89.6</td><td>54.3</td><td>74.7</td><td>89.4</td><td>53.2</td><td>77.2</td></tr><tr><td rowspan="4">e. #Global</td><td>*1. g = 0</td><td>87.2</td><td>37.6</td><td>70.7</td><td>87.5</td><td>37.6</td><td>72.7</td></tr><tr><td>*2. g = 1</td><td>87.1</td><td>42.1</td><td>70.8</td><td>86.8</td><td>42.1</td><td>73.0</td></tr><tr><td>*3. g = 3 (Ours)</td><td>89.6</td><td>54.3</td><td>74.7</td><td>89.4</td><td>53.2</td><td>77.2</td></tr><tr><td>*4. g = 5</td><td>87.1</td><td>52.5</td><td>72.1</td><td>87.0</td><td>52.4</td><td>73.3</td></tr></table>
243
+
244
+ Table 4. Ablation study on (rotated) 3DLoMatch. 5,000 points/correspondences are leveraged. * indicates the methods with intrinsic rotation invariance. See the Appendix for the results on (rotated) 3DMatch.
245
+
246
+ has less effect on an intrinsically rotation-invariant method, more data is required for training a larger model.
247
+
248
+ # 5. Conclusion
249
+
250
+ We introduced RoITr - an intrinsically rotation-invariant model for point cloud matching. We proposed PAM (PPF Attention Mechanism) that embeds PPF-based local coordinates to encode rotation-invariant geometry. This design lies at the core of AAL (Attention Abstraction Layer), PAL (PPF Attention Layer), and TUL (Transition Up Layer) which are consecutively stacked to compose PPFTrans (PPF Transformer) for representative and pose-agnostic geometry description. We further enhanced features by introducing a novel global transformer architecture, which ensures the rotation-invariant cross-frame spatial awareness. Extensive experiments are conducted on both rigid and non-rigid benchmarks to demonstrate the superiority of our approach, especially the remarkable robustness against arbitrary rotations. Limitations are discussed in the Appendix.
251
+
252
+ Acknowledgment. This paper is supported by the National Natural Science Foundation of China under Grant No. 62025208. We appreciate the help from Lennart Bastian, Mert Karaoglu, Ning Liu, and Zhiying Leng.
253
+
254
+ # References
255
+
256
+ [1] Sheng Ao, Qingyong Hu, Bo Yang, Andrew Markham, and Yulan Guo. Spinnet: Learning a general surface descriptor for 3d point cloud registration. In CVPR, 2021. 1, 2, 6, 7
257
+ [2] Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hinton. Layer normalization. arXiv preprint arXiv:1607.06450, 2016. 4
258
+ [3] Xuyang Bai, Zixin Luo, Lei Zhou, Hongbo Fu, Long Quan, and Chiew-Lan Tai. D3feat: Joint learning of dense detection and description of 3d local features. In CVPR, 2020. 2
259
+ [4] Axel Barroso-Laguna, Yannick Verdie, Benjamin Busam, and Krystian Mikolajczyk. Hdd-net: Hybrid detector descriptor with mutual interactive learning. In Proceedings of the Asian Conference on Computer Vision, 2020. 2
260
+ [5] Tolga Birdal and Slobodan Ilic. Point pair features based object detection and pose estimation revisited. In 3DV, 2015. 4
261
+ [6] Christopher Choy, JunYoung Gwak, and Silvio Savarese. 4d spatio-temporal convnets: Minkowski convolutional neural networks. In CVPR, 2019. 1
262
+ [7] Christopher Choy, Jaesik Park, and Vladlen Koltun. Fully convolutional geometric features. In ICCV, 2019. 1, 2
263
+ [8] Chin Seng Chua and Ray Jarvis. Point signatures: A new representation for 3d object recognition. IJCV, 1997. 2
264
+ [9] Haowen Deng, Tolga Birdal, and Slobodan Ilic. Ppf-foldnet: Unsupervised learning of rotation invariant 3d local descriptors. In ECCV, 2018. 1, 2, 4
265
+ [10] Haowen Deng, Tolga Birdal, and Slobodan Ilic. Ppfnet: Global context aware local features for robust 3d point matching. In CVPR, 2018. 1, 2
266
+ [11] Bertram Drost, Markus Ulrich, Nassir Navab, and Slobodan Ilic. Model globally, match locally: Efficient and robust 3d object recognition. In CVPR, 2010. 1, 2, 3, 4
267
+ [12] Martin A Fischler and Robert C Bolles. Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. Communications of the ACM, 1981. 6
268
+ [13] Zan Gojcic, Caifa Zhou, Jan D Wegner, and Andreas Wieser. The perfect match: 3d point cloud matching with smoothed densities. In CVPR, 2019. 1, 2
269
+ [14] Yulan Guo, Ferdous Sohel, Mohammed Bennamoun, Min Lu, and Jianwei Wan. Rotational projection statistics for 3d local surface description and object recognition. IJCV, 2013. 2
270
+ [15] Ji Hou, Benjamin Graham, Matthias Nießner, and Saining Xie. Exploring data-efficient 3d scene understanding with contrastive scene contexts. In CVPR, 2021. 1
271
+ [16] Ji Hou, Saining Xie, Benjamin Graham, Angela Dai, and Matthias Nießner. Pri3d: Can 3d priors help 2d representation learning? In ICCV, 2021. 1
272
+ [17] Chun-Hao Huang, Benjamin Allain, Jean-Sébastien Franco, Nassir Navab, Slobodan Ilic, and Edmond Boyer. Volumetric 3d tracking by detection. In CVPR, 2016. 1
273
+ [18] Chun-Hao Paul Huang, Benjamin Allain, Edmond Boyer, Jean-Sébastien Franco, Federico Tombari, Nassir Navab, and Slobodan Ilic. Tracking-by-detection of 3d human shapes: from surfaces to volumes. TPAMI, 2017. 1
274
+
275
+ [19] Shengyu Huang, Zan Gojcic, Mikhail Usvyatsov, Andreas Wieser, and Konrad Schindler. Predator: Registration of 3d point clouds with low overlap. In CVPR, 2021. 1, 2, 5, 6, 7, 8
276
+ [20] Yang Li and Tatsuya Harada. Lepard: Learning partial point cloud matching in rigid and deformable scenes. In CVPR, 2022. 1, 2, 6, 7, 8
277
+ [21] Yang Li, Hikari Takehara, Takafumi Taketomi, Bo Zheng, and Matthias Nießner. 4dcomplete: Non-rigid motion estimation beyond the observable surface. In ICCV, 2021. 7
278
+ [22] Richard A Newcombe, Dieter Fox, and Steven M Seitz. Dynamicfusion: Reconstruction and tracking of non-rigid scenes in real-time. In CVPR, 2015. 1
279
+ [23] Richard A Newcombe, Shahram Izadi, Otmar Hilliges, David Molyneaux, David Kim, Andrew J Davison, Pushmeet Kohi, Jamie Shotton, Steve Hodges, and Andrew Fitzgibbon. Kinectfusion: Real-time dense surface mapping and tracking. In IEEE International Symposium on Mixed and Augmented Reality, 2011. 1
280
+ [24] Gilles Puy, Alexandre Boulch, and Renaud Marlet. Flot: Scene flow on point clouds guided by optimal transport. In ECCV, 2020. 7
281
+ [25] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In CVPR, 2017. 1, 2, 8
282
+ [26] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. NeurIPS, 2017. 1, 4, 8
283
+ [27] Zheng Qin, Hao Yu, Changjian Wang, Yulan Guo, Yuxing Peng, and Kai Xu. Geometric transformer for fast and robust point cloud registration. In CVPR, 2022. 1, 2, 3, 5, 6, 7, 8
284
+ [28] Zheng Qin, Hao Yu, Changjian Wang, Yuxing Peng, and Kai Xu. Deep graph-based spatial consistency for robust non-rigid point cloud registration. arXiv preprint arXiv:2303.09950, 2023. 2
285
+ [29] Ignacio Rocco, Mircea Cimpoi, Relja Arandjelović, Akihiko Torii, Tomas Pajdla, and Josef Sivic. Neighbourhood consensus networks. *NeurIPS*, 2018. 6
286
+ [30] Radu Bogdan Rusu, Nico Blodow, and Michael Beetz. Fast point feature histograms (fpfh) for 3d registration. In ICRA, 2009. 1, 2
287
+ [31] Radu Bogdan Rusu, Nico Blodow, Zoltan Csaba Marton, and Michael Beetz. Aligning point cloud views using persistent feature histograms. In IROS, 2008. 1, 2
288
+ [32] Mahdi Saleh, Shervin Dehghani, Benjamin Busam, Nassir Navab, and Federico Tombari. Graphite: Graph-induced feature extraction for point cloud registration. In 3DV, 2020. 1, 2
289
+ [33] Mahdi Saleh, Yige Wang, Nassir Navab, Benjamin Busam, and Federico Tombari. Cloudattention: Efficient multi-scale attention scheme for 3d point cloud learning. arXiv preprint arXiv:2208.00524, 2022. 1
290
+ [34] Mahdi Saleh, Shun-Cheng Wu, Luca Cosmo, Nassir Navab, Benjamin Busam, and Federico Tombari. Bending graphs: Hierarchical shape matching using gated optimal transport. In CVPR, 2022. 1, 2
291
+
292
+ [35] Paul-Edouard Sarlin, Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superglue: Learning feature matching with graph neural networks. In CVPR, 2020. 6
293
+ [36] Richard Sinkhorn and Paul Knopp. Concerning nonnegative matrices and doubly stochastic matrices. Pacific Journal of Mathematics, 1967. 6
294
+ [37] Jiaming Sun, Zehong Shen, Yuang Wang, Hujun Bao, and Xiaowei Zhou. Loftr: Detector-free local feature matching with transformers. In CVPR, 2021. 6
295
+ [38] Jiapeng Tang, Lev Markhasin, Bi Wang, Justus Thies, and Matthias Nießner. Neural shape deformation priors. arXiv preprint arXiv:2210.05616, 2022. 2
296
+ [39] Jiapeng Tang, Dan Xu, Kui Jia, and Lei Zhang. Learning parallel dense correspondence from spatio-temporal descriptors for efficient and robust 4d reconstruction. In CVPR, 2021. 1
297
+ [40] Hugues Thomas, Charles R Qi, Jean-Emmanuel Deschaud, Beatrix Marcotegui, François Goulette, and Leonidas J Guibas. Kpconv: Flexible and deformable convolution for point clouds. In ICCV, 2019. 1, 8
298
+ [41] Federico Tombari, Samuele Salti, and Luigi Di Stefano. Unique signatures of histograms for local surface description. In ECCV, 2010. 1, 2
299
+ [42] Giovanni Trappolini, Luca Cosmo, Luca Moschella, Ricardo Marin, Simone Melzi, and Emanuele Rodola. Shape registration in the time of transformers. NeurIPS, 2021. 2
300
+ [43] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. NeurIPS, 2017. 3, 5, 7
301
+ [44] Haiping Wang, Yuan Liu, Zhen Dong, and Wenping Wang. You only hypothesize once: Point cloud registration with rotation-equivariant descriptors. In ACM MM, 2022. 1, 2, 6, 7
302
+ [45] Wenxuan Wu, Zhiyuan Wang, Zhuwen Li, Wei Liu, and Li Fuxin. Pointpwc-net: A coarse-to-fine network for supervised and self-supervised scene flow estimation on 3d point clouds. arXiv preprint arXiv:1911.12408, 2019. 7
303
+ [46] Saining Xie, Jiatao Gu, Demi Guo, Charles R Qi, Leonidas Guibas, and Or Litany. Pointcontrast: Unsupervised pretraining for 3d point cloud understanding. In ECCV, 2020. 1
304
+ [47] Yaoqing Yang, Chen Feng, Yiru Shen, and Dong Tian. Foldingnet: Point cloud auto-encoder via deep grid deformation. In CVPR, 2018. 2
305
+ [48] Zi Jian Yew and Gim Hee Lee. Regtr: End-to-end point cloud correspondences with transformers. In CVPR, 2022. 1, 2, 5
306
+ [49] Hao Yu, Ji Hou, Zheng Qin, Mahdi Saleh, Ivan Shugurov, Kai Wang, Benjamin Busam, and Slobodan Ilic. Riga: Rotation-invariant and globally-aware descriptors for point cloud registration. arXiv preprint arXiv:2209.13252, 2022. 1, 2, 5, 6, 7
307
+ [50] Hao Yu, Fu Li, Mahdi Saleh, Benjamin Busam, and Slobodan Ilic. Cofinet: Reliable coarse-to-fine correspondences for robust pointcloud registration. NeurIPS, 2021. 1, 2, 3, 5, 6, 7, 8
308
+
309
+ [51] Andy Zeng, Shuran Song, Matthias Nießner, Matthew Fisher, Jianxiong Xiao, and Thomas Funkhouser. 3dmatch: Learning local geometric descriptors from rgb-d reconstructions. In CVPR, 2017. 1, 2, 6
310
+ [52] Yu Zhang, Junle Yu, Xiaolin Huang, Wenhui Zhou, and Ji Hou. Pcr-cg: Point cloud registration via deep explicit color and geometry. In ECCV, 2022. 1
311
+ [53] Hengshuang Zhao, Li Jiang, Jiaya Jia, Philip HS Torr, and Vladlen Koltun. Point transformer. In ICCV, 2021. 1, 2, 7, 8
2023/Rotation-Invariant Transformer for Point Cloud Matching/images.zip ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:edea9ea66ad35ff734125a192dbdbc47b7ba617a136267678ea0346d59f66b46
3
+ size 761189
2023/Rotation-Invariant Transformer for Point Cloud Matching/layout.json ADDED
The diff for this file is too large to render. See raw diff
 
2023/Run, Don't Walk_ Chasing Higher FLOPS for Faster Neural Networks/6a2198d1-cecd-4fd5-b373-fd2aeb690abb_content_list.json ADDED
@@ -0,0 +1,1585 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ [
2
+ {
3
+ "type": "text",
4
+ "text": "Run, Don’t Walk: Chasing Higher FLOPS for Faster Neural Networks",
5
+ "text_level": 1,
6
+ "bbox": [
7
+ 127,
8
+ 130,
9
+ 841,
10
+ 151
11
+ ],
12
+ "page_idx": 0
13
+ },
14
+ {
15
+ "type": "text",
16
+ "text": "Jierun Chen $^{1}$ , Shiu-hong Kao $^{1}$ , Hao He $^{1}$ \nWeipeng Zhuo $^{1}$ , Song Wen $^{2}$ , Chul-Ho Lee $^{3}$ , S.-H. Gary Chan $^{1}$ $^{1}$ HKUST, $^{2}$ Rutgers University, $^{3}$ Texas State University",
17
+ "bbox": [
18
+ 236,
19
+ 179,
20
+ 728,
21
+ 234
22
+ ],
23
+ "page_idx": 0
24
+ },
25
+ {
26
+ "type": "text",
27
+ "text": "Abstract",
28
+ "text_level": 1,
29
+ "bbox": [
30
+ 233,
31
+ 268,
32
+ 313,
33
+ 284
34
+ ],
35
+ "page_idx": 0
36
+ },
37
+ {
38
+ "type": "text",
39
+ "text": "To design fast neural networks, many works have been focusing on reducing the number of floating-point operations (FLOPs). We observe that such reduction in FLOPs, however, does not necessarily lead to a similar level of reduction in latency. This mainly stems from inefficiently low floating-point operations per second (FLOPS). To achieve faster networks, we revisit popular operators and demonstrate that such low FLOPS is mainly due to frequent memory access of the operators, especially the depthwise convolution. We hence propose a novel partial convolution (PCnv) that extracts spatial features more efficiently, by cutting down redundant computation and memory access simultaneously. Building upon our PCnv, we further propose FasterNet, a new family of neural networks, which attains substantially higher running speed than others on a wide range of devices, without compromising on accuracy for various vision tasks. For example, on ImageNet-1k, our tiny FasterNet-T0 is $2.8 \\times 3.3 \\times$ and $2.4 \\times$ faster than MobileViT-XXS on GPU, CPU, and ARM processors, respectively, while being $2.9\\%$ more accurate. Our large FasterNet-L achieves impressive $83.5\\%$ top-1 accuracy, on par with the emerging Swin-B, while having $36\\%$ higher inference throughput on GPU, as well as saving $37\\%$ compute time on CPU. Code is available at https://github.com/JierunChen/FasterNet.",
40
+ "bbox": [
41
+ 75,
42
+ 300,
43
+ 473,
44
+ 679
45
+ ],
46
+ "page_idx": 0
47
+ },
48
+ {
49
+ "type": "text",
50
+ "text": "1. Introduction",
51
+ "text_level": 1,
52
+ "bbox": [
53
+ 76,
54
+ 708,
55
+ 209,
56
+ 723
57
+ ],
58
+ "page_idx": 0
59
+ },
60
+ {
61
+ "type": "text",
62
+ "text": "Neural networks have undergone rapid development in various computer vision tasks such as image classification, detection and segmentation. While their impressive performance has powered many applications, a roaring trend is to pursue fast neural networks with low latency and high throughput for great user experiences, instant responses, safety reasons, etc.",
63
+ "bbox": [
64
+ 75,
65
+ 733,
66
+ 468,
67
+ 839
68
+ ],
69
+ "page_idx": 0
70
+ },
71
+ {
72
+ "type": "text",
73
+ "text": "How to be fast? Instead of asking for more costly computing devices, researchers and practitioners prefer to design cost-effective fast neural networks with reduced computational complexity, mainly measured in the number of",
74
+ "bbox": [
75
+ 75,
76
+ 839,
77
+ 470,
78
+ 900
79
+ ],
80
+ "page_idx": 0
81
+ },
82
+ {
83
+ "type": "image",
84
+ "img_path": "images/c8cfeebc3fde8d06bfd86b330dae02078ebc45c79a1a14f32c35eaaaa1393820.jpg",
85
+ "image_caption": [
86
+ "(a) Convolution"
87
+ ],
88
+ "image_footnote": [],
89
+ "bbox": [
90
+ 501,
91
+ 267,
92
+ 692,
93
+ 321
94
+ ],
95
+ "page_idx": 0
96
+ },
97
+ {
98
+ "type": "image",
99
+ "img_path": "images/1d015ffc47d7943c325b92090435436c20108bf9c1713ea630e73d7b6ad1b0b6.jpg",
100
+ "image_caption": [
101
+ "(b) Depthwise/Group Convolution"
102
+ ],
103
+ "image_footnote": [],
104
+ "bbox": [
105
+ 700,
106
+ 267,
107
+ 890,
108
+ 321
109
+ ],
110
+ "page_idx": 0
111
+ },
112
+ {
113
+ "type": "image",
114
+ "img_path": "images/ef34393040525a2e7dacb9ed8654cc27ceb71d56a7f92940dd88d05170faf8d2.jpg",
115
+ "image_caption": [
116
+ "(c) Partial Convolution (ours)",
117
+ "Figure 1. Our partial convolution (PConv) is fast and efficient by applying filters on only a few input channels while leaving the remaining ones untouched. PConv obtains lower FLOPs than the regular convolution and higher FLOPS than the depthwise/group convolution."
118
+ ],
119
+ "image_footnote": [],
120
+ "bbox": [
121
+ 547,
122
+ 344,
123
+ 736,
124
+ 392
125
+ ],
126
+ "page_idx": 0
127
+ },
128
+ {
129
+ "type": "list",
130
+ "sub_type": "text",
131
+ "list_items": [
132
+ "Input/output",
133
+ "Filter",
134
+ "\\*Convolution",
135
+ "$\\rightarrow$ Identity"
136
+ ],
137
+ "bbox": [
138
+ 769,
139
+ 359,
140
+ 852,
141
+ 402
142
+ ],
143
+ "page_idx": 0
144
+ },
145
+ {
146
+ "type": "text",
147
+ "text": "floating-point operations (FLOPs) $^1$ . MobileNets [20, 21, 47], ShuffleNets [40, 76] and GhostNet [13], among others, leverage the depthwise convolution (DWConv) [48] and/or group convolution (GConv) [27] to extract spatial features. However, in the effort to reduce FLOPs, the operators often suffer from the side effect of increased memory access. MicroNet [29] further decomposes and sparsifies the network to push its FLOPs to an extremely low level. Despite its improvement in FLOPs, this approach experiences inefficient fragmented computation. Besides, the above networks are often accompanied by additional data manipulations, such as concatenation, shuffling, and pooling, whose running time tends to be significant for tiny models.",
148
+ "bbox": [
149
+ 496,
150
+ 508,
151
+ 890,
152
+ 705
153
+ ],
154
+ "page_idx": 0
155
+ },
156
+ {
157
+ "type": "text",
158
+ "text": "Apart from the above pure convolutional neural networks (CNNs), there is an emerging interest in making vision transformers (ViTs) [11] and multilayer perceptrons (MLPs) architectures [57] smaller and faster. For example, MobileViTs [42, 43, 63] and MobileFormer [6] reduce the computational complexity by combining DWConv with a modified attention mechanism. However, they still suffer from the aforementioned issue with DWConv and also need dedicated hardware support for the modified attention mechanism. The use of advanced yet time-consuming nor",
159
+ "bbox": [
160
+ 496,
161
+ 707,
162
+ 890,
163
+ 859
164
+ ],
165
+ "page_idx": 0
166
+ },
167
+ {
168
+ "type": "header",
169
+ "text": "CVF",
170
+ "bbox": [
171
+ 106,
172
+ 2,
173
+ 181,
174
+ 42
175
+ ],
176
+ "page_idx": 0
177
+ },
178
+ {
179
+ "type": "header",
180
+ "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.",
181
+ "bbox": [
182
+ 236,
183
+ 0,
184
+ 810,
185
+ 44
186
+ ],
187
+ "page_idx": 0
188
+ },
189
+ {
190
+ "type": "page_footnote",
191
+ "text": "1We follow a widely adopted definition of FLOPs, as the number of multiply-adds [36, 76].",
192
+ "bbox": [
193
+ 500,
194
+ 875,
195
+ 893,
196
+ 900
197
+ ],
198
+ "page_idx": 0
199
+ },
200
+ {
201
+ "type": "page_number",
202
+ "text": "12021",
203
+ "bbox": [
204
+ 480,
205
+ 924,
206
+ 517,
207
+ 936
208
+ ],
209
+ "page_idx": 0
210
+ },
211
+ {
212
+ "type": "image",
213
+ "img_path": "images/bc689bdb3a1a6f8205702f0379a1f5c1f8be0e664dd4c0e302b8927ca1ec1b3b.jpg",
214
+ "image_caption": [],
215
+ "image_footnote": [],
216
+ "bbox": [
217
+ 81,
218
+ 89,
219
+ 202,
220
+ 239
221
+ ],
222
+ "page_idx": 1
223
+ },
224
+ {
225
+ "type": "image",
226
+ "img_path": "images/bcfdc95e5623ea475b3f4225deff2acec6dc75fdab39cada93fdc6eb8e26e653.jpg",
227
+ "image_caption": [
228
+ "(a)"
229
+ ],
230
+ "image_footnote": [],
231
+ "bbox": [
232
+ 205,
233
+ 88,
234
+ 436,
235
+ 260
236
+ ],
237
+ "page_idx": 1
238
+ },
239
+ {
240
+ "type": "image",
241
+ "img_path": "images/52013b4b7475c1d42f589de8b8dc903a733f7e78b03a0098e6384b5eb0496566.jpg",
242
+ "image_caption": [],
243
+ "image_footnote": [],
244
+ "bbox": [
245
+ 446,
246
+ 90,
247
+ 648,
248
+ 241
249
+ ],
250
+ "page_idx": 1
251
+ },
252
+ {
253
+ "type": "image",
254
+ "img_path": "images/173d56cec53757d48dc7e1af33b12d573a1492f15e75fd81fcccb8649b4d37eb.jpg",
255
+ "image_caption": [
256
+ "(b)",
257
+ "Figure 2. (a) FLOPS under varied FLOPs on CPU. Many existing neural networks suffer from low computational speed issues. Their effective FLOPS are lower than the popular ResNet50. By contrast, our FasterNet attains higher FLOPS. (b) Latency under varied FLOPs on CPU. Our FasterNet obtains lower latency than others with the same amount of FLOPs."
258
+ ],
259
+ "image_footnote": [],
260
+ "bbox": [
261
+ 648,
262
+ 90,
263
+ 890,
264
+ 262
265
+ ],
266
+ "page_idx": 1
267
+ },
268
+ {
269
+ "type": "text",
270
+ "text": "malization and activation layers may also limit their speed on devices.",
271
+ "bbox": [
272
+ 75,
273
+ 337,
274
+ 468,
275
+ 366
276
+ ],
277
+ "page_idx": 1
278
+ },
279
+ {
280
+ "type": "text",
281
+ "text": "All these issues together lead to the following question: Are these \"fast\" neural networks really fast? To answer this, we examine the relationship between latency and FLOPs, which is captured by",
282
+ "bbox": [
283
+ 75,
284
+ 367,
285
+ 468,
286
+ 428
287
+ ],
288
+ "page_idx": 1
289
+ },
290
+ {
291
+ "type": "equation",
292
+ "text": "\n$$\nL a t e n c y = \\frac {F L O P s}{F L O P S}, \\tag {1}\n$$\n",
293
+ "text_format": "latex",
294
+ "bbox": [
295
+ 194,
296
+ 438,
297
+ 468,
298
+ 469
299
+ ],
300
+ "page_idx": 1
301
+ },
302
+ {
303
+ "type": "text",
304
+ "text": "where FLOPS is short for floating-point operations per second, as a measure of the effective computational speed. While there are many attempts to reduce FLOPs, they seldom consider optimizing FLOPS at the same time to achieve truly low latency. To better understand the situation, we compare the FLOPS of typical neural networks on an Intel CPU. The results in Fig. 2 show that many existing neural networks suffer from low FLOPS, and their FLOPS is generally lower than the popular ResNet50. With such low FLOPS, these \"fast\" neural networks are actually not fast enough. Their reduction in FLOPs cannot be translated into the exact amount of reduction in latency. In some cases, there is no improvement, and it even leads to worse latency. For example, CycleMLP-B1 [5] has half of FLOPs of ResNet50 [16] but runs more slowly (i.e., CycleMLP-B1 vs. ResNet50: 116.1ms vs. 73.0ms). Note that this discrepancy between FLOPs and latency has also been noticed in previous works [40, 42] but remains unresolved partially because they employ the DWConv/GConv and various data manipulations with low FLOPS. It is deemed there are no better alternatives available.",
305
+ "bbox": [
306
+ 75,
307
+ 477,
308
+ 468,
309
+ 792
310
+ ],
311
+ "page_idx": 1
312
+ },
313
+ {
314
+ "type": "text",
315
+ "text": "This paper aims to eliminate the discrepancy by developing a simple yet fast and effective operator that maintains high FLOPS with reduced FLOPs. Specifically, we reexamine existing operators, particularly DWConv, in terms of the computational speed - FLOPS. We uncover that the main reason causing the low FLOPS issue is frequent memory access. We then propose a novel partial convolution (PCov)",
316
+ "bbox": [
317
+ 75,
318
+ 795,
319
+ 468,
320
+ 901
321
+ ],
322
+ "page_idx": 1
323
+ },
324
+ {
325
+ "type": "text",
326
+ "text": "as a competitive alternative that reduces the computational redundancy as well as the number of memory access. Fig. 1 illustrates the design of our PConv. It takes advantage of redundancy within the feature maps and systematically applies a regular convolution (Conv) on only a part of the input channels while leaving the remaining ones untouched. By nature, PConv has lower FLOPs than the regular Conv while having higher FLOPS than the DWConv/GConv. In other words, PConv better exploits the on-device computational capacity. PConv is also effective in extracting spatial features as empirically validated later in the paper.",
327
+ "bbox": [
328
+ 496,
329
+ 337,
330
+ 890,
331
+ 502
332
+ ],
333
+ "page_idx": 1
334
+ },
335
+ {
336
+ "type": "text",
337
+ "text": "We further introduce FasterNet, which is primarily built upon our PConv, as a new family of networks that run highly fast on various devices. In particular, our FasterNet achieves state-of-the-art performance for classification, detection, and segmentation tasks while having much lower latency and higher throughput. For example, our tiny FasterNet-T0 is $2.8 \\times 3.3 \\times$ and $2.4 \\times$ faster than MobileViT-XXS [42] on GPU, CPU, and ARM processors, respectively, while being $2.9\\%$ more accurate on ImageNet-1k. Our large FasterNet-L achieves $83.5\\%$ top-1 accuracy, on par with the emerging Swin-B [35], while offering $36\\%$ higher throughput on GPU and saving $37\\%$ compute time on CPU. To summarize, our contributions are as follows:",
338
+ "bbox": [
339
+ 496,
340
+ 503,
341
+ 892,
342
+ 699
343
+ ],
344
+ "page_idx": 1
345
+ },
346
+ {
347
+ "type": "list",
348
+ "sub_type": "text",
349
+ "list_items": [
350
+ "- We point out the importance of achieving higher FLOPS beyond simply reducing FLOPs for faster neural networks.",
351
+ "- We introduce a simple yet fast and effective operator called PConv, which has a high potential to replace the existing go-to choice, DWConv.",
352
+ "- We introduce FasterNet which runs favorably and universally fast on a variety of devices such as GPU, CPU, and ARM processors.",
353
+ "- We conduct extensive experiments on various tasks and validate the high speed and effectiveness of our PConv and FasterNet."
354
+ ],
355
+ "bbox": [
356
+ 517,
357
+ 705,
358
+ 890,
359
+ 898
360
+ ],
361
+ "page_idx": 1
362
+ },
363
+ {
364
+ "type": "page_number",
365
+ "text": "12022",
366
+ "bbox": [
367
+ 478,
368
+ 924,
369
+ 517,
370
+ 936
371
+ ],
372
+ "page_idx": 1
373
+ },
374
+ {
375
+ "type": "text",
376
+ "text": "2. Related Work",
377
+ "text_level": 1,
378
+ "bbox": [
379
+ 76,
380
+ 89,
381
+ 218,
382
+ 104
383
+ ],
384
+ "page_idx": 2
385
+ },
386
+ {
387
+ "type": "text",
388
+ "text": "We briefly review prior works on fast and efficient neural networks and differentiate this work from them.",
389
+ "bbox": [
390
+ 76,
391
+ 114,
392
+ 468,
393
+ 145
394
+ ],
395
+ "page_idx": 2
396
+ },
397
+ {
398
+ "type": "text",
399
+ "text": "CNN. CNNs are the mainstream architecture in the computer vision field, especially when it comes to deployment in practice, where being fast is as important as being accurate. Though there have been numerous studies [7, 8, 17, 29, 48, 49, 75, 78] to achieve higher efficiency, the rationale behind them is more or less to perform a low-rank approximation. Specifically, the group convolution [27] and the depthwise separable convolution [48] (consisting of depthwise and pointwise convolutions) are probably the most popular ones. They have been widely adopted in mobile/edge-oriented networks, such as MobileNets [20, 21, 47], ShuffleNets [40, 76], GhostNet [13], EfficientNets [54, 55], TinyNet [14], Xception [8], CondenseNet [23, 70], TVConv [4], MnasNet [53], and FBNet [67]. While they exploit the redundancy in filters to reduce the number of parameters and FLOPs, they suffer from increased memory access when increasing the network width to compensate for the accuracy drop. By contrast, we consider the redundancy in feature maps and propose a partial convolution to reduce FLOPs and memory access simultaneously.",
400
+ "bbox": [
401
+ 76,
402
+ 155,
403
+ 472,
404
+ 473
405
+ ],
406
+ "page_idx": 2
407
+ },
408
+ {
409
+ "type": "text",
410
+ "text": "ViT, MLP, and variants. There is a growing interest in studying ViT ever since Dosovitskiy et al. [11] expanded the application scope of transformers [62] from machine translation [62] or forecasting [66] to the computer vision field. Many follow-up works have attempted to improve ViT in terms of training setting [51, 58, 59] and model design [12, 34, 35, 65, 77]. One notable trend is to pursue a better accuracy-latency trade-off by reducing the complexity of the attention operator [1, 25, 39, 56, 61], incorporating convolution into ViTs [6, 10, 50], or doing both [3, 30, 43, 45]. Besides, other studies [5, 31, 57] propose to replace the attention with simple MLP-based operators. However, they often evolve to be CNN-like [33]. In this paper, we focus on analyzing the convolution operations, particularly DWConv, due to the following reasons: First, the advantage of attention over convolution is unclear or debatable [36, 64]. Second, the attention-based mechanism generally runs slower than its convolutional counterparts and thus becomes less favorable for the current industry [22, 42]. Finally, DW-Conv is still a popular choice in many hybrid models, so it is worth a careful examination.",
411
+ "bbox": [
412
+ 76,
413
+ 481,
414
+ 472,
415
+ 797
416
+ ],
417
+ "page_idx": 2
418
+ },
419
+ {
420
+ "type": "text",
421
+ "text": "3. Design of PConv and FasterNet",
422
+ "text_level": 1,
423
+ "bbox": [
424
+ 76,
425
+ 814,
426
+ 364,
427
+ 830
428
+ ],
429
+ "page_idx": 2
430
+ },
431
+ {
432
+ "type": "text",
433
+ "text": "In this section, we first revisit DWConv and analyze the issue with its frequent memory access. We then introduce PConv as a competitive alternative operator to resolve the issue. After that, we introduce FasterNet and explain its",
434
+ "bbox": [
435
+ 76,
436
+ 839,
437
+ 468,
438
+ 900
439
+ ],
440
+ "page_idx": 2
441
+ },
442
+ {
443
+ "type": "image",
444
+ "img_path": "images/43b8b037deb68c10cea574d7d56f635729de216ab6d0f26d1b74b15ba6828430.jpg",
445
+ "image_caption": [
446
+ "Figure 3. Visualization of feature maps in an intermediate layer of a pre-trained ResNet50, with the top-left image as the input. Qualitatively, we can see the high redundancies across different channels."
447
+ ],
448
+ "image_footnote": [],
449
+ "bbox": [
450
+ 506,
451
+ 90,
452
+ 887,
453
+ 239
454
+ ],
455
+ "page_idx": 2
456
+ },
457
+ {
458
+ "type": "text",
459
+ "text": "details, including design considerations.",
460
+ "bbox": [
461
+ 500,
462
+ 320,
463
+ 764,
464
+ 335
465
+ ],
466
+ "page_idx": 2
467
+ },
468
+ {
469
+ "type": "text",
470
+ "text": "3.1. Preliminary",
471
+ "text_level": 1,
472
+ "bbox": [
473
+ 500,
474
+ 344,
475
+ 629,
476
+ 361
477
+ ],
478
+ "page_idx": 2
479
+ },
480
+ {
481
+ "type": "text",
482
+ "text": "DWConv is a popular variant of Conv and has been widely adopted as a key building block for many neural networks. For an input $\\mathbf{I} \\in \\mathbb{R}^{c \\times h \\times w}$ , DWConv applies $c$ filters $\\mathbf{W} \\in \\mathbb{R}^{k \\times k}$ to compute the output $\\mathbf{O} \\in \\mathbb{R}^{c \\times h \\times w}$ . As shown in Fig. 1(b), each filter slides spatially on one input channel and contributes to one output channel. This depthwise computation makes DWConv have as low FLOPs as $h \\times w \\times k^2 \\times c$ compared to a regular Conv with $h \\times w \\times k^2 \\times c^2$ . While effective in reducing FLOPs, a DWConv, which is typically followed by a pointwise convolution, or PWConv, cannot be simply used to replace a regular Conv as it would incur a severe accuracy drop. Thus, in practice the channel number $c$ (or the network width) of DWConv is increased to $c'(c' > c)$ to compensate the accuracy drop, e.g., the width is expanded by six times for the DWConv in the inverted residual blocks [47]. This, however, results in much higher memory access that can cause non-negligible delay and slow down the overall computation, especially for I/O-bound devices. In particular, the number of memory access now escalates to",
483
+ "bbox": [
484
+ 496,
485
+ 367,
486
+ 893,
487
+ 671
488
+ ],
489
+ "page_idx": 2
490
+ },
491
+ {
492
+ "type": "equation",
493
+ "text": "\n$$\nh \\times w \\times 2 c ^ {\\prime} + k ^ {2} \\times c ^ {\\prime} \\approx h \\times w \\times 2 c ^ {\\prime}, \\tag {2}\n$$\n",
494
+ "text_format": "latex",
495
+ "bbox": [
496
+ 563,
497
+ 680,
498
+ 890,
499
+ 698
500
+ ],
501
+ "page_idx": 2
502
+ },
503
+ {
504
+ "type": "text",
505
+ "text": "which is higher than that of a regular Conv, i.e.,",
506
+ "bbox": [
507
+ 500,
508
+ 708,
509
+ 815,
510
+ 723
511
+ ],
512
+ "page_idx": 2
513
+ },
514
+ {
515
+ "type": "equation",
516
+ "text": "\n$$\nh \\times w \\times 2 c + k ^ {2} \\times c ^ {2} \\approx h \\times w \\times 2 c. \\tag {3}\n$$\n",
517
+ "text_format": "latex",
518
+ "bbox": [
519
+ 566,
520
+ 734,
521
+ 890,
522
+ 751
523
+ ],
524
+ "page_idx": 2
525
+ },
526
+ {
527
+ "type": "text",
528
+ "text": "Note that the $h \\times w \\times 2c'$ memory access is spent on the I/O operation, which is deemed to be already the minimum cost and hard to optimize further.",
529
+ "bbox": [
530
+ 498,
531
+ 762,
532
+ 893,
533
+ 808
534
+ ],
535
+ "page_idx": 2
536
+ },
537
+ {
538
+ "type": "text",
539
+ "text": "3.2. Partial convolution as a basic operator",
540
+ "text_level": 1,
541
+ "bbox": [
542
+ 500,
543
+ 816,
544
+ 831,
545
+ 834
546
+ ],
547
+ "page_idx": 2
548
+ },
549
+ {
550
+ "type": "text",
551
+ "text": "We below demonstrate that the cost can be further optimized by leveraging the feature maps' redundancy. As visualized in Fig. 3, the feature maps share high similarities among different channels. This redundancy has also",
552
+ "bbox": [
553
+ 498,
554
+ 840,
555
+ 893,
556
+ 900
557
+ ],
558
+ "page_idx": 2
559
+ },
560
+ {
561
+ "type": "page_number",
562
+ "text": "12023",
563
+ "bbox": [
564
+ 480,
565
+ 925,
566
+ 517,
567
+ 936
568
+ ],
569
+ "page_idx": 2
570
+ },
571
+ {
572
+ "type": "text",
573
+ "text": "been covered in many other works [13, 74], but few of them make full use of it in a simple yet effective way.",
574
+ "bbox": [
575
+ 75,
576
+ 90,
577
+ 468,
578
+ 119
579
+ ],
580
+ "page_idx": 3
581
+ },
582
+ {
583
+ "type": "text",
584
+ "text": "Specifically, we propose a simple PConv to reduce computational redundancy and memory access simultaneously. The bottom-left corner in Fig. 4 illustrates how our PConv works. It simply applies a regular Conv on only a part of the input channels for spatial feature extraction and leaves the remaining channels untouched. For contiguous or regular memory access, we consider the first or last consecutive $c_p$ channels as the representatives of the whole feature maps for computation. Without loss of generality, we consider the input and output feature maps to have the same number of channels. Therefore, the FLOPs of a PConv are only",
585
+ "bbox": [
586
+ 75,
587
+ 121,
588
+ 470,
589
+ 287
590
+ ],
591
+ "page_idx": 3
592
+ },
593
+ {
594
+ "type": "equation",
595
+ "text": "\n$$\nh \\times w \\times k ^ {2} \\times c _ {p} ^ {2}. \\tag {4}\n$$\n",
596
+ "text_format": "latex",
597
+ "bbox": [
598
+ 210,
599
+ 297,
600
+ 468,
601
+ 319
602
+ ],
603
+ "page_idx": 3
604
+ },
605
+ {
606
+ "type": "text",
607
+ "text": "With a typical partial ratio $r = \\frac{c_p}{c} = \\frac{1}{4}$ , the FLOPs of a PConv is only $\\frac{1}{16}$ of a regular Conv. Besides, PConv has a smaller amount of memory access, i.e.,",
608
+ "bbox": [
609
+ 76,
610
+ 330,
611
+ 468,
612
+ 377
613
+ ],
614
+ "page_idx": 3
615
+ },
616
+ {
617
+ "type": "equation",
618
+ "text": "\n$$\nh \\times w \\times 2 c _ {p} + k ^ {2} \\times c _ {p} ^ {2} \\approx h \\times w \\times 2 c _ {p}, \\tag {5}\n$$\n",
619
+ "text_format": "latex",
620
+ "bbox": [
621
+ 135,
622
+ 387,
623
+ 468,
624
+ 407
625
+ ],
626
+ "page_idx": 3
627
+ },
628
+ {
629
+ "type": "text",
630
+ "text": "which is only $\\frac{1}{4}$ of a regular Conv for $r = \\frac{1}{4}$ .",
631
+ "bbox": [
632
+ 76,
633
+ 417,
634
+ 372,
635
+ 434
636
+ ],
637
+ "page_idx": 3
638
+ },
639
+ {
640
+ "type": "text",
641
+ "text": "Since there are only $c_p$ channels utilized for spatial feature extraction, one may ask if we can simply remove the remaining $(c - c_p)$ channels? If so, PConv would degrade to a regular Conv with fewer channels, which deviates from our objective to reduce redundancy. Note that we keep the remaining channels untouched instead of removing them from the feature maps. It is because they are useful for a subsequent PWConv layer, which allows the feature information to flow through all channels.",
642
+ "bbox": [
643
+ 75,
644
+ 435,
645
+ 468,
646
+ 570
647
+ ],
648
+ "page_idx": 3
649
+ },
650
+ {
651
+ "type": "text",
652
+ "text": "3.3. PConv followed by PWConv",
653
+ "text_level": 1,
654
+ "bbox": [
655
+ 76,
656
+ 580,
657
+ 331,
658
+ 595
659
+ ],
660
+ "page_idx": 3
661
+ },
662
+ {
663
+ "type": "text",
664
+ "text": "To fully and efficiently leverage the information from all channels, we further append a pointwise convolution (PW-Conv) to our PConv. Their effective receptive field together on the input feature maps looks like a T-shaped Conv, which focuses more on the center position compared to a regular Conv uniformly processing a patch, as shown in Fig. 5. To justify this T-shaped receptive field, we first evaluate the importance of each position by calculating the position-wise Frobenius norm. We assume that a position tends to be more important if it has a larger Frobenius norm than other positions. For a regular Conv filter $\\mathbf{F} \\in \\mathbb{R}^{k^2 \\times c}$ , the Frobenius norm at position $i$ is calculated by $\\| \\mathbf{F}_i \\| = \\sqrt{\\sum_{j=1}^{c} |f_{ij}|^2}$ , for $i = 1, 2, 3\\ldots, k^2$ . We consider a salient position to be the one with the maximum Frobenius norm. We then collectively examine each filter in a pre-trained ResNet18, find out their salient positions, and plot a histogram of the salient positions. Results in Fig. 6 show that the center position turns out to be the salient position most frequently among the filters. In other words, the center position weighs more",
665
+ "bbox": [
666
+ 75,
667
+ 603,
668
+ 470,
669
+ 900
670
+ ],
671
+ "page_idx": 3
672
+ },
673
+ {
674
+ "type": "text",
675
+ "text": "than its surrounding neighbors. This is consistent with the T-shaped computation which concentrates on the center position.",
676
+ "bbox": [
677
+ 498,
678
+ 90,
679
+ 890,
680
+ 135
681
+ ],
682
+ "page_idx": 3
683
+ },
684
+ {
685
+ "type": "text",
686
+ "text": "While the T-shaped Conv can be directly used for efficient computation, we show that it is better to decompose the T-shaped Conv into a PConv and a PWConv because the decomposition exploits the inter-filter redundancy and further saves FLOPs. For the same input $\\mathbf{I} \\in \\mathbb{R}^{c \\times h \\times w}$ and output $\\mathbf{O} \\in \\mathbb{R}^{c \\times h \\times w}$ , a T-shaped Conv's FLOPs can be calculated as",
687
+ "bbox": [
688
+ 496,
689
+ 136,
690
+ 890,
691
+ 242
692
+ ],
693
+ "page_idx": 3
694
+ },
695
+ {
696
+ "type": "equation",
697
+ "text": "\n$$\nh \\times w \\times \\left(k ^ {2} \\times c _ {p} \\times c + c \\times \\left(c - c _ {p}\\right)\\right), \\tag {6}\n$$\n",
698
+ "text_format": "latex",
699
+ "bbox": [
700
+ 560,
701
+ 253,
702
+ 890,
703
+ 272
704
+ ],
705
+ "page_idx": 3
706
+ },
707
+ {
708
+ "type": "text",
709
+ "text": "which is higher than the FLOPs of a PConv and a PWConv, i.e.,",
710
+ "bbox": [
711
+ 496,
712
+ 282,
713
+ 890,
714
+ 311
715
+ ],
716
+ "page_idx": 3
717
+ },
718
+ {
719
+ "type": "equation",
720
+ "text": "\n$$\nh \\times w \\times \\left(k ^ {2} \\times c _ {p} ^ {2} + c \\times c _ {p}\\right), \\tag {7}\n$$\n",
721
+ "text_format": "latex",
722
+ "bbox": [
723
+ 596,
724
+ 313,
725
+ 890,
726
+ 332
727
+ ],
728
+ "page_idx": 3
729
+ },
730
+ {
731
+ "type": "text",
732
+ "text": "where $c > c_p$ and $c - c_p > c_p$ (e.g. when $c_p = \\frac{c}{4}$ ). Besides, we can readily leverage the regular Conv for the two-step implementation.",
733
+ "bbox": [
734
+ 496,
735
+ 338,
736
+ 890,
737
+ 383
738
+ ],
739
+ "page_idx": 3
740
+ },
741
+ {
742
+ "type": "text",
743
+ "text": "3.4. FasterNet as a general backbone",
744
+ "text_level": 1,
745
+ "bbox": [
746
+ 498,
747
+ 393,
748
+ 785,
749
+ 409
750
+ ],
751
+ "page_idx": 3
752
+ },
753
+ {
754
+ "type": "text",
755
+ "text": "Given our novel PConv and off-the-shelf PWConv as the primary building operators, we further propose FasterNet, a new family of neural networks that runs favorably fast and is highly effective for many vision tasks. We aim to keep the architecture as simple as possible, without bells and whistles, to make it hardware-friendly in general.",
756
+ "bbox": [
757
+ 496,
758
+ 416,
759
+ 890,
760
+ 506
761
+ ],
762
+ "page_idx": 3
763
+ },
764
+ {
765
+ "type": "text",
766
+ "text": "We present the overall architecture in Fig. 4. It has four hierarchical stages, each of which is preceded by an embedding layer (a regular Conv $4 \\times 4$ with stride 4) or a merging layer (a regular Conv $2 \\times 2$ with stride 2) for spatial downsampling and channel number expanding. Each stage has a stack of FasterNet blocks. We observe that the blocks in the last two stages consume less memory access and tend to have higher FLOPS, as empirically validated in Tab. 1. Thus, we put more FasterNet blocks and correspondingly assign more computations to the last two stages. Each FasterNet block has a PConv layer followed by two PWConv (or Conv $1 \\times 1$ ) layers. Together, they appear as inverted residual blocks where the middle layer has an expanded number of channels, and a shortcut connection is placed to reuse the input features.",
767
+ "bbox": [
768
+ 496,
769
+ 508,
770
+ 890,
771
+ 734
772
+ ],
773
+ "page_idx": 3
774
+ },
775
+ {
776
+ "type": "text",
777
+ "text": "In addition to the above operators, the normalization and activation layers are also indispensable for high-performing neural networks. Many prior works [13, 16, 47], however, overuse such layers throughout the network, which may limit the feature diversity and thus hurt the performance. It can also slow down the overall computation. By contrast, we put them only after each middle PWConv to preserve the feature diversity and achieve lower latency. Besides, we use the batch normalization (BN) [26] instead of other alternative ones [2, 60, 68]. The benefit of BN is that it can be merged into its adjacent Conv layers for faster inference",
778
+ "bbox": [
779
+ 496,
780
+ 734,
781
+ 890,
782
+ 900
783
+ ],
784
+ "page_idx": 3
785
+ },
786
+ {
787
+ "type": "page_number",
788
+ "text": "12024",
789
+ "bbox": [
790
+ 480,
791
+ 924,
792
+ 517,
793
+ 936
794
+ ],
795
+ "page_idx": 3
796
+ },
797
+ {
798
+ "type": "image",
799
+ "img_path": "images/4cbebcfcba214df5a0773b9aac408ea26fe033a4af1473a805836ce83676b38c.jpg",
800
+ "image_caption": [],
801
+ "image_footnote": [],
802
+ "bbox": [
803
+ 145,
804
+ 83,
805
+ 823,
806
+ 305
807
+ ],
808
+ "page_idx": 4
809
+ },
810
+ {
811
+ "type": "image",
812
+ "img_path": "images/70be33816e6e05deda16ab9e60a97129c8caba1c4290915d28b7400b1ac8fd84.jpg",
813
+ "image_caption": [
814
+ "(a) PConv + PWConv"
815
+ ],
816
+ "image_footnote": [],
817
+ "bbox": [
818
+ 96,
819
+ 373,
820
+ 215,
821
+ 443
822
+ ],
823
+ "page_idx": 4
824
+ },
825
+ {
826
+ "type": "image",
827
+ "img_path": "images/1cc9a62bb3d40acb040421f220c81886e43d4bbeade06a31e06c9d39b0107c51.jpg",
828
+ "image_caption": [
829
+ "(b) T-shaped Conv"
830
+ ],
831
+ "image_footnote": [],
832
+ "bbox": [
833
+ 238,
834
+ 375,
835
+ 328,
836
+ 441
837
+ ],
838
+ "page_idx": 4
839
+ },
840
+ {
841
+ "type": "image",
842
+ "img_path": "images/fb412d9f49483344d931ecb60a4c7b368f1ddac0868fdaca07ac7ead2459edc1.jpg",
843
+ "image_caption": [
844
+ "(c) regular Conv",
845
+ "Figure 5. Comparison of convolutional variants. A PConv followed by a PWConv (a) resembles a T-shaped Conv (b), which spends more computation on the center position compared to a regular Conv (c)."
846
+ ],
847
+ "image_footnote": [],
848
+ "bbox": [
849
+ 357,
850
+ 373,
851
+ 450,
852
+ 446
853
+ ],
854
+ "page_idx": 4
855
+ },
856
+ {
857
+ "type": "text",
858
+ "text": "while being as effective as the others. As for the activation layers, we empirically choose GELU [18] for smaller FasterNet variants and ReLU [44] for bigger FasterNet variants, considering both running time and effectiveness. The last three layers, i.e. a global average pooling, a Conv $1 \\times 1$ , and a fully-connected layer, are used together for feature transformation and classification.",
859
+ "bbox": [
860
+ 75,
861
+ 536,
862
+ 468,
863
+ 641
864
+ ],
865
+ "page_idx": 4
866
+ },
867
+ {
868
+ "type": "text",
869
+ "text": "To serve a wide range of applications under different computational budgets, we provide tiny, small, medium, and large variants of FasterNet, referred to as FasterNet-T0/1/2, FasterNet-S, FasterNet-M, and FasterNet-L, respectively. They share a similar architecture but vary in depth and width. Detailed architecture specifications are provided in the appendix.",
870
+ "bbox": [
871
+ 75,
872
+ 643,
873
+ 468,
874
+ 750
875
+ ],
876
+ "page_idx": 4
877
+ },
878
+ {
879
+ "type": "text",
880
+ "text": "4. Experimental Results",
881
+ "text_level": 1,
882
+ "bbox": [
883
+ 76,
884
+ 766,
885
+ 282,
886
+ 782
887
+ ],
888
+ "page_idx": 4
889
+ },
890
+ {
891
+ "type": "text",
892
+ "text": "We first examine the computational speed of our PConv and its effectiveness when combined with a PWConv. We then comprehensively evaluate the performance of our FasterNet for classification, detection, and segmentation tasks. Finally, we conduct a brief ablation study.",
893
+ "bbox": [
894
+ 75,
895
+ 792,
896
+ 468,
897
+ 868
898
+ ],
899
+ "page_idx": 4
900
+ },
901
+ {
902
+ "type": "text",
903
+ "text": "To benchmark the latency and throughput, we choose the following three typical processors, which cover a wide",
904
+ "bbox": [
905
+ 76,
906
+ 869,
907
+ 468,
908
+ 900
909
+ ],
910
+ "page_idx": 4
911
+ },
912
+ {
913
+ "type": "image",
914
+ "img_path": "images/c3e91ad083f5845c8ed7743c6664f13f9076194653b00f9c257d1e8f0f658866.jpg",
915
+ "image_caption": [
916
+ "Figure 4. Overall architecture of our FasterNet. It has four hierarchical stages, each with a stack of FasterNet blocks and preceded by an embedding or merging layer. The last three layers are used for feature classification. Within each FasterNet block, a PConv layer is followed by two PWConv layers. We put normalization and activation layers only after the middle layer to preserve the feature diversity and achieve lower latency.",
917
+ "Figure 6. Histogram of salient position distribution for the regular Conv $3 \\times 3$ filters in a pre-trained ResNet18. The histogram contains four kinds of bars, corresponding to different stages in the network. In all stages, the center position (position 5) appears as a salient position most frequently."
918
+ ],
919
+ "image_footnote": [],
920
+ "bbox": [
921
+ 500,
922
+ 359,
923
+ 887,
924
+ 462
925
+ ],
926
+ "page_idx": 4
927
+ },
928
+ {
929
+ "type": "text",
930
+ "text": "range of computational capacity: GPU (2080Ti), CPU (Intel i9-9900X, using a single thread), and ARM (Cortex-A72, using a single thread). We report their latency for inputs with a batch size of 1 and throughput for inputs with a batch size of 32. During inference, the BN layers are merged to their adjacent layers wherever applicable.",
931
+ "bbox": [
932
+ 498,
933
+ 551,
934
+ 890,
935
+ 642
936
+ ],
937
+ "page_idx": 4
938
+ },
939
+ {
940
+ "type": "text",
941
+ "text": "4.1. PConv is fast with high FLOPS",
942
+ "text_level": 1,
943
+ "bbox": [
944
+ 500,
945
+ 650,
946
+ 774,
947
+ 667
948
+ ],
949
+ "page_idx": 4
950
+ },
951
+ {
952
+ "type": "text",
953
+ "text": "We below show that our PConv is fast and better exploits the on-device computational capacity. Specifically, we stack 10 layers of pure PConv and take feature maps of typical dimensions as inputs. We then measure FLOPs and latency/throughput on GPU, CPU, and ARM processors, which also allow us to further compute FLOPS. We repeat the same procedure for other convolutional variants and make comparisons.",
954
+ "bbox": [
955
+ 496,
956
+ 674,
957
+ 890,
958
+ 794
959
+ ],
960
+ "page_idx": 4
961
+ },
962
+ {
963
+ "type": "text",
964
+ "text": "Results in Tab. 1 show that PConv is overall an appealing choice for high FLOPS with reduced FLOPs. It has only $\\frac{1}{16}$ FLOPs of a regular Conv and achieves $10.5\\times$ , $6.2\\times$ , and $22.8\\times$ higher FLOPS than the DWConv on GPU, CPU, and ARM, respectively. We are unsurprised to see that the regular Conv has the highest FLOPS as it has been constantly optimized for years. However, its total FLOPs and",
965
+ "bbox": [
966
+ 496,
967
+ 795,
968
+ 890,
969
+ 900
970
+ ],
971
+ "page_idx": 4
972
+ },
973
+ {
974
+ "type": "page_number",
975
+ "text": "12025",
976
+ "bbox": [
977
+ 480,
978
+ 925,
979
+ 517,
980
+ 936
981
+ ],
982
+ "page_idx": 4
983
+ },
984
+ {
985
+ "type": "table",
986
+ "img_path": "images/57398df037efacea461ed9e36ffe849a481051795f3dd9e649ed57795f9f4b0b.jpg",
987
+ "table_caption": [],
988
+ "table_footnote": [],
989
+ "table_body": "<table><tr><td>Operator</td><td>Feature map size</td><td>FLOPs (M), ×10 layers</td><td>GPU Throughput (fps)</td><td>FLOPS (G/s)</td><td>CPU Latency (ms)</td><td>FLOPS (G/s)</td><td>ARM Latency (ms)</td><td>FLOPS (G/s)</td></tr><tr><td rowspan=\"5\">Conv 3×3</td><td>96×56×56</td><td>2601</td><td>3010</td><td>7824</td><td>35.67</td><td>72.90</td><td>779.57</td><td>3.33</td></tr><tr><td>192×28×28</td><td>2601</td><td>4893</td><td>12717</td><td>28.41</td><td>91.53</td><td>619.64</td><td>4.19</td></tr><tr><td>384×14×14</td><td>2601</td><td>4558</td><td>11854</td><td>31.85</td><td>81.66</td><td>595.09</td><td>4.37</td></tr><tr><td>768×7×7</td><td>2601</td><td>3159</td><td>8212</td><td>62.71</td><td>41.47</td><td>662.17</td><td>3.92</td></tr><tr><td colspan=\"2\">Average</td><td>-</td><td>10151</td><td>-</td><td>71.89</td><td>-</td><td>3.95</td></tr><tr><td rowspan=\"5\">GConv 3×3 (16 groups)</td><td>96×56×56</td><td>162</td><td>2888</td><td>469</td><td>21.90</td><td>7.42</td><td>166.30</td><td>0.97</td></tr><tr><td>192×28×28</td><td>162</td><td>10811</td><td>1754</td><td>7.58</td><td>21.44</td><td>96.22</td><td>1.68</td></tr><tr><td>384×14×14</td><td>162</td><td>15534</td><td>2514</td><td>4.40</td><td>36.88</td><td>63.57</td><td>2.55</td></tr><tr><td>768×7×7</td><td>162</td><td>16000</td><td>2598</td><td>4.28</td><td>37.97</td><td>65.20</td><td>2.49</td></tr><tr><td colspan=\"2\">Average</td><td>-</td><td>1833</td><td>-</td><td>25.93</td><td>-</td><td>1.92</td></tr><tr><td rowspan=\"5\">DWConv 3×3</td><td>96×56×56</td><td>27.09</td><td>11940</td><td>323</td><td>3.59</td><td>7.52</td><td>108.70</td><td>0.24</td></tr><tr><td>192×28×28</td><td>13.54</td><td>23358</td><td>315</td><td>1.97</td><td>6.86</td><td>82.01</td><td>0.16</td></tr><tr><td>384×14×14</td><td>6.77</td><td>46377</td><td>313</td><td>1.06</td><td>6.35</td><td>94.89</td><td>0.07</td></tr><tr><td>768×7×7</td><td>3.38</td><td>88889</td><td>302</td><td>0.68</td><td>4.93</td><td>150.89</td><td>0.02</td></tr><tr><td colspan=\"2\">Average</td><td>-</td><td>313</td><td>-</td><td>6.42</td><td>-</td><td>0.12</td></tr><tr><td rowspan=\"5\">PConv 3×3 (ours, with r = 1/4)</td><td>96×56×56</td><td>162</td><td>9215</td><td>1493</td><td>5.46</td><td>29.67</td><td>85.30</td><td>1.90</td></tr><tr><td>192×28×28</td><td>162</td><td>14360</td><td>2326</td><td>3.09</td><td>52.43</td><td>66.46</td><td>2.44</td></tr><tr><td>384×14×14</td><td>162</td><td>24408</td><td>3954</td><td>3.58</td><td>45.25</td><td>49.98</td><td>3.24</td></tr><tr><td>768×7×7</td><td>162</td><td>32866</td><td>5324</td><td>5.02</td><td>32.27</td><td>48.30</td><td>3.35</td></tr><tr><td colspan=\"2\">Average</td><td>-</td><td>3274</td><td>-</td><td>39.91</td><td>-</td><td>2.73</td></tr></table>",
990
+ "bbox": [
991
+ 151,
992
+ 88,
993
+ 818,
994
+ 380
995
+ ],
996
+ "page_idx": 5
997
+ },
998
+ {
999
+ "type": "text",
1000
+ "text": "latency/throughput are unaffordable. GConv and DWConv, despite their significant reduction in FLOPs, suffer from a drastic decrease in FLOPS. In addition, they tend to increase the number of channels to compensate for the performance drop, which, however, increase their latency.",
1001
+ "bbox": [
1002
+ 75,
1003
+ 411,
1004
+ 470,
1005
+ 488
1006
+ ],
1007
+ "page_idx": 5
1008
+ },
1009
+ {
1010
+ "type": "text",
1011
+ "text": "4.2. PConv is effective together with PWConv",
1012
+ "text_level": 1,
1013
+ "bbox": [
1014
+ 76,
1015
+ 496,
1016
+ 429,
1017
+ 512
1018
+ ],
1019
+ "page_idx": 5
1020
+ },
1021
+ {
1022
+ "type": "text",
1023
+ "text": "We next show that a PConv followed by a PWConv is effective in approximating a regular Conv to transform the feature maps. To this end, we first build four datasets by feeding the ImageNet-1k val split images into a pre-trained ResNet50, and extract the feature maps before and after the first Conv $3 \\times 3$ in each of the four stages. Each feature map dataset is further spilt into the train $(70\\%)$ , val $(10\\%)$ , and test $(20\\%)$ subsets. We then build a simple network consisting of a PConv followed by a PWConv and train it on the feature map datasets with a mean squared error loss. For comparison, we also build and train networks for DWConv + PWConv and GConv + PWConv under the same setting.",
1024
+ "bbox": [
1025
+ 75,
1026
+ 518,
1027
+ 468,
1028
+ 700
1029
+ ],
1030
+ "page_idx": 5
1031
+ },
1032
+ {
1033
+ "type": "text",
1034
+ "text": "Tab. 2 shows that PConv + PWConv achieve the lowest test loss, meaning that they better approximate a regular Conv in feature transformation. The results also suggest that it is sufficient and efficient to capture spatial features from only a part of the feature maps. PConv shows a great potential to be the new go-to choice in designing fast and effective neural networks.",
1035
+ "bbox": [
1036
+ 75,
1037
+ 700,
1038
+ 470,
1039
+ 806
1040
+ ],
1041
+ "page_idx": 5
1042
+ },
1043
+ {
1044
+ "type": "text",
1045
+ "text": "4.3. FasterNet on ImageNet-1k classification",
1046
+ "text_level": 1,
1047
+ "bbox": [
1048
+ 76,
1049
+ 816,
1050
+ 419,
1051
+ 832
1052
+ ],
1053
+ "page_idx": 5
1054
+ },
1055
+ {
1056
+ "type": "text",
1057
+ "text": "To verify the effectiveness and efficiency of our FasterNet, we first conduct experiments on the large-scale ImageNet-1k classification dataset [46]. It covers 1k categories of common objects and contains about 1.3M la",
1058
+ "bbox": [
1059
+ 75,
1060
+ 839,
1061
+ 470,
1062
+ 902
1063
+ ],
1064
+ "page_idx": 5
1065
+ },
1066
+ {
1067
+ "type": "table",
1068
+ "img_path": "images/6e4606f3000e463c4f2ccd6c994236e18415b170c6a41f56fadf70af6dfdbf6c.jpg",
1069
+ "table_caption": [
1070
+ "Table 1. On-device FLOPS for different operations. PConv appears as an appealing choice for high FLOPS with reduced FLOPs."
1071
+ ],
1072
+ "table_footnote": [],
1073
+ "table_body": "<table><tr><td>Stage</td><td>DWConv+PWConv</td><td>GConv+PWConv(16 groups)</td><td>PConv+PWConvr = 1/4</td></tr><tr><td>1</td><td>0.0089</td><td>0.0065</td><td>0.0069</td></tr><tr><td>2</td><td>0.0158</td><td>0.0137</td><td>0.0136</td></tr><tr><td>3</td><td>0.0214</td><td>0.0202</td><td>0.0172</td></tr><tr><td>4</td><td>0.0130</td><td>0.0128</td><td>0.0115</td></tr><tr><td>Average</td><td>0.0148</td><td>0.0133</td><td>0.0123</td></tr></table>",
1074
+ "bbox": [
1075
+ 522,
1076
+ 412,
1077
+ 870,
1078
+ 521
1079
+ ],
1080
+ "page_idx": 5
1081
+ },
1082
+ {
1083
+ "type": "text",
1084
+ "text": "Table 2. A PConv followed by a PWConv well approximates the regular Conv $3 \\times 3$ at different stages of a pre-trained ResNet50. PConv + PWConv together have the lowest test loss on average.",
1085
+ "bbox": [
1086
+ 498,
1087
+ 526,
1088
+ 890,
1089
+ 570
1090
+ ],
1091
+ "page_idx": 5
1092
+ },
1093
+ {
1094
+ "type": "text",
1095
+ "text": "beled images for training and 50k labeled images for validation. We train our models for 300 epochs using AdamW optimizer [38]. We set the batch size to 2048 for the FasterNet-M/L and 4096 for other variants. We use cosine learning rate scheduler [37] with a peak value of $0.001 \\cdot$ batch size/1024 and a 20-epoch linear warmup. We apply commonly-used regularization and augmentation techniques, including Weight Decay [28], Stochastic Depth [24], Label Smoothing [52], Mixup [73], Cutmix [72] and Rand Augment [9], with varying magnitudes for different FasterNet variants. To reduce the training time, we use $192 \\times 192$ resolution for the first 280 training epochs and $224 \\times 224$ for the remaining 20 epochs. For fair comparison, we do not use knowledge distillation [19] and neural architecture search [79]. We report our top-1 accuracy on the validation set with a center crop at $224 \\times 224$ resolution and a 0.9 crop ratio. Detailed training and validation settings are provided in the appendix.",
1096
+ "bbox": [
1097
+ 496,
1098
+ 582,
1099
+ 890,
1100
+ 853
1101
+ ],
1102
+ "page_idx": 5
1103
+ },
1104
+ {
1105
+ "type": "text",
1106
+ "text": "Fig. 7 and Tab. 3 demonstrate the superiority of our FasterNet over state-of-the-art classification models. The trade-off curves in Fig. 7 clearly show that FasterNet sets",
1107
+ "bbox": [
1108
+ 498,
1109
+ 854,
1110
+ 893,
1111
+ 900
1112
+ ],
1113
+ "page_idx": 5
1114
+ },
1115
+ {
1116
+ "type": "page_number",
1117
+ "text": "12026",
1118
+ "bbox": [
1119
+ 480,
1120
+ 925,
1121
+ 519,
1122
+ 936
1123
+ ],
1124
+ "page_idx": 5
1125
+ },
1126
+ {
1127
+ "type": "image",
1128
+ "img_path": "images/815b3c3a86d1a130a402adcae33bb24083901f97882f97150fb926cc712b2e97.jpg",
1129
+ "image_caption": [],
1130
+ "image_footnote": [],
1131
+ "bbox": [
1132
+ 80,
1133
+ 85,
1134
+ 467,
1135
+ 301
1136
+ ],
1137
+ "page_idx": 6
1138
+ },
1139
+ {
1140
+ "type": "image",
1141
+ "img_path": "images/9ee0963af0ff702c4f7440623b2fef107742aa7aa992df772c6ff8f22d9f3054.jpg",
1142
+ "image_caption": [],
1143
+ "image_footnote": [],
1144
+ "bbox": [
1145
+ 80,
1146
+ 303,
1147
+ 467,
1148
+ 455
1149
+ ],
1150
+ "page_idx": 6
1151
+ },
1152
+ {
1153
+ "type": "image",
1154
+ "img_path": "images/5959c7303514d8b8e2d1673f69f30c473f3e2009cb5c8ddc69a8bd14351a9e35.jpg",
1155
+ "image_caption": [
1156
+ "Figure 7. FasterNet has the highest efficiency in balancing accuracy-throughput and accuracy-latency trade-offs for different devices. To save space and make the plots more proportionate, we showcase network variants within a certain range of latency. Full plots can be found in the appendix, which show consistent results."
1157
+ ],
1158
+ "image_footnote": [],
1159
+ "bbox": [
1160
+ 80,
1161
+ 459,
1162
+ 465,
1163
+ 611
1164
+ ],
1165
+ "page_idx": 6
1166
+ },
1167
+ {
1168
+ "type": "text",
1169
+ "text": "the new state-of-the-art in balancing accuracy and latency/throughput among all the networks examined. From another perspective, FasterNet runs faster than various CNN, ViT and MLP models on a wide range of devices, when having similar top-1 accuracy. As quantitatively shown in Tab. 3, FasterNet-T0 is $2.8 \\times 3.3 \\times$ and $2.4 \\times$ faster than MobileViT-XXS [42] on GPU, CPU, and ARM processors, respectively, while being $2.9\\%$ more accurate. Our large FasterNet-L achieves $83.5\\%$ top-1 accuracy, comparable to the emerging Swin-B [35] and ConvNeXt-B [36] while having $36\\%$ and $28\\%$ higher inference throughput on GPU, as well as saving $37\\%$ and $15\\%$ compute time on CPU. Given such promising results, we highlight that our FasterNet is",
1170
+ "bbox": [
1171
+ 75,
1172
+ 704,
1173
+ 468,
1174
+ 900
1175
+ ],
1176
+ "page_idx": 6
1177
+ },
1178
+ {
1179
+ "type": "table",
1180
+ "img_path": "images/5d23304daed3ad28732bde0f3b7a1fd6fb4b2f5cad8e2e0f0c42d4ab6692b4f0.jpg",
1181
+ "table_caption": [],
1182
+ "table_footnote": [],
1183
+ "table_body": "<table><tr><td>Network</td><td>Params (M)</td><td>FLOPs (G)</td><td>Throughput on GPU (fps) ↑</td><td>Latency on CPU (ms) ↓</td><td>Latency on ARM (ms) ↓</td><td>Acc. (%)</td></tr><tr><td>ShuffleNetV2 ×1.5 [40]</td><td>3.5</td><td>0.30</td><td>4878</td><td>12.1</td><td>266</td><td>72.6</td></tr><tr><td>MobileNetV2 [47]</td><td>3.5</td><td>0.31</td><td>4198</td><td>12.2</td><td>442</td><td>72.0</td></tr><tr><td>MobileViT-XXS [42]</td><td>1.3</td><td>0.42</td><td>2393</td><td>30.8</td><td>348</td><td>69.0</td></tr><tr><td>EdgeNeXt-XXS [41]</td><td>1.3</td><td>0.26</td><td>2765</td><td>15.7</td><td>239</td><td>71.2</td></tr><tr><td>FasterNet-T0</td><td>3.9</td><td>0.34</td><td>6807</td><td>9.2</td><td>143</td><td>71.9</td></tr><tr><td>GhostNet ×1.3 [13]</td><td>7.4</td><td>0.24</td><td>2988</td><td>17.9</td><td>481</td><td>75.7</td></tr><tr><td>ShuffleNetV2 ×2 [40]</td><td>7.4</td><td>0.59</td><td>3339</td><td>17.8</td><td>403</td><td>74.9</td></tr><tr><td>MobileNetV2 ×1.4 [47]</td><td>6.1</td><td>0.60</td><td>2711</td><td>22.6</td><td>650</td><td>74.7</td></tr><tr><td>MobileViT-XS [42]</td><td>2.3</td><td>1.05</td><td>1392</td><td>40.8</td><td>648</td><td>74.8</td></tr><tr><td>EdgeNeXt-XS [41]</td><td>2.3</td><td>0.54</td><td>1738</td><td>24.4</td><td>434</td><td>75.0</td></tr><tr><td>PVT-Tiny [65]</td><td>13.2</td><td>1.94</td><td>1266</td><td>55.6</td><td>708</td><td>75.1</td></tr><tr><td>FasterNet-T1</td><td>7.6</td><td>0.85</td><td>3782</td><td>17.7</td><td>285</td><td>76.2</td></tr><tr><td>CycleMLP-B1 [5]</td><td>15.2</td><td>2.10</td><td>865</td><td>116.1</td><td>892</td><td>79.1</td></tr><tr><td>PoolFormer-S12 [71]</td><td>11.9</td><td>1.82</td><td>1439</td><td>49.0</td><td>665</td><td>77.2</td></tr><tr><td>MobileViT-S [42]</td><td>5.6</td><td>2.03</td><td>1039</td><td>56.7</td><td>941</td><td>78.4</td></tr><tr><td>EdgeNeXt-S [41]</td><td>5.6</td><td>1.26</td><td>1128</td><td>39.2</td><td>743</td><td>79.4</td></tr><tr><td>ResNet50 [16, 36]</td><td>25.6</td><td>4.11</td><td>959</td><td>73.0</td><td>1131</td><td>78.8</td></tr><tr><td>FasterNet-T2</td><td>15.0</td><td>1.91</td><td>1991</td><td>33.5</td><td>497</td><td>78.9</td></tr><tr><td>CycleMLP-B2 [5]</td><td>26.8</td><td>3.90</td><td>528</td><td>186.3</td><td>1502</td><td>81.6</td></tr><tr><td>PoolFormer-S24 [71]</td><td>21.4</td><td>3.41</td><td>748</td><td>92.8</td><td>1261</td><td>80.3</td></tr><tr><td>PoolFormer-S36 [71]</td><td>30.9</td><td>5.00</td><td>507</td><td>138.0</td><td>1860</td><td>81.4</td></tr><tr><td>ConvNeXt-T [36]</td><td>28.6</td><td>4.47</td><td>657</td><td>86.3</td><td>1889</td><td>82.1</td></tr><tr><td>Swin-T [35]</td><td>28.3</td><td>4.51</td><td>609</td><td>122.2</td><td>1424</td><td>81.3</td></tr><tr><td>PVT-Small [65]</td><td>24.5</td><td>3.83</td><td>689</td><td>89.6</td><td>1345</td><td>79.8</td></tr><tr><td>PVT-Medium [65]</td><td>44.2</td><td>6.69</td><td>438</td><td>143.6</td><td>2142</td><td>81.2</td></tr><tr><td>FasterNet-S</td><td>31.1</td><td>4.56</td><td>1029</td><td>71.2</td><td>1103</td><td>81.3</td></tr><tr><td>PoolFormer-M36 [71]</td><td>56.2</td><td>8.80</td><td>320</td><td>215.0</td><td>2979</td><td>82.1</td></tr><tr><td>ConvNeXt-S [36]</td><td>50.2</td><td>8.71</td><td>377</td><td>153.2</td><td>3484</td><td>83.1</td></tr><tr><td>Swin-S [35]</td><td>49.6</td><td>8.77</td><td>348</td><td>224.2</td><td>2613</td><td>83.0</td></tr><tr><td>PVT-Large [65]</td><td>61.4</td><td>9.85</td><td>306</td><td>203.4</td><td>3101</td><td>81.7</td></tr><tr><td>FasterNet-M</td><td>53.5</td><td>8.74</td><td>500</td><td>129.5</td><td>2092</td><td>83.0</td></tr><tr><td>PoolFormer-M48 [71]</td><td>73.5</td><td>11.59</td><td>242</td><td>281.8</td><td>OOM</td><td>82.5</td></tr><tr><td>ConvNeXt-B [36]</td><td>88.6</td><td>15.38</td><td>253</td><td>257.1</td><td>OOM</td><td>83.8</td></tr><tr><td>Swin-B [35]</td><td>87.8</td><td>15.47</td><td>237</td><td>349.2</td><td>OOM</td><td>83.5</td></tr><tr><td>FasterNet-L</td><td>93.5</td><td>15.52</td><td>323</td><td>219.5</td><td>OOM</td><td>83.5</td></tr></table>",
1184
+ "bbox": [
1185
+ 501,
1186
+ 84,
1187
+ 890,
1188
+ 588
1189
+ ],
1190
+ "page_idx": 6
1191
+ },
1192
+ {
1193
+ "type": "text",
1194
+ "text": "Table 3. Comparison on ImageNet-1k benchmark. Models with similar top-1 accuracy are grouped together. For each group, our FasterNet achieves the highest throughput on GPU and the lowest latency on CPU and ARM. All models are evaluated at $224 \\times 224$ resolution except for the MobileViT and EdgeNeXt with $256 \\times 256$ . OOM is short for out of memory.",
1195
+ "bbox": [
1196
+ 496,
1197
+ 592,
1198
+ 890,
1199
+ 676
1200
+ ],
1201
+ "page_idx": 6
1202
+ },
1203
+ {
1204
+ "type": "text",
1205
+ "text": "much simpler than many other models in terms of architectural design, which showcases the feasibility of designing simple yet powerful neural networks.",
1206
+ "bbox": [
1207
+ 496,
1208
+ 696,
1209
+ 890,
1210
+ 742
1211
+ ],
1212
+ "page_idx": 6
1213
+ },
1214
+ {
1215
+ "type": "text",
1216
+ "text": "4.4. FasterNet on downstream tasks",
1217
+ "text_level": 1,
1218
+ "bbox": [
1219
+ 498,
1220
+ 755,
1221
+ 777,
1222
+ 768
1223
+ ],
1224
+ "page_idx": 6
1225
+ },
1226
+ {
1227
+ "type": "text",
1228
+ "text": "To further evaluate the generalization ability of FasterNet, we conduct experiments on the challenging COCO dataset [32] for object detection and instance segmentation. As a common practice, we employ the ImageNet pre-trained FasterNet as a backbone and equip it with the popular Mask R-CNN detector [15]. To highlight the effectiveness of the backbone itself, we simply follow PoolFormer [71] and adopt an AdamW optimizer, a $1\\times$ training",
1229
+ "bbox": [
1230
+ 496,
1231
+ 779,
1232
+ 890,
1233
+ 900
1234
+ ],
1235
+ "page_idx": 6
1236
+ },
1237
+ {
1238
+ "type": "page_number",
1239
+ "text": "12027",
1240
+ "bbox": [
1241
+ 480,
1242
+ 925,
1243
+ 517,
1244
+ 936
1245
+ ],
1246
+ "page_idx": 6
1247
+ },
1248
+ {
1249
+ "type": "table",
1250
+ "img_path": "images/c5f70244aef2e32f5512124b593580c255e602f3bc0a47d09897dd295233bd84.jpg",
1251
+ "table_caption": [],
1252
+ "table_footnote": [],
1253
+ "table_body": "<table><tr><td>Backbone</td><td>Params (M)</td><td>FLOPs (G)</td><td>Latency on GPU (ms)</td><td>\\(AP^b\\)</td><td>\\(AP_{50}^b\\)</td><td>\\(AP_{75}^b\\)</td><td>\\(AP^m\\)</td><td>\\(AP_{50}^m\\)</td><td>\\(AP_{75}^m\\)</td></tr><tr><td>ResNet50 [16]</td><td>44.2</td><td>253</td><td>54.9</td><td>38.0</td><td>58.6</td><td>41.4</td><td>34.4</td><td>55.1</td><td>36.7</td></tr><tr><td>PoolFormer-S24 [71]</td><td>41.0</td><td>233</td><td>111.0</td><td>40.1</td><td>62.2</td><td>43.4</td><td>37.0</td><td>59.1</td><td>39.6</td></tr><tr><td>PVT-Small [65]</td><td>44.1</td><td>238</td><td>89.5</td><td>40.4</td><td>62.9</td><td>43.8</td><td>37.8</td><td>60.1</td><td>40.3</td></tr><tr><td>FasterNet-S</td><td>49.0</td><td>258</td><td>54.3</td><td>39.9</td><td>61.2</td><td>43.6</td><td>36.9</td><td>58.1</td><td>39.7</td></tr><tr><td>ResNet101 [16]</td><td>63.2</td><td>329</td><td>68.9</td><td>40.4</td><td>61.1</td><td>44.2</td><td>36.4</td><td>57.7</td><td>38.8</td></tr><tr><td>ResNeXt101-32×4d [69]</td><td>62.8</td><td>333</td><td>80.5</td><td>41.9</td><td>62.5</td><td>45.9</td><td>37.5</td><td>59.4</td><td>40.2</td></tr><tr><td>PoolFormer-S36 [71]</td><td>50.5</td><td>266</td><td>146.9</td><td>41.0</td><td>63.1</td><td>44.8</td><td>37.7</td><td>60.1</td><td>40.0</td></tr><tr><td>PVT-Medium [65]</td><td>63.9</td><td>295</td><td>117.3</td><td>42.0</td><td>64.4</td><td>45.6</td><td>39.0</td><td>61.6</td><td>42.1</td></tr><tr><td>FasterNet-M</td><td>71.2</td><td>344</td><td>71.4</td><td>43.0</td><td>64.4</td><td>47.4</td><td>39.1</td><td>61.5</td><td>42.3</td></tr><tr><td>ResNeXt101-64×4d [69]</td><td>101.9</td><td>487</td><td>112.9</td><td>42.8</td><td>63.8</td><td>47.3</td><td>38.4</td><td>60.6</td><td>41.3</td></tr><tr><td>PVT-Large [65]</td><td>81.0</td><td>358</td><td>152.2</td><td>42.9</td><td>65.0</td><td>46.6</td><td>39.5</td><td>61.9</td><td>42.5</td></tr><tr><td>FasterNet-L</td><td>110.9</td><td>484</td><td>93.8</td><td>44.0</td><td>65.6</td><td>48.2</td><td>39.9</td><td>62.3</td><td>43.0</td></tr></table>",
1254
+ "bbox": [
1255
+ 119,
1256
+ 88,
1257
+ 849,
1258
+ 295
1259
+ ],
1260
+ "page_idx": 7
1261
+ },
1262
+ {
1263
+ "type": "table",
1264
+ "img_path": "images/7038d17176173fd3f66cbaf3bc12fe53d4593c585ef6a012831af2873370e950.jpg",
1265
+ "table_caption": [
1266
+ "Table 4. Results on COCO object detection and instance segmentation benchmarks. FLOPs are calculated with image size (1280, 800)."
1267
+ ],
1268
+ "table_footnote": [
1269
+ "Table 5. Ablation on the partial ratio, normalization, and activation of FasterNet. Rows highlighted in grey are the default settings. $\\mathrm{T0}^*$ denotes T0 variants with modified network width and depth."
1270
+ ],
1271
+ "table_body": "<table><tr><td>Ablation</td><td>Variant</td><td>Throughput on GPU (fps)</td><td>Latency on CPU (ms)</td><td>Latency on ARM (ms)</td><td>Acc. (%)</td></tr><tr><td rowspan=\"3\">Partial ratio</td><td>T0* w/ r = 1/2</td><td>6626</td><td>9.6</td><td>145</td><td>71.7</td></tr><tr><td>T0 w/ r = 1/4</td><td>6807</td><td>9.2</td><td>143</td><td>71.9</td></tr><tr><td>T0* w/ r = 1/8</td><td>6204</td><td>8.9</td><td>140</td><td>71.3</td></tr><tr><td rowspan=\"2\">Normalization</td><td>T0 w/ BN</td><td>6807</td><td>9.2</td><td>143</td><td>71.9</td></tr><tr><td>T0 w/ LN</td><td>5515</td><td>10.7</td><td>159</td><td>71.9</td></tr><tr><td rowspan=\"5\">Activation</td><td>T0 w/ ReLU</td><td>6929</td><td>8.2</td><td>114</td><td>71.3</td></tr><tr><td>T0* w/ ReLU</td><td>5866</td><td>9.3</td><td>143</td><td>71.7</td></tr><tr><td>T0 w/ GELU</td><td>6807</td><td>9.2</td><td>143</td><td>71.9</td></tr><tr><td>T2 w/ ReLU</td><td>1991</td><td>33.5</td><td>497</td><td>78.9</td></tr><tr><td>T2 w/ GELU</td><td>1985</td><td>35.4</td><td>557</td><td>78.7</td></tr></table>",
1272
+ "bbox": [
1273
+ 81,
1274
+ 329,
1275
+ 468,
1276
+ 534
1277
+ ],
1278
+ "page_idx": 7
1279
+ },
1280
+ {
1281
+ "type": "text",
1282
+ "text": "schedule (12 epochs), a batch size of 16, and other training settings without further hyper-parameter tuning.",
1283
+ "bbox": [
1284
+ 75,
1285
+ 595,
1286
+ 468,
1287
+ 626
1288
+ ],
1289
+ "page_idx": 7
1290
+ },
1291
+ {
1292
+ "type": "text",
1293
+ "text": "Tab. 4 shows the results for comparison between FasterNet and representative models. FasterNet consistently outperforms ResNet and ResNext by having higher average precision (AP) with similar latency. Specifically, FasterNet-S yields $+1.9$ higher box AP and $+2.4$ higher mask AP compared to the standard baseline ResNet50. FasterNet is also competitive against the ViT variants. Under similar FLOPs, FasterNet-L reduces PVT-Large's latency by $38\\%$ , i.e., from $152.2\\mathrm{ms}$ to $93.8\\mathrm{ms}$ on GPU, and achieves $+1.1$ higher box AP and $+0.4$ higher mask AP.",
1294
+ "bbox": [
1295
+ 75,
1296
+ 627,
1297
+ 468,
1298
+ 777
1299
+ ],
1300
+ "page_idx": 7
1301
+ },
1302
+ {
1303
+ "type": "text",
1304
+ "text": "4.5. Ablation study",
1305
+ "text_level": 1,
1306
+ "bbox": [
1307
+ 76,
1308
+ 786,
1309
+ 225,
1310
+ 801
1311
+ ],
1312
+ "page_idx": 7
1313
+ },
1314
+ {
1315
+ "type": "text",
1316
+ "text": "We conduct a brief ablation study on the value of partial ratio $r$ and the choices of activation and normalization layers. We compare different variants in terms of ImageNet top-1 accuracy and on-device latency/throughput. Results are summarized in Tab. 5. For the partial ratio $r$ , we set it to $\\frac{1}{4}$ for all FasterNet variants by default, which achieves",
1317
+ "bbox": [
1318
+ 75,
1319
+ 809,
1320
+ 468,
1321
+ 904
1322
+ ],
1323
+ "page_idx": 7
1324
+ },
1325
+ {
1326
+ "type": "text",
1327
+ "text": "higher accuracy, higher throughput, and lower latency at similar complexity. A too large partial ratio $r$ would make PConv degrade to a regular Conv, while a too small value would render PConv less effective in capturing the spatial features. For the normalization layers, we choose BatchNorm over LayerNorm because BatchNorm can be merged into its adjacent convolutional layers for faster inference while it is as effective as LayerNorm in our experiment. For the activation function, interestingly, we empirically found that GELU fits FasterNet-T0/T1 models more efficiently than ReLU. It, however, becomes opposite for FasterNet-T2/S/M/L. Here we only show two examples in Tab. 5 due to space constraint. We conjecture that GELU strengthens FasterNet-T0/T1 by having higher non-linearity, while the benefit fades away for larger FasterNet variants.",
1328
+ "bbox": [
1329
+ 496,
1330
+ 330,
1331
+ 890,
1332
+ 556
1333
+ ],
1334
+ "page_idx": 7
1335
+ },
1336
+ {
1337
+ "type": "text",
1338
+ "text": "5. Conclusion",
1339
+ "text_level": 1,
1340
+ "bbox": [
1341
+ 500,
1342
+ 575,
1343
+ 617,
1344
+ 590
1345
+ ],
1346
+ "page_idx": 7
1347
+ },
1348
+ {
1349
+ "type": "text",
1350
+ "text": "In this paper, we have investigated the common and unresolved issue that many established neural networks suffer from low floating-point operations per second (FLOPS). We have revisited a bottleneck operator, DWConv, and analyzed its main cause for a slowdown - frequent memory access. To overcome the issue and achieve faster neural networks, we have proposed a simple yet fast and effective operator, PConv, that can be readily plugged into many existing networks. We have further introduced our general-purpose FasterNet, built upon our PConv, that achieves state-of-the-art speed and accuracy trade-off on various devices and vision tasks. We hope that our PConv and FasterNet would inspire more research on simple yet effective neural networks, going beyond academia to impact the industry and community directly.",
1351
+ "bbox": [
1352
+ 496,
1353
+ 602,
1354
+ 890,
1355
+ 829
1356
+ ],
1357
+ "page_idx": 7
1358
+ },
1359
+ {
1360
+ "type": "text",
1361
+ "text": "Acknowledgement This work was supported, in part, by Hong Kong General Research Fund under grant number 16200120. The work of C.-H. Lee was supported, in part, by the NSF under Grant IIS-2209921.",
1362
+ "bbox": [
1363
+ 496,
1364
+ 839,
1365
+ 890,
1366
+ 900
1367
+ ],
1368
+ "page_idx": 7
1369
+ },
1370
+ {
1371
+ "type": "page_number",
1372
+ "text": "12028",
1373
+ "bbox": [
1374
+ 480,
1375
+ 925,
1376
+ 517,
1377
+ 936
1378
+ ],
1379
+ "page_idx": 7
1380
+ },
1381
+ {
1382
+ "type": "text",
1383
+ "text": "References",
1384
+ "text_level": 1,
1385
+ "bbox": [
1386
+ 78,
1387
+ 89,
1388
+ 173,
1389
+ 104
1390
+ ],
1391
+ "page_idx": 8
1392
+ },
1393
+ {
1394
+ "type": "list",
1395
+ "sub_type": "ref_text",
1396
+ "list_items": [
1397
+ "[1] Alaaeldin Ali, Hugo Touvron, Mathilde Caron, Piotr Bojanowski, Matthijs Douze, Armand Joulin, Ivan Laptev, Natalia Neverova, Gabriel Synnaeve, Jakob Verbeek, et al. Xcit: Cross-covariance image transformers. Advances in neural information processing systems, 34:20014-20027, 2021. 3",
1398
+ "[2] Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hinton. Layer normalization. arXiv preprint arXiv:1607.06450, 2016. 4",
1399
+ "[3] Han Cai, Chuang Gan, and Song Han. Efficientvit: Enhanced linear attention for high-resolution low-computation visual recognition. arXiv preprint arXiv:2205.14756, 2022. 3",
1400
+ "[4] Jierun Chen, Tianlang He, Weipeng Zhuo, Li Ma, Sangtae Ha, and S-H Gary Chan. Tvconv: Efficient translation variant convolution for layout-aware visual processing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12548-12558, 2022. 3",
1401
+ "[5] Shoufa Chen, Enze Xie, Chongjian Ge, Ding Liang, and Ping Luo. Cycle MLP: A mlp-like architecture for dense prediction. arXiv preprint arXiv:2107.10224, 2021. 2, 3, 7",
1402
+ "[6] Yinpeng Chen, Xiyang Dai, Dongdong Chen, Mengchen Liu, Xiaoyi Dong, Lu Yuan, and Zicheng Liu. Mobileformer: Bridging mobilenet and transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5270-5279, 2022. 1, 3",
1403
+ "[7] Yunpeng Chen, Haoqi Fan, Bing Xu, Zhicheng Yan, Yannis Kalantidis, Marcus Rohrbach, Shuicheng Yan, and Ji-ashi Feng. Drop an octave: Reducing spatial redundancy in convolutional neural networks with octave convolution. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3435–3444, 2019. 3",
1404
+ "[8] François Chollet. Xception: Deep learning with depthwise separable convolutions. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1251-1258, 2017. 3",
1405
+ "[9] Ekin D Cubuk, Barret Zoph, Jonathon Shlens, and Quoc V Le. Randaugment: Practical automated data augmentation with a reduced search space. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 702-703, 2020. 6",
1406
+ "[10] Zihang Dai, Hanxiao Liu, Quoc V Le, and Mingxing Tan. Coatnet: Marrying convolution and attention for all data sizes. Advances in Neural Information Processing Systems, 34:3965-3977, 2021. 3",
1407
+ "[11] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 1, 3",
1408
+ "[12] Benjamin Graham, Alaaeldin El-Nouby, Hugo Touvron, Pierre Stock, Armand Joulin, Hervé Jégou, and Matthijs Douze. Levit: a vision transformer in convnet's clothing for faster inference. In Proceedings of the IEEE/CVF international conference on computer vision, pages 12259-12269, 2021. 3"
1409
+ ],
1410
+ "bbox": [
1411
+ 78,
1412
+ 114,
1413
+ 470,
1414
+ 898
1415
+ ],
1416
+ "page_idx": 8
1417
+ },
1418
+ {
1419
+ "type": "list",
1420
+ "sub_type": "ref_text",
1421
+ "list_items": [
1422
+ "[13] Kai Han, Yunhe Wang, Qi Tian, Jianyuan Guo, Chunjing Xu, and Chang Xu. Ghostnet: More features from cheap operations. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1580-1589, 2020. 1, 3, 4, 7",
1423
+ "[14] Kai Han, Yunhe Wang, Qiulin Zhang, Wei Zhang, Chunjing Xu, and Tong Zhang. Model rubik's cube: Twisting resolution, depth and width for tinynets. Advances in Neural Information Processing Systems, 33:19353-19364, 2020. 3",
1424
+ "[15] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961-2969, 2017. 7",
1425
+ "[16] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 2, 4, 7, 8",
1426
+ "[17] Tianlang He, Jiajie Tan, Weipeng Zhuo, Maximilian Printz, and S-H Gary Chan. Tackling multipath and biased training data for imu-assisted ble proximity detection. In IEEE INFOCOM 2022-IEEE Conference on Computer Communications, pages 1259-1268. IEEE, 2022. 3",
1427
+ "[18] Dan Hendrycks and Kevin Gimpel. Gaussian error linear units (gelus). arXiv preprint arXiv:1606.08415, 2016. 5",
1428
+ "[19] Geoffrey Hinton, Oriol Vinyals, Jeff Dean, et al. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2(7), 2015. 6",
1429
+ "[20] Andrew Howard, Mark Sandler, Grace Chu, Liang-Chieh Chen, Bo Chen, Mingxing Tan, Weijun Wang, Yukun Zhu, Ruoming Pang, Vijay Vasudevan, et al. Searching for mobilenetv3. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1314-1324, 2019. 1, 3",
1430
+ "[21] Andrew G Howard, Menglong Zhu, Bo Chen, Dmitry Kalenichenko, Weijun Wang, Tobias Weyand, Marco Andretto, and Hartwig Adam. Mobilenets: Efficient convolutional neural networks for mobile vision applications. arXiv preprint arXiv:1704.04861, 2017. 1, 3",
1431
+ "[22] Han Hu, Zheng Zhang, Zhenda Xie, and Stephen Lin. Local relation networks for image recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3464-3473, 2019. 3",
1432
+ "[23] Gao Huang, Shichen Liu, Laurens Van der Maaten, and Kilian Q Weinberger. Condensenet: An efficient densenet using learned group convolutions. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2752-2761, 2018. 3",
1433
+ "[24] Gao Huang, Yu Sun, Zhuang Liu, Daniel Sedra, and Kilian Q Weinberger. Deep networks with stochastic depth. In European conference on computer vision, pages 646-661. Springer, 2016. 6",
1434
+ "[25] Tao Huang, Lang Huang, Shan You, Fei Wang, Chen Qian, and Chang Xu. Lightvit: Towards light-weight convolution-free vision transformers. arXiv preprint arXiv:2207.05557, 2022.3",
1435
+ "[26] Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. In International conference on machine learning, pages 448-456. PMLR, 2015. 4"
1436
+ ],
1437
+ "bbox": [
1438
+ 501,
1439
+ 92,
1440
+ 890,
1441
+ 900
1442
+ ],
1443
+ "page_idx": 8
1444
+ },
1445
+ {
1446
+ "type": "page_number",
1447
+ "text": "12029",
1448
+ "bbox": [
1449
+ 480,
1450
+ 926,
1451
+ 519,
1452
+ 936
1453
+ ],
1454
+ "page_idx": 8
1455
+ },
1456
+ {
1457
+ "type": "list",
1458
+ "sub_type": "ref_text",
1459
+ "list_items": [
1460
+ "[27] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. Advances in neural information processing systems, 25, 2012. 1, 3",
1461
+ "[28] Anders Krogh and John Hertz. A simple weight decay can improve generalization. Advances in neural information processing systems, 4, 1991. 6",
1462
+ "[29] Yunsheng Li, Yinpeng Chen, Xiyang Dai, Dongdong Chen, Mengchen Liu, Lu Yuan, Zicheng Liu, Lei Zhang, and Nuno Vasconcelos. Micronet: Improving image recognition with extremely low flops. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 468-477, 2021. 1, 3",
1463
+ "[30] Yanyu Li, Geng Yuan, Yang Wen, Eric Hu, Georgios Evangelidis, Sergey Tulyakov, Yanzhi Wang, and Jian Ren. Efficientformer: Vision transformers at mobilenet speed. arXiv preprint arXiv:2206.01191, 2022. 3",
1464
+ "[31] Dongze Lian, Zehao Yu, Xing Sun, and Shenghua Gao. Asmlp: An axial shifted mlp architecture for vision. arXiv preprint arXiv:2107.08391, 2021. 3",
1465
+ "[32] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014. 7",
1466
+ "[33] Ruiyang Liu, Yinghui Li, Linmi Tao, Dun Liang, and Haitao Zheng. Are we ready for a new paradigm shift? a survey on visual deep mlp. Patterns, 3(7):100520, 2022. 3",
1467
+ "[34] Ze Liu, Han Hu, Yutong Lin, Zhuliang Yao, Zhenda Xie, Yixuan Wei, Jia Ning, Yue Cao, Zheng Zhang, Li Dong, et al. Swin transformer v2: Scaling up capacity and resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12009-12019, 2022. 3",
1468
+ "[35] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10012-10022, 2021. 2, 3, 7",
1469
+ "[36] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11976-11986, 2022. 1, 3, 7",
1470
+ "[37] Ilya Loshchilov and Frank Hutter. Sgdr: Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983, 2016. 6",
1471
+ "[38] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6",
1472
+ "[39] Jiachen Lu, Jinghan Yao, Junge Zhang, Xiatian Zhu, Hang Xu, Weiguo Gao, Chunjing Xu, Tao Xiang, and Li Zhang. Soft: softmax-free transformer with linear complexity. Advances in Neural Information Processing Systems, 34:21297-21309, 2021. 3",
1473
+ "[40] Ningning Ma, Xiangyu Zhang, Hai-Tao Zheng, and Jian Sun. Shufflenet v2: Practical guidelines for efficient cnn architecture design. In Proceedings of the European conference on computer vision (ECCV), pages 116-131, 2018. 1, 2, 3, 7"
1474
+ ],
1475
+ "bbox": [
1476
+ 78,
1477
+ 90,
1478
+ 470,
1479
+ 900
1480
+ ],
1481
+ "page_idx": 9
1482
+ },
1483
+ {
1484
+ "type": "list",
1485
+ "sub_type": "ref_text",
1486
+ "list_items": [
1487
+ "[41] Muhammad Maaz, Abdelrahman Shaker, Hisham Cholakkal, Salman Khan, Syed Waqas Zamir, Rao Muhammad Anwer, and Fahad Shahbaz Khan. Edgenext: efficiently amalgamated cnn-transformer architecture for mobile vision applications. arXiv preprint arXiv:2206.10589, 2022. 7",
1488
+ "[42] Sachin Mehta and Mohammad Rastegari. Mobilevit: lightweight, general-purpose, and mobile-friendly vision transformer. arXiv preprint arXiv:2110.02178, 2021. 1, 2, 3, 7",
1489
+ "[43] Sachin Mehta and Mohammad Rastegari. Separable self-attention for mobile vision transformers. arXiv preprint arXiv:2206.02680, 2022. 1, 3",
1490
+ "[44] Vinod Nair and Geoffrey E Hinton. Rectified linear units improve restricted boltzmann machines. In Icml, 2010. 5",
1491
+ "[45] Junting Pan, Adrian Bulat, Fuwen Tan, Xiatian Zhu, Lukasz Dudziak, Hongsheng Li, Georgios Tzimiropoulos, and Brais Martinez. Edgevits: Competing light-weight cnns on mobile devices with vision transformers. arXiv preprint arXiv:2205.03436, pages 1-6, 2022. 3",
1492
+ "[46] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. International journal of computer vision, 115(3):211-252, 2015. 6",
1493
+ "[47] Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zh-moginov, and Liang-Chieh Chen. Mobilenetv2: Inverted residuals and linear bottlenecks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4510-4520, 2018. 1, 3, 4, 7",
1494
+ "[48] Laurent Sifre and Stéphane Mallat. Rigid-motion scattering for texture classification. arXiv preprint arXiv:1403.1687, 2014. 1, 3",
1495
+ "[49] Pravendra Singh, Vinay Kumar Verma, Piyush Rai, and Vinay P Namboodiri. Hetconv: Heterogeneous kernel-based convolutions for deep cnns. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4835-4844, 2019. 3",
1496
+ "[50] Aravind Srinivas, Tsung-Yi Lin, Niki Parmar, Jonathon Shlens, Pieter Abbeel, and Ashish Vaswani. Bottleneck transformers for visual recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16519-16529, 2021. 3",
1497
+ "[51] Andreas Steiner, Alexander Kolesnikov, Xiaohua Zhai, Ross Wightman, Jakob Uszkoreit, and Lucas Beyer. How to train your vit? data, augmentation, and regularization in vision transformers. arXiv preprint arXiv:2106.10270, 2021. 3",
1498
+ "[52] Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jon Shlens, and Zbigniew Wojna. Rethinking the inception architecture for computer vision. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2818-2826, 2016. 6",
1499
+ "[53] Mingxing Tan, Bo Chen, Ruoming Pang, Vijay Vasudevan, Mark Sandler, Andrew Howard, and Quoc V Le. Mnasnet: Platform-aware neural architecture search for mobile. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2820-2828, 2019. 3",
1500
+ "[54] Mingxing Tan and Quoc Le. Efficientnet: Rethinking model scaling for convolutional neural networks. In International"
1501
+ ],
1502
+ "bbox": [
1503
+ 501,
1504
+ 92,
1505
+ 890,
1506
+ 900
1507
+ ],
1508
+ "page_idx": 9
1509
+ },
1510
+ {
1511
+ "type": "page_number",
1512
+ "text": "12030",
1513
+ "bbox": [
1514
+ 480,
1515
+ 926,
1516
+ 519,
1517
+ 936
1518
+ ],
1519
+ "page_idx": 9
1520
+ },
1521
+ {
1522
+ "type": "list",
1523
+ "sub_type": "ref_text",
1524
+ "list_items": [
1525
+ "conference on machine learning, pages 6105-6114. PMLR, 2019.3",
1526
+ "[55] Mingxing Tan and Quoc Le. Efficientnetv2: Smaller models and faster training. In International Conference on Machine Learning, pages 10096-10106. PMLR, 2021. 3",
1527
+ "[56] Shitao Tang, Jiahui Zhang, Siyu Zhu, and Ping Tan. Quadtree attention for vision transformers. arXiv preprint arXiv:2201.02767, 2022. 3",
1528
+ "[57] Ilya O Tolstikhin, Neil Houlsby, Alexander Kolesnikov, Lucas Beyer, Xiaohua Zhai, Thomas Unterthiner, Jessica Yung, Andreas Steiner, Daniel Keysers, Jakob Uszkoreit, et al. Mlp-mixer: An all-mlp architecture for vision. Advances in Neural Information Processing Systems, 34:24261-24272, 2021. 1, 3",
1529
+ "[58] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In International Conference on Machine Learning, pages 10347-10357. PMLR, 2021. 3",
1530
+ "[59] Hugo Touvron, Matthieu Cord, and Hervé Jégou. Deit iii: Revenge of the vit. arXiv preprint arXiv:2204.07118, 2022. 3",
1531
+ "[60] Dmitry Ulyanov, Andrea Vedaldi, and Victor Lempitsky. Instance normalization: The missing ingredient for fast stylization. arXiv preprint arXiv:1607.08022, 2016. 4",
1532
+ "[61] Ashish Vaswani, Prajit Ramachandran, Aravind Srinivas, Niki Parmar, Blake Hechtman, and Jonathon Shlens. Scaling local self-attention for parameter efficient visual backbones. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12894–12904, 2021. 3",
1533
+ "[62] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 3",
1534
+ "[63] Shakti N Wadekar and Abhishek Chaurasia. Mobilevitv3: Mobile-friendly vision transformer with simple and effective fusion of local, global and input features. arXiv preprint arXiv:2209.15159, 2022. 1",
1535
+ "[64] Guangting Wang, Yucheng Zhao, Chuanxin Tang, Chong Luo, and Wenjun Zeng. When shift operation meets vision transformer: An extremely simple alternative to attention mechanism. arXiv preprint arXiv:2201.10801, 2022. 3",
1536
+ "[65] Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, and Ling Shao. Pyramid vision transformer: A versatile backbone for dense prediction without convolutions. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 568-578, 2021. 3, 7, 8",
1537
+ "[66] Song Wen, Hao Wang, and Dimitris Metaxas. Social ode: Multi-agent trajectory forecasting with neural ordinary differential equations. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXII, pages 217-233. Springer, 2022. 3",
1538
+ "[67] Bichen Wu, Xiaoliang Dai, Peizhao Zhang, Yanghan Wang, Fei Sun, Yiming Wu, Yuandong Tian, Peter Vajda, Yangqing Jia, and Kurt Keutzer. Fbnet: Hardware-aware efficient convnet design via differentiable neural architecture search. In"
1539
+ ],
1540
+ "bbox": [
1541
+ 78,
1542
+ 90,
1543
+ 468,
1544
+ 898
1545
+ ],
1546
+ "page_idx": 10
1547
+ },
1548
+ {
1549
+ "type": "list",
1550
+ "sub_type": "ref_text",
1551
+ "list_items": [
1552
+ "Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10734-10742, 2019. 3",
1553
+ "[68] Yuxin Wu and Kaiming He. Group normalization. In Proceedings of the European conference on computer vision (ECCV), pages 3-19, 2018. 4",
1554
+ "[69] Saining Xie, Ross Girshick, Piotr Dólár, Zhuowen Tu, and Kaiming He. Aggregated residual transformations for deep neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1492-1500, 2017. 8",
1555
+ "[70] Le Yang, Haojun Jiang, Ruojin Cai, Yulin Wang, Shiji Song, Gao Huang, and Qi Tian. Condensenet v2: Sparse feature reactivation for deep networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3569-3578, 2021. 3",
1556
+ "[71] Weihao Yu, Mi Luo, Pan Zhou, Chenyang Si, Yichen Zhou, Xinchao Wang, Jiashi Feng, and Shuicheng Yan. Metaformer is actually what you need for vision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10819-10829, 2022. 7, 8",
1557
+ "[72] Sangdoo Yun, Dongyoon Han, Seong Joon Oh, Sanghyuk Chun, Junsuk Choe, and Youngjoon Yoo. Cutmix: Regularization strategy to train strong classifiers with localizable features. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6023-6032, 2019. 6",
1558
+ "[73] Hongyi Zhang, Moustapha Cisse, Yann N Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization. arXiv preprint arXiv:1710.09412, 2017. 6",
1559
+ "[74] Qiulin Zhang, Zhuqing Jiang, Qishuo Lu, Jia'nan Han, Zhengxin Zeng, Shang-Hua Gao, and Aidong Men. Split to be slim: An overlooked redundancy in vanilla convolution. arXiv preprint arXiv:2006.12085, 2020. 4",
1560
+ "[75] Ting Zhang, Guo-Jun Qi, Bin Xiao, and Jingdong Wang. Interleaved group convolutions. In Proceedings of the IEEE international conference on computer vision, pages 4373-4382, 2017. 3",
1561
+ "[76] Xiangyu Zhang, Xinyu Zhou, Mengxiao Lin, and Jian Sun. Shufflenet: An extremely efficient convolutional neural network for mobile devices. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6848-6856, 2018. 1, 3",
1562
+ "[77] Shuhan Zhong, Sizhe Song, Guanyao Li, and S-H Gary Chan. A tree-based structure-aware transformer decoder for image-to-markup generation. In Proceedings of the 30th ACM International Conference on Multimedia, pages 5751-5760, 2022. 3",
1563
+ "[78] Weipeng Zhuo, Ka Ho Chiu, Jierun Chen, Jiajie Tan, Edmund Sumpena, S-H Gary Chan, Sangtae Ha, and Chul-Ho Lee. Semi-supervised learning with network embedding on ambient rf signals for geofencing services. arXiv preprint arXiv:2210.07889, 2022. 3",
1564
+ "[79] Barret Zoph and Quoc V Le. Neural architecture search with reinforcement learning. arXiv preprint arXiv:1611.01578, 2016. 6"
1565
+ ],
1566
+ "bbox": [
1567
+ 501,
1568
+ 92,
1569
+ 890,
1570
+ 853
1571
+ ],
1572
+ "page_idx": 10
1573
+ },
1574
+ {
1575
+ "type": "page_number",
1576
+ "text": "12031",
1577
+ "bbox": [
1578
+ 480,
1579
+ 926,
1580
+ 517,
1581
+ 936
1582
+ ],
1583
+ "page_idx": 10
1584
+ }
1585
+ ]
2023/Run, Don't Walk_ Chasing Higher FLOPS for Faster Neural Networks/6a2198d1-cecd-4fd5-b373-fd2aeb690abb_model.json ADDED
The diff for this file is too large to render. See raw diff
 
2023/Run, Don't Walk_ Chasing Higher FLOPS for Faster Neural Networks/6a2198d1-cecd-4fd5-b373-fd2aeb690abb_origin.pdf ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:a7ad310892017e2f6407eb1484103c5e9c047b9cb693af5eddd3da32449bacd7
3
+ size 886667
2023/Run, Don't Walk_ Chasing Higher FLOPS for Faster Neural Networks/full.md ADDED
@@ -0,0 +1,338 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Run, Don’t Walk: Chasing Higher FLOPS for Faster Neural Networks
2
+
3
+ Jierun Chen $^{1}$ , Shiu-hong Kao $^{1}$ , Hao He $^{1}$
4
+ Weipeng Zhuo $^{1}$ , Song Wen $^{2}$ , Chul-Ho Lee $^{3}$ , S.-H. Gary Chan $^{1}$ $^{1}$ HKUST, $^{2}$ Rutgers University, $^{3}$ Texas State University
5
+
6
+ # Abstract
7
+
8
+ To design fast neural networks, many works have been focusing on reducing the number of floating-point operations (FLOPs). We observe that such reduction in FLOPs, however, does not necessarily lead to a similar level of reduction in latency. This mainly stems from inefficiently low floating-point operations per second (FLOPS). To achieve faster networks, we revisit popular operators and demonstrate that such low FLOPS is mainly due to frequent memory access of the operators, especially the depthwise convolution. We hence propose a novel partial convolution (PCnv) that extracts spatial features more efficiently, by cutting down redundant computation and memory access simultaneously. Building upon our PCnv, we further propose FasterNet, a new family of neural networks, which attains substantially higher running speed than others on a wide range of devices, without compromising on accuracy for various vision tasks. For example, on ImageNet-1k, our tiny FasterNet-T0 is $2.8 \times 3.3 \times$ and $2.4 \times$ faster than MobileViT-XXS on GPU, CPU, and ARM processors, respectively, while being $2.9\%$ more accurate. Our large FasterNet-L achieves impressive $83.5\%$ top-1 accuracy, on par with the emerging Swin-B, while having $36\%$ higher inference throughput on GPU, as well as saving $37\%$ compute time on CPU. Code is available at https://github.com/JierunChen/FasterNet.
9
+
10
+ # 1. Introduction
11
+
12
+ Neural networks have undergone rapid development in various computer vision tasks such as image classification, detection and segmentation. While their impressive performance has powered many applications, a roaring trend is to pursue fast neural networks with low latency and high throughput for great user experiences, instant responses, safety reasons, etc.
13
+
14
+ How to be fast? Instead of asking for more costly computing devices, researchers and practitioners prefer to design cost-effective fast neural networks with reduced computational complexity, mainly measured in the number of
15
+
16
+ ![](images/c8cfeebc3fde8d06bfd86b330dae02078ebc45c79a1a14f32c35eaaaa1393820.jpg)
17
+ (a) Convolution
18
+
19
+ ![](images/1d015ffc47d7943c325b92090435436c20108bf9c1713ea630e73d7b6ad1b0b6.jpg)
20
+ (b) Depthwise/Group Convolution
21
+
22
+ ![](images/ef34393040525a2e7dacb9ed8654cc27ceb71d56a7f92940dd88d05170faf8d2.jpg)
23
+ (c) Partial Convolution (ours)
24
+ Figure 1. Our partial convolution (PConv) is fast and efficient by applying filters on only a few input channels while leaving the remaining ones untouched. PConv obtains lower FLOPs than the regular convolution and higher FLOPS than the depthwise/group convolution.
25
+
26
+ Input/output
27
+ Filter
28
+ \*Convolution
29
+ $\rightarrow$ Identity
30
+
31
+ floating-point operations (FLOPs) $^1$ . MobileNets [20, 21, 47], ShuffleNets [40, 76] and GhostNet [13], among others, leverage the depthwise convolution (DWConv) [48] and/or group convolution (GConv) [27] to extract spatial features. However, in the effort to reduce FLOPs, the operators often suffer from the side effect of increased memory access. MicroNet [29] further decomposes and sparsifies the network to push its FLOPs to an extremely low level. Despite its improvement in FLOPs, this approach experiences inefficient fragmented computation. Besides, the above networks are often accompanied by additional data manipulations, such as concatenation, shuffling, and pooling, whose running time tends to be significant for tiny models.
32
+
33
+ Apart from the above pure convolutional neural networks (CNNs), there is an emerging interest in making vision transformers (ViTs) [11] and multilayer perceptrons (MLPs) architectures [57] smaller and faster. For example, MobileViTs [42, 43, 63] and MobileFormer [6] reduce the computational complexity by combining DWConv with a modified attention mechanism. However, they still suffer from the aforementioned issue with DWConv and also need dedicated hardware support for the modified attention mechanism. The use of advanced yet time-consuming nor
34
+
35
+ ![](images/bc689bdb3a1a6f8205702f0379a1f5c1f8be0e664dd4c0e302b8927ca1ec1b3b.jpg)
36
+
37
+ ![](images/bcfdc95e5623ea475b3f4225deff2acec6dc75fdab39cada93fdc6eb8e26e653.jpg)
38
+ (a)
39
+
40
+ ![](images/52013b4b7475c1d42f589de8b8dc903a733f7e78b03a0098e6384b5eb0496566.jpg)
41
+
42
+ ![](images/173d56cec53757d48dc7e1af33b12d573a1492f15e75fd81fcccb8649b4d37eb.jpg)
43
+ (b)
44
+ Figure 2. (a) FLOPS under varied FLOPs on CPU. Many existing neural networks suffer from low computational speed issues. Their effective FLOPS are lower than the popular ResNet50. By contrast, our FasterNet attains higher FLOPS. (b) Latency under varied FLOPs on CPU. Our FasterNet obtains lower latency than others with the same amount of FLOPs.
45
+
46
+ malization and activation layers may also limit their speed on devices.
47
+
48
+ All these issues together lead to the following question: Are these "fast" neural networks really fast? To answer this, we examine the relationship between latency and FLOPs, which is captured by
49
+
50
+ $$
51
+ L a t e n c y = \frac {F L O P s}{F L O P S}, \tag {1}
52
+ $$
53
+
54
+ where FLOPS is short for floating-point operations per second, as a measure of the effective computational speed. While there are many attempts to reduce FLOPs, they seldom consider optimizing FLOPS at the same time to achieve truly low latency. To better understand the situation, we compare the FLOPS of typical neural networks on an Intel CPU. The results in Fig. 2 show that many existing neural networks suffer from low FLOPS, and their FLOPS is generally lower than the popular ResNet50. With such low FLOPS, these "fast" neural networks are actually not fast enough. Their reduction in FLOPs cannot be translated into the exact amount of reduction in latency. In some cases, there is no improvement, and it even leads to worse latency. For example, CycleMLP-B1 [5] has half of FLOPs of ResNet50 [16] but runs more slowly (i.e., CycleMLP-B1 vs. ResNet50: 116.1ms vs. 73.0ms). Note that this discrepancy between FLOPs and latency has also been noticed in previous works [40, 42] but remains unresolved partially because they employ the DWConv/GConv and various data manipulations with low FLOPS. It is deemed there are no better alternatives available.
55
+
56
+ This paper aims to eliminate the discrepancy by developing a simple yet fast and effective operator that maintains high FLOPS with reduced FLOPs. Specifically, we reexamine existing operators, particularly DWConv, in terms of the computational speed - FLOPS. We uncover that the main reason causing the low FLOPS issue is frequent memory access. We then propose a novel partial convolution (PCov)
57
+
58
+ as a competitive alternative that reduces the computational redundancy as well as the number of memory access. Fig. 1 illustrates the design of our PConv. It takes advantage of redundancy within the feature maps and systematically applies a regular convolution (Conv) on only a part of the input channels while leaving the remaining ones untouched. By nature, PConv has lower FLOPs than the regular Conv while having higher FLOPS than the DWConv/GConv. In other words, PConv better exploits the on-device computational capacity. PConv is also effective in extracting spatial features as empirically validated later in the paper.
59
+
60
+ We further introduce FasterNet, which is primarily built upon our PConv, as a new family of networks that run highly fast on various devices. In particular, our FasterNet achieves state-of-the-art performance for classification, detection, and segmentation tasks while having much lower latency and higher throughput. For example, our tiny FasterNet-T0 is $2.8 \times 3.3 \times$ and $2.4 \times$ faster than MobileViT-XXS [42] on GPU, CPU, and ARM processors, respectively, while being $2.9\%$ more accurate on ImageNet-1k. Our large FasterNet-L achieves $83.5\%$ top-1 accuracy, on par with the emerging Swin-B [35], while offering $36\%$ higher throughput on GPU and saving $37\%$ compute time on CPU. To summarize, our contributions are as follows:
61
+
62
+ - We point out the importance of achieving higher FLOPS beyond simply reducing FLOPs for faster neural networks.
63
+ - We introduce a simple yet fast and effective operator called PConv, which has a high potential to replace the existing go-to choice, DWConv.
64
+ - We introduce FasterNet which runs favorably and universally fast on a variety of devices such as GPU, CPU, and ARM processors.
65
+ - We conduct extensive experiments on various tasks and validate the high speed and effectiveness of our PConv and FasterNet.
66
+
67
+ # 2. Related Work
68
+
69
+ We briefly review prior works on fast and efficient neural networks and differentiate this work from them.
70
+
71
+ CNN. CNNs are the mainstream architecture in the computer vision field, especially when it comes to deployment in practice, where being fast is as important as being accurate. Though there have been numerous studies [7, 8, 17, 29, 48, 49, 75, 78] to achieve higher efficiency, the rationale behind them is more or less to perform a low-rank approximation. Specifically, the group convolution [27] and the depthwise separable convolution [48] (consisting of depthwise and pointwise convolutions) are probably the most popular ones. They have been widely adopted in mobile/edge-oriented networks, such as MobileNets [20, 21, 47], ShuffleNets [40, 76], GhostNet [13], EfficientNets [54, 55], TinyNet [14], Xception [8], CondenseNet [23, 70], TVConv [4], MnasNet [53], and FBNet [67]. While they exploit the redundancy in filters to reduce the number of parameters and FLOPs, they suffer from increased memory access when increasing the network width to compensate for the accuracy drop. By contrast, we consider the redundancy in feature maps and propose a partial convolution to reduce FLOPs and memory access simultaneously.
72
+
73
+ ViT, MLP, and variants. There is a growing interest in studying ViT ever since Dosovitskiy et al. [11] expanded the application scope of transformers [62] from machine translation [62] or forecasting [66] to the computer vision field. Many follow-up works have attempted to improve ViT in terms of training setting [51, 58, 59] and model design [12, 34, 35, 65, 77]. One notable trend is to pursue a better accuracy-latency trade-off by reducing the complexity of the attention operator [1, 25, 39, 56, 61], incorporating convolution into ViTs [6, 10, 50], or doing both [3, 30, 43, 45]. Besides, other studies [5, 31, 57] propose to replace the attention with simple MLP-based operators. However, they often evolve to be CNN-like [33]. In this paper, we focus on analyzing the convolution operations, particularly DWConv, due to the following reasons: First, the advantage of attention over convolution is unclear or debatable [36, 64]. Second, the attention-based mechanism generally runs slower than its convolutional counterparts and thus becomes less favorable for the current industry [22, 42]. Finally, DW-Conv is still a popular choice in many hybrid models, so it is worth a careful examination.
74
+
75
+ # 3. Design of PConv and FasterNet
76
+
77
+ In this section, we first revisit DWConv and analyze the issue with its frequent memory access. We then introduce PConv as a competitive alternative operator to resolve the issue. After that, we introduce FasterNet and explain its
78
+
79
+ ![](images/43b8b037deb68c10cea574d7d56f635729de216ab6d0f26d1b74b15ba6828430.jpg)
80
+ Figure 3. Visualization of feature maps in an intermediate layer of a pre-trained ResNet50, with the top-left image as the input. Qualitatively, we can see the high redundancies across different channels.
81
+
82
+ details, including design considerations.
83
+
84
+ # 3.1. Preliminary
85
+
86
+ DWConv is a popular variant of Conv and has been widely adopted as a key building block for many neural networks. For an input $\mathbf{I} \in \mathbb{R}^{c \times h \times w}$ , DWConv applies $c$ filters $\mathbf{W} \in \mathbb{R}^{k \times k}$ to compute the output $\mathbf{O} \in \mathbb{R}^{c \times h \times w}$ . As shown in Fig. 1(b), each filter slides spatially on one input channel and contributes to one output channel. This depthwise computation makes DWConv have as low FLOPs as $h \times w \times k^2 \times c$ compared to a regular Conv with $h \times w \times k^2 \times c^2$ . While effective in reducing FLOPs, a DWConv, which is typically followed by a pointwise convolution, or PWConv, cannot be simply used to replace a regular Conv as it would incur a severe accuracy drop. Thus, in practice the channel number $c$ (or the network width) of DWConv is increased to $c'(c' > c)$ to compensate the accuracy drop, e.g., the width is expanded by six times for the DWConv in the inverted residual blocks [47]. This, however, results in much higher memory access that can cause non-negligible delay and slow down the overall computation, especially for I/O-bound devices. In particular, the number of memory access now escalates to
87
+
88
+ $$
89
+ h \times w \times 2 c ^ {\prime} + k ^ {2} \times c ^ {\prime} \approx h \times w \times 2 c ^ {\prime}, \tag {2}
90
+ $$
91
+
92
+ which is higher than that of a regular Conv, i.e.,
93
+
94
+ $$
95
+ h \times w \times 2 c + k ^ {2} \times c ^ {2} \approx h \times w \times 2 c. \tag {3}
96
+ $$
97
+
98
+ Note that the $h \times w \times 2c'$ memory access is spent on the I/O operation, which is deemed to be already the minimum cost and hard to optimize further.
99
+
100
+ # 3.2. Partial convolution as a basic operator
101
+
102
+ We below demonstrate that the cost can be further optimized by leveraging the feature maps' redundancy. As visualized in Fig. 3, the feature maps share high similarities among different channels. This redundancy has also
103
+
104
+ been covered in many other works [13, 74], but few of them make full use of it in a simple yet effective way.
105
+
106
+ Specifically, we propose a simple PConv to reduce computational redundancy and memory access simultaneously. The bottom-left corner in Fig. 4 illustrates how our PConv works. It simply applies a regular Conv on only a part of the input channels for spatial feature extraction and leaves the remaining channels untouched. For contiguous or regular memory access, we consider the first or last consecutive $c_p$ channels as the representatives of the whole feature maps for computation. Without loss of generality, we consider the input and output feature maps to have the same number of channels. Therefore, the FLOPs of a PConv are only
107
+
108
+ $$
109
+ h \times w \times k ^ {2} \times c _ {p} ^ {2}. \tag {4}
110
+ $$
111
+
112
+ With a typical partial ratio $r = \frac{c_p}{c} = \frac{1}{4}$ , the FLOPs of a PConv is only $\frac{1}{16}$ of a regular Conv. Besides, PConv has a smaller amount of memory access, i.e.,
113
+
114
+ $$
115
+ h \times w \times 2 c _ {p} + k ^ {2} \times c _ {p} ^ {2} \approx h \times w \times 2 c _ {p}, \tag {5}
116
+ $$
117
+
118
+ which is only $\frac{1}{4}$ of a regular Conv for $r = \frac{1}{4}$ .
119
+
120
+ Since there are only $c_p$ channels utilized for spatial feature extraction, one may ask if we can simply remove the remaining $(c - c_p)$ channels? If so, PConv would degrade to a regular Conv with fewer channels, which deviates from our objective to reduce redundancy. Note that we keep the remaining channels untouched instead of removing them from the feature maps. It is because they are useful for a subsequent PWConv layer, which allows the feature information to flow through all channels.
121
+
122
+ # 3.3. PConv followed by PWConv
123
+
124
+ To fully and efficiently leverage the information from all channels, we further append a pointwise convolution (PW-Conv) to our PConv. Their effective receptive field together on the input feature maps looks like a T-shaped Conv, which focuses more on the center position compared to a regular Conv uniformly processing a patch, as shown in Fig. 5. To justify this T-shaped receptive field, we first evaluate the importance of each position by calculating the position-wise Frobenius norm. We assume that a position tends to be more important if it has a larger Frobenius norm than other positions. For a regular Conv filter $\mathbf{F} \in \mathbb{R}^{k^2 \times c}$ , the Frobenius norm at position $i$ is calculated by $\| \mathbf{F}_i \| = \sqrt{\sum_{j=1}^{c} |f_{ij}|^2}$ , for $i = 1, 2, 3\ldots, k^2$ . We consider a salient position to be the one with the maximum Frobenius norm. We then collectively examine each filter in a pre-trained ResNet18, find out their salient positions, and plot a histogram of the salient positions. Results in Fig. 6 show that the center position turns out to be the salient position most frequently among the filters. In other words, the center position weighs more
125
+
126
+ than its surrounding neighbors. This is consistent with the T-shaped computation which concentrates on the center position.
127
+
128
+ While the T-shaped Conv can be directly used for efficient computation, we show that it is better to decompose the T-shaped Conv into a PConv and a PWConv because the decomposition exploits the inter-filter redundancy and further saves FLOPs. For the same input $\mathbf{I} \in \mathbb{R}^{c \times h \times w}$ and output $\mathbf{O} \in \mathbb{R}^{c \times h \times w}$ , a T-shaped Conv's FLOPs can be calculated as
129
+
130
+ $$
131
+ h \times w \times \left(k ^ {2} \times c _ {p} \times c + c \times \left(c - c _ {p}\right)\right), \tag {6}
132
+ $$
133
+
134
+ which is higher than the FLOPs of a PConv and a PWConv, i.e.,
135
+
136
+ $$
137
+ h \times w \times \left(k ^ {2} \times c _ {p} ^ {2} + c \times c _ {p}\right), \tag {7}
138
+ $$
139
+
140
+ where $c > c_p$ and $c - c_p > c_p$ (e.g. when $c_p = \frac{c}{4}$ ). Besides, we can readily leverage the regular Conv for the two-step implementation.
141
+
142
+ # 3.4. FasterNet as a general backbone
143
+
144
+ Given our novel PConv and off-the-shelf PWConv as the primary building operators, we further propose FasterNet, a new family of neural networks that runs favorably fast and is highly effective for many vision tasks. We aim to keep the architecture as simple as possible, without bells and whistles, to make it hardware-friendly in general.
145
+
146
+ We present the overall architecture in Fig. 4. It has four hierarchical stages, each of which is preceded by an embedding layer (a regular Conv $4 \times 4$ with stride 4) or a merging layer (a regular Conv $2 \times 2$ with stride 2) for spatial downsampling and channel number expanding. Each stage has a stack of FasterNet blocks. We observe that the blocks in the last two stages consume less memory access and tend to have higher FLOPS, as empirically validated in Tab. 1. Thus, we put more FasterNet blocks and correspondingly assign more computations to the last two stages. Each FasterNet block has a PConv layer followed by two PWConv (or Conv $1 \times 1$ ) layers. Together, they appear as inverted residual blocks where the middle layer has an expanded number of channels, and a shortcut connection is placed to reuse the input features.
147
+
148
+ In addition to the above operators, the normalization and activation layers are also indispensable for high-performing neural networks. Many prior works [13, 16, 47], however, overuse such layers throughout the network, which may limit the feature diversity and thus hurt the performance. It can also slow down the overall computation. By contrast, we put them only after each middle PWConv to preserve the feature diversity and achieve lower latency. Besides, we use the batch normalization (BN) [26] instead of other alternative ones [2, 60, 68]. The benefit of BN is that it can be merged into its adjacent Conv layers for faster inference
149
+
150
+ ![](images/4cbebcfcba214df5a0773b9aac408ea26fe033a4af1473a805836ce83676b38c.jpg)
151
+
152
+ ![](images/70be33816e6e05deda16ab9e60a97129c8caba1c4290915d28b7400b1ac8fd84.jpg)
153
+ (a) PConv + PWConv
154
+
155
+ ![](images/1cc9a62bb3d40acb040421f220c81886e43d4bbeade06a31e06c9d39b0107c51.jpg)
156
+ (b) T-shaped Conv
157
+
158
+ ![](images/fb412d9f49483344d931ecb60a4c7b368f1ddac0868fdaca07ac7ead2459edc1.jpg)
159
+ (c) regular Conv
160
+ Figure 5. Comparison of convolutional variants. A PConv followed by a PWConv (a) resembles a T-shaped Conv (b), which spends more computation on the center position compared to a regular Conv (c).
161
+
162
+ while being as effective as the others. As for the activation layers, we empirically choose GELU [18] for smaller FasterNet variants and ReLU [44] for bigger FasterNet variants, considering both running time and effectiveness. The last three layers, i.e. a global average pooling, a Conv $1 \times 1$ , and a fully-connected layer, are used together for feature transformation and classification.
163
+
164
+ To serve a wide range of applications under different computational budgets, we provide tiny, small, medium, and large variants of FasterNet, referred to as FasterNet-T0/1/2, FasterNet-S, FasterNet-M, and FasterNet-L, respectively. They share a similar architecture but vary in depth and width. Detailed architecture specifications are provided in the appendix.
165
+
166
+ # 4. Experimental Results
167
+
168
+ We first examine the computational speed of our PConv and its effectiveness when combined with a PWConv. We then comprehensively evaluate the performance of our FasterNet for classification, detection, and segmentation tasks. Finally, we conduct a brief ablation study.
169
+
170
+ To benchmark the latency and throughput, we choose the following three typical processors, which cover a wide
171
+
172
+ ![](images/c3e91ad083f5845c8ed7743c6664f13f9076194653b00f9c257d1e8f0f658866.jpg)
173
+ Figure 4. Overall architecture of our FasterNet. It has four hierarchical stages, each with a stack of FasterNet blocks and preceded by an embedding or merging layer. The last three layers are used for feature classification. Within each FasterNet block, a PConv layer is followed by two PWConv layers. We put normalization and activation layers only after the middle layer to preserve the feature diversity and achieve lower latency.
174
+ Figure 6. Histogram of salient position distribution for the regular Conv $3 \times 3$ filters in a pre-trained ResNet18. The histogram contains four kinds of bars, corresponding to different stages in the network. In all stages, the center position (position 5) appears as a salient position most frequently.
175
+
176
+ range of computational capacity: GPU (2080Ti), CPU (Intel i9-9900X, using a single thread), and ARM (Cortex-A72, using a single thread). We report their latency for inputs with a batch size of 1 and throughput for inputs with a batch size of 32. During inference, the BN layers are merged to their adjacent layers wherever applicable.
177
+
178
+ # 4.1. PConv is fast with high FLOPS
179
+
180
+ We below show that our PConv is fast and better exploits the on-device computational capacity. Specifically, we stack 10 layers of pure PConv and take feature maps of typical dimensions as inputs. We then measure FLOPs and latency/throughput on GPU, CPU, and ARM processors, which also allow us to further compute FLOPS. We repeat the same procedure for other convolutional variants and make comparisons.
181
+
182
+ Results in Tab. 1 show that PConv is overall an appealing choice for high FLOPS with reduced FLOPs. It has only $\frac{1}{16}$ FLOPs of a regular Conv and achieves $10.5\times$ , $6.2\times$ , and $22.8\times$ higher FLOPS than the DWConv on GPU, CPU, and ARM, respectively. We are unsurprised to see that the regular Conv has the highest FLOPS as it has been constantly optimized for years. However, its total FLOPs and
183
+
184
+ <table><tr><td>Operator</td><td>Feature map size</td><td>FLOPs (M), ×10 layers</td><td>GPU Throughput (fps)</td><td>FLOPS (G/s)</td><td>CPU Latency (ms)</td><td>FLOPS (G/s)</td><td>ARM Latency (ms)</td><td>FLOPS (G/s)</td></tr><tr><td rowspan="5">Conv 3×3</td><td>96×56×56</td><td>2601</td><td>3010</td><td>7824</td><td>35.67</td><td>72.90</td><td>779.57</td><td>3.33</td></tr><tr><td>192×28×28</td><td>2601</td><td>4893</td><td>12717</td><td>28.41</td><td>91.53</td><td>619.64</td><td>4.19</td></tr><tr><td>384×14×14</td><td>2601</td><td>4558</td><td>11854</td><td>31.85</td><td>81.66</td><td>595.09</td><td>4.37</td></tr><tr><td>768×7×7</td><td>2601</td><td>3159</td><td>8212</td><td>62.71</td><td>41.47</td><td>662.17</td><td>3.92</td></tr><tr><td colspan="2">Average</td><td>-</td><td>10151</td><td>-</td><td>71.89</td><td>-</td><td>3.95</td></tr><tr><td rowspan="5">GConv 3×3 (16 groups)</td><td>96×56×56</td><td>162</td><td>2888</td><td>469</td><td>21.90</td><td>7.42</td><td>166.30</td><td>0.97</td></tr><tr><td>192×28×28</td><td>162</td><td>10811</td><td>1754</td><td>7.58</td><td>21.44</td><td>96.22</td><td>1.68</td></tr><tr><td>384×14×14</td><td>162</td><td>15534</td><td>2514</td><td>4.40</td><td>36.88</td><td>63.57</td><td>2.55</td></tr><tr><td>768×7×7</td><td>162</td><td>16000</td><td>2598</td><td>4.28</td><td>37.97</td><td>65.20</td><td>2.49</td></tr><tr><td colspan="2">Average</td><td>-</td><td>1833</td><td>-</td><td>25.93</td><td>-</td><td>1.92</td></tr><tr><td rowspan="5">DWConv 3×3</td><td>96×56×56</td><td>27.09</td><td>11940</td><td>323</td><td>3.59</td><td>7.52</td><td>108.70</td><td>0.24</td></tr><tr><td>192×28×28</td><td>13.54</td><td>23358</td><td>315</td><td>1.97</td><td>6.86</td><td>82.01</td><td>0.16</td></tr><tr><td>384×14×14</td><td>6.77</td><td>46377</td><td>313</td><td>1.06</td><td>6.35</td><td>94.89</td><td>0.07</td></tr><tr><td>768×7×7</td><td>3.38</td><td>88889</td><td>302</td><td>0.68</td><td>4.93</td><td>150.89</td><td>0.02</td></tr><tr><td colspan="2">Average</td><td>-</td><td>313</td><td>-</td><td>6.42</td><td>-</td><td>0.12</td></tr><tr><td rowspan="5">PConv 3×3 (ours, with r = 1/4)</td><td>96×56×56</td><td>162</td><td>9215</td><td>1493</td><td>5.46</td><td>29.67</td><td>85.30</td><td>1.90</td></tr><tr><td>192×28×28</td><td>162</td><td>14360</td><td>2326</td><td>3.09</td><td>52.43</td><td>66.46</td><td>2.44</td></tr><tr><td>384×14×14</td><td>162</td><td>24408</td><td>3954</td><td>3.58</td><td>45.25</td><td>49.98</td><td>3.24</td></tr><tr><td>768×7×7</td><td>162</td><td>32866</td><td>5324</td><td>5.02</td><td>32.27</td><td>48.30</td><td>3.35</td></tr><tr><td colspan="2">Average</td><td>-</td><td>3274</td><td>-</td><td>39.91</td><td>-</td><td>2.73</td></tr></table>
185
+
186
+ latency/throughput are unaffordable. GConv and DWConv, despite their significant reduction in FLOPs, suffer from a drastic decrease in FLOPS. In addition, they tend to increase the number of channels to compensate for the performance drop, which, however, increase their latency.
187
+
188
+ # 4.2. PConv is effective together with PWConv
189
+
190
+ We next show that a PConv followed by a PWConv is effective in approximating a regular Conv to transform the feature maps. To this end, we first build four datasets by feeding the ImageNet-1k val split images into a pre-trained ResNet50, and extract the feature maps before and after the first Conv $3 \times 3$ in each of the four stages. Each feature map dataset is further spilt into the train $(70\%)$ , val $(10\%)$ , and test $(20\%)$ subsets. We then build a simple network consisting of a PConv followed by a PWConv and train it on the feature map datasets with a mean squared error loss. For comparison, we also build and train networks for DWConv + PWConv and GConv + PWConv under the same setting.
191
+
192
+ Tab. 2 shows that PConv + PWConv achieve the lowest test loss, meaning that they better approximate a regular Conv in feature transformation. The results also suggest that it is sufficient and efficient to capture spatial features from only a part of the feature maps. PConv shows a great potential to be the new go-to choice in designing fast and effective neural networks.
193
+
194
+ # 4.3. FasterNet on ImageNet-1k classification
195
+
196
+ To verify the effectiveness and efficiency of our FasterNet, we first conduct experiments on the large-scale ImageNet-1k classification dataset [46]. It covers 1k categories of common objects and contains about 1.3M la
197
+
198
+ Table 1. On-device FLOPS for different operations. PConv appears as an appealing choice for high FLOPS with reduced FLOPs.
199
+
200
+ <table><tr><td>Stage</td><td>DWConv+PWConv</td><td>GConv+PWConv(16 groups)</td><td>PConv+PWConvr = 1/4</td></tr><tr><td>1</td><td>0.0089</td><td>0.0065</td><td>0.0069</td></tr><tr><td>2</td><td>0.0158</td><td>0.0137</td><td>0.0136</td></tr><tr><td>3</td><td>0.0214</td><td>0.0202</td><td>0.0172</td></tr><tr><td>4</td><td>0.0130</td><td>0.0128</td><td>0.0115</td></tr><tr><td>Average</td><td>0.0148</td><td>0.0133</td><td>0.0123</td></tr></table>
201
+
202
+ Table 2. A PConv followed by a PWConv well approximates the regular Conv $3 \times 3$ at different stages of a pre-trained ResNet50. PConv + PWConv together have the lowest test loss on average.
203
+
204
+ beled images for training and 50k labeled images for validation. We train our models for 300 epochs using AdamW optimizer [38]. We set the batch size to 2048 for the FasterNet-M/L and 4096 for other variants. We use cosine learning rate scheduler [37] with a peak value of $0.001 \cdot$ batch size/1024 and a 20-epoch linear warmup. We apply commonly-used regularization and augmentation techniques, including Weight Decay [28], Stochastic Depth [24], Label Smoothing [52], Mixup [73], Cutmix [72] and Rand Augment [9], with varying magnitudes for different FasterNet variants. To reduce the training time, we use $192 \times 192$ resolution for the first 280 training epochs and $224 \times 224$ for the remaining 20 epochs. For fair comparison, we do not use knowledge distillation [19] and neural architecture search [79]. We report our top-1 accuracy on the validation set with a center crop at $224 \times 224$ resolution and a 0.9 crop ratio. Detailed training and validation settings are provided in the appendix.
205
+
206
+ Fig. 7 and Tab. 3 demonstrate the superiority of our FasterNet over state-of-the-art classification models. The trade-off curves in Fig. 7 clearly show that FasterNet sets
207
+
208
+ ![](images/815b3c3a86d1a130a402adcae33bb24083901f97882f97150fb926cc712b2e97.jpg)
209
+
210
+ ![](images/9ee0963af0ff702c4f7440623b2fef107742aa7aa992df772c6ff8f22d9f3054.jpg)
211
+
212
+ ![](images/5959c7303514d8b8e2d1673f69f30c473f3e2009cb5c8ddc69a8bd14351a9e35.jpg)
213
+ Figure 7. FasterNet has the highest efficiency in balancing accuracy-throughput and accuracy-latency trade-offs for different devices. To save space and make the plots more proportionate, we showcase network variants within a certain range of latency. Full plots can be found in the appendix, which show consistent results.
214
+
215
+ the new state-of-the-art in balancing accuracy and latency/throughput among all the networks examined. From another perspective, FasterNet runs faster than various CNN, ViT and MLP models on a wide range of devices, when having similar top-1 accuracy. As quantitatively shown in Tab. 3, FasterNet-T0 is $2.8 \times 3.3 \times$ and $2.4 \times$ faster than MobileViT-XXS [42] on GPU, CPU, and ARM processors, respectively, while being $2.9\%$ more accurate. Our large FasterNet-L achieves $83.5\%$ top-1 accuracy, comparable to the emerging Swin-B [35] and ConvNeXt-B [36] while having $36\%$ and $28\%$ higher inference throughput on GPU, as well as saving $37\%$ and $15\%$ compute time on CPU. Given such promising results, we highlight that our FasterNet is
216
+
217
+ <table><tr><td>Network</td><td>Params (M)</td><td>FLOPs (G)</td><td>Throughput on GPU (fps) ↑</td><td>Latency on CPU (ms) ↓</td><td>Latency on ARM (ms) ↓</td><td>Acc. (%)</td></tr><tr><td>ShuffleNetV2 ×1.5 [40]</td><td>3.5</td><td>0.30</td><td>4878</td><td>12.1</td><td>266</td><td>72.6</td></tr><tr><td>MobileNetV2 [47]</td><td>3.5</td><td>0.31</td><td>4198</td><td>12.2</td><td>442</td><td>72.0</td></tr><tr><td>MobileViT-XXS [42]</td><td>1.3</td><td>0.42</td><td>2393</td><td>30.8</td><td>348</td><td>69.0</td></tr><tr><td>EdgeNeXt-XXS [41]</td><td>1.3</td><td>0.26</td><td>2765</td><td>15.7</td><td>239</td><td>71.2</td></tr><tr><td>FasterNet-T0</td><td>3.9</td><td>0.34</td><td>6807</td><td>9.2</td><td>143</td><td>71.9</td></tr><tr><td>GhostNet ×1.3 [13]</td><td>7.4</td><td>0.24</td><td>2988</td><td>17.9</td><td>481</td><td>75.7</td></tr><tr><td>ShuffleNetV2 ×2 [40]</td><td>7.4</td><td>0.59</td><td>3339</td><td>17.8</td><td>403</td><td>74.9</td></tr><tr><td>MobileNetV2 ×1.4 [47]</td><td>6.1</td><td>0.60</td><td>2711</td><td>22.6</td><td>650</td><td>74.7</td></tr><tr><td>MobileViT-XS [42]</td><td>2.3</td><td>1.05</td><td>1392</td><td>40.8</td><td>648</td><td>74.8</td></tr><tr><td>EdgeNeXt-XS [41]</td><td>2.3</td><td>0.54</td><td>1738</td><td>24.4</td><td>434</td><td>75.0</td></tr><tr><td>PVT-Tiny [65]</td><td>13.2</td><td>1.94</td><td>1266</td><td>55.6</td><td>708</td><td>75.1</td></tr><tr><td>FasterNet-T1</td><td>7.6</td><td>0.85</td><td>3782</td><td>17.7</td><td>285</td><td>76.2</td></tr><tr><td>CycleMLP-B1 [5]</td><td>15.2</td><td>2.10</td><td>865</td><td>116.1</td><td>892</td><td>79.1</td></tr><tr><td>PoolFormer-S12 [71]</td><td>11.9</td><td>1.82</td><td>1439</td><td>49.0</td><td>665</td><td>77.2</td></tr><tr><td>MobileViT-S [42]</td><td>5.6</td><td>2.03</td><td>1039</td><td>56.7</td><td>941</td><td>78.4</td></tr><tr><td>EdgeNeXt-S [41]</td><td>5.6</td><td>1.26</td><td>1128</td><td>39.2</td><td>743</td><td>79.4</td></tr><tr><td>ResNet50 [16, 36]</td><td>25.6</td><td>4.11</td><td>959</td><td>73.0</td><td>1131</td><td>78.8</td></tr><tr><td>FasterNet-T2</td><td>15.0</td><td>1.91</td><td>1991</td><td>33.5</td><td>497</td><td>78.9</td></tr><tr><td>CycleMLP-B2 [5]</td><td>26.8</td><td>3.90</td><td>528</td><td>186.3</td><td>1502</td><td>81.6</td></tr><tr><td>PoolFormer-S24 [71]</td><td>21.4</td><td>3.41</td><td>748</td><td>92.8</td><td>1261</td><td>80.3</td></tr><tr><td>PoolFormer-S36 [71]</td><td>30.9</td><td>5.00</td><td>507</td><td>138.0</td><td>1860</td><td>81.4</td></tr><tr><td>ConvNeXt-T [36]</td><td>28.6</td><td>4.47</td><td>657</td><td>86.3</td><td>1889</td><td>82.1</td></tr><tr><td>Swin-T [35]</td><td>28.3</td><td>4.51</td><td>609</td><td>122.2</td><td>1424</td><td>81.3</td></tr><tr><td>PVT-Small [65]</td><td>24.5</td><td>3.83</td><td>689</td><td>89.6</td><td>1345</td><td>79.8</td></tr><tr><td>PVT-Medium [65]</td><td>44.2</td><td>6.69</td><td>438</td><td>143.6</td><td>2142</td><td>81.2</td></tr><tr><td>FasterNet-S</td><td>31.1</td><td>4.56</td><td>1029</td><td>71.2</td><td>1103</td><td>81.3</td></tr><tr><td>PoolFormer-M36 [71]</td><td>56.2</td><td>8.80</td><td>320</td><td>215.0</td><td>2979</td><td>82.1</td></tr><tr><td>ConvNeXt-S [36]</td><td>50.2</td><td>8.71</td><td>377</td><td>153.2</td><td>3484</td><td>83.1</td></tr><tr><td>Swin-S [35]</td><td>49.6</td><td>8.77</td><td>348</td><td>224.2</td><td>2613</td><td>83.0</td></tr><tr><td>PVT-Large [65]</td><td>61.4</td><td>9.85</td><td>306</td><td>203.4</td><td>3101</td><td>81.7</td></tr><tr><td>FasterNet-M</td><td>53.5</td><td>8.74</td><td>500</td><td>129.5</td><td>2092</td><td>83.0</td></tr><tr><td>PoolFormer-M48 [71]</td><td>73.5</td><td>11.59</td><td>242</td><td>281.8</td><td>OOM</td><td>82.5</td></tr><tr><td>ConvNeXt-B [36]</td><td>88.6</td><td>15.38</td><td>253</td><td>257.1</td><td>OOM</td><td>83.8</td></tr><tr><td>Swin-B [35]</td><td>87.8</td><td>15.47</td><td>237</td><td>349.2</td><td>OOM</td><td>83.5</td></tr><tr><td>FasterNet-L</td><td>93.5</td><td>15.52</td><td>323</td><td>219.5</td><td>OOM</td><td>83.5</td></tr></table>
218
+
219
+ Table 3. Comparison on ImageNet-1k benchmark. Models with similar top-1 accuracy are grouped together. For each group, our FasterNet achieves the highest throughput on GPU and the lowest latency on CPU and ARM. All models are evaluated at $224 \times 224$ resolution except for the MobileViT and EdgeNeXt with $256 \times 256$ . OOM is short for out of memory.
220
+
221
+ much simpler than many other models in terms of architectural design, which showcases the feasibility of designing simple yet powerful neural networks.
222
+
223
+ # 4.4. FasterNet on downstream tasks
224
+
225
+ To further evaluate the generalization ability of FasterNet, we conduct experiments on the challenging COCO dataset [32] for object detection and instance segmentation. As a common practice, we employ the ImageNet pre-trained FasterNet as a backbone and equip it with the popular Mask R-CNN detector [15]. To highlight the effectiveness of the backbone itself, we simply follow PoolFormer [71] and adopt an AdamW optimizer, a $1\times$ training
226
+
227
+ <table><tr><td>Backbone</td><td>Params (M)</td><td>FLOPs (G)</td><td>Latency on GPU (ms)</td><td>\(AP^b\)</td><td>\(AP_{50}^b\)</td><td>\(AP_{75}^b\)</td><td>\(AP^m\)</td><td>\(AP_{50}^m\)</td><td>\(AP_{75}^m\)</td></tr><tr><td>ResNet50 [16]</td><td>44.2</td><td>253</td><td>54.9</td><td>38.0</td><td>58.6</td><td>41.4</td><td>34.4</td><td>55.1</td><td>36.7</td></tr><tr><td>PoolFormer-S24 [71]</td><td>41.0</td><td>233</td><td>111.0</td><td>40.1</td><td>62.2</td><td>43.4</td><td>37.0</td><td>59.1</td><td>39.6</td></tr><tr><td>PVT-Small [65]</td><td>44.1</td><td>238</td><td>89.5</td><td>40.4</td><td>62.9</td><td>43.8</td><td>37.8</td><td>60.1</td><td>40.3</td></tr><tr><td>FasterNet-S</td><td>49.0</td><td>258</td><td>54.3</td><td>39.9</td><td>61.2</td><td>43.6</td><td>36.9</td><td>58.1</td><td>39.7</td></tr><tr><td>ResNet101 [16]</td><td>63.2</td><td>329</td><td>68.9</td><td>40.4</td><td>61.1</td><td>44.2</td><td>36.4</td><td>57.7</td><td>38.8</td></tr><tr><td>ResNeXt101-32×4d [69]</td><td>62.8</td><td>333</td><td>80.5</td><td>41.9</td><td>62.5</td><td>45.9</td><td>37.5</td><td>59.4</td><td>40.2</td></tr><tr><td>PoolFormer-S36 [71]</td><td>50.5</td><td>266</td><td>146.9</td><td>41.0</td><td>63.1</td><td>44.8</td><td>37.7</td><td>60.1</td><td>40.0</td></tr><tr><td>PVT-Medium [65]</td><td>63.9</td><td>295</td><td>117.3</td><td>42.0</td><td>64.4</td><td>45.6</td><td>39.0</td><td>61.6</td><td>42.1</td></tr><tr><td>FasterNet-M</td><td>71.2</td><td>344</td><td>71.4</td><td>43.0</td><td>64.4</td><td>47.4</td><td>39.1</td><td>61.5</td><td>42.3</td></tr><tr><td>ResNeXt101-64×4d [69]</td><td>101.9</td><td>487</td><td>112.9</td><td>42.8</td><td>63.8</td><td>47.3</td><td>38.4</td><td>60.6</td><td>41.3</td></tr><tr><td>PVT-Large [65]</td><td>81.0</td><td>358</td><td>152.2</td><td>42.9</td><td>65.0</td><td>46.6</td><td>39.5</td><td>61.9</td><td>42.5</td></tr><tr><td>FasterNet-L</td><td>110.9</td><td>484</td><td>93.8</td><td>44.0</td><td>65.6</td><td>48.2</td><td>39.9</td><td>62.3</td><td>43.0</td></tr></table>
228
+
229
+ Table 4. Results on COCO object detection and instance segmentation benchmarks. FLOPs are calculated with image size (1280, 800).
230
+
231
+ <table><tr><td>Ablation</td><td>Variant</td><td>Throughput on GPU (fps)</td><td>Latency on CPU (ms)</td><td>Latency on ARM (ms)</td><td>Acc. (%)</td></tr><tr><td rowspan="3">Partial ratio</td><td>T0* w/ r = 1/2</td><td>6626</td><td>9.6</td><td>145</td><td>71.7</td></tr><tr><td>T0 w/ r = 1/4</td><td>6807</td><td>9.2</td><td>143</td><td>71.9</td></tr><tr><td>T0* w/ r = 1/8</td><td>6204</td><td>8.9</td><td>140</td><td>71.3</td></tr><tr><td rowspan="2">Normalization</td><td>T0 w/ BN</td><td>6807</td><td>9.2</td><td>143</td><td>71.9</td></tr><tr><td>T0 w/ LN</td><td>5515</td><td>10.7</td><td>159</td><td>71.9</td></tr><tr><td rowspan="5">Activation</td><td>T0 w/ ReLU</td><td>6929</td><td>8.2</td><td>114</td><td>71.3</td></tr><tr><td>T0* w/ ReLU</td><td>5866</td><td>9.3</td><td>143</td><td>71.7</td></tr><tr><td>T0 w/ GELU</td><td>6807</td><td>9.2</td><td>143</td><td>71.9</td></tr><tr><td>T2 w/ ReLU</td><td>1991</td><td>33.5</td><td>497</td><td>78.9</td></tr><tr><td>T2 w/ GELU</td><td>1985</td><td>35.4</td><td>557</td><td>78.7</td></tr></table>
232
+
233
+ Table 5. Ablation on the partial ratio, normalization, and activation of FasterNet. Rows highlighted in grey are the default settings. $\mathrm{T0}^*$ denotes T0 variants with modified network width and depth.
234
+
235
+ schedule (12 epochs), a batch size of 16, and other training settings without further hyper-parameter tuning.
236
+
237
+ Tab. 4 shows the results for comparison between FasterNet and representative models. FasterNet consistently outperforms ResNet and ResNext by having higher average precision (AP) with similar latency. Specifically, FasterNet-S yields $+1.9$ higher box AP and $+2.4$ higher mask AP compared to the standard baseline ResNet50. FasterNet is also competitive against the ViT variants. Under similar FLOPs, FasterNet-L reduces PVT-Large's latency by $38\%$ , i.e., from $152.2\mathrm{ms}$ to $93.8\mathrm{ms}$ on GPU, and achieves $+1.1$ higher box AP and $+0.4$ higher mask AP.
238
+
239
+ # 4.5. Ablation study
240
+
241
+ We conduct a brief ablation study on the value of partial ratio $r$ and the choices of activation and normalization layers. We compare different variants in terms of ImageNet top-1 accuracy and on-device latency/throughput. Results are summarized in Tab. 5. For the partial ratio $r$ , we set it to $\frac{1}{4}$ for all FasterNet variants by default, which achieves
242
+
243
+ higher accuracy, higher throughput, and lower latency at similar complexity. A too large partial ratio $r$ would make PConv degrade to a regular Conv, while a too small value would render PConv less effective in capturing the spatial features. For the normalization layers, we choose BatchNorm over LayerNorm because BatchNorm can be merged into its adjacent convolutional layers for faster inference while it is as effective as LayerNorm in our experiment. For the activation function, interestingly, we empirically found that GELU fits FasterNet-T0/T1 models more efficiently than ReLU. It, however, becomes opposite for FasterNet-T2/S/M/L. Here we only show two examples in Tab. 5 due to space constraint. We conjecture that GELU strengthens FasterNet-T0/T1 by having higher non-linearity, while the benefit fades away for larger FasterNet variants.
244
+
245
+ # 5. Conclusion
246
+
247
+ In this paper, we have investigated the common and unresolved issue that many established neural networks suffer from low floating-point operations per second (FLOPS). We have revisited a bottleneck operator, DWConv, and analyzed its main cause for a slowdown - frequent memory access. To overcome the issue and achieve faster neural networks, we have proposed a simple yet fast and effective operator, PConv, that can be readily plugged into many existing networks. We have further introduced our general-purpose FasterNet, built upon our PConv, that achieves state-of-the-art speed and accuracy trade-off on various devices and vision tasks. We hope that our PConv and FasterNet would inspire more research on simple yet effective neural networks, going beyond academia to impact the industry and community directly.
248
+
249
+ Acknowledgement This work was supported, in part, by Hong Kong General Research Fund under grant number 16200120. The work of C.-H. Lee was supported, in part, by the NSF under Grant IIS-2209921.
250
+
251
+ # References
252
+
253
+ [1] Alaaeldin Ali, Hugo Touvron, Mathilde Caron, Piotr Bojanowski, Matthijs Douze, Armand Joulin, Ivan Laptev, Natalia Neverova, Gabriel Synnaeve, Jakob Verbeek, et al. Xcit: Cross-covariance image transformers. Advances in neural information processing systems, 34:20014-20027, 2021. 3
254
+ [2] Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hinton. Layer normalization. arXiv preprint arXiv:1607.06450, 2016. 4
255
+ [3] Han Cai, Chuang Gan, and Song Han. Efficientvit: Enhanced linear attention for high-resolution low-computation visual recognition. arXiv preprint arXiv:2205.14756, 2022. 3
256
+ [4] Jierun Chen, Tianlang He, Weipeng Zhuo, Li Ma, Sangtae Ha, and S-H Gary Chan. Tvconv: Efficient translation variant convolution for layout-aware visual processing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12548-12558, 2022. 3
257
+ [5] Shoufa Chen, Enze Xie, Chongjian Ge, Ding Liang, and Ping Luo. Cycle MLP: A mlp-like architecture for dense prediction. arXiv preprint arXiv:2107.10224, 2021. 2, 3, 7
258
+ [6] Yinpeng Chen, Xiyang Dai, Dongdong Chen, Mengchen Liu, Xiaoyi Dong, Lu Yuan, and Zicheng Liu. Mobileformer: Bridging mobilenet and transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5270-5279, 2022. 1, 3
259
+ [7] Yunpeng Chen, Haoqi Fan, Bing Xu, Zhicheng Yan, Yannis Kalantidis, Marcus Rohrbach, Shuicheng Yan, and Ji-ashi Feng. Drop an octave: Reducing spatial redundancy in convolutional neural networks with octave convolution. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3435–3444, 2019. 3
260
+ [8] François Chollet. Xception: Deep learning with depthwise separable convolutions. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1251-1258, 2017. 3
261
+ [9] Ekin D Cubuk, Barret Zoph, Jonathon Shlens, and Quoc V Le. Randaugment: Practical automated data augmentation with a reduced search space. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 702-703, 2020. 6
262
+ [10] Zihang Dai, Hanxiao Liu, Quoc V Le, and Mingxing Tan. Coatnet: Marrying convolution and attention for all data sizes. Advances in Neural Information Processing Systems, 34:3965-3977, 2021. 3
263
+ [11] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 1, 3
264
+ [12] Benjamin Graham, Alaaeldin El-Nouby, Hugo Touvron, Pierre Stock, Armand Joulin, Hervé Jégou, and Matthijs Douze. Levit: a vision transformer in convnet's clothing for faster inference. In Proceedings of the IEEE/CVF international conference on computer vision, pages 12259-12269, 2021. 3
265
+
266
+ [13] Kai Han, Yunhe Wang, Qi Tian, Jianyuan Guo, Chunjing Xu, and Chang Xu. Ghostnet: More features from cheap operations. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1580-1589, 2020. 1, 3, 4, 7
267
+ [14] Kai Han, Yunhe Wang, Qiulin Zhang, Wei Zhang, Chunjing Xu, and Tong Zhang. Model rubik's cube: Twisting resolution, depth and width for tinynets. Advances in Neural Information Processing Systems, 33:19353-19364, 2020. 3
268
+ [15] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961-2969, 2017. 7
269
+ [16] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 2, 4, 7, 8
270
+ [17] Tianlang He, Jiajie Tan, Weipeng Zhuo, Maximilian Printz, and S-H Gary Chan. Tackling multipath and biased training data for imu-assisted ble proximity detection. In IEEE INFOCOM 2022-IEEE Conference on Computer Communications, pages 1259-1268. IEEE, 2022. 3
271
+ [18] Dan Hendrycks and Kevin Gimpel. Gaussian error linear units (gelus). arXiv preprint arXiv:1606.08415, 2016. 5
272
+ [19] Geoffrey Hinton, Oriol Vinyals, Jeff Dean, et al. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2(7), 2015. 6
273
+ [20] Andrew Howard, Mark Sandler, Grace Chu, Liang-Chieh Chen, Bo Chen, Mingxing Tan, Weijun Wang, Yukun Zhu, Ruoming Pang, Vijay Vasudevan, et al. Searching for mobilenetv3. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1314-1324, 2019. 1, 3
274
+ [21] Andrew G Howard, Menglong Zhu, Bo Chen, Dmitry Kalenichenko, Weijun Wang, Tobias Weyand, Marco Andretto, and Hartwig Adam. Mobilenets: Efficient convolutional neural networks for mobile vision applications. arXiv preprint arXiv:1704.04861, 2017. 1, 3
275
+ [22] Han Hu, Zheng Zhang, Zhenda Xie, and Stephen Lin. Local relation networks for image recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3464-3473, 2019. 3
276
+ [23] Gao Huang, Shichen Liu, Laurens Van der Maaten, and Kilian Q Weinberger. Condensenet: An efficient densenet using learned group convolutions. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2752-2761, 2018. 3
277
+ [24] Gao Huang, Yu Sun, Zhuang Liu, Daniel Sedra, and Kilian Q Weinberger. Deep networks with stochastic depth. In European conference on computer vision, pages 646-661. Springer, 2016. 6
278
+ [25] Tao Huang, Lang Huang, Shan You, Fei Wang, Chen Qian, and Chang Xu. Lightvit: Towards light-weight convolution-free vision transformers. arXiv preprint arXiv:2207.05557, 2022.3
279
+ [26] Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. In International conference on machine learning, pages 448-456. PMLR, 2015. 4
280
+
281
+ [27] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. Advances in neural information processing systems, 25, 2012. 1, 3
282
+ [28] Anders Krogh and John Hertz. A simple weight decay can improve generalization. Advances in neural information processing systems, 4, 1991. 6
283
+ [29] Yunsheng Li, Yinpeng Chen, Xiyang Dai, Dongdong Chen, Mengchen Liu, Lu Yuan, Zicheng Liu, Lei Zhang, and Nuno Vasconcelos. Micronet: Improving image recognition with extremely low flops. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 468-477, 2021. 1, 3
284
+ [30] Yanyu Li, Geng Yuan, Yang Wen, Eric Hu, Georgios Evangelidis, Sergey Tulyakov, Yanzhi Wang, and Jian Ren. Efficientformer: Vision transformers at mobilenet speed. arXiv preprint arXiv:2206.01191, 2022. 3
285
+ [31] Dongze Lian, Zehao Yu, Xing Sun, and Shenghua Gao. Asmlp: An axial shifted mlp architecture for vision. arXiv preprint arXiv:2107.08391, 2021. 3
286
+ [32] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014. 7
287
+ [33] Ruiyang Liu, Yinghui Li, Linmi Tao, Dun Liang, and Haitao Zheng. Are we ready for a new paradigm shift? a survey on visual deep mlp. Patterns, 3(7):100520, 2022. 3
288
+ [34] Ze Liu, Han Hu, Yutong Lin, Zhuliang Yao, Zhenda Xie, Yixuan Wei, Jia Ning, Yue Cao, Zheng Zhang, Li Dong, et al. Swin transformer v2: Scaling up capacity and resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12009-12019, 2022. 3
289
+ [35] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10012-10022, 2021. 2, 3, 7
290
+ [36] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11976-11986, 2022. 1, 3, 7
291
+ [37] Ilya Loshchilov and Frank Hutter. Sgdr: Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983, 2016. 6
292
+ [38] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6
293
+ [39] Jiachen Lu, Jinghan Yao, Junge Zhang, Xiatian Zhu, Hang Xu, Weiguo Gao, Chunjing Xu, Tao Xiang, and Li Zhang. Soft: softmax-free transformer with linear complexity. Advances in Neural Information Processing Systems, 34:21297-21309, 2021. 3
294
+ [40] Ningning Ma, Xiangyu Zhang, Hai-Tao Zheng, and Jian Sun. Shufflenet v2: Practical guidelines for efficient cnn architecture design. In Proceedings of the European conference on computer vision (ECCV), pages 116-131, 2018. 1, 2, 3, 7
295
+
296
+ [41] Muhammad Maaz, Abdelrahman Shaker, Hisham Cholakkal, Salman Khan, Syed Waqas Zamir, Rao Muhammad Anwer, and Fahad Shahbaz Khan. Edgenext: efficiently amalgamated cnn-transformer architecture for mobile vision applications. arXiv preprint arXiv:2206.10589, 2022. 7
297
+ [42] Sachin Mehta and Mohammad Rastegari. Mobilevit: lightweight, general-purpose, and mobile-friendly vision transformer. arXiv preprint arXiv:2110.02178, 2021. 1, 2, 3, 7
298
+ [43] Sachin Mehta and Mohammad Rastegari. Separable self-attention for mobile vision transformers. arXiv preprint arXiv:2206.02680, 2022. 1, 3
299
+ [44] Vinod Nair and Geoffrey E Hinton. Rectified linear units improve restricted boltzmann machines. In Icml, 2010. 5
300
+ [45] Junting Pan, Adrian Bulat, Fuwen Tan, Xiatian Zhu, Lukasz Dudziak, Hongsheng Li, Georgios Tzimiropoulos, and Brais Martinez. Edgevits: Competing light-weight cnns on mobile devices with vision transformers. arXiv preprint arXiv:2205.03436, pages 1-6, 2022. 3
301
+ [46] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. International journal of computer vision, 115(3):211-252, 2015. 6
302
+ [47] Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zh-moginov, and Liang-Chieh Chen. Mobilenetv2: Inverted residuals and linear bottlenecks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4510-4520, 2018. 1, 3, 4, 7
303
+ [48] Laurent Sifre and Stéphane Mallat. Rigid-motion scattering for texture classification. arXiv preprint arXiv:1403.1687, 2014. 1, 3
304
+ [49] Pravendra Singh, Vinay Kumar Verma, Piyush Rai, and Vinay P Namboodiri. Hetconv: Heterogeneous kernel-based convolutions for deep cnns. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4835-4844, 2019. 3
305
+ [50] Aravind Srinivas, Tsung-Yi Lin, Niki Parmar, Jonathon Shlens, Pieter Abbeel, and Ashish Vaswani. Bottleneck transformers for visual recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16519-16529, 2021. 3
306
+ [51] Andreas Steiner, Alexander Kolesnikov, Xiaohua Zhai, Ross Wightman, Jakob Uszkoreit, and Lucas Beyer. How to train your vit? data, augmentation, and regularization in vision transformers. arXiv preprint arXiv:2106.10270, 2021. 3
307
+ [52] Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jon Shlens, and Zbigniew Wojna. Rethinking the inception architecture for computer vision. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2818-2826, 2016. 6
308
+ [53] Mingxing Tan, Bo Chen, Ruoming Pang, Vijay Vasudevan, Mark Sandler, Andrew Howard, and Quoc V Le. Mnasnet: Platform-aware neural architecture search for mobile. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2820-2828, 2019. 3
309
+ [54] Mingxing Tan and Quoc Le. Efficientnet: Rethinking model scaling for convolutional neural networks. In International
310
+
311
+ conference on machine learning, pages 6105-6114. PMLR, 2019.3
312
+ [55] Mingxing Tan and Quoc Le. Efficientnetv2: Smaller models and faster training. In International Conference on Machine Learning, pages 10096-10106. PMLR, 2021. 3
313
+ [56] Shitao Tang, Jiahui Zhang, Siyu Zhu, and Ping Tan. Quadtree attention for vision transformers. arXiv preprint arXiv:2201.02767, 2022. 3
314
+ [57] Ilya O Tolstikhin, Neil Houlsby, Alexander Kolesnikov, Lucas Beyer, Xiaohua Zhai, Thomas Unterthiner, Jessica Yung, Andreas Steiner, Daniel Keysers, Jakob Uszkoreit, et al. Mlp-mixer: An all-mlp architecture for vision. Advances in Neural Information Processing Systems, 34:24261-24272, 2021. 1, 3
315
+ [58] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In International Conference on Machine Learning, pages 10347-10357. PMLR, 2021. 3
316
+ [59] Hugo Touvron, Matthieu Cord, and Hervé Jégou. Deit iii: Revenge of the vit. arXiv preprint arXiv:2204.07118, 2022. 3
317
+ [60] Dmitry Ulyanov, Andrea Vedaldi, and Victor Lempitsky. Instance normalization: The missing ingredient for fast stylization. arXiv preprint arXiv:1607.08022, 2016. 4
318
+ [61] Ashish Vaswani, Prajit Ramachandran, Aravind Srinivas, Niki Parmar, Blake Hechtman, and Jonathon Shlens. Scaling local self-attention for parameter efficient visual backbones. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12894–12904, 2021. 3
319
+ [62] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 3
320
+ [63] Shakti N Wadekar and Abhishek Chaurasia. Mobilevitv3: Mobile-friendly vision transformer with simple and effective fusion of local, global and input features. arXiv preprint arXiv:2209.15159, 2022. 1
321
+ [64] Guangting Wang, Yucheng Zhao, Chuanxin Tang, Chong Luo, and Wenjun Zeng. When shift operation meets vision transformer: An extremely simple alternative to attention mechanism. arXiv preprint arXiv:2201.10801, 2022. 3
322
+ [65] Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, and Ling Shao. Pyramid vision transformer: A versatile backbone for dense prediction without convolutions. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 568-578, 2021. 3, 7, 8
323
+ [66] Song Wen, Hao Wang, and Dimitris Metaxas. Social ode: Multi-agent trajectory forecasting with neural ordinary differential equations. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXII, pages 217-233. Springer, 2022. 3
324
+ [67] Bichen Wu, Xiaoliang Dai, Peizhao Zhang, Yanghan Wang, Fei Sun, Yiming Wu, Yuandong Tian, Peter Vajda, Yangqing Jia, and Kurt Keutzer. Fbnet: Hardware-aware efficient convnet design via differentiable neural architecture search. In
325
+
326
+ Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10734-10742, 2019. 3
327
+ [68] Yuxin Wu and Kaiming He. Group normalization. In Proceedings of the European conference on computer vision (ECCV), pages 3-19, 2018. 4
328
+ [69] Saining Xie, Ross Girshick, Piotr Dólár, Zhuowen Tu, and Kaiming He. Aggregated residual transformations for deep neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1492-1500, 2017. 8
329
+ [70] Le Yang, Haojun Jiang, Ruojin Cai, Yulin Wang, Shiji Song, Gao Huang, and Qi Tian. Condensenet v2: Sparse feature reactivation for deep networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3569-3578, 2021. 3
330
+ [71] Weihao Yu, Mi Luo, Pan Zhou, Chenyang Si, Yichen Zhou, Xinchao Wang, Jiashi Feng, and Shuicheng Yan. Metaformer is actually what you need for vision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10819-10829, 2022. 7, 8
331
+ [72] Sangdoo Yun, Dongyoon Han, Seong Joon Oh, Sanghyuk Chun, Junsuk Choe, and Youngjoon Yoo. Cutmix: Regularization strategy to train strong classifiers with localizable features. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6023-6032, 2019. 6
332
+ [73] Hongyi Zhang, Moustapha Cisse, Yann N Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization. arXiv preprint arXiv:1710.09412, 2017. 6
333
+ [74] Qiulin Zhang, Zhuqing Jiang, Qishuo Lu, Jia'nan Han, Zhengxin Zeng, Shang-Hua Gao, and Aidong Men. Split to be slim: An overlooked redundancy in vanilla convolution. arXiv preprint arXiv:2006.12085, 2020. 4
334
+ [75] Ting Zhang, Guo-Jun Qi, Bin Xiao, and Jingdong Wang. Interleaved group convolutions. In Proceedings of the IEEE international conference on computer vision, pages 4373-4382, 2017. 3
335
+ [76] Xiangyu Zhang, Xinyu Zhou, Mengxiao Lin, and Jian Sun. Shufflenet: An extremely efficient convolutional neural network for mobile devices. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6848-6856, 2018. 1, 3
336
+ [77] Shuhan Zhong, Sizhe Song, Guanyao Li, and S-H Gary Chan. A tree-based structure-aware transformer decoder for image-to-markup generation. In Proceedings of the 30th ACM International Conference on Multimedia, pages 5751-5760, 2022. 3
337
+ [78] Weipeng Zhuo, Ka Ho Chiu, Jierun Chen, Jiajie Tan, Edmund Sumpena, S-H Gary Chan, Sangtae Ha, and Chul-Ho Lee. Semi-supervised learning with network embedding on ambient rf signals for geofencing services. arXiv preprint arXiv:2210.07889, 2022. 3
338
+ [79] Barret Zoph and Quoc V Le. Neural architecture search with reinforcement learning. arXiv preprint arXiv:1611.01578, 2016. 6
2023/Run, Don't Walk_ Chasing Higher FLOPS for Faster Neural Networks/images.zip ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:31fd3710058a41879ca616962308fd5ed2d81475a53ed1f00d5f026e237e985a
3
+ size 767289
2023/Run, Don't Walk_ Chasing Higher FLOPS for Faster Neural Networks/layout.json ADDED
The diff for this file is too large to render. See raw diff
 
2023/S3C_ Semi-Supervised VQA Natural Language Explanation via Self-Critical Learning/6ead34e8-75d6-4d08-a4ab-1da1402619c2_content_list.json ADDED
@@ -0,0 +1,1506 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ [
2
+ {
3
+ "type": "text",
4
+ "text": "$\\mathrm{S}^{3}\\mathrm{C}$ : Semi-Supervised VQA Natural Language Explanation via Self-Critical Learning",
5
+ "text_level": 1,
6
+ "bbox": [
7
+ 163,
8
+ 128,
9
+ 805,
10
+ 175
11
+ ],
12
+ "page_idx": 0
13
+ },
14
+ {
15
+ "type": "text",
16
+ "text": "Wei Suo $^{1*}$ , Mengyang Sun $^{2*}$ , Weisong Liu $^{1}$ , Yiqi Gao $^{1}$ , Peng Wang $^{1\\dagger}$ , Yanning Zhang $^{1\\dagger}$ , Qi Wu $^{3}$ $^{1}$ School of Computer Science and Ningbo Institute, Northwestern Polytechnical University, China. \n $^{2}$ School of Cybersecurity, Northwestern Polytechnical University, China. \n $^{3}$ University of Adelaide, Australia.",
17
+ "bbox": [
18
+ 96,
19
+ 202,
20
+ 872,
21
+ 273
22
+ ],
23
+ "page_idx": 0
24
+ },
25
+ {
26
+ "type": "text",
27
+ "text": "{suowei1994,sunmenmian,liuweisong,gyqjz}@mail.nwpu.edu.cn",
28
+ "bbox": [
29
+ 228,
30
+ 276,
31
+ 736,
32
+ 292
33
+ ],
34
+ "page_idx": 0
35
+ },
36
+ {
37
+ "type": "text",
38
+ "text": "{peng.wang, ynzhang}@nwpu.edu.cn, qi.wu01@adelaide.edu.au",
39
+ "bbox": [
40
+ 228,
41
+ 295,
42
+ 723,
43
+ 309
44
+ ],
45
+ "page_idx": 0
46
+ },
47
+ {
48
+ "type": "text",
49
+ "text": "Abstract",
50
+ "text_level": 1,
51
+ "bbox": [
52
+ 233,
53
+ 344,
54
+ 313,
55
+ 359
56
+ ],
57
+ "page_idx": 0
58
+ },
59
+ {
60
+ "type": "text",
61
+ "text": "VQA Natural Language Explanation (VQA-NLE) task aims to explain the decision-making process of VQA models in natural language. Unlike traditional attention or gradient analysis, free-text rationales can be easier to understand and gain users' trust. Existing methods mostly use post-hoc or self-rationalization models to obtain a plausible explanation. However, these frameworks are bottlenecked by the following challenges: 1) the reasoning process cannot be faithfully responded to and suffer from the problem of logical inconsistency. 2) Human-annotated explanations are expensive and time-consuming to collect. In this paper, we propose a new Semi-Supervised VQA-NLE via Self-Critical Learning $(S^3 C)$ , which evaluates the candidate explanations by answering rewards to improve the logical consistency between answers and rationales. With a semi-supervised learning framework, the $S^3 C$ can benefit from a tremendous amount of samples without human-annotated explanations. A large number of automatic measures and human evaluations all show the effectiveness of our method. Meanwhile, the framework achieves a new state-of-the-art performance on the two VQA-NLE datasets.",
62
+ "bbox": [
63
+ 75,
64
+ 375,
65
+ 473,
66
+ 693
67
+ ],
68
+ "page_idx": 0
69
+ },
70
+ {
71
+ "type": "text",
72
+ "text": "1. Introduction",
73
+ "text_level": 1,
74
+ "bbox": [
75
+ 76,
76
+ 734,
77
+ 207,
78
+ 750
79
+ ],
80
+ "page_idx": 0
81
+ },
82
+ {
83
+ "type": "text",
84
+ "text": "Deep neural networks have enabled significant breakthroughs in a variety of vision-language (VL) tasks such as image captioning [10, 47] and visual question answering (VQA) [2, 39]. Unfortunately, most of them are black box systems, which makes it challenging to gain users' trust [20]. Explaining the decision-making process of deep VL models is a long-standing and essential problem.",
85
+ "bbox": [
86
+ 75,
87
+ 761,
88
+ 468,
89
+ 867
90
+ ],
91
+ "page_idx": 0
92
+ },
93
+ {
94
+ "type": "image",
95
+ "img_path": "images/af310065e8d5f1bdc8df7d8c31b4609c45d4ba16f61dad695af4654d8d7b5216.jpg",
96
+ "image_caption": [],
97
+ "image_footnote": [],
98
+ "bbox": [
99
+ 506,
100
+ 345,
101
+ 897,
102
+ 421
103
+ ],
104
+ "page_idx": 0
105
+ },
106
+ {
107
+ "type": "image",
108
+ "img_path": "images/c5759907e61c6aea5ef5031fae6ab82cc0a7fb498d64b875941375cdd035402c.jpg",
109
+ "image_caption": [],
110
+ "image_footnote": [],
111
+ "bbox": [
112
+ 504,
113
+ 422,
114
+ 897,
115
+ 494
116
+ ],
117
+ "page_idx": 0
118
+ },
119
+ {
120
+ "type": "image",
121
+ "img_path": "images/c16a8892e5fba59ebcfaa7daacb1445d0ba698e7ddee1afd6e25253591b95f64.jpg",
122
+ "image_caption": [
123
+ "Figure 1. Paradigm comparison of different VQA-NLE methods. (a) Post-hoc explanation method adopts two independent models to predict answers and explanations respectively. (b) Self-rationalization method uses a united VL model to simultaneously generate answers and explanations. (c) Our self-critical strategy utilizes answer scores as rewards and obtains more reliable rationales with semi-supervised learning."
124
+ ],
125
+ "image_footnote": [],
126
+ "bbox": [
127
+ 504,
128
+ 498,
129
+ 897,
130
+ 580
131
+ ],
132
+ "page_idx": 0
133
+ },
134
+ {
135
+ "type": "text",
136
+ "text": "Some approaches depend on attention mechanisms [2, 30] or gradient-based localization [50] to acquire visual explanations, which can highlight some contributing image regions for the predicted answers. However, simple visualization cannot explain how these areas support the answers and they are also hard to comprehend [20, 48]. Conversely, Natural Language Explanation (NLE) task [6, 38] can explain the decision-making process of a model by generating a natural language sentence. The language-based explanations are more accessible for users to understand, and they can also help researchers optimize the structure of models [34].",
137
+ "bbox": [
138
+ 496,
139
+ 718,
140
+ 890,
141
+ 883
142
+ ],
143
+ "page_idx": 0
144
+ },
145
+ {
146
+ "type": "text",
147
+ "text": "Recently, some models of NLE in the VL commu",
148
+ "bbox": [
149
+ 517,
150
+ 885,
151
+ 890,
152
+ 900
153
+ ],
154
+ "page_idx": 0
155
+ },
156
+ {
157
+ "type": "header",
158
+ "text": "CVF",
159
+ "bbox": [
160
+ 106,
161
+ 2,
162
+ 181,
163
+ 42
164
+ ],
165
+ "page_idx": 0
166
+ },
167
+ {
168
+ "type": "header",
169
+ "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.",
170
+ "bbox": [
171
+ 236,
172
+ 0,
173
+ 807,
174
+ 46
175
+ ],
176
+ "page_idx": 0
177
+ },
178
+ {
179
+ "type": "page_footnote",
180
+ "text": "*These authors contributed equally to this work.",
181
+ "bbox": [
182
+ 94,
183
+ 875,
184
+ 351,
185
+ 887
186
+ ],
187
+ "page_idx": 0
188
+ },
189
+ {
190
+ "type": "page_footnote",
191
+ "text": "† Corresponding authors.",
192
+ "bbox": [
193
+ 96,
194
+ 887,
195
+ 225,
196
+ 898
197
+ ],
198
+ "page_idx": 0
199
+ },
200
+ {
201
+ "type": "page_number",
202
+ "text": "2646",
203
+ "bbox": [
204
+ 482,
205
+ 944,
206
+ 514,
207
+ 955
208
+ ],
209
+ "page_idx": 0
210
+ },
211
+ {
212
+ "type": "text",
213
+ "text": "nity have achieved pretty-well results, especially for VQANLE [20,34,41,48,58]. They can guide models to generate natural language sentences and interpret how the models get answers. Specifically, the first research line usually treats VQA-NLE as a predict-then-explain task [20, 34, 41, 58], namely post-hoc explanations method. As shown in Fig. 1 (a), these methods first depend on pre-trained VL models (such as UNITER [8] or Oscar [25]) to gain answers. Then the fused multi-modal features and the predicted answers are fed into a separated language model (e.g., LSTM [16] or Transformer [54]) to generate corresponding explanations. As shown in Fig. 1 (b), the other line [48] relies on a united VL model while generating both answers and explanations, which is known as the self-rationalization method. This framework can simultaneously predict an answer and generate a rationale by formulating the answer as a text-generation task along with the explanation.",
214
+ "bbox": [
215
+ 75,
216
+ 90,
217
+ 472,
218
+ 347
219
+ ],
220
+ "page_idx": 1
221
+ },
222
+ {
223
+ "type": "text",
224
+ "text": "Though significant progress has been made, the two paradigms are still restricted by the following challenges: 1) For the first paradigm, since the decision-making model and interpretation part are two separate modules, it would inevitably lead to unfaithful responses to the reasoning process of the decision models. 2) Due to the lack of explicitly logical relationship modeling, previous work [19] has proved that the straightforward self-rationalization frameworks suffer from the problem of logical inconsistency. 3) The above strategies all require an amount of human-annotated explanations, which are expensive and time-consuming to collect [62].",
225
+ "bbox": [
226
+ 75,
227
+ 348,
228
+ 472,
229
+ 527
230
+ ],
231
+ "page_idx": 1
232
+ },
233
+ {
234
+ "type": "text",
235
+ "text": "To solve the above challenges, inspired by [5, 51], we argue that a reasonable rationale can assist the model in obtaining a correct answer, and vice versa, the answer can be converted as an evaluation criterion for possible explanations. In this paper, we propose a new Semi-Supervised VQA-NLE method with Self-Critical learning, which is called $S^3 C$ for short. As shown in Fig.1 (c), given images and related questions, we first leverage a prompting mechanism to construct answer and explanation templates, which can guide the pre-trained VL model to generate answers and multiple candidate explanations based on sequence sampling [2]. Then we design a new self-critical method that converts the answer scores as rewards and encourages the model to generate the explanations which contribute to improving the answer scores. In particular, to reduce the dependency on expensive human annotations, we further extend our method to the semi-supervised version, which utilizes the unlabelled samples<sup>1</sup> (i.e., conventional VQA data [4, 36]) to significantly enhance the self-interpretability of the model. With the self-critical strategy and the semi-supervised learning, our method effectively models the logical relationships and promotes the logical",
236
+ "bbox": [
237
+ 75,
238
+ 530,
239
+ 472,
240
+ 864
241
+ ],
242
+ "page_idx": 1
243
+ },
244
+ {
245
+ "type": "text",
246
+ "text": "consistency between answer-explanation pairs. According to automatic measures and human evaluations, the $S^3 C$ outperforms the state-of-the-art models for the VQA-NLE task on the widely used two datasets and provides a new paradigm for our community. In summary, we make the following contributions:",
247
+ "bbox": [
248
+ 496,
249
+ 90,
250
+ 893,
251
+ 181
252
+ ],
253
+ "page_idx": 1
254
+ },
255
+ {
256
+ "type": "list",
257
+ "sub_type": "text",
258
+ "list_items": [
259
+ "1) We propose a new self-critical VQA-NLE method that can model the logical relationships between answer-explanation pairs and evaluate the generated rationales by answering rewards. This strategy effectively improves the logical consistency and the reliability of the interpretations.",
260
+ "2) We develop an advanced semi-supervised learning framework for VQA-NLE, which utilizes amounts of samples without human-annotated explanations to boost the self-interpretability of the model further. To the best of our knowledge, we are the first to explore semi-supervised learning on the VQA Natural Language Explanation.",
261
+ "3) The proposed $S^3 C$ achieves new state-of-the-art performance on VQA-X [13] and A-OKVQA [49] benchmark datasets. Meanwhile, automatic measures and human evaluations all show the effectiveness of our method."
262
+ ],
263
+ "bbox": [
264
+ 496,
265
+ 184,
266
+ 893,
267
+ 415
268
+ ],
269
+ "page_idx": 1
270
+ },
271
+ {
272
+ "type": "text",
273
+ "text": "2. Related work",
274
+ "text_level": 1,
275
+ "bbox": [
276
+ 500,
277
+ 438,
278
+ 638,
279
+ 454
280
+ ],
281
+ "page_idx": 1
282
+ },
283
+ {
284
+ "type": "text",
285
+ "text": "2.1. Explainability in Visual Question Answering",
286
+ "text_level": 1,
287
+ "bbox": [
288
+ 500,
289
+ 465,
290
+ 879,
291
+ 483
292
+ ],
293
+ "page_idx": 1
294
+ },
295
+ {
296
+ "type": "text",
297
+ "text": "The visual question answering (VQA) is firstly proposed by [33] that requires an intelligent agent to generate an answer by giving an image and a question. Many approaches have been introduced such as joint embedding [13, 61], attention mechanisms [3, 31], memory networks [32, 59] and graph neural networks [23, 56]. Although the VQA task has been well studied, the reasoning process of the models is always agnostic. Some methods apply visualization technologies to achieve visual explanation, such as Grad-CAM [50] and U-CAM [42]. However, because image visualization cannot support the answer based on the attended areas [57], in this paper, we focus on improving free-text explanations that are more convenient and easier for users to understand. In this topic, the early work is proposed by [41]. The paper conducts the VQA-X dataset and utilizes human annotations to inspire the decision-making process of VQA models. [20] designs a new model that combines a pre-trained language model and a VL model to generate free-text explanations. [60] combines stronger pre-trained VL model (i.e. Oscar [25]) and generation model (i.e. GPT-2 [46]) to obtain better results. Recently, [48] proposes a unified model which can simultaneously predict answers and explanations based on a pre-trained caption model. Unlike previous methods, we introduce a new self-critical strategy to model the logical relationships between answers and rationales. It can encourage the model to enhance logical consistency and generate more reasonable explanations.",
298
+ "bbox": [
299
+ 496,
300
+ 493,
301
+ 893,
302
+ 902
303
+ ],
304
+ "page_idx": 1
305
+ },
306
+ {
307
+ "type": "page_footnote",
308
+ "text": "<sup>1</sup>In this paper, we use \"unlabelled samples\" and \"labelled samples\" to indicate the question-answer (QA) pairs without/with human explanations.",
309
+ "bbox": [
310
+ 76,
311
+ 875,
312
+ 470,
313
+ 902
314
+ ],
315
+ "page_idx": 1
316
+ },
317
+ {
318
+ "type": "page_number",
319
+ "text": "2647",
320
+ "bbox": [
321
+ 482,
322
+ 944,
323
+ 516,
324
+ 955
325
+ ],
326
+ "page_idx": 1
327
+ },
328
+ {
329
+ "type": "text",
330
+ "text": "2.2. Pre-trained models and Prompt learning",
331
+ "text_level": 1,
332
+ "bbox": [
333
+ 76,
334
+ 90,
335
+ 426,
336
+ 107
337
+ ],
338
+ "page_idx": 2
339
+ },
340
+ {
341
+ "type": "text",
342
+ "text": "Pre-trained models have been applied in many fields, such as various NLP tasks [12,46] and VL tasks [29,52,65]. Most of these pre-trained models utilize a stack of Transformer structures as the backbone. To generalize the pretrained models to other downstream tasks, previous works mostly fine-tune whole models on each downstream VL task. However, the ability of pre-trained models would be limited due to the mismatch between pre-trained tasks and downstream tasks. Hence, prompt learning [9, 18, 45, 64] is proposed, and it can keep the optimization consistency. [45] designs the templates to transfer the knowledge to downstream tasks. [9] uses a textual generation framework for uniform optimization. In this paper, we use the pre-trained model based on the image caption task as our backbone. It has been proved that these additional pre-train tasks (e.g., image feature regression or mask language tokens) cannot significantly improve the performance for explanations [48]. Meanwhile, we apply different language prompt templates to motivate the model to generate corresponding answers and rationales.",
343
+ "bbox": [
344
+ 75,
345
+ 113,
346
+ 472,
347
+ 417
348
+ ],
349
+ "page_idx": 2
350
+ },
351
+ {
352
+ "type": "text",
353
+ "text": "2.3. Semi-supervised learning",
354
+ "text_level": 1,
355
+ "bbox": [
356
+ 76,
357
+ 426,
358
+ 308,
359
+ 444
360
+ ],
361
+ "page_idx": 2
362
+ },
363
+ {
364
+ "type": "text",
365
+ "text": "The development of deep learning depends on a large number of labelled data, while there are many cases in that only a small amount of data can be obtained [44]. To solve this challenge, Semi-supervised learning is proposed [44], which aims to train models using a small number of labelled data and amounts of unlabelled data. For example, [22] proposes a generatively semi-supervised framework based on variational autoencoders, and it can jointly optimize the model and variational parameters. Recently, [63] simultaneously leverages self-supervised and semi-supervised learning to address image classification tasks. To our knowledge, our work is the first semi-supervised learning framework for the VQA-NLE task, which effectively alleviates the reliance on expensive human annotations and further boost the self-interpretability of the model.",
366
+ "bbox": [
367
+ 75,
368
+ 450,
369
+ 472,
370
+ 680
371
+ ],
372
+ "page_idx": 2
373
+ },
374
+ {
375
+ "type": "text",
376
+ "text": "3. Method",
377
+ "text_level": 1,
378
+ "bbox": [
379
+ 76,
380
+ 693,
381
+ 169,
382
+ 709
383
+ ],
384
+ "page_idx": 2
385
+ },
386
+ {
387
+ "type": "text",
388
+ "text": "In this section, we introduce our Semi-Supervised VQANLE method with the Self-Critical learning $(S^3 C)$ framework. Our aim is to strengthen the logical consistency between answer-explanation pairs and improve the reliability of the rationales. As shown in Fig. 2, the $S^3 C$ comprises an \"Answer-Explanation Prompt\" module and a \"Self-Critical Reinforcement\" module. Unlike previous approaches, our method first uses a prompting mechanism to generate answers and candidate explanations. Then, we design a new self-critical module that converts the answer scores as rewards to evaluate these reasons. Furthermore, this strategy can conveniently apply the unlabelled QA pairs to enrich",
389
+ "bbox": [
390
+ 75,
391
+ 719,
392
+ 472,
393
+ 902
394
+ ],
395
+ "page_idx": 2
396
+ },
397
+ {
398
+ "type": "text",
399
+ "text": "training data and enhance the self-interpretability. Next, we describe the components of our model in detail.",
400
+ "bbox": [
401
+ 496,
402
+ 90,
403
+ 890,
404
+ 121
405
+ ],
406
+ "page_idx": 2
407
+ },
408
+ {
409
+ "type": "text",
410
+ "text": "3.1. Pre-trained Vision-Language Backbone",
411
+ "text_level": 1,
412
+ "bbox": [
413
+ 498,
414
+ 132,
415
+ 841,
416
+ 148
417
+ ],
418
+ "page_idx": 2
419
+ },
420
+ {
421
+ "type": "text",
422
+ "text": "Given an image $I \\in \\mathbb{R}^{W \\times H \\times 3}$ and a natural language question $Q = \\{q_t\\}_{t=1}^T$ , where $q_t$ represents the $t$ -th word, $T$ is the length of the question and $W \\times H \\times 3$ denotes the size of the image. Our goal is to predict the answer and generate a corresponding free-text rationale. Following previous works [48, 49], we adopt the CLIP vision encoder [45] and a pre-trained image caption model (i.e., ClipCap [37]) as our basic backbone. During pre-training, the VL model uses image embeddings from the CLIP as prefixes and fine-tunes a language model (i.e., GPT-2 [46]) to generate image captions. We refer readers to [37] for more information about the pre-trained model. In this paper, we consider the image $I$ and question $Q$ as the prefixes of the answer-explanation sequences. Specifically, we first apply ViT-B [14] and \"classification token\" from the CLIP to obtain the image features. Then, a group of light and simple Multi-Layer Perceptron (MLP) is utilized for transforming the image features to the $V = \\{v_s\\}_{s=1}^S$ , $v_s \\in \\mathbb{R}^c$ , where the dimension size $c = 768$ and the image sequence length $S = 10$ . The above computations can be formulated as follows:",
423
+ "bbox": [
424
+ 496,
425
+ 154,
426
+ 890,
427
+ 470
428
+ ],
429
+ "page_idx": 2
430
+ },
431
+ {
432
+ "type": "equation",
433
+ "text": "\n$$\nv _ {1}, v _ {2}, \\dots , v _ {S} = \\operatorname {M L P} (\\operatorname {C L I P} (I)). \\tag {1}\n$$\n",
434
+ "text_format": "latex",
435
+ "bbox": [
436
+ 576,
437
+ 473,
438
+ 890,
439
+ 491
440
+ ],
441
+ "page_idx": 2
442
+ },
443
+ {
444
+ "type": "text",
445
+ "text": "Note that we only update the mapping network (i.e., MLP) during training, while the original visual encoder parameters from the CLIP would remain frozen. For question $Q$ , each word $q_{t}$ would be mapped to the corresponding word embedding $e_{t} \\in \\mathbb{R}^{c}$ by the pre-trained caption model. Finally, we obtain the image and question sequences $Z$ :",
446
+ "bbox": [
447
+ 496,
448
+ 497,
449
+ 890,
450
+ 589
451
+ ],
452
+ "page_idx": 2
453
+ },
454
+ {
455
+ "type": "equation",
456
+ "text": "\n$$\nZ = \\overbrace {\\left[ v _ {1} , v _ {2} , \\cdots , v _ {S} \\right.} ^ {\\text {i m a g e}} \\underbrace {e _ {1} , e _ {2} , \\cdots , e _ {T}} _ {\\substack {\\text {q u e s t i o n} \\\\ \\text {e m b d e d i n g}}} \\bigg ], \\tag{2}\n$$\n",
457
+ "text_format": "latex",
458
+ "bbox": [
459
+ 571,
460
+ 601,
461
+ 890,
462
+ 657
463
+ ],
464
+ "page_idx": 2
465
+ },
466
+ {
467
+ "type": "text",
468
+ "text": "where $Z$ is the concatenated multi-modal prefixes.",
469
+ "bbox": [
470
+ 500,
471
+ 669,
472
+ 833,
473
+ 684
474
+ ],
475
+ "page_idx": 2
476
+ },
477
+ {
478
+ "type": "text",
479
+ "text": "3.2. Answer-Explanation Prompt Module",
480
+ "text_level": 1,
481
+ "bbox": [
482
+ 500,
483
+ 695,
484
+ 821,
485
+ 710
486
+ ],
487
+ "page_idx": 2
488
+ },
489
+ {
490
+ "type": "text",
491
+ "text": "It has been proved that the prompting mechanism can maintain the same optimization objectives between pretrained and downstream tasks [18, 28]. Considering the convenience and explainability, we leverage hand-crafted prompts as templates to enable the model to generate answers or explanations. As shown in Fig. 2, the $S^3 C$ includes two different kinds of templates, which are introduced respectively as follows.",
492
+ "bbox": [
493
+ 496,
494
+ 719,
495
+ 890,
496
+ 839
497
+ ],
498
+ "page_idx": 2
499
+ },
500
+ {
501
+ "type": "text",
502
+ "text": "Base Answer Template. The core idea of our $S^3 C$ is to use the answer scores as evaluation criteria. Hence, we first establish a basic answer template to acquire base answer scores. Different from previous works [20, 41], our",
503
+ "bbox": [
504
+ 496,
505
+ 839,
506
+ 893,
507
+ 901
508
+ ],
509
+ "page_idx": 2
510
+ },
511
+ {
512
+ "type": "page_number",
513
+ "text": "2648",
514
+ "bbox": [
515
+ 482,
516
+ 944,
517
+ 514,
518
+ 955
519
+ ],
520
+ "page_idx": 2
521
+ },
522
+ {
523
+ "type": "image",
524
+ "img_path": "images/6c9a1ae1014099259c66fe6acd5d5ccb3871a086fe23b1181827247d177149a2.jpg",
525
+ "image_caption": [
526
+ "Figure 2. Overview of our Semi-Supervised VQA-NLE method via Self-Critical learning $(S^3 C)$ framework. Given images and corresponding questions for labelled and unlabelled samples, we first use Answer-Explanation Prompt module to obtain the base answer scores and candidate explanations with a pre-trained VL model. Then these reasons are reorganized and fed back into the model to capture the explanatory answer score. Further, our Self-Critical Reinforcement module evaluates the generated explanations and returns the rewards to improve the self-interpretability of the model."
527
+ ],
528
+ "image_footnote": [],
529
+ "bbox": [
530
+ 86,
531
+ 94,
532
+ 885,
533
+ 340
534
+ ],
535
+ "page_idx": 3
536
+ },
537
+ {
538
+ "type": "text",
539
+ "text": "method treats the visual question answering task as a generation task so that the model can produce answers without a predefined answer space. Specifically, for a given image and question sequence $Z$ , we use natural language tokens \"the answer is\" to inspire the model to generate proper answers. By concatenating the language prompt, the base answer template $Z_{a} = [Z; \\langle \\text{answer} \\rangle]$ can be obtained, where $[; ]$ and $\\langle \\text{answer} \\rangle$ indicate concatenation operation and the specific language prompts, respectively. During training, the base answer template and ground-true answer label $A = \\{a_{n}\\}_{n=1}^{N}$ are fed into the VL model, where $a_{n}$ is the $n$ -th answer token and $N$ denotes the length of the answer label. Next, we compute the answer loss conditioned on the $Z_{a}$ in an autoregressive fashion:",
540
+ "bbox": [
541
+ 75,
542
+ 450,
543
+ 472,
544
+ 664
545
+ ],
546
+ "page_idx": 3
547
+ },
548
+ {
549
+ "type": "equation",
550
+ "text": "\n$$\nL _ {a} = - \\frac {1}{N} \\sum_ {n = 1} ^ {N} \\log P \\left(a _ {n} \\mid Z _ {a}, a _ {1}, a _ {2}, \\dots , a _ {n - 1}; \\theta\\right), \\tag {3}\n$$\n",
551
+ "text_format": "latex",
552
+ "bbox": [
553
+ 94,
554
+ 671,
555
+ 468,
556
+ 712
557
+ ],
558
+ "page_idx": 3
559
+ },
560
+ {
561
+ "type": "text",
562
+ "text": "where $\\theta$ denotes the parameters of the VL model. Moreover, based on the indexes of ground-true answers, we can acquire average probability $p_{\\theta}^{b}(A)$ as our base answer score.",
563
+ "bbox": [
564
+ 76,
565
+ 719,
566
+ 468,
567
+ 765
568
+ ],
569
+ "page_idx": 3
570
+ },
571
+ {
572
+ "type": "text",
573
+ "text": "Explanation Generation Template. To produce a reasonable rationale, we leverage the language prompt \"the reason is\" to motivate the model to generate free-text explanations. Like the base answer template, an explanation generation template $Z_{e} = [Z; \\langle \\text{reason} \\rangle]$ is constructed, where $\\langle \\text{reason} \\rangle$ is the natural language tokens. For labelled samples, we follow Eq. 3 in the autoregressive fashion and the cross-entropy loss to compute explanation loss $L_{e}$ . Because there are no human-annotated explanations available, we",
574
+ "bbox": [
575
+ 75,
576
+ 763,
577
+ 470,
578
+ 901
579
+ ],
580
+ "page_idx": 3
581
+ },
582
+ {
583
+ "type": "text",
584
+ "text": "would not compute any loss for unlabelled QA samples.",
585
+ "bbox": [
586
+ 500,
587
+ 450,
588
+ 870,
589
+ 467
590
+ ],
591
+ "page_idx": 3
592
+ },
593
+ {
594
+ "type": "text",
595
+ "text": "3.3. Self-Critical Reinforcement Module",
596
+ "text_level": 1,
597
+ "bbox": [
598
+ 500,
599
+ 481,
600
+ 812,
601
+ 496
602
+ ],
603
+ "page_idx": 3
604
+ },
605
+ {
606
+ "type": "text",
607
+ "text": "To gain logically consistent rationales, in this module, we expand the searching space by introducing sequence sampling algorithm [2] and generate a set of candidate explanations. Besides, the answer scores are treated as rewards to encourage the model to output more detailed interpretations. It's worth noting that the above operations would be implemented on both labelled and unlabelled samples.",
608
+ "bbox": [
609
+ 496,
610
+ 506,
611
+ 890,
612
+ 611
613
+ ],
614
+ "page_idx": 3
615
+ },
616
+ {
617
+ "type": "text",
618
+ "text": "Candidate Explanation Generation. Benefiting from the language-based prompting strategy and the pre-trained VL model, our explanation generation template can easily guide the model to produce human-readable sentences. Hence, for unlabelled QA samples, we directly apply the explanation generation template $Z_{e}$ to generate the candidate rationales. To be more specific, we utilize beam search [2] to sample the top- $K$ words from the VL model probability distribution at each time step and maintain these sequences with the highest probability. Then, these generated sentences are integrated into candidate explanations $R_{s} = \\{r_{k}^{s}\\}_{k = 1}^{K}$ for each QA pair without human-annotated explanations, where $r_k^s$ and $K$ indicate $k$ -th rationale and the size of beam search, respectively. Moreover, for labelled samples, a similar mechanism is used to build the corresponding explanation set $R_{f} = \\{r_{k}^{f}\\}_{k = 1}^{K}$ . The above operations are implemented in all samples for the following reasons: 1) Larger search space. As shown in Fig. 2, \"frying food is unhealthy\" and \"it includes many calories\"",
619
+ "bbox": [
620
+ 496,
621
+ 613,
622
+ 892,
623
+ 902
624
+ ],
625
+ "page_idx": 3
626
+ },
627
+ {
628
+ "type": "page_number",
629
+ "text": "2649",
630
+ "bbox": [
631
+ 482,
632
+ 944,
633
+ 514,
634
+ 955
635
+ ],
636
+ "page_idx": 3
637
+ },
638
+ {
639
+ "type": "text",
640
+ "text": "can both be rational explanations for the answer \"no\". The expanded search space provides more possibilities for generating reliable rationales. 2) Avoid overfitting. For labelled samples, although we can depend on human explanations to train the model, these labels are still one-sided and subjective. Using the sequence sampling strategy can prevent the model from overfitting these specific annotations [2,47]. In the end, because the labelled and unlabelled samples would be integrated into a mini-batch during training, we simplify $R_{f}$ and $R_{s}$ into $R = \\{r_k\\}_{k=1}^K$ to indicate the candidate explanations for each sample.",
641
+ "bbox": [
642
+ 75,
643
+ 90,
644
+ 472,
645
+ 256
646
+ ],
647
+ "page_idx": 4
648
+ },
649
+ {
650
+ "type": "text",
651
+ "text": "Self-Critical Reward. Inspired by [5, 51], an ideal rationale can help the model to infer the answer better. Based on this insight, we argue that the answer scores can be converted as self-critical rewards to evaluate these candidate explanations. Considering that this is a non-differentiable operation, we adopt reinforcement learning method [47] to achieve end-to-end training. In particular, given a sample and corresponding candidate explanations $R$ , we design a new input template:",
652
+ "bbox": [
653
+ 75,
654
+ 257,
655
+ 472,
656
+ 393
657
+ ],
658
+ "page_idx": 4
659
+ },
660
+ {
661
+ "type": "equation",
662
+ "text": "\n$$\nZ _ {r} ^ {k} = [ Z; \\langle r e a s o n \\rangle ; r _ {k}; \\langle a n s w e r \\rangle ], \\tag {4}\n$$\n",
663
+ "text_format": "latex",
664
+ "bbox": [
665
+ 153,
666
+ 402,
667
+ 468,
668
+ 421
669
+ ],
670
+ "page_idx": 4
671
+ },
672
+ {
673
+ "type": "text",
674
+ "text": "where $Z_{r}^{k}$ denotes the template of adding possible rationale $r_{k}$ . Then, this template is fed back into the model and obtains average probability $p_{\\theta}^{r}(A)$ about the answer, namely explanatory answer scores, where $A$ is the ground-truth answer for the sample. Meanwhile, we use the average probability $p_{\\theta}^{b}(A)$ from the output of the base answer template as the base scores. By applying reinforcement learning, the gradient is calculated by:",
675
+ "bbox": [
676
+ 75,
677
+ 431,
678
+ 472,
679
+ 555
680
+ ],
681
+ "page_idx": 4
682
+ },
683
+ {
684
+ "type": "equation",
685
+ "text": "\n$$\n\\nabla_ {\\theta} L _ {r} (\\theta) = - \\frac {1}{K} \\sum_ {k = 1} ^ {K} \\left(p _ {\\theta} ^ {r} (A) - p _ {\\theta} ^ {b} (A)\\right) \\nabla_ {\\theta} \\operatorname {l o g} p _ {\\theta} \\left(r _ {k}\\right), \\tag {5}\n$$\n",
686
+ "text_format": "latex",
687
+ "bbox": [
688
+ 86,
689
+ 565,
690
+ 468,
691
+ 606
692
+ ],
693
+ "page_idx": 4
694
+ },
695
+ {
696
+ "type": "text",
697
+ "text": "where $p_{\\theta}(r_k)$ is the probability of $k$ -th explanations. Based on the above computation, this gradient would tend to increase the probability of $k$ -th rationales when the answer score $p_{\\theta}^{r}(A)$ higher than the scores $p_{\\theta}^{b}(A)$ from the base answer template. Finally, for labelled samples, we append the human-annotated rationales to the candidate explanation $R$ and predict the answers by cross-entropy loss, namely $L_{ea}$ .",
698
+ "bbox": [
699
+ 75,
700
+ 617,
701
+ 468,
702
+ 724
703
+ ],
704
+ "page_idx": 4
705
+ },
706
+ {
707
+ "type": "text",
708
+ "text": "3.4. Loss",
709
+ "text_level": 1,
710
+ "bbox": [
711
+ 76,
712
+ 732,
713
+ 148,
714
+ 746
715
+ ],
716
+ "page_idx": 4
717
+ },
718
+ {
719
+ "type": "text",
720
+ "text": "During the process of training, the overall loss function can be represented as follows:",
721
+ "bbox": [
722
+ 75,
723
+ 755,
724
+ 468,
725
+ 786
726
+ ],
727
+ "page_idx": 4
728
+ },
729
+ {
730
+ "type": "equation",
731
+ "text": "\n$$\nL = L _ {a} + L _ {e} + L _ {e a} + \\lambda L _ {r}, \\tag {6}\n$$\n",
732
+ "text_format": "latex",
733
+ "bbox": [
734
+ 176,
735
+ 797,
736
+ 468,
737
+ 813
738
+ ],
739
+ "page_idx": 4
740
+ },
741
+ {
742
+ "type": "text",
743
+ "text": "where $\\lambda$ is used for balancing these two different types of losses (i.e., cross-entropy loss and reinforcement loss). At inference time, our model would first generate the rationales about QA pairs, then we use the explanatory template (i.e., Eq. 4) to obtain the corresponding answers.",
744
+ "bbox": [
745
+ 75,
746
+ 824,
747
+ 468,
748
+ 901
749
+ ],
750
+ "page_idx": 4
751
+ },
752
+ {
753
+ "type": "text",
754
+ "text": "4. Experiment",
755
+ "text_level": 1,
756
+ "bbox": [
757
+ 500,
758
+ 90,
759
+ 624,
760
+ 107
761
+ ],
762
+ "page_idx": 4
763
+ },
764
+ {
765
+ "type": "text",
766
+ "text": "4.1. Experimental setting",
767
+ "text_level": 1,
768
+ "bbox": [
769
+ 500,
770
+ 114,
771
+ 696,
772
+ 131
773
+ ],
774
+ "page_idx": 4
775
+ },
776
+ {
777
+ "type": "text",
778
+ "text": "Datasets. Following previous methods [48, 49], we mainly carry out the experiments on the two different VQANLE datasets: VQA-X [41] and A-OKVQA [49]. Meanwhile, since explain annotations are expensive and time-consuming, we also utilize large-scale VQA v2.0 [15] and OK-VQA [36] datasets to build the semi-supervised learning paradigm. Next, we will introduce the four datasets.",
779
+ "bbox": [
780
+ 496,
781
+ 138,
782
+ 890,
783
+ 243
784
+ ],
785
+ "page_idx": 4
786
+ },
787
+ {
788
+ "type": "text",
789
+ "text": "VQA-X. It is a vision-language dataset that provides explanations for justifying the answers. VQA-X is collected from the Visual Question Answering (VQA) dataset [4] where the images are obtained from the MSCOCO [27]. It consists of 28K images and 33K QA pairs, split into 29K/1.4K/1.9K for training, validation and testing. Meanwhile, VQA-X constructs complementary pairs which provide a question and two semantically similar images with different answers.",
790
+ "bbox": [
791
+ 496,
792
+ 244,
793
+ 890,
794
+ 378
795
+ ],
796
+ "page_idx": 4
797
+ },
798
+ {
799
+ "type": "text",
800
+ "text": "A-OKVQA. Compared to the VQA-X, the questions of A-OKVQA generally are required commonsense reasoning about the scene described in the images. It includes 24,903 Question/Answer/Rationale triplets, split into $17.1\\mathrm{K} / 1.1\\mathrm{K} / 6.7\\mathrm{K}$ for training, validation and testing. It collects images from the COCO 2017 [7] dataset, and is further filtered to obtain $23.7\\mathrm{K}$ unique images. Compared with previous datasets, the A-OKVQA has richer questions and requires broader areas of knowledge for reasoning.",
801
+ "bbox": [
802
+ 496,
803
+ 381,
804
+ 890,
805
+ 517
806
+ ],
807
+ "page_idx": 4
808
+ },
809
+ {
810
+ "type": "text",
811
+ "text": "VQA & OK-VQA. We use VQA v2.0 and OK-VQA to provide large-scale unlabelled datasets for semi-supervised learning. The VQA v2.0 dataset is widely used for many previous works [2, 52]. It consists of 443k questions and 195k images. To select explainable questions instead of some obvious cases (e.g., How many...?, What color...?), we filter out these questions from VQA v2.0 and obtain the $\\sim 90\\mathrm{k}$ additional questions based on the rules [41]. On the other hand, we also use knowledge-based OK-VQA [36] dataset to provide unlabelled knowledge-based QA pairs. It contains a total of 14k questions on 14k images.",
812
+ "bbox": [
813
+ 496,
814
+ 517,
815
+ 890,
816
+ 683
817
+ ],
818
+ "page_idx": 4
819
+ },
820
+ {
821
+ "type": "text",
822
+ "text": "Implementation Details. Following [37], each image is first pre-processed (such as including image resize, center crop, and normalize) by CLIP [45]. Meanwhile, we fix the ViT-B weights from the CLIP visual encoder to accelerate the training speed. For the mapping network, the image sequence length $S = 10$ and the embedding size is 768. The AdamW [21] is used as our optimizer with the weight decay 1e-5, and the batch size and beam size $K$ are set to 4 and 2. The weight coefficient $\\lambda$ is set to 10. We train all models on the four 1080Ti GPUs for 30 epochs with a learning rate of 1e-5.",
823
+ "bbox": [
824
+ 496,
825
+ 684,
826
+ 890,
827
+ 849
828
+ ],
829
+ "page_idx": 4
830
+ },
831
+ {
832
+ "type": "text",
833
+ "text": "4.2. Evaluation Measures",
834
+ "text_level": 1,
835
+ "bbox": [
836
+ 500,
837
+ 861,
838
+ 697,
839
+ 876
840
+ ],
841
+ "page_idx": 4
842
+ },
843
+ {
844
+ "type": "text",
845
+ "text": "Automatic Metering. Following [48, 60], we use the auto",
846
+ "bbox": [
847
+ 500,
848
+ 885,
849
+ 890,
850
+ 900
851
+ ],
852
+ "page_idx": 4
853
+ },
854
+ {
855
+ "type": "page_number",
856
+ "text": "2650",
857
+ "bbox": [
858
+ 482,
859
+ 944,
860
+ 514,
861
+ 955
862
+ ],
863
+ "page_idx": 4
864
+ },
865
+ {
866
+ "type": "table",
867
+ "img_path": "images/e6eda03b43de5d0d1ce034a25bc5d1f3741e2e12f4632b9b28221ee9ae4f7ef4.jpg",
868
+ "table_caption": [
869
+ "Table 1. Comparison with the state-of-the-art methods on the VQA-X. Note that these results are unfiltered scores. $S^3 C^*$ denotes the model without unlabelled samples."
870
+ ],
871
+ "table_footnote": [],
872
+ "table_body": "<table><tr><td></td><td colspan=\"7\">VQA-X</td></tr><tr><td>Approach</td><td>B4</td><td>M</td><td>R</td><td>S</td><td>C</td><td>Acc</td><td>Human</td></tr><tr><td>CAPS [41]</td><td>5.9</td><td>12.6</td><td>26.3</td><td>11.9</td><td>35.2</td><td>68.6</td><td>-</td></tr><tr><td>PJ-X [41]</td><td>19.5</td><td>18.2</td><td>43.4</td><td>15.1</td><td>71.3</td><td>76.4</td><td>65.4</td></tr><tr><td>FME [58]</td><td>24.4</td><td>19.5</td><td>47.7</td><td>17.9</td><td>88.8</td><td>75.5</td><td>-</td></tr><tr><td>NLX-GPT [48]</td><td>25.6</td><td>21.5</td><td>48.7</td><td>20.2</td><td>97.2</td><td>83.1</td><td>70.2</td></tr><tr><td>S3C*(ours)</td><td>26.5</td><td>22.0</td><td>49.0</td><td>20.9</td><td>100.5</td><td>83.7</td><td>73.9</td></tr><tr><td>S3C(ours)</td><td>27.8</td><td>22.8</td><td>50.7</td><td>21.5</td><td>104.4</td><td>85.6</td><td>77.4</td></tr></table>",
873
+ "bbox": [
874
+ 81,
875
+ 142,
876
+ 444,
877
+ 237
878
+ ],
879
+ "page_idx": 5
880
+ },
881
+ {
882
+ "type": "table",
883
+ "img_path": "images/9d9db038b4ead71b9d70515ea371f9b815f8f90b17037e6a262483fd9d5503c4.jpg",
884
+ "table_caption": [
885
+ "Table 2. Comparison with the state-of-the-art methods on the A-OKVQA. Note that these results are unfiltered scores. $S^3 C^*$ denotes the model without unlabelled samples."
886
+ ],
887
+ "table_footnote": [],
888
+ "table_body": "<table><tr><td></td><td colspan=\"8\">AOKVQA</td></tr><tr><td>Approach</td><td>B4</td><td>M</td><td>R</td><td>S</td><td>C</td><td>Acc val</td><td>Acc test</td><td>Human</td></tr><tr><td>ViLBERT [29]</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>30.6</td><td>25.9</td><td>-</td></tr><tr><td>LXMERT [52]</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>30.7</td><td>25.9</td><td>-</td></tr><tr><td>KRISP [35]</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>33.7</td><td>27.1</td><td>-</td></tr><tr><td>Clipcap [49]</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>30.8</td><td>25.9</td><td>-</td></tr><tr><td>e-UG [20]</td><td>15.1</td><td>18.1</td><td>42.4</td><td>14.9</td><td>51.5</td><td>30.5</td><td>25.6</td><td>44.1</td></tr><tr><td>NLX-GPT [48]</td><td>20.1</td><td>17.0</td><td>46.3</td><td>15.8</td><td>65.4</td><td>32.7</td><td>28.7</td><td>46.9</td></tr><tr><td>\\( S^2C^{*&#x27;} \\) (ours)</td><td>21.8</td><td>17.9</td><td>47.3</td><td>17.3</td><td>70.6</td><td>33.0</td><td>29.6</td><td>49.4</td></tr><tr><td>\\( S^3C \\) (ours)</td><td>22.5</td><td>18.5</td><td>48.4</td><td>18.1</td><td>74.4</td><td>34.2</td><td>33.5</td><td>54.7</td></tr></table>",
889
+ "bbox": [
890
+ 496,
891
+ 142,
892
+ 883,
893
+ 248
894
+ ],
895
+ "page_idx": 5
896
+ },
897
+ {
898
+ "type": "table",
899
+ "img_path": "images/5938ddc4399a3bab2f4a2431c7b8a6f600490b95bdaced85fe09bba08461188a.jpg",
900
+ "table_caption": [
901
+ "Table 3. Comparison with the state-of-the-art methods on the VQA-X. Note that these results are filtered scores. $S^3 C^*$ denotes the model without unlabelled samples."
902
+ ],
903
+ "table_footnote": [],
904
+ "table_body": "<table><tr><td></td><td>B1</td><td>B2</td><td>B3</td><td>B4</td><td>M</td><td>R</td><td>S</td><td>C</td><td>Acc</td><td>Human</td></tr><tr><td>RVT [34]</td><td>51.9</td><td>37.0</td><td>25.6</td><td>17.4</td><td>19.2</td><td>42.1</td><td>15.8</td><td>52.5</td><td>68.6</td><td>60.5</td></tr><tr><td>PJ-X [41]</td><td>57.4</td><td>42.2</td><td>30.9</td><td>22.7</td><td>19.7</td><td>46.0</td><td>17.1</td><td>82.7</td><td>76.4</td><td>69.3</td></tr><tr><td>FME [58]</td><td>59.1</td><td>43.4</td><td>31.7</td><td>23.1</td><td>20.4</td><td>47.1</td><td>18.4</td><td>87.0</td><td>75.5</td><td>-</td></tr><tr><td>QA-only [20]</td><td>51.0</td><td>36.4</td><td>25.3</td><td>17.3</td><td>18.6</td><td>41.9</td><td>14.9</td><td>49.9</td><td>-</td><td>-</td></tr><tr><td>e-UG [20]</td><td>57.3</td><td>42.7</td><td>31.4</td><td>23.2</td><td>22.1</td><td>45.7</td><td>20.1</td><td>74.1</td><td>80.5</td><td>71.4</td></tr><tr><td>NLX-GPT [48]</td><td>64.2</td><td>49.5</td><td>37.6</td><td>28.5</td><td>23.1</td><td>51.5</td><td>22.1</td><td>110.6</td><td>83.2</td><td>73.7</td></tr><tr><td>S3C*(ours)</td><td>64.4</td><td>49.9</td><td>38.0</td><td>29.1</td><td>23.4</td><td>51.9</td><td>22.7</td><td>112.1</td><td>83.7</td><td>75.9</td></tr><tr><td>S3C(ours)</td><td>64.7</td><td>50.5</td><td>38.8</td><td>30.7</td><td>23.9</td><td>52.1</td><td>23.0</td><td>116.7</td><td>85.6</td><td>79.2</td></tr></table>",
905
+ "bbox": [
906
+ 200,
907
+ 301,
908
+ 767,
909
+ 429
910
+ ],
911
+ "page_idx": 5
912
+ },
913
+ {
914
+ "type": "text",
915
+ "text": "matic metrics BLEU [40],METEOR[11],ROUGE-L[26], SPICE[1] and CIDEr[55] to evaluate generated explanations.For the evaluation of predicted answers, we follow [48,60] to compute the VQA accuracy.",
916
+ "bbox": [
917
+ 75,
918
+ 454,
919
+ 467,
920
+ 515
921
+ ],
922
+ "page_idx": 5
923
+ },
924
+ {
925
+ "type": "text",
926
+ "text": "Human Evaluation. Automatic VQA-NLE measures do not always reflect the correctness and logicality of the explanations [20, 53], thus we also build human evaluations. The process is similar to [20, 34]. Specifically, for each explanation, three human evaluators are required to decide whether an explanation can justify the answer and select an option (including \"yes, weak yes, weak no and no\"). The selection will be mapped to scores $(1, \\frac{2}{3}, \\frac{1}{3}$ and 0). The final scores are computed by averaging among all test samples. Meanwhile, these evaluators are asked to choose the reasons for unqualified explanations. We follow [20] to define three kinds of aspects to evaluate the explanations: irrelevant explanations, insufficient explanations and meaningless explanations. First, the irrelevant explanations may not match the image, for example, \"have a long neck\" is a good explanation for the answer \"giraffes\" when asked \"what animals are these?\", but the image may display cows. Second, the insufficient explanations only describe the image but cannot corroborate the answers. For example, the sentence \"there are some people\" does not sufficiently justify the question \"do people have a party?\". Lastly, some nonsensical sentences could be judged as contradictory explanations, such as \"a man is a woman\". For each sample, the human evaluators can select multiple shortcomings. More details of the human evaluation can be found in the [20].",
927
+ "bbox": [
928
+ 75,
929
+ 523,
930
+ 467,
931
+ 900
932
+ ],
933
+ "page_idx": 5
934
+ },
935
+ {
936
+ "type": "text",
937
+ "text": "4.3. Quantitative evaluation.",
938
+ "text_level": 1,
939
+ "bbox": [
940
+ 500,
941
+ 454,
942
+ 718,
943
+ 469
944
+ ],
945
+ "page_idx": 5
946
+ },
947
+ {
948
+ "type": "text",
949
+ "text": "Automatic Evaluation. We compare our method with the state-of-the-art models on the VQA-X and A-OKVQA datasets in Table 1-3. The B4, M, R, S, C, Acc and Human are short for BLEU-4, METEOR, ROUGE-L, SPICE, CIDEr, Answer precision and Human evaluation. We use \"unfiltered\" to indicate that the explanations are evaluated regardless of whether the answer is true or false. While \"filtered\" is to only consider the explanations which have correct answers. From Table 1, the $S^3 C^*$ denotes that we only apply the self-critical framework without using unlabelled samples, while the row $S^3 C$ indicates our method in the semi-supervised setting. We observe that the proposed framework outperforms both the post-hoc explanation methods [41, 58] and the self-rationalization method [48]. Meanwhile, it's worth noting that although the NLX-GPT uses a more powerful pre-trained VL model with larger pre-trained datasets (i.e., COCO captions [27], Flicker30k [43] and Visual Genome [24]), our method still obtains 3.3 absolute gains on the CIDEr indicator with fewer pre-trained data (only using COCO captions). Furthermore, when we utilize our proposed semi-supervised paradigm, the results are further improved by 7.2 points. These results show that our model can generate more reliable explanations and it can benefit from the amount of data without human-explanation labels. We also report the accuracy of answers in the column of \"Acc\". It can be observed that our self-critical method with semi-supervised learning can simultaneously boost the precision of answers and cor",
950
+ "bbox": [
951
+ 496,
952
+ 477,
953
+ 890,
954
+ 900
955
+ ],
956
+ "page_idx": 5
957
+ },
958
+ {
959
+ "type": "page_number",
960
+ "text": "2651",
961
+ "bbox": [
962
+ 482,
963
+ 944,
964
+ 513,
965
+ 955
966
+ ],
967
+ "page_idx": 5
968
+ },
969
+ {
970
+ "type": "table",
971
+ "img_path": "images/da21cabd57c16c24a7e2169ce89c73de4852738715211d6249d7ca49aab58162.jpg",
972
+ "table_caption": [
973
+ "Table 4. Main shortcomings. The main shortcomings of unqualified explanations on the VQA-X dataset. For each sample, human evaluators can select multiple shortcomings."
974
+ ],
975
+ "table_footnote": [],
976
+ "table_body": "<table><tr><td>Model</td><td>Irrelevant explanations</td><td>Insufficient explanations</td><td>Meaningless explanations</td></tr><tr><td>RVT [34]</td><td>25.7%</td><td>33.5%</td><td>11.4%</td></tr><tr><td>PJ-X [41]</td><td>21.1%</td><td>28.4%</td><td>9.2%</td></tr><tr><td>e-UG [20]</td><td>22.8%</td><td>25.4%</td><td>8.7%</td></tr><tr><td>NLX-GPT [48]</td><td>20.3%</td><td>22.2%</td><td>9.1%</td></tr><tr><td>S3C (ours)</td><td>17.3%</td><td>18.9%</td><td>8.2%</td></tr></table>",
977
+ "bbox": [
978
+ 81,
979
+ 141,
980
+ 468,
981
+ 243
982
+ ],
983
+ "page_idx": 6
984
+ },
985
+ {
986
+ "type": "table",
987
+ "img_path": "images/9384a847859a37b77b0e223920bb38a1ba6644d8f4857f0151c7dcc6ca18dda1.jpg",
988
+ "table_caption": [
989
+ "Table 5. Cross-dataset testing. We alternately use the VQA-X and A-OKVQA as source dataset and target dataset to test the generalization of our framework."
990
+ ],
991
+ "table_footnote": [],
992
+ "table_body": "<table><tr><td></td><td colspan=\"6\">VQA-X→A-OKVQA</td></tr><tr><td>Approach</td><td>B4</td><td>M</td><td>R</td><td>S</td><td>C</td><td>Acc</td></tr><tr><td>NLX-GPT [48]</td><td>10.7</td><td>12.7</td><td>34.2</td><td>10.7</td><td>35.4</td><td>10.4</td></tr><tr><td>S3C(ours)</td><td>12.0</td><td>13.3</td><td>34.3</td><td>12.5</td><td>45.3</td><td>18.8</td></tr><tr><td></td><td colspan=\"6\">A-OKVQA→VQA-X</td></tr><tr><td>Approach</td><td>B4</td><td>M</td><td>R</td><td>S</td><td>C</td><td>Acc</td></tr><tr><td>NLX-GPT [48]</td><td>9.1</td><td>13.6</td><td>32.8</td><td>9.1</td><td>33.2</td><td>42.4</td></tr><tr><td>S3C(ours)</td><td>10.9</td><td>15.0</td><td>34.1</td><td>10.4</td><td>38.6</td><td>43.8</td></tr></table>",
993
+ "bbox": [
994
+ 519,
995
+ 142,
996
+ 880,
997
+ 253
998
+ ],
999
+ "page_idx": 6
1000
+ },
1001
+ {
1002
+ "type": "table",
1003
+ "img_path": "images/e1eac8ada72bf5d72678ba027f7e008e2317a7cf3997c08c3e1cd6c72cd59876.jpg",
1004
+ "table_caption": [
1005
+ "Table 6. Ablation study. We ablate key components to demonstrate the effectiveness of our method. SCR and Semi are Self-Critical Reinforcement module and Semi-supervised learning paradigm respectively."
1006
+ ],
1007
+ "table_footnote": [],
1008
+ "table_body": "<table><tr><td></td><td>question</td><td>image</td><td>answer</td><td>explanation</td><td>SCR</td><td>Semi</td><td>B4</td><td>M</td><td>R</td><td>S</td><td>C</td><td>Acc</td></tr><tr><td>1</td><td>✓</td><td>✓</td><td>✓</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>80.1</td></tr><tr><td>2</td><td>✓</td><td>✓</td><td>-</td><td>✓</td><td>-</td><td>-</td><td>24.4</td><td>20.7</td><td>47.3</td><td>19.5</td><td>90.4</td><td>-</td></tr><tr><td>3</td><td>✓</td><td>✓</td><td>✓</td><td>✓</td><td>-</td><td>-</td><td>27.5</td><td>22.9</td><td>50.4</td><td>21.9</td><td>109.1</td><td>82.2</td></tr><tr><td>4</td><td>✓</td><td>✓</td><td>✓</td><td>✓</td><td>✓</td><td>-</td><td>29.1</td><td>23.4</td><td>51.9</td><td>22.7</td><td>112.1</td><td>83.7</td></tr><tr><td>5</td><td>✓</td><td>✓</td><td>✓</td><td>✓</td><td>✓</td><td>✓</td><td>30.7</td><td>23.9</td><td>52.1</td><td>23.0</td><td>116.7</td><td>85.6</td></tr></table>",
1009
+ "bbox": [
1010
+ 140,
1011
+ 306,
1012
+ 823,
1013
+ 398
1014
+ ],
1015
+ "page_idx": 6
1016
+ },
1017
+ {
1018
+ "type": "text",
1019
+ "text": "responding explanations.",
1020
+ "bbox": [
1021
+ 76,
1022
+ 424,
1023
+ 243,
1024
+ 439
1025
+ ],
1026
+ "page_idx": 6
1027
+ },
1028
+ {
1029
+ "type": "text",
1030
+ "text": "In Table 2, we evaluate our method on the A-OKVQA dataset. The results show that the $S^3 C$ can outperform all the previous works [17, 29, 35, 48, 49, 52]. Especially, compared to the SOTA model [48], there are 9.0 points and $4.8\\%$ improvements in the CIDEr and answer accuracy based on our model. It demonstrates that our model can benefit from semi-supervised learning and generate explanations about commonsense reasoning.",
1031
+ "bbox": [
1032
+ 75,
1033
+ 446,
1034
+ 468,
1035
+ 568
1036
+ ],
1037
+ "page_idx": 6
1038
+ },
1039
+ {
1040
+ "type": "text",
1041
+ "text": "In addition, to prove the algorithm's validity, we follow [20, 48] to report the filtered scores for the VQAX dataset in Table 3. Through filtering the correct answers, our method can outperform whatever post-hoc [20, 34, 41, 58] and self-rationalization [48] methods. Meanwhile, the proposed $S^3 C$ achieves a new state-of-the-art with the CIDEr score improved by 6.1 points to 116.7 and boost the answer accuracy by $2.4\\%$ to 85.6.",
1042
+ "bbox": [
1043
+ 75,
1044
+ 575,
1045
+ 468,
1046
+ 696
1047
+ ],
1048
+ "page_idx": 6
1049
+ },
1050
+ {
1051
+ "type": "text",
1052
+ "text": "Human Evaluation. To evaluate the faithfulness and correctness of these generated explanations, we conduct the human evaluation that is shown in the column of \"Human\" in Table 1-3. The experimental results further prove our method has better self-interpretability for the VQA-NLE task. Moreover, we also ask the human evaluators to select the shortcomings for each unqualified explanation on the VQA-X dataset. As shown in Table 4, we build three shortcoming options (i.e., irrelevant explanations, insufficient explanations and meaningless explanations) with a multi-choice process. These results indicate that the $S^3 C$ can obtain relatively better rationales and empirically confirm the effectiveness of our method.",
1053
+ "bbox": [
1054
+ 75,
1055
+ 703,
1056
+ 470,
1057
+ 900
1058
+ ],
1059
+ "page_idx": 6
1060
+ },
1061
+ {
1062
+ "type": "text",
1063
+ "text": "4.4. Cross-dataset Performance.",
1064
+ "text_level": 1,
1065
+ "bbox": [
1066
+ 500,
1067
+ 422,
1068
+ 748,
1069
+ 439
1070
+ ],
1071
+ "page_idx": 6
1072
+ },
1073
+ {
1074
+ "type": "text",
1075
+ "text": "We propose cross-dataset experiments to measure the generalization and reliability of our framework. For all we know, there is very little research on cross-dataset explanation generation. As shown in the first two rows of Table 5, we use the VQA-X and the A-OKVQA as the source dataset and target dataset to test the generalization ability of our framework. Specifically, we use the weights from the state-of-the-art method [48], which are trained on the VQA-X dataset and tested on the A-OKVQA dataset to obtain the results of the NLX-GPT. From the comparisons of six indicators, we find that our self-critical model has better generalization from the general VQA to the knowledge-based VQA. This suggests that $S^3 C$ can learn commonsense from visual question answering and generate better explanations and answers. Meanwhile, when we exchange the source dataset and target dataset as shown in the last two rows of Table 5, our framework can still surpass the existing SOTA model on all evaluation measures. These quantitative results effectively demonstrate the generalization and stability of our model.",
1076
+ "bbox": [
1077
+ 496,
1078
+ 446,
1079
+ 890,
1080
+ 748
1081
+ ],
1082
+ "page_idx": 6
1083
+ },
1084
+ {
1085
+ "type": "text",
1086
+ "text": "4.5. Ablation Studies.",
1087
+ "text_level": 1,
1088
+ "bbox": [
1089
+ 500,
1090
+ 756,
1091
+ 668,
1092
+ 771
1093
+ ],
1094
+ "page_idx": 6
1095
+ },
1096
+ {
1097
+ "type": "text",
1098
+ "text": "As shown in Table 6, we conduct several ablation studies on the VQA-X to demonstrate the effectiveness of our method. In particular, we first build the experiments in the 1-3 rows of Table 6, which use our Answer-Explanation Prompt module with cross-entropy loss as the baseline models. It can be found that when we simultaneously generate both answers and interpretations, the answer accuracy and CIDEr can achieve significant improvements. Meanwhile,",
1099
+ "bbox": [
1100
+ 496,
1101
+ 779,
1102
+ 890,
1103
+ 900
1104
+ ],
1105
+ "page_idx": 6
1106
+ },
1107
+ {
1108
+ "type": "page_number",
1109
+ "text": "2652",
1110
+ "bbox": [
1111
+ 482,
1112
+ 944,
1113
+ 514,
1114
+ 955
1115
+ ],
1116
+ "page_idx": 6
1117
+ },
1118
+ {
1119
+ "type": "image",
1120
+ "img_path": "images/7e01d8549bb6df38c2dafa143287e46634835e7efea7ff386f6c873aac75eb11.jpg",
1121
+ "image_caption": [
1122
+ "(a) Is this man dressed formal?"
1123
+ ],
1124
+ "image_footnote": [],
1125
+ "bbox": [
1126
+ 111,
1127
+ 97,
1128
+ 308,
1129
+ 191
1130
+ ],
1131
+ "page_idx": 7
1132
+ },
1133
+ {
1134
+ "type": "image",
1135
+ "img_path": "images/ca253b406d47da5ce9e8dd0643f14741dc101b45433029313006510d4e8be934.jpg",
1136
+ "image_caption": [
1137
+ "Ours: [yes] <he is wearing a dress shirt and tie>",
1138
+ "(b) Is he in a swimming pool?",
1139
+ "NLX: [yes] <he is on a surfboard>",
1140
+ "Ours: [no] <he is surfing in the ocean>"
1141
+ ],
1142
+ "image_footnote": [],
1143
+ "bbox": [
1144
+ 387,
1145
+ 97,
1146
+ 584,
1147
+ 191
1148
+ ],
1149
+ "page_idx": 7
1150
+ },
1151
+ {
1152
+ "type": "image",
1153
+ "img_path": "images/39ae1e8ba51b4a97e658d08b9090d5b5fe7801fe50db87ef8fd67f5798d0bfac.jpg",
1154
+ "image_caption": [
1155
+ "(c) What is the elephant doing?",
1156
+ "NLX: [drinking] <it is in the water>",
1157
+ "Ours: [drinking]<the trunk is in the water>"
1158
+ ],
1159
+ "image_footnote": [],
1160
+ "bbox": [
1161
+ 653,
1162
+ 97,
1163
+ 848,
1164
+ 191
1165
+ ],
1166
+ "page_idx": 7
1167
+ },
1168
+ {
1169
+ "type": "image",
1170
+ "img_path": "images/c4f860296fc1546b20f04255e426a9cdbc3d82adb6d7857051de46cd6ac72882.jpg",
1171
+ "image_caption": [
1172
+ "NLX: [no] <he is wearing a tie>",
1173
+ "(d) Are the people having a party?",
1174
+ "NLX: [no] <there are no people present>",
1175
+ "Ours: [yes] <they are laughing and having a good time>",
1176
+ "GT: [yes] <they are standing in a kitchen",
1177
+ "with drinks in hand and laughing>",
1178
+ "Figure 3. Some generation results on the VQA-X datasets. The [ ] and $<<$ indicate answers and explanations respectively. We show the results of the state-of-the-art NLX-GPT [48], our method and ground-truth (they are short for NLX, Ours and GT)."
1179
+ ],
1180
+ "image_footnote": [],
1181
+ "bbox": [
1182
+ 111,
1183
+ 251,
1184
+ 305,
1185
+ 347
1186
+ ],
1187
+ "page_idx": 7
1188
+ },
1189
+ {
1190
+ "type": "image",
1191
+ "img_path": "images/245f9a6f2d0c2769e88489f6a21f838f96135dde6cd9756a73f082c8f5875f4d.jpg",
1192
+ "image_caption": [
1193
+ "GT: [yes] <he's wearing a crisp dress shirt and a tie>",
1194
+ "GT: [no] <the man is riding an ocean wave>",
1195
+ "(e) Does this animal appear to be looking",
1196
+ "at the camera?",
1197
+ "NLX: [yes] <it is looking at something>",
1198
+ "Ours: [no] <it is looking away from the camera>",
1199
+ "GT: [no] <the animal is looking off to the side>"
1200
+ ],
1201
+ "image_footnote": [],
1202
+ "bbox": [
1203
+ 387,
1204
+ 251,
1205
+ 584,
1206
+ 347
1207
+ ],
1208
+ "page_idx": 7
1209
+ },
1210
+ {
1211
+ "type": "image",
1212
+ "img_path": "images/7bec96df02b62f6838cf074149ae4a727fd068f7ab72ee06b34553bac058bac9.jpg",
1213
+ "image_caption": [
1214
+ "GT: [drinking]<his trunk is in the water>",
1215
+ "(f) Is this a bar?",
1216
+ "NLX: [yes] <there are people sitting at tables>",
1217
+ "Ours: [yes] <the men are sitting at a bar with drinks in front of them>",
1218
+ "GT: [yes] <there are drinks on a long countertop with people eating and drinking>"
1219
+ ],
1220
+ "image_footnote": [],
1221
+ "bbox": [
1222
+ 653,
1223
+ 251,
1224
+ 849,
1225
+ 347
1226
+ ],
1227
+ "page_idx": 7
1228
+ },
1229
+ {
1230
+ "type": "text",
1231
+ "text": "these results also prove that the answer-explanation pairs have inherent consistency and they can promote each other. Further, we establish the Self-Critical Reinforcement module as shown in the fourth row, and the result of accuracy and CIDEr are boosted by $1.5\\%$ and 3.0 points. It suggests that our self-critical method with sequence sampling strategy can encourage the model to generate more reliable explanations and correct answers. Finally, when we extend our model to the semi-supervised setting, the performance improves to 85.6 on the accuracy and 116.7 on the CIDEr over the baseline model (i.e., the row 5 of Table 6) by $3.4\\%$ and 7.6 points. The results demonstrate that our framework is remarkably effective in improving the logical consistency and the self-interpretability of the model.",
1232
+ "bbox": [
1233
+ 75,
1234
+ 484,
1235
+ 472,
1236
+ 696
1237
+ ],
1238
+ "page_idx": 7
1239
+ },
1240
+ {
1241
+ "type": "text",
1242
+ "text": "4.6. Qualitative evaluation.",
1243
+ "text_level": 1,
1244
+ "bbox": [
1245
+ 76,
1246
+ 709,
1247
+ 287,
1248
+ 724
1249
+ ],
1250
+ "page_idx": 7
1251
+ },
1252
+ {
1253
+ "type": "text",
1254
+ "text": "As shown in Fig. 3, we show some qualitative results from NLX-GPT [48] and our $S^3 C$ method on the VQA-X dataset. Through overall comparison, our model achieves better logical consistency between answers and explanations. For example, in Fig. 3 (a), although the SOTA method [48] correctly identifies the significant symbol of formal dress (i.e., tie), the predicted answer is \"no\" that is contradictory to the explanation. On the contrary, our method not only obtains a more complete and faithful rationale but also generates a logically consistent answer. Additionally, the $S^3 C$ can also produce more persuasive and",
1255
+ "bbox": [
1256
+ 75,
1257
+ 734,
1258
+ 472,
1259
+ 902
1260
+ ],
1261
+ "page_idx": 7
1262
+ },
1263
+ {
1264
+ "type": "text",
1265
+ "text": "trusty sentences. For instance, in Fig. 3 (b)-(f), the generated sentences contain scene description \"having a good time\" and fine-grained rationale \"the trunk is in the water\".",
1266
+ "bbox": [
1267
+ 496,
1268
+ 484,
1269
+ 890,
1270
+ 531
1271
+ ],
1272
+ "page_idx": 7
1273
+ },
1274
+ {
1275
+ "type": "text",
1276
+ "text": "5. Conclusion",
1277
+ "text_level": 1,
1278
+ "bbox": [
1279
+ 500,
1280
+ 549,
1281
+ 617,
1282
+ 564
1283
+ ],
1284
+ "page_idx": 7
1285
+ },
1286
+ {
1287
+ "type": "text",
1288
+ "text": "In this paper, we propose a new Semi-Supervised VQANLE method via Self-Critical Learning $(S^3 C)$ . Different from previous works, our method first utilizes the prompting mechanism to motivate the model to generate answers and candidate explanations. Meanwhile, we design a novel Self-Critical Reinforcement module, which converts the answer scores as rewards to evaluate these possible rationales. Furthermore, our framework can benefit from an abundance of question-answer pairs without human-annotated explanations and further boost the self-interpretability of the model. With the automatic measures and human evaluations, our $S^3 C$ achieves a new state-of-the-art on multiple benchmarks and provides a new paradigm for our community.",
1289
+ "bbox": [
1290
+ 496,
1291
+ 575,
1292
+ 890,
1293
+ 773
1294
+ ],
1295
+ "page_idx": 7
1296
+ },
1297
+ {
1298
+ "type": "text",
1299
+ "text": "Acknowledgements",
1300
+ "text_level": 1,
1301
+ "bbox": [
1302
+ 500,
1303
+ 790,
1304
+ 666,
1305
+ 806
1306
+ ],
1307
+ "page_idx": 7
1308
+ },
1309
+ {
1310
+ "type": "text",
1311
+ "text": "This work was supported by National Key R&D Program of China (No.2020AAA0106900), the National Natural Science Foundation of China (No.U19B2037), Shaanxi Provincial Key RD Program (No.2021KWZ-03), and Natural Science Basic Research Program of Shaanxi (No.2021JCW-03).",
1312
+ "bbox": [
1313
+ 496,
1314
+ 816,
1315
+ 890,
1316
+ 900
1317
+ ],
1318
+ "page_idx": 7
1319
+ },
1320
+ {
1321
+ "type": "page_number",
1322
+ "text": "2653",
1323
+ "bbox": [
1324
+ 482,
1325
+ 944,
1326
+ 514,
1327
+ 955
1328
+ ],
1329
+ "page_idx": 7
1330
+ },
1331
+ {
1332
+ "type": "text",
1333
+ "text": "References",
1334
+ "text_level": 1,
1335
+ "bbox": [
1336
+ 78,
1337
+ 89,
1338
+ 173,
1339
+ 104
1340
+ ],
1341
+ "page_idx": 8
1342
+ },
1343
+ {
1344
+ "type": "list",
1345
+ "sub_type": "ref_text",
1346
+ "list_items": [
1347
+ "[1] Peter Anderson, Basura Fernando, Mark Johnson, and Stephen Gould. Spice: Semantic propositional image caption evaluation. In European conference on computer vision, pages 382-398. Springer, 2016.",
1348
+ "[2] Peter Anderson, Xiaodong He, Chris Buehler, Damien Teney, Mark Johnson, Stephen Gould, and Lei Zhang. Bottom-up and top-down attention for image captioning and visual question answering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6077-6086, 2018.",
1349
+ "[3] Peter Anderson, Xiaodong He, Chris Buehler, Damien Teney, Mark Johnson, Stephen Gould, and Lei Zhang. Bottom-up and top-down attention for image captioning and visual question answering. In CVPR, pages 6077-6086, 2018.",
1350
+ "[4] Stanislaw Antol, Aishwarya Agrawal, Jiasen Lu, Margaret Mitchell, Dhruv Batra, C Lawrence Zitnick, and Devi Parikh. Vqa: Visual question answering. In Proceedings of the IEEE international conference on computer vision, pages 2425-2433, 2015.",
1351
+ "[5] Ermanno Bencivenga. Free logics. In Handbook of philosophical logic, pages 147-196. Springer, 2002.",
1352
+ "[6] Oana-Maria Camburu, Tim Rocktäschel, Thomas Lukasiewicz, and Phil Blunsom. e-snli: Natural language inference with natural language explanations. Advances in Neural Information Processing Systems, 31, 2018.",
1353
+ "[7] Xinlei Chen, Hao Fang, Tsung-Yi Lin, Ramakrishna Vedantam, Saurabh Gupta, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco captions: Data collection and evaluation server. arXiv preprint arXiv:1504.00325, 2015.",
1354
+ "[8] Yen-Chun Chen, Linjie Li, Licheng Yu, Ahmed El Kholy, Faisal Ahmed, Zhe Gan, Yu Cheng, and Jingjing Liu. Uniter: Universal image-text representation learning. In European conference on computer vision, pages 104-120. Springer, 2020.",
1355
+ "[9] Jaemin Cho, Jie Lei, Hao Tan, and Mohit Bansal. Unifying vision-and-language tasks via text generation. In International Conference on Machine Learning, pages 1931-1942. PMLR, 2021.",
1356
+ "[10] Marcella Cornia, Matteo Stefanini, Lorenzo Baraldi, and Rita Cucchiara. Meshed-memory transformer for image captioning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10578-10587, 2020.",
1357
+ "[11] Michael Denkowski and Alon Lavie. Meteor universal: Language specific translation evaluation for any target language. In Proceedings of the ninth workshop on statistical machine translation, pages 376-380, 2014.",
1358
+ "[12] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018.",
1359
+ "[13] Jianfeng Dong, Xirong Li, and Cees G M Snoek. Predicting visual features from text for image and video caption retrieval. IEEE TMM, 20(12):3377-3388, 2018."
1360
+ ],
1361
+ "bbox": [
1362
+ 76,
1363
+ 114,
1364
+ 470,
1365
+ 900
1366
+ ],
1367
+ "page_idx": 8
1368
+ },
1369
+ {
1370
+ "type": "list",
1371
+ "sub_type": "ref_text",
1372
+ "list_items": [
1373
+ "[14] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020.",
1374
+ "[15] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Batra, and Devi Parikh. Making the v in vqa matter: Elevating the role of image understanding in visual question answering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6904-6913, 2017.",
1375
+ "[16] Sepp Hochreiter and Jürgen Schmidhuber. Long short-term memory. Neural computation, 9(8):1735-1780, 1997.",
1376
+ "[17] Yu Jiang, Vivek Natarajan, Xinlei Chen, Marcus Rohrbach, Dhruv Batra, and Devi Parikh. Pythia v0. 1: the winning entry to the vqa challenge 2018. arXiv preprint arXiv:1807.09956, 2018.",
1377
+ "[18] Woojeong Jin, Yu Cheng, Yelong Shen, Weizhu Chen, and Xiang Ren. A good prompt is worth millions of parameters: Low-resource prompt-based learning for vision-language models. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 2763-2775, 2022.",
1378
+ "[19] Jaehun Jung, Lianhui Qin, Sean Welleck, Faeze Brahman, Chandra Bhagavatula, Ronan Le Bras, and Yejin Choi. Maieutic prompting: Logically consistent reasoning with recursive explanations. arXiv preprint arXiv:2205.11822, 2022.",
1379
+ "[20] Maxime Kayser, Oana-Maria Camburu, Leonard Salewski, Cornelius Emde, Virginie Do, Zeynep Akata, and Thomas Lukasiewicz. e-vil: A dataset and benchmark for natural language explanations in vision-language tasks. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1244-1254, 2021.",
1380
+ "[21] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014.",
1381
+ "[22] Durk P Kingma, Shakir Mohamed, Danilo Jimenez Rezende, and Max Welling. Semi-supervised learning with deep generative models. Advances in neural information processing systems, 27, 2014.",
1382
+ "[23] Thomas N. Kipf and Max Welling. Semi-supervised classification with graph convolutional networks. In ICLR, 2017.",
1383
+ "[24] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome: Connecting language and vision using crowdsourced dense image annotations. International journal of computer vision, 123(1):32-73, 2017.",
1384
+ "[25] Xiujun Li, Xi Yin, Chunyuan Li, Pengchuan Zhang, Xiaowei Hu, Lei Zhang, Lijuan Wang, Houdong Hu, Li Dong, Furu Wei, et al. Oscar: Object-semantics aligned pre-training for vision-language tasks. In European Conference on Computer Vision, pages 121-137. Springer, 2020.",
1385
+ "[26] Chin-Yew Lin. Rouge: A package for automatic evaluation of summaries. In Text summarization branches out, pages 74-81, 2004."
1386
+ ],
1387
+ "bbox": [
1388
+ 501,
1389
+ 92,
1390
+ 893,
1391
+ 898
1392
+ ],
1393
+ "page_idx": 8
1394
+ },
1395
+ {
1396
+ "type": "page_number",
1397
+ "text": "2654",
1398
+ "bbox": [
1399
+ 482,
1400
+ 944,
1401
+ 514,
1402
+ 955
1403
+ ],
1404
+ "page_idx": 8
1405
+ },
1406
+ {
1407
+ "type": "list",
1408
+ "sub_type": "ref_text",
1409
+ "list_items": [
1410
+ "[27] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014.",
1411
+ "[28] Yuhang Liu, Wei Wei, Daowan Peng, and Feida Zhu. Declaration-based prompt tuning for visual question answering. arXiv preprint arXiv:2205.02456, 2022.",
1412
+ "[29] Jiasen Lu, Dhruv Batra, Devi Parikh, and Stefan Lee. Vilbert: Pretraining task-agnostic visiolinguistic representations for vision-and-language tasks. Advances in neural information processing systems, 32, 2019.",
1413
+ "[30] Jiasen Lu, Jianwei Yang, Dhruv Batra, and Devi Parikh. Hierarchical question-image co-attention for visual question answering. Advances in neural information processing systems, 29, 2016.",
1414
+ "[31] Jiasen Lu, Jianwei Yang, Dhruv Batra, and Devi Parikh. Hierarchical question-image co-attention for visual question answering. In NeurIPS, pages 289–297, 2016.",
1415
+ "[32] Chao Ma, Chunhua Shen, Anthony R. Dick, Qi Wu, Peng Wang, Anton van den Hengel, and Ian D. Reid. Visual question answering with memory-augmented networks. In CVPR, pages 6975-6984, 2018.",
1416
+ "[33] Mateusz Malinowski and Mario Fritz. A multi-world approach to question answering about real-world scenes based on uncertain input. Advances in neural information processing systems, 27, 2014.",
1417
+ "[34] Ana Marasović, Chandra Bhagavatula, Jae Sung Park, Ronan Le Bras, Noah A Smith, and Yejin Choi. Natural language rationales with full-stack visual reasoning: From pixels to semantic frames to commonsense graphs. arXiv preprint arXiv:2010.07526, 2020.",
1418
+ "[35] Kenneth Marino, Xinlei Chen, Devi Parikh, Abhinav Gupta, and Marcus Rohrbach. Krisp: Integrating implicit and symbolic knowledge for open-domain knowledge-based vqa. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14111-14121, 2021.",
1419
+ "[36] Kenneth Marino, Mohammad Rastegari, Ali Farhadi, and Roozbeh Mottaghi. Ok-vqa: A visual question answering benchmark requiring external knowledge. In Proceedings of the IEEE/cvf conference on computer vision and pattern recognition, pages 3195-3204, 2019.",
1420
+ "[37] Ron Mokady, Amir Hertz, and Amit H Bermano. Clipcap: Clip prefix for image captioning. arXiv preprint arXiv:2111.09734, 2021.",
1421
+ "[38] Sharan Narang, Colin Raffel, Katherine Lee, Adam Roberts, Noah Fiedel, and Karishma Malkan. Wt5?! training text-to-text models to explain their predictions. arXiv preprint arXiv:2004.14546, 2020.",
1422
+ "[39] Yingwei Pan, Ting Yao, Yehao Li, and Tao Mei. X-linear attention networks for image captioning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10971-10980, 2020.",
1423
+ "[40] Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. Bleu: a method for automatic evaluation of machine translation. In Proceedings of the 40th annual meeting of the Association for Computational Linguistics, pages 311-318, 2002."
1424
+ ],
1425
+ "bbox": [
1426
+ 78,
1427
+ 90,
1428
+ 468,
1429
+ 898
1430
+ ],
1431
+ "page_idx": 9
1432
+ },
1433
+ {
1434
+ "type": "list",
1435
+ "sub_type": "ref_text",
1436
+ "list_items": [
1437
+ "[41] Dong Huk Park, Lisa Anne Hendricks, Zeynep Akata, Anna Rohrbach, Bernt Schiele, Trevor Darrell, and Marcus Rohrbach. Multimodal explanations: Justifying decisions and pointing to the evidence. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8779-8788, 2018.",
1438
+ "[42] Badri N Patro, Mayank Lunayach, Shivansh Patel, and Vinay P Namboodiri. U-cam: Visual explanation using uncertainty based class activation maps. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7444-7453, 2019.",
1439
+ "[43] Bryan A Plummer, Liwei Wang, Chris M Cervantes, Juan C Caicedo, Julia Hockenmaier, and Svetlana Lazebnik. Flickr30k entities: Collecting region-to-phrase correspondences for richer image-to-sentence models. In Proceedings of the IEEE international conference on computer vision, pages 2641–2649, 2015.",
1440
+ "[44] Guo-Jun Qi and Jiebo Luo. Small data challenges in big data era: A survey of recent progress on unsupervised and semi-supervised methods. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2020.",
1441
+ "[45] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021.",
1442
+ "[46] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9, 2019.",
1443
+ "[47] Steven J Rennie, Etienne Marcheret, Youssef Mroueh, Jerret Ross, and Vaibhava Goel. Self-critical sequence training for image captioning. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7008-7024, 2017.",
1444
+ "[48] Fawaz Sammani, Tanmoy Mukherjee, and Nikos Deligiannis. Nlx-gpt: A model for natural language explanations in vision and vision-language tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8322-8332, 2022.",
1445
+ "[49] Dustin Schwenk, Apoorv Khandelwal, Christopher Clark, Kenneth Marino, and Roozbeh Mottaghi. A-okvqa: A benchmark for visual question answering using world knowledge. arXiv preprint arXiv:2206.01718, 2022.",
1446
+ "[50] Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In Proceedings of the IEEE international conference on computer vision, pages 618-626, 2017.",
1447
+ "[51] Suzanne Sia, Anton Belyy, Amjad Almahairi, Madian Khabsa, Luke Zettlemoyer, and Lambert Mathias. Logical satisfiability of counterfactuals for faithful explanations in nli. arXiv preprint arXiv:2205.12469, 2022.",
1448
+ "[52] Hao Tan and Mohit Bansal. Lxmert: Learning cross-modality encoder representations from transformers. arXiv preprint arXiv:1908.07490, 2019."
1449
+ ],
1450
+ "bbox": [
1451
+ 501,
1452
+ 92,
1453
+ 890,
1454
+ 900
1455
+ ],
1456
+ "page_idx": 9
1457
+ },
1458
+ {
1459
+ "type": "page_number",
1460
+ "text": "2655",
1461
+ "bbox": [
1462
+ 482,
1463
+ 945,
1464
+ 514,
1465
+ 955
1466
+ ],
1467
+ "page_idx": 9
1468
+ },
1469
+ {
1470
+ "type": "list",
1471
+ "sub_type": "ref_text",
1472
+ "list_items": [
1473
+ "[53] Rakesh Vaideeswaran, Feng Gao, ABHINAV MATHUR, and Govind Thattai. Towards reasoning-aware explainable vqa. In Workshop on Trustworthy and Socially Responsible Machine Learning, NeurIPS 2022.",
1474
+ "[54] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017.",
1475
+ "[55] Ramakrishna Vedantam, C Lawrence Zitnick, and Devi Parikh. Cider: Consensus-based image description evaluation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4566-4575, 2015.",
1476
+ "[56] Petar Velickovic, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Liò, and Yoshua Bengio. Graph attention networks. In ICLR, 2018.",
1477
+ "[57] Jialin Wu and Raymond Mooney. Self-critical reasoning for robust visual question answering. Advances in Neural Information Processing Systems, 32, 2019.",
1478
+ "[58] Jialin Wu and Raymond J Mooney. Faithful multimodal explanation for visual question answering. arXiv preprint arXiv:1809.02805, 2018.",
1479
+ "[59] Caiming Xiong, Stephen Merity, and Richard Socher. Dynamic memory networks for visual and textual question answering. pages 2397-2406, 2016.",
1480
+ "[60] Qian Yang, Yunxin Li, Baotian Hu, Lin Ma, Yuxin Ding, and Min Zhang. Chunk-aware alignment and lexical constraint for visual entailment with natural language explanations. In Proceedings of the 30th ACM International Conference on Multimedia, pages 3587-3597, 2022.",
1481
+ "[61] Ting Yao, Yingwei Pan, Yehao Li, and Tao Mei. Hierarchy parsing for image captioning. In IJCV, pages 2621-2629, 2019.",
1482
+ "[62] Yordan Yordanov, Vid Kocijan, Thomas Lukasiewicz, and Oana-Maria Camburu. Few-shot out-of-domain transfer learning of natural language explanations. arXiv preprint arXiv:2112.06204, 2021.",
1483
+ "[63] Xiaohua Zhai, Avital Oliver, Alexander Kolesnikov, and Lucas Beyer. S4l: Self-supervised semi-supervised learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1476-1485, 2019.",
1484
+ "[64] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Conditional prompt learning for vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16816-16825, 2022.",
1485
+ "[65] Luowei Zhou, Hamid Palangi, Lei Zhang, Houdong Hu, Jason Corso, and Jianfeng Gao. Unified vision-language pretraining for image captioning and vqa. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 13041-13049, 2020."
1486
+ ],
1487
+ "bbox": [
1488
+ 78,
1489
+ 90,
1490
+ 470,
1491
+ 799
1492
+ ],
1493
+ "page_idx": 10
1494
+ },
1495
+ {
1496
+ "type": "page_number",
1497
+ "text": "2656",
1498
+ "bbox": [
1499
+ 482,
1500
+ 945,
1501
+ 514,
1502
+ 955
1503
+ ],
1504
+ "page_idx": 10
1505
+ }
1506
+ ]
2023/S3C_ Semi-Supervised VQA Natural Language Explanation via Self-Critical Learning/6ead34e8-75d6-4d08-a4ab-1da1402619c2_model.json ADDED
The diff for this file is too large to render. See raw diff
 
2023/S3C_ Semi-Supervised VQA Natural Language Explanation via Self-Critical Learning/6ead34e8-75d6-4d08-a4ab-1da1402619c2_origin.pdf ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:4feb930de5f1f5810c4ed84e42b5370ea67e835f64c0bced3680625240924bbb
3
+ size 1390444
2023/S3C_ Semi-Supervised VQA Natural Language Explanation via Self-Critical Learning/full.md ADDED
@@ -0,0 +1,322 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # $\mathrm{S}^{3}\mathrm{C}$ : Semi-Supervised VQA Natural Language Explanation via Self-Critical Learning
2
+
3
+ Wei Suo $^{1*}$ , Mengyang Sun $^{2*}$ , Weisong Liu $^{1}$ , Yiqi Gao $^{1}$ , Peng Wang $^{1\dagger}$ , Yanning Zhang $^{1\dagger}$ , Qi Wu $^{3}$ $^{1}$ School of Computer Science and Ningbo Institute, Northwestern Polytechnical University, China.
4
+ $^{2}$ School of Cybersecurity, Northwestern Polytechnical University, China.
5
+ $^{3}$ University of Adelaide, Australia.
6
+
7
+ {suowei1994,sunmenmian,liuweisong,gyqjz}@mail.nwpu.edu.cn
8
+
9
+ {peng.wang, ynzhang}@nwpu.edu.cn, qi.wu01@adelaide.edu.au
10
+
11
+ # Abstract
12
+
13
+ VQA Natural Language Explanation (VQA-NLE) task aims to explain the decision-making process of VQA models in natural language. Unlike traditional attention or gradient analysis, free-text rationales can be easier to understand and gain users' trust. Existing methods mostly use post-hoc or self-rationalization models to obtain a plausible explanation. However, these frameworks are bottlenecked by the following challenges: 1) the reasoning process cannot be faithfully responded to and suffer from the problem of logical inconsistency. 2) Human-annotated explanations are expensive and time-consuming to collect. In this paper, we propose a new Semi-Supervised VQA-NLE via Self-Critical Learning $(S^3 C)$ , which evaluates the candidate explanations by answering rewards to improve the logical consistency between answers and rationales. With a semi-supervised learning framework, the $S^3 C$ can benefit from a tremendous amount of samples without human-annotated explanations. A large number of automatic measures and human evaluations all show the effectiveness of our method. Meanwhile, the framework achieves a new state-of-the-art performance on the two VQA-NLE datasets.
14
+
15
+ # 1. Introduction
16
+
17
+ Deep neural networks have enabled significant breakthroughs in a variety of vision-language (VL) tasks such as image captioning [10, 47] and visual question answering (VQA) [2, 39]. Unfortunately, most of them are black box systems, which makes it challenging to gain users' trust [20]. Explaining the decision-making process of deep VL models is a long-standing and essential problem.
18
+
19
+ ![](images/af310065e8d5f1bdc8df7d8c31b4609c45d4ba16f61dad695af4654d8d7b5216.jpg)
20
+
21
+ ![](images/c5759907e61c6aea5ef5031fae6ab82cc0a7fb498d64b875941375cdd035402c.jpg)
22
+
23
+ ![](images/c16a8892e5fba59ebcfaa7daacb1445d0ba698e7ddee1afd6e25253591b95f64.jpg)
24
+ Figure 1. Paradigm comparison of different VQA-NLE methods. (a) Post-hoc explanation method adopts two independent models to predict answers and explanations respectively. (b) Self-rationalization method uses a united VL model to simultaneously generate answers and explanations. (c) Our self-critical strategy utilizes answer scores as rewards and obtains more reliable rationales with semi-supervised learning.
25
+
26
+ Some approaches depend on attention mechanisms [2, 30] or gradient-based localization [50] to acquire visual explanations, which can highlight some contributing image regions for the predicted answers. However, simple visualization cannot explain how these areas support the answers and they are also hard to comprehend [20, 48]. Conversely, Natural Language Explanation (NLE) task [6, 38] can explain the decision-making process of a model by generating a natural language sentence. The language-based explanations are more accessible for users to understand, and they can also help researchers optimize the structure of models [34].
27
+
28
+ Recently, some models of NLE in the VL commu
29
+
30
+ nity have achieved pretty-well results, especially for VQANLE [20,34,41,48,58]. They can guide models to generate natural language sentences and interpret how the models get answers. Specifically, the first research line usually treats VQA-NLE as a predict-then-explain task [20, 34, 41, 58], namely post-hoc explanations method. As shown in Fig. 1 (a), these methods first depend on pre-trained VL models (such as UNITER [8] or Oscar [25]) to gain answers. Then the fused multi-modal features and the predicted answers are fed into a separated language model (e.g., LSTM [16] or Transformer [54]) to generate corresponding explanations. As shown in Fig. 1 (b), the other line [48] relies on a united VL model while generating both answers and explanations, which is known as the self-rationalization method. This framework can simultaneously predict an answer and generate a rationale by formulating the answer as a text-generation task along with the explanation.
31
+
32
+ Though significant progress has been made, the two paradigms are still restricted by the following challenges: 1) For the first paradigm, since the decision-making model and interpretation part are two separate modules, it would inevitably lead to unfaithful responses to the reasoning process of the decision models. 2) Due to the lack of explicitly logical relationship modeling, previous work [19] has proved that the straightforward self-rationalization frameworks suffer from the problem of logical inconsistency. 3) The above strategies all require an amount of human-annotated explanations, which are expensive and time-consuming to collect [62].
33
+
34
+ To solve the above challenges, inspired by [5, 51], we argue that a reasonable rationale can assist the model in obtaining a correct answer, and vice versa, the answer can be converted as an evaluation criterion for possible explanations. In this paper, we propose a new Semi-Supervised VQA-NLE method with Self-Critical learning, which is called $S^3 C$ for short. As shown in Fig.1 (c), given images and related questions, we first leverage a prompting mechanism to construct answer and explanation templates, which can guide the pre-trained VL model to generate answers and multiple candidate explanations based on sequence sampling [2]. Then we design a new self-critical method that converts the answer scores as rewards and encourages the model to generate the explanations which contribute to improving the answer scores. In particular, to reduce the dependency on expensive human annotations, we further extend our method to the semi-supervised version, which utilizes the unlabelled samples<sup>1</sup> (i.e., conventional VQA data [4, 36]) to significantly enhance the self-interpretability of the model. With the self-critical strategy and the semi-supervised learning, our method effectively models the logical relationships and promotes the logical
35
+
36
+ consistency between answer-explanation pairs. According to automatic measures and human evaluations, the $S^3 C$ outperforms the state-of-the-art models for the VQA-NLE task on the widely used two datasets and provides a new paradigm for our community. In summary, we make the following contributions:
37
+
38
+ 1) We propose a new self-critical VQA-NLE method that can model the logical relationships between answer-explanation pairs and evaluate the generated rationales by answering rewards. This strategy effectively improves the logical consistency and the reliability of the interpretations.
39
+ 2) We develop an advanced semi-supervised learning framework for VQA-NLE, which utilizes amounts of samples without human-annotated explanations to boost the self-interpretability of the model further. To the best of our knowledge, we are the first to explore semi-supervised learning on the VQA Natural Language Explanation.
40
+ 3) The proposed $S^3 C$ achieves new state-of-the-art performance on VQA-X [13] and A-OKVQA [49] benchmark datasets. Meanwhile, automatic measures and human evaluations all show the effectiveness of our method.
41
+
42
+ # 2. Related work
43
+
44
+ # 2.1. Explainability in Visual Question Answering
45
+
46
+ The visual question answering (VQA) is firstly proposed by [33] that requires an intelligent agent to generate an answer by giving an image and a question. Many approaches have been introduced such as joint embedding [13, 61], attention mechanisms [3, 31], memory networks [32, 59] and graph neural networks [23, 56]. Although the VQA task has been well studied, the reasoning process of the models is always agnostic. Some methods apply visualization technologies to achieve visual explanation, such as Grad-CAM [50] and U-CAM [42]. However, because image visualization cannot support the answer based on the attended areas [57], in this paper, we focus on improving free-text explanations that are more convenient and easier for users to understand. In this topic, the early work is proposed by [41]. The paper conducts the VQA-X dataset and utilizes human annotations to inspire the decision-making process of VQA models. [20] designs a new model that combines a pre-trained language model and a VL model to generate free-text explanations. [60] combines stronger pre-trained VL model (i.e. Oscar [25]) and generation model (i.e. GPT-2 [46]) to obtain better results. Recently, [48] proposes a unified model which can simultaneously predict answers and explanations based on a pre-trained caption model. Unlike previous methods, we introduce a new self-critical strategy to model the logical relationships between answers and rationales. It can encourage the model to enhance logical consistency and generate more reasonable explanations.
47
+
48
+ # 2.2. Pre-trained models and Prompt learning
49
+
50
+ Pre-trained models have been applied in many fields, such as various NLP tasks [12,46] and VL tasks [29,52,65]. Most of these pre-trained models utilize a stack of Transformer structures as the backbone. To generalize the pretrained models to other downstream tasks, previous works mostly fine-tune whole models on each downstream VL task. However, the ability of pre-trained models would be limited due to the mismatch between pre-trained tasks and downstream tasks. Hence, prompt learning [9, 18, 45, 64] is proposed, and it can keep the optimization consistency. [45] designs the templates to transfer the knowledge to downstream tasks. [9] uses a textual generation framework for uniform optimization. In this paper, we use the pre-trained model based on the image caption task as our backbone. It has been proved that these additional pre-train tasks (e.g., image feature regression or mask language tokens) cannot significantly improve the performance for explanations [48]. Meanwhile, we apply different language prompt templates to motivate the model to generate corresponding answers and rationales.
51
+
52
+ # 2.3. Semi-supervised learning
53
+
54
+ The development of deep learning depends on a large number of labelled data, while there are many cases in that only a small amount of data can be obtained [44]. To solve this challenge, Semi-supervised learning is proposed [44], which aims to train models using a small number of labelled data and amounts of unlabelled data. For example, [22] proposes a generatively semi-supervised framework based on variational autoencoders, and it can jointly optimize the model and variational parameters. Recently, [63] simultaneously leverages self-supervised and semi-supervised learning to address image classification tasks. To our knowledge, our work is the first semi-supervised learning framework for the VQA-NLE task, which effectively alleviates the reliance on expensive human annotations and further boost the self-interpretability of the model.
55
+
56
+ # 3. Method
57
+
58
+ In this section, we introduce our Semi-Supervised VQANLE method with the Self-Critical learning $(S^3 C)$ framework. Our aim is to strengthen the logical consistency between answer-explanation pairs and improve the reliability of the rationales. As shown in Fig. 2, the $S^3 C$ comprises an "Answer-Explanation Prompt" module and a "Self-Critical Reinforcement" module. Unlike previous approaches, our method first uses a prompting mechanism to generate answers and candidate explanations. Then, we design a new self-critical module that converts the answer scores as rewards to evaluate these reasons. Furthermore, this strategy can conveniently apply the unlabelled QA pairs to enrich
59
+
60
+ training data and enhance the self-interpretability. Next, we describe the components of our model in detail.
61
+
62
+ # 3.1. Pre-trained Vision-Language Backbone
63
+
64
+ Given an image $I \in \mathbb{R}^{W \times H \times 3}$ and a natural language question $Q = \{q_t\}_{t=1}^T$ , where $q_t$ represents the $t$ -th word, $T$ is the length of the question and $W \times H \times 3$ denotes the size of the image. Our goal is to predict the answer and generate a corresponding free-text rationale. Following previous works [48, 49], we adopt the CLIP vision encoder [45] and a pre-trained image caption model (i.e., ClipCap [37]) as our basic backbone. During pre-training, the VL model uses image embeddings from the CLIP as prefixes and fine-tunes a language model (i.e., GPT-2 [46]) to generate image captions. We refer readers to [37] for more information about the pre-trained model. In this paper, we consider the image $I$ and question $Q$ as the prefixes of the answer-explanation sequences. Specifically, we first apply ViT-B [14] and "classification token" from the CLIP to obtain the image features. Then, a group of light and simple Multi-Layer Perceptron (MLP) is utilized for transforming the image features to the $V = \{v_s\}_{s=1}^S$ , $v_s \in \mathbb{R}^c$ , where the dimension size $c = 768$ and the image sequence length $S = 10$ . The above computations can be formulated as follows:
65
+
66
+ $$
67
+ v _ {1}, v _ {2}, \dots , v _ {S} = \operatorname {M L P} (\operatorname {C L I P} (I)). \tag {1}
68
+ $$
69
+
70
+ Note that we only update the mapping network (i.e., MLP) during training, while the original visual encoder parameters from the CLIP would remain frozen. For question $Q$ , each word $q_{t}$ would be mapped to the corresponding word embedding $e_{t} \in \mathbb{R}^{c}$ by the pre-trained caption model. Finally, we obtain the image and question sequences $Z$ :
71
+
72
+ $$
73
+ Z = \overbrace {\left[ v _ {1} , v _ {2} , \cdots , v _ {S} \right.} ^ {\text {i m a g e}} \underbrace {e _ {1} , e _ {2} , \cdots , e _ {T}} _ {\substack {\text {q u e s t i o n} \\ \text {e m b d e d i n g}}} \bigg ], \tag{2}
74
+ $$
75
+
76
+ where $Z$ is the concatenated multi-modal prefixes.
77
+
78
+ # 3.2. Answer-Explanation Prompt Module
79
+
80
+ It has been proved that the prompting mechanism can maintain the same optimization objectives between pretrained and downstream tasks [18, 28]. Considering the convenience and explainability, we leverage hand-crafted prompts as templates to enable the model to generate answers or explanations. As shown in Fig. 2, the $S^3 C$ includes two different kinds of templates, which are introduced respectively as follows.
81
+
82
+ Base Answer Template. The core idea of our $S^3 C$ is to use the answer scores as evaluation criteria. Hence, we first establish a basic answer template to acquire base answer scores. Different from previous works [20, 41], our
83
+
84
+ ![](images/6c9a1ae1014099259c66fe6acd5d5ccb3871a086fe23b1181827247d177149a2.jpg)
85
+ Figure 2. Overview of our Semi-Supervised VQA-NLE method via Self-Critical learning $(S^3 C)$ framework. Given images and corresponding questions for labelled and unlabelled samples, we first use Answer-Explanation Prompt module to obtain the base answer scores and candidate explanations with a pre-trained VL model. Then these reasons are reorganized and fed back into the model to capture the explanatory answer score. Further, our Self-Critical Reinforcement module evaluates the generated explanations and returns the rewards to improve the self-interpretability of the model.
86
+
87
+ method treats the visual question answering task as a generation task so that the model can produce answers without a predefined answer space. Specifically, for a given image and question sequence $Z$ , we use natural language tokens "the answer is" to inspire the model to generate proper answers. By concatenating the language prompt, the base answer template $Z_{a} = [Z; \langle \text{answer} \rangle]$ can be obtained, where $[; ]$ and $\langle \text{answer} \rangle$ indicate concatenation operation and the specific language prompts, respectively. During training, the base answer template and ground-true answer label $A = \{a_{n}\}_{n=1}^{N}$ are fed into the VL model, where $a_{n}$ is the $n$ -th answer token and $N$ denotes the length of the answer label. Next, we compute the answer loss conditioned on the $Z_{a}$ in an autoregressive fashion:
88
+
89
+ $$
90
+ L _ {a} = - \frac {1}{N} \sum_ {n = 1} ^ {N} \log P \left(a _ {n} \mid Z _ {a}, a _ {1}, a _ {2}, \dots , a _ {n - 1}; \theta\right), \tag {3}
91
+ $$
92
+
93
+ where $\theta$ denotes the parameters of the VL model. Moreover, based on the indexes of ground-true answers, we can acquire average probability $p_{\theta}^{b}(A)$ as our base answer score.
94
+
95
+ Explanation Generation Template. To produce a reasonable rationale, we leverage the language prompt "the reason is" to motivate the model to generate free-text explanations. Like the base answer template, an explanation generation template $Z_{e} = [Z; \langle \text{reason} \rangle]$ is constructed, where $\langle \text{reason} \rangle$ is the natural language tokens. For labelled samples, we follow Eq. 3 in the autoregressive fashion and the cross-entropy loss to compute explanation loss $L_{e}$ . Because there are no human-annotated explanations available, we
96
+
97
+ would not compute any loss for unlabelled QA samples.
98
+
99
+ # 3.3. Self-Critical Reinforcement Module
100
+
101
+ To gain logically consistent rationales, in this module, we expand the searching space by introducing sequence sampling algorithm [2] and generate a set of candidate explanations. Besides, the answer scores are treated as rewards to encourage the model to output more detailed interpretations. It's worth noting that the above operations would be implemented on both labelled and unlabelled samples.
102
+
103
+ Candidate Explanation Generation. Benefiting from the language-based prompting strategy and the pre-trained VL model, our explanation generation template can easily guide the model to produce human-readable sentences. Hence, for unlabelled QA samples, we directly apply the explanation generation template $Z_{e}$ to generate the candidate rationales. To be more specific, we utilize beam search [2] to sample the top- $K$ words from the VL model probability distribution at each time step and maintain these sequences with the highest probability. Then, these generated sentences are integrated into candidate explanations $R_{s} = \{r_{k}^{s}\}_{k = 1}^{K}$ for each QA pair without human-annotated explanations, where $r_k^s$ and $K$ indicate $k$ -th rationale and the size of beam search, respectively. Moreover, for labelled samples, a similar mechanism is used to build the corresponding explanation set $R_{f} = \{r_{k}^{f}\}_{k = 1}^{K}$ . The above operations are implemented in all samples for the following reasons: 1) Larger search space. As shown in Fig. 2, "frying food is unhealthy" and "it includes many calories"
104
+
105
+ can both be rational explanations for the answer "no". The expanded search space provides more possibilities for generating reliable rationales. 2) Avoid overfitting. For labelled samples, although we can depend on human explanations to train the model, these labels are still one-sided and subjective. Using the sequence sampling strategy can prevent the model from overfitting these specific annotations [2,47]. In the end, because the labelled and unlabelled samples would be integrated into a mini-batch during training, we simplify $R_{f}$ and $R_{s}$ into $R = \{r_k\}_{k=1}^K$ to indicate the candidate explanations for each sample.
106
+
107
+ Self-Critical Reward. Inspired by [5, 51], an ideal rationale can help the model to infer the answer better. Based on this insight, we argue that the answer scores can be converted as self-critical rewards to evaluate these candidate explanations. Considering that this is a non-differentiable operation, we adopt reinforcement learning method [47] to achieve end-to-end training. In particular, given a sample and corresponding candidate explanations $R$ , we design a new input template:
108
+
109
+ $$
110
+ Z _ {r} ^ {k} = [ Z; \langle r e a s o n \rangle ; r _ {k}; \langle a n s w e r \rangle ], \tag {4}
111
+ $$
112
+
113
+ where $Z_{r}^{k}$ denotes the template of adding possible rationale $r_{k}$ . Then, this template is fed back into the model and obtains average probability $p_{\theta}^{r}(A)$ about the answer, namely explanatory answer scores, where $A$ is the ground-truth answer for the sample. Meanwhile, we use the average probability $p_{\theta}^{b}(A)$ from the output of the base answer template as the base scores. By applying reinforcement learning, the gradient is calculated by:
114
+
115
+ $$
116
+ \nabla_ {\theta} L _ {r} (\theta) = - \frac {1}{K} \sum_ {k = 1} ^ {K} \left(p _ {\theta} ^ {r} (A) - p _ {\theta} ^ {b} (A)\right) \nabla_ {\theta} \operatorname {l o g} p _ {\theta} \left(r _ {k}\right), \tag {5}
117
+ $$
118
+
119
+ where $p_{\theta}(r_k)$ is the probability of $k$ -th explanations. Based on the above computation, this gradient would tend to increase the probability of $k$ -th rationales when the answer score $p_{\theta}^{r}(A)$ higher than the scores $p_{\theta}^{b}(A)$ from the base answer template. Finally, for labelled samples, we append the human-annotated rationales to the candidate explanation $R$ and predict the answers by cross-entropy loss, namely $L_{ea}$ .
120
+
121
+ # 3.4. Loss
122
+
123
+ During the process of training, the overall loss function can be represented as follows:
124
+
125
+ $$
126
+ L = L _ {a} + L _ {e} + L _ {e a} + \lambda L _ {r}, \tag {6}
127
+ $$
128
+
129
+ where $\lambda$ is used for balancing these two different types of losses (i.e., cross-entropy loss and reinforcement loss). At inference time, our model would first generate the rationales about QA pairs, then we use the explanatory template (i.e., Eq. 4) to obtain the corresponding answers.
130
+
131
+ # 4. Experiment
132
+
133
+ # 4.1. Experimental setting
134
+
135
+ Datasets. Following previous methods [48, 49], we mainly carry out the experiments on the two different VQANLE datasets: VQA-X [41] and A-OKVQA [49]. Meanwhile, since explain annotations are expensive and time-consuming, we also utilize large-scale VQA v2.0 [15] and OK-VQA [36] datasets to build the semi-supervised learning paradigm. Next, we will introduce the four datasets.
136
+
137
+ VQA-X. It is a vision-language dataset that provides explanations for justifying the answers. VQA-X is collected from the Visual Question Answering (VQA) dataset [4] where the images are obtained from the MSCOCO [27]. It consists of 28K images and 33K QA pairs, split into 29K/1.4K/1.9K for training, validation and testing. Meanwhile, VQA-X constructs complementary pairs which provide a question and two semantically similar images with different answers.
138
+
139
+ A-OKVQA. Compared to the VQA-X, the questions of A-OKVQA generally are required commonsense reasoning about the scene described in the images. It includes 24,903 Question/Answer/Rationale triplets, split into $17.1\mathrm{K} / 1.1\mathrm{K} / 6.7\mathrm{K}$ for training, validation and testing. It collects images from the COCO 2017 [7] dataset, and is further filtered to obtain $23.7\mathrm{K}$ unique images. Compared with previous datasets, the A-OKVQA has richer questions and requires broader areas of knowledge for reasoning.
140
+
141
+ VQA & OK-VQA. We use VQA v2.0 and OK-VQA to provide large-scale unlabelled datasets for semi-supervised learning. The VQA v2.0 dataset is widely used for many previous works [2, 52]. It consists of 443k questions and 195k images. To select explainable questions instead of some obvious cases (e.g., How many...?, What color...?), we filter out these questions from VQA v2.0 and obtain the $\sim 90\mathrm{k}$ additional questions based on the rules [41]. On the other hand, we also use knowledge-based OK-VQA [36] dataset to provide unlabelled knowledge-based QA pairs. It contains a total of 14k questions on 14k images.
142
+
143
+ Implementation Details. Following [37], each image is first pre-processed (such as including image resize, center crop, and normalize) by CLIP [45]. Meanwhile, we fix the ViT-B weights from the CLIP visual encoder to accelerate the training speed. For the mapping network, the image sequence length $S = 10$ and the embedding size is 768. The AdamW [21] is used as our optimizer with the weight decay 1e-5, and the batch size and beam size $K$ are set to 4 and 2. The weight coefficient $\lambda$ is set to 10. We train all models on the four 1080Ti GPUs for 30 epochs with a learning rate of 1e-5.
144
+
145
+ # 4.2. Evaluation Measures
146
+
147
+ Automatic Metering. Following [48, 60], we use the auto
148
+
149
+ Table 1. Comparison with the state-of-the-art methods on the VQA-X. Note that these results are unfiltered scores. $S^3 C^*$ denotes the model without unlabelled samples.
150
+
151
+ <table><tr><td></td><td colspan="7">VQA-X</td></tr><tr><td>Approach</td><td>B4</td><td>M</td><td>R</td><td>S</td><td>C</td><td>Acc</td><td>Human</td></tr><tr><td>CAPS [41]</td><td>5.9</td><td>12.6</td><td>26.3</td><td>11.9</td><td>35.2</td><td>68.6</td><td>-</td></tr><tr><td>PJ-X [41]</td><td>19.5</td><td>18.2</td><td>43.4</td><td>15.1</td><td>71.3</td><td>76.4</td><td>65.4</td></tr><tr><td>FME [58]</td><td>24.4</td><td>19.5</td><td>47.7</td><td>17.9</td><td>88.8</td><td>75.5</td><td>-</td></tr><tr><td>NLX-GPT [48]</td><td>25.6</td><td>21.5</td><td>48.7</td><td>20.2</td><td>97.2</td><td>83.1</td><td>70.2</td></tr><tr><td>S3C*(ours)</td><td>26.5</td><td>22.0</td><td>49.0</td><td>20.9</td><td>100.5</td><td>83.7</td><td>73.9</td></tr><tr><td>S3C(ours)</td><td>27.8</td><td>22.8</td><td>50.7</td><td>21.5</td><td>104.4</td><td>85.6</td><td>77.4</td></tr></table>
152
+
153
+ Table 2. Comparison with the state-of-the-art methods on the A-OKVQA. Note that these results are unfiltered scores. $S^3 C^*$ denotes the model without unlabelled samples.
154
+
155
+ <table><tr><td></td><td colspan="8">AOKVQA</td></tr><tr><td>Approach</td><td>B4</td><td>M</td><td>R</td><td>S</td><td>C</td><td>Acc val</td><td>Acc test</td><td>Human</td></tr><tr><td>ViLBERT [29]</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>30.6</td><td>25.9</td><td>-</td></tr><tr><td>LXMERT [52]</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>30.7</td><td>25.9</td><td>-</td></tr><tr><td>KRISP [35]</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>33.7</td><td>27.1</td><td>-</td></tr><tr><td>Clipcap [49]</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>30.8</td><td>25.9</td><td>-</td></tr><tr><td>e-UG [20]</td><td>15.1</td><td>18.1</td><td>42.4</td><td>14.9</td><td>51.5</td><td>30.5</td><td>25.6</td><td>44.1</td></tr><tr><td>NLX-GPT [48]</td><td>20.1</td><td>17.0</td><td>46.3</td><td>15.8</td><td>65.4</td><td>32.7</td><td>28.7</td><td>46.9</td></tr><tr><td>\( S^2C^{*&#x27;} \) (ours)</td><td>21.8</td><td>17.9</td><td>47.3</td><td>17.3</td><td>70.6</td><td>33.0</td><td>29.6</td><td>49.4</td></tr><tr><td>\( S^3C \) (ours)</td><td>22.5</td><td>18.5</td><td>48.4</td><td>18.1</td><td>74.4</td><td>34.2</td><td>33.5</td><td>54.7</td></tr></table>
156
+
157
+ Table 3. Comparison with the state-of-the-art methods on the VQA-X. Note that these results are filtered scores. $S^3 C^*$ denotes the model without unlabelled samples.
158
+
159
+ <table><tr><td></td><td>B1</td><td>B2</td><td>B3</td><td>B4</td><td>M</td><td>R</td><td>S</td><td>C</td><td>Acc</td><td>Human</td></tr><tr><td>RVT [34]</td><td>51.9</td><td>37.0</td><td>25.6</td><td>17.4</td><td>19.2</td><td>42.1</td><td>15.8</td><td>52.5</td><td>68.6</td><td>60.5</td></tr><tr><td>PJ-X [41]</td><td>57.4</td><td>42.2</td><td>30.9</td><td>22.7</td><td>19.7</td><td>46.0</td><td>17.1</td><td>82.7</td><td>76.4</td><td>69.3</td></tr><tr><td>FME [58]</td><td>59.1</td><td>43.4</td><td>31.7</td><td>23.1</td><td>20.4</td><td>47.1</td><td>18.4</td><td>87.0</td><td>75.5</td><td>-</td></tr><tr><td>QA-only [20]</td><td>51.0</td><td>36.4</td><td>25.3</td><td>17.3</td><td>18.6</td><td>41.9</td><td>14.9</td><td>49.9</td><td>-</td><td>-</td></tr><tr><td>e-UG [20]</td><td>57.3</td><td>42.7</td><td>31.4</td><td>23.2</td><td>22.1</td><td>45.7</td><td>20.1</td><td>74.1</td><td>80.5</td><td>71.4</td></tr><tr><td>NLX-GPT [48]</td><td>64.2</td><td>49.5</td><td>37.6</td><td>28.5</td><td>23.1</td><td>51.5</td><td>22.1</td><td>110.6</td><td>83.2</td><td>73.7</td></tr><tr><td>S3C*(ours)</td><td>64.4</td><td>49.9</td><td>38.0</td><td>29.1</td><td>23.4</td><td>51.9</td><td>22.7</td><td>112.1</td><td>83.7</td><td>75.9</td></tr><tr><td>S3C(ours)</td><td>64.7</td><td>50.5</td><td>38.8</td><td>30.7</td><td>23.9</td><td>52.1</td><td>23.0</td><td>116.7</td><td>85.6</td><td>79.2</td></tr></table>
160
+
161
+ matic metrics BLEU [40],METEOR[11],ROUGE-L[26], SPICE[1] and CIDEr[55] to evaluate generated explanations.For the evaluation of predicted answers, we follow [48,60] to compute the VQA accuracy.
162
+
163
+ Human Evaluation. Automatic VQA-NLE measures do not always reflect the correctness and logicality of the explanations [20, 53], thus we also build human evaluations. The process is similar to [20, 34]. Specifically, for each explanation, three human evaluators are required to decide whether an explanation can justify the answer and select an option (including "yes, weak yes, weak no and no"). The selection will be mapped to scores $(1, \frac{2}{3}, \frac{1}{3}$ and 0). The final scores are computed by averaging among all test samples. Meanwhile, these evaluators are asked to choose the reasons for unqualified explanations. We follow [20] to define three kinds of aspects to evaluate the explanations: irrelevant explanations, insufficient explanations and meaningless explanations. First, the irrelevant explanations may not match the image, for example, "have a long neck" is a good explanation for the answer "giraffes" when asked "what animals are these?", but the image may display cows. Second, the insufficient explanations only describe the image but cannot corroborate the answers. For example, the sentence "there are some people" does not sufficiently justify the question "do people have a party?". Lastly, some nonsensical sentences could be judged as contradictory explanations, such as "a man is a woman". For each sample, the human evaluators can select multiple shortcomings. More details of the human evaluation can be found in the [20].
164
+
165
+ # 4.3. Quantitative evaluation.
166
+
167
+ Automatic Evaluation. We compare our method with the state-of-the-art models on the VQA-X and A-OKVQA datasets in Table 1-3. The B4, M, R, S, C, Acc and Human are short for BLEU-4, METEOR, ROUGE-L, SPICE, CIDEr, Answer precision and Human evaluation. We use "unfiltered" to indicate that the explanations are evaluated regardless of whether the answer is true or false. While "filtered" is to only consider the explanations which have correct answers. From Table 1, the $S^3 C^*$ denotes that we only apply the self-critical framework without using unlabelled samples, while the row $S^3 C$ indicates our method in the semi-supervised setting. We observe that the proposed framework outperforms both the post-hoc explanation methods [41, 58] and the self-rationalization method [48]. Meanwhile, it's worth noting that although the NLX-GPT uses a more powerful pre-trained VL model with larger pre-trained datasets (i.e., COCO captions [27], Flicker30k [43] and Visual Genome [24]), our method still obtains 3.3 absolute gains on the CIDEr indicator with fewer pre-trained data (only using COCO captions). Furthermore, when we utilize our proposed semi-supervised paradigm, the results are further improved by 7.2 points. These results show that our model can generate more reliable explanations and it can benefit from the amount of data without human-explanation labels. We also report the accuracy of answers in the column of "Acc". It can be observed that our self-critical method with semi-supervised learning can simultaneously boost the precision of answers and cor
168
+
169
+ Table 4. Main shortcomings. The main shortcomings of unqualified explanations on the VQA-X dataset. For each sample, human evaluators can select multiple shortcomings.
170
+
171
+ <table><tr><td>Model</td><td>Irrelevant explanations</td><td>Insufficient explanations</td><td>Meaningless explanations</td></tr><tr><td>RVT [34]</td><td>25.7%</td><td>33.5%</td><td>11.4%</td></tr><tr><td>PJ-X [41]</td><td>21.1%</td><td>28.4%</td><td>9.2%</td></tr><tr><td>e-UG [20]</td><td>22.8%</td><td>25.4%</td><td>8.7%</td></tr><tr><td>NLX-GPT [48]</td><td>20.3%</td><td>22.2%</td><td>9.1%</td></tr><tr><td>S3C (ours)</td><td>17.3%</td><td>18.9%</td><td>8.2%</td></tr></table>
172
+
173
+ Table 5. Cross-dataset testing. We alternately use the VQA-X and A-OKVQA as source dataset and target dataset to test the generalization of our framework.
174
+
175
+ <table><tr><td></td><td colspan="6">VQA-X→A-OKVQA</td></tr><tr><td>Approach</td><td>B4</td><td>M</td><td>R</td><td>S</td><td>C</td><td>Acc</td></tr><tr><td>NLX-GPT [48]</td><td>10.7</td><td>12.7</td><td>34.2</td><td>10.7</td><td>35.4</td><td>10.4</td></tr><tr><td>S3C(ours)</td><td>12.0</td><td>13.3</td><td>34.3</td><td>12.5</td><td>45.3</td><td>18.8</td></tr><tr><td></td><td colspan="6">A-OKVQA→VQA-X</td></tr><tr><td>Approach</td><td>B4</td><td>M</td><td>R</td><td>S</td><td>C</td><td>Acc</td></tr><tr><td>NLX-GPT [48]</td><td>9.1</td><td>13.6</td><td>32.8</td><td>9.1</td><td>33.2</td><td>42.4</td></tr><tr><td>S3C(ours)</td><td>10.9</td><td>15.0</td><td>34.1</td><td>10.4</td><td>38.6</td><td>43.8</td></tr></table>
176
+
177
+ Table 6. Ablation study. We ablate key components to demonstrate the effectiveness of our method. SCR and Semi are Self-Critical Reinforcement module and Semi-supervised learning paradigm respectively.
178
+
179
+ <table><tr><td></td><td>question</td><td>image</td><td>answer</td><td>explanation</td><td>SCR</td><td>Semi</td><td>B4</td><td>M</td><td>R</td><td>S</td><td>C</td><td>Acc</td></tr><tr><td>1</td><td>✓</td><td>✓</td><td>✓</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>80.1</td></tr><tr><td>2</td><td>✓</td><td>✓</td><td>-</td><td>✓</td><td>-</td><td>-</td><td>24.4</td><td>20.7</td><td>47.3</td><td>19.5</td><td>90.4</td><td>-</td></tr><tr><td>3</td><td>✓</td><td>✓</td><td>✓</td><td>✓</td><td>-</td><td>-</td><td>27.5</td><td>22.9</td><td>50.4</td><td>21.9</td><td>109.1</td><td>82.2</td></tr><tr><td>4</td><td>✓</td><td>✓</td><td>✓</td><td>✓</td><td>✓</td><td>-</td><td>29.1</td><td>23.4</td><td>51.9</td><td>22.7</td><td>112.1</td><td>83.7</td></tr><tr><td>5</td><td>✓</td><td>✓</td><td>✓</td><td>✓</td><td>✓</td><td>✓</td><td>30.7</td><td>23.9</td><td>52.1</td><td>23.0</td><td>116.7</td><td>85.6</td></tr></table>
180
+
181
+ responding explanations.
182
+
183
+ In Table 2, we evaluate our method on the A-OKVQA dataset. The results show that the $S^3 C$ can outperform all the previous works [17, 29, 35, 48, 49, 52]. Especially, compared to the SOTA model [48], there are 9.0 points and $4.8\%$ improvements in the CIDEr and answer accuracy based on our model. It demonstrates that our model can benefit from semi-supervised learning and generate explanations about commonsense reasoning.
184
+
185
+ In addition, to prove the algorithm's validity, we follow [20, 48] to report the filtered scores for the VQAX dataset in Table 3. Through filtering the correct answers, our method can outperform whatever post-hoc [20, 34, 41, 58] and self-rationalization [48] methods. Meanwhile, the proposed $S^3 C$ achieves a new state-of-the-art with the CIDEr score improved by 6.1 points to 116.7 and boost the answer accuracy by $2.4\%$ to 85.6.
186
+
187
+ Human Evaluation. To evaluate the faithfulness and correctness of these generated explanations, we conduct the human evaluation that is shown in the column of "Human" in Table 1-3. The experimental results further prove our method has better self-interpretability for the VQA-NLE task. Moreover, we also ask the human evaluators to select the shortcomings for each unqualified explanation on the VQA-X dataset. As shown in Table 4, we build three shortcoming options (i.e., irrelevant explanations, insufficient explanations and meaningless explanations) with a multi-choice process. These results indicate that the $S^3 C$ can obtain relatively better rationales and empirically confirm the effectiveness of our method.
188
+
189
+ # 4.4. Cross-dataset Performance.
190
+
191
+ We propose cross-dataset experiments to measure the generalization and reliability of our framework. For all we know, there is very little research on cross-dataset explanation generation. As shown in the first two rows of Table 5, we use the VQA-X and the A-OKVQA as the source dataset and target dataset to test the generalization ability of our framework. Specifically, we use the weights from the state-of-the-art method [48], which are trained on the VQA-X dataset and tested on the A-OKVQA dataset to obtain the results of the NLX-GPT. From the comparisons of six indicators, we find that our self-critical model has better generalization from the general VQA to the knowledge-based VQA. This suggests that $S^3 C$ can learn commonsense from visual question answering and generate better explanations and answers. Meanwhile, when we exchange the source dataset and target dataset as shown in the last two rows of Table 5, our framework can still surpass the existing SOTA model on all evaluation measures. These quantitative results effectively demonstrate the generalization and stability of our model.
192
+
193
+ # 4.5. Ablation Studies.
194
+
195
+ As shown in Table 6, we conduct several ablation studies on the VQA-X to demonstrate the effectiveness of our method. In particular, we first build the experiments in the 1-3 rows of Table 6, which use our Answer-Explanation Prompt module with cross-entropy loss as the baseline models. It can be found that when we simultaneously generate both answers and interpretations, the answer accuracy and CIDEr can achieve significant improvements. Meanwhile,
196
+
197
+ ![](images/7e01d8549bb6df38c2dafa143287e46634835e7efea7ff386f6c873aac75eb11.jpg)
198
+ (a) Is this man dressed formal?
199
+
200
+ ![](images/ca253b406d47da5ce9e8dd0643f14741dc101b45433029313006510d4e8be934.jpg)
201
+ Ours: [yes] <he is wearing a dress shirt and tie>
202
+ (b) Is he in a swimming pool?
203
+ NLX: [yes] <he is on a surfboard>
204
+ Ours: [no] <he is surfing in the ocean>
205
+
206
+ ![](images/39ae1e8ba51b4a97e658d08b9090d5b5fe7801fe50db87ef8fd67f5798d0bfac.jpg)
207
+ (c) What is the elephant doing?
208
+ NLX: [drinking] <it is in the water>
209
+ Ours: [drinking]<the trunk is in the water>
210
+
211
+ ![](images/c4f860296fc1546b20f04255e426a9cdbc3d82adb6d7857051de46cd6ac72882.jpg)
212
+ NLX: [no] <he is wearing a tie>
213
+ (d) Are the people having a party?
214
+ NLX: [no] <there are no people present>
215
+ Ours: [yes] <they are laughing and having a good time>
216
+ GT: [yes] <they are standing in a kitchen
217
+ with drinks in hand and laughing>
218
+ Figure 3. Some generation results on the VQA-X datasets. The [ ] and $<<$ indicate answers and explanations respectively. We show the results of the state-of-the-art NLX-GPT [48], our method and ground-truth (they are short for NLX, Ours and GT).
219
+
220
+ ![](images/245f9a6f2d0c2769e88489f6a21f838f96135dde6cd9756a73f082c8f5875f4d.jpg)
221
+ GT: [yes] <he's wearing a crisp dress shirt and a tie>
222
+ GT: [no] <the man is riding an ocean wave>
223
+ (e) Does this animal appear to be looking
224
+ at the camera?
225
+ NLX: [yes] <it is looking at something>
226
+ Ours: [no] <it is looking away from the camera>
227
+ GT: [no] <the animal is looking off to the side>
228
+
229
+ ![](images/7bec96df02b62f6838cf074149ae4a727fd068f7ab72ee06b34553bac058bac9.jpg)
230
+ GT: [drinking]<his trunk is in the water>
231
+ (f) Is this a bar?
232
+ NLX: [yes] <there are people sitting at tables>
233
+ Ours: [yes] <the men are sitting at a bar with drinks in front of them>
234
+ GT: [yes] <there are drinks on a long countertop with people eating and drinking>
235
+
236
+ these results also prove that the answer-explanation pairs have inherent consistency and they can promote each other. Further, we establish the Self-Critical Reinforcement module as shown in the fourth row, and the result of accuracy and CIDEr are boosted by $1.5\%$ and 3.0 points. It suggests that our self-critical method with sequence sampling strategy can encourage the model to generate more reliable explanations and correct answers. Finally, when we extend our model to the semi-supervised setting, the performance improves to 85.6 on the accuracy and 116.7 on the CIDEr over the baseline model (i.e., the row 5 of Table 6) by $3.4\%$ and 7.6 points. The results demonstrate that our framework is remarkably effective in improving the logical consistency and the self-interpretability of the model.
237
+
238
+ # 4.6. Qualitative evaluation.
239
+
240
+ As shown in Fig. 3, we show some qualitative results from NLX-GPT [48] and our $S^3 C$ method on the VQA-X dataset. Through overall comparison, our model achieves better logical consistency between answers and explanations. For example, in Fig. 3 (a), although the SOTA method [48] correctly identifies the significant symbol of formal dress (i.e., tie), the predicted answer is "no" that is contradictory to the explanation. On the contrary, our method not only obtains a more complete and faithful rationale but also generates a logically consistent answer. Additionally, the $S^3 C$ can also produce more persuasive and
241
+
242
+ trusty sentences. For instance, in Fig. 3 (b)-(f), the generated sentences contain scene description "having a good time" and fine-grained rationale "the trunk is in the water".
243
+
244
+ # 5. Conclusion
245
+
246
+ In this paper, we propose a new Semi-Supervised VQANLE method via Self-Critical Learning $(S^3 C)$ . Different from previous works, our method first utilizes the prompting mechanism to motivate the model to generate answers and candidate explanations. Meanwhile, we design a novel Self-Critical Reinforcement module, which converts the answer scores as rewards to evaluate these possible rationales. Furthermore, our framework can benefit from an abundance of question-answer pairs without human-annotated explanations and further boost the self-interpretability of the model. With the automatic measures and human evaluations, our $S^3 C$ achieves a new state-of-the-art on multiple benchmarks and provides a new paradigm for our community.
247
+
248
+ # Acknowledgements
249
+
250
+ This work was supported by National Key R&D Program of China (No.2020AAA0106900), the National Natural Science Foundation of China (No.U19B2037), Shaanxi Provincial Key RD Program (No.2021KWZ-03), and Natural Science Basic Research Program of Shaanxi (No.2021JCW-03).
251
+
252
+ # References
253
+
254
+ [1] Peter Anderson, Basura Fernando, Mark Johnson, and Stephen Gould. Spice: Semantic propositional image caption evaluation. In European conference on computer vision, pages 382-398. Springer, 2016.
255
+ [2] Peter Anderson, Xiaodong He, Chris Buehler, Damien Teney, Mark Johnson, Stephen Gould, and Lei Zhang. Bottom-up and top-down attention for image captioning and visual question answering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6077-6086, 2018.
256
+ [3] Peter Anderson, Xiaodong He, Chris Buehler, Damien Teney, Mark Johnson, Stephen Gould, and Lei Zhang. Bottom-up and top-down attention for image captioning and visual question answering. In CVPR, pages 6077-6086, 2018.
257
+ [4] Stanislaw Antol, Aishwarya Agrawal, Jiasen Lu, Margaret Mitchell, Dhruv Batra, C Lawrence Zitnick, and Devi Parikh. Vqa: Visual question answering. In Proceedings of the IEEE international conference on computer vision, pages 2425-2433, 2015.
258
+ [5] Ermanno Bencivenga. Free logics. In Handbook of philosophical logic, pages 147-196. Springer, 2002.
259
+ [6] Oana-Maria Camburu, Tim Rocktäschel, Thomas Lukasiewicz, and Phil Blunsom. e-snli: Natural language inference with natural language explanations. Advances in Neural Information Processing Systems, 31, 2018.
260
+ [7] Xinlei Chen, Hao Fang, Tsung-Yi Lin, Ramakrishna Vedantam, Saurabh Gupta, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco captions: Data collection and evaluation server. arXiv preprint arXiv:1504.00325, 2015.
261
+ [8] Yen-Chun Chen, Linjie Li, Licheng Yu, Ahmed El Kholy, Faisal Ahmed, Zhe Gan, Yu Cheng, and Jingjing Liu. Uniter: Universal image-text representation learning. In European conference on computer vision, pages 104-120. Springer, 2020.
262
+ [9] Jaemin Cho, Jie Lei, Hao Tan, and Mohit Bansal. Unifying vision-and-language tasks via text generation. In International Conference on Machine Learning, pages 1931-1942. PMLR, 2021.
263
+ [10] Marcella Cornia, Matteo Stefanini, Lorenzo Baraldi, and Rita Cucchiara. Meshed-memory transformer for image captioning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10578-10587, 2020.
264
+ [11] Michael Denkowski and Alon Lavie. Meteor universal: Language specific translation evaluation for any target language. In Proceedings of the ninth workshop on statistical machine translation, pages 376-380, 2014.
265
+ [12] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018.
266
+ [13] Jianfeng Dong, Xirong Li, and Cees G M Snoek. Predicting visual features from text for image and video caption retrieval. IEEE TMM, 20(12):3377-3388, 2018.
267
+
268
+ [14] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020.
269
+ [15] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Batra, and Devi Parikh. Making the v in vqa matter: Elevating the role of image understanding in visual question answering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6904-6913, 2017.
270
+ [16] Sepp Hochreiter and Jürgen Schmidhuber. Long short-term memory. Neural computation, 9(8):1735-1780, 1997.
271
+ [17] Yu Jiang, Vivek Natarajan, Xinlei Chen, Marcus Rohrbach, Dhruv Batra, and Devi Parikh. Pythia v0. 1: the winning entry to the vqa challenge 2018. arXiv preprint arXiv:1807.09956, 2018.
272
+ [18] Woojeong Jin, Yu Cheng, Yelong Shen, Weizhu Chen, and Xiang Ren. A good prompt is worth millions of parameters: Low-resource prompt-based learning for vision-language models. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 2763-2775, 2022.
273
+ [19] Jaehun Jung, Lianhui Qin, Sean Welleck, Faeze Brahman, Chandra Bhagavatula, Ronan Le Bras, and Yejin Choi. Maieutic prompting: Logically consistent reasoning with recursive explanations. arXiv preprint arXiv:2205.11822, 2022.
274
+ [20] Maxime Kayser, Oana-Maria Camburu, Leonard Salewski, Cornelius Emde, Virginie Do, Zeynep Akata, and Thomas Lukasiewicz. e-vil: A dataset and benchmark for natural language explanations in vision-language tasks. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1244-1254, 2021.
275
+ [21] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014.
276
+ [22] Durk P Kingma, Shakir Mohamed, Danilo Jimenez Rezende, and Max Welling. Semi-supervised learning with deep generative models. Advances in neural information processing systems, 27, 2014.
277
+ [23] Thomas N. Kipf and Max Welling. Semi-supervised classification with graph convolutional networks. In ICLR, 2017.
278
+ [24] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome: Connecting language and vision using crowdsourced dense image annotations. International journal of computer vision, 123(1):32-73, 2017.
279
+ [25] Xiujun Li, Xi Yin, Chunyuan Li, Pengchuan Zhang, Xiaowei Hu, Lei Zhang, Lijuan Wang, Houdong Hu, Li Dong, Furu Wei, et al. Oscar: Object-semantics aligned pre-training for vision-language tasks. In European Conference on Computer Vision, pages 121-137. Springer, 2020.
280
+ [26] Chin-Yew Lin. Rouge: A package for automatic evaluation of summaries. In Text summarization branches out, pages 74-81, 2004.
281
+
282
+ [27] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014.
283
+ [28] Yuhang Liu, Wei Wei, Daowan Peng, and Feida Zhu. Declaration-based prompt tuning for visual question answering. arXiv preprint arXiv:2205.02456, 2022.
284
+ [29] Jiasen Lu, Dhruv Batra, Devi Parikh, and Stefan Lee. Vilbert: Pretraining task-agnostic visiolinguistic representations for vision-and-language tasks. Advances in neural information processing systems, 32, 2019.
285
+ [30] Jiasen Lu, Jianwei Yang, Dhruv Batra, and Devi Parikh. Hierarchical question-image co-attention for visual question answering. Advances in neural information processing systems, 29, 2016.
286
+ [31] Jiasen Lu, Jianwei Yang, Dhruv Batra, and Devi Parikh. Hierarchical question-image co-attention for visual question answering. In NeurIPS, pages 289–297, 2016.
287
+ [32] Chao Ma, Chunhua Shen, Anthony R. Dick, Qi Wu, Peng Wang, Anton van den Hengel, and Ian D. Reid. Visual question answering with memory-augmented networks. In CVPR, pages 6975-6984, 2018.
288
+ [33] Mateusz Malinowski and Mario Fritz. A multi-world approach to question answering about real-world scenes based on uncertain input. Advances in neural information processing systems, 27, 2014.
289
+ [34] Ana Marasović, Chandra Bhagavatula, Jae Sung Park, Ronan Le Bras, Noah A Smith, and Yejin Choi. Natural language rationales with full-stack visual reasoning: From pixels to semantic frames to commonsense graphs. arXiv preprint arXiv:2010.07526, 2020.
290
+ [35] Kenneth Marino, Xinlei Chen, Devi Parikh, Abhinav Gupta, and Marcus Rohrbach. Krisp: Integrating implicit and symbolic knowledge for open-domain knowledge-based vqa. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14111-14121, 2021.
291
+ [36] Kenneth Marino, Mohammad Rastegari, Ali Farhadi, and Roozbeh Mottaghi. Ok-vqa: A visual question answering benchmark requiring external knowledge. In Proceedings of the IEEE/cvf conference on computer vision and pattern recognition, pages 3195-3204, 2019.
292
+ [37] Ron Mokady, Amir Hertz, and Amit H Bermano. Clipcap: Clip prefix for image captioning. arXiv preprint arXiv:2111.09734, 2021.
293
+ [38] Sharan Narang, Colin Raffel, Katherine Lee, Adam Roberts, Noah Fiedel, and Karishma Malkan. Wt5?! training text-to-text models to explain their predictions. arXiv preprint arXiv:2004.14546, 2020.
294
+ [39] Yingwei Pan, Ting Yao, Yehao Li, and Tao Mei. X-linear attention networks for image captioning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10971-10980, 2020.
295
+ [40] Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. Bleu: a method for automatic evaluation of machine translation. In Proceedings of the 40th annual meeting of the Association for Computational Linguistics, pages 311-318, 2002.
296
+
297
+ [41] Dong Huk Park, Lisa Anne Hendricks, Zeynep Akata, Anna Rohrbach, Bernt Schiele, Trevor Darrell, and Marcus Rohrbach. Multimodal explanations: Justifying decisions and pointing to the evidence. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8779-8788, 2018.
298
+ [42] Badri N Patro, Mayank Lunayach, Shivansh Patel, and Vinay P Namboodiri. U-cam: Visual explanation using uncertainty based class activation maps. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7444-7453, 2019.
299
+ [43] Bryan A Plummer, Liwei Wang, Chris M Cervantes, Juan C Caicedo, Julia Hockenmaier, and Svetlana Lazebnik. Flickr30k entities: Collecting region-to-phrase correspondences for richer image-to-sentence models. In Proceedings of the IEEE international conference on computer vision, pages 2641–2649, 2015.
300
+ [44] Guo-Jun Qi and Jiebo Luo. Small data challenges in big data era: A survey of recent progress on unsupervised and semi-supervised methods. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2020.
301
+ [45] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021.
302
+ [46] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9, 2019.
303
+ [47] Steven J Rennie, Etienne Marcheret, Youssef Mroueh, Jerret Ross, and Vaibhava Goel. Self-critical sequence training for image captioning. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7008-7024, 2017.
304
+ [48] Fawaz Sammani, Tanmoy Mukherjee, and Nikos Deligiannis. Nlx-gpt: A model for natural language explanations in vision and vision-language tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8322-8332, 2022.
305
+ [49] Dustin Schwenk, Apoorv Khandelwal, Christopher Clark, Kenneth Marino, and Roozbeh Mottaghi. A-okvqa: A benchmark for visual question answering using world knowledge. arXiv preprint arXiv:2206.01718, 2022.
306
+ [50] Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In Proceedings of the IEEE international conference on computer vision, pages 618-626, 2017.
307
+ [51] Suzanne Sia, Anton Belyy, Amjad Almahairi, Madian Khabsa, Luke Zettlemoyer, and Lambert Mathias. Logical satisfiability of counterfactuals for faithful explanations in nli. arXiv preprint arXiv:2205.12469, 2022.
308
+ [52] Hao Tan and Mohit Bansal. Lxmert: Learning cross-modality encoder representations from transformers. arXiv preprint arXiv:1908.07490, 2019.
309
+
310
+ [53] Rakesh Vaideeswaran, Feng Gao, ABHINAV MATHUR, and Govind Thattai. Towards reasoning-aware explainable vqa. In Workshop on Trustworthy and Socially Responsible Machine Learning, NeurIPS 2022.
311
+ [54] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017.
312
+ [55] Ramakrishna Vedantam, C Lawrence Zitnick, and Devi Parikh. Cider: Consensus-based image description evaluation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4566-4575, 2015.
313
+ [56] Petar Velickovic, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Liò, and Yoshua Bengio. Graph attention networks. In ICLR, 2018.
314
+ [57] Jialin Wu and Raymond Mooney. Self-critical reasoning for robust visual question answering. Advances in Neural Information Processing Systems, 32, 2019.
315
+ [58] Jialin Wu and Raymond J Mooney. Faithful multimodal explanation for visual question answering. arXiv preprint arXiv:1809.02805, 2018.
316
+ [59] Caiming Xiong, Stephen Merity, and Richard Socher. Dynamic memory networks for visual and textual question answering. pages 2397-2406, 2016.
317
+ [60] Qian Yang, Yunxin Li, Baotian Hu, Lin Ma, Yuxin Ding, and Min Zhang. Chunk-aware alignment and lexical constraint for visual entailment with natural language explanations. In Proceedings of the 30th ACM International Conference on Multimedia, pages 3587-3597, 2022.
318
+ [61] Ting Yao, Yingwei Pan, Yehao Li, and Tao Mei. Hierarchy parsing for image captioning. In IJCV, pages 2621-2629, 2019.
319
+ [62] Yordan Yordanov, Vid Kocijan, Thomas Lukasiewicz, and Oana-Maria Camburu. Few-shot out-of-domain transfer learning of natural language explanations. arXiv preprint arXiv:2112.06204, 2021.
320
+ [63] Xiaohua Zhai, Avital Oliver, Alexander Kolesnikov, and Lucas Beyer. S4l: Self-supervised semi-supervised learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1476-1485, 2019.
321
+ [64] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Conditional prompt learning for vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16816-16825, 2022.
322
+ [65] Luowei Zhou, Hamid Palangi, Lei Zhang, Houdong Hu, Jason Corso, and Jianfeng Gao. Unified vision-language pretraining for image captioning and vqa. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 13041-13049, 2020.
2023/S3C_ Semi-Supervised VQA Natural Language Explanation via Self-Critical Learning/images.zip ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:3f335d4d895a34fa726146d318628d48e91af5684c81551043765c1db7a38a32
3
+ size 488485
2023/S3C_ Semi-Supervised VQA Natural Language Explanation via Self-Critical Learning/layout.json ADDED
The diff for this file is too large to render. See raw diff
 
2023/SAP-DETR_ Bridging the Gap Between Salient Points and Queries-Based Transformer Detector for Fast Model Convergency/909bbbe4-4e26-472f-ab37-1d60d8bab4b6_content_list.json ADDED
@@ -0,0 +1,1378 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ [
2
+ {
3
+ "type": "text",
4
+ "text": "SAP-DETR: Bridging the Gap between Salient Points and Queries-Based Transformer Detector for Fast Model Convergence",
5
+ "text_level": 1,
6
+ "bbox": [
7
+ 114,
8
+ 128,
9
+ 852,
10
+ 176
11
+ ],
12
+ "page_idx": 0
13
+ },
14
+ {
15
+ "type": "text",
16
+ "text": "Yang Liu $^{1,3}$ * Yao Zhang $^{1,3}$ * Yixin Wang $^{2}$ * Yang Zhang $^{4}$ Jiang Tian $^{4}$ Zhongchao Shi $^{4}$ Jianping Fan $^{4}$ Zhiqiang He $^{1,3,5\\dagger}$",
17
+ "bbox": [
18
+ 184,
19
+ 200,
20
+ 782,
21
+ 239
22
+ ],
23
+ "page_idx": 0
24
+ },
25
+ {
26
+ "type": "text",
27
+ "text": "$^{1}$ Institute of Computing Technology (ICT), Chinese Academy of Sciences $^{2}$ Stanford University $^{3}$ University of Chinese Academy of Sciences $^{4}$ AI Lab, Lenovo Research $^{5}$ Lenovo Ltd.",
28
+ "bbox": [
29
+ 99,
30
+ 238,
31
+ 869,
32
+ 273
33
+ ],
34
+ "page_idx": 0
35
+ },
36
+ {
37
+ "type": "text",
38
+ "text": "{liuyang20c,zhangyao215}@mails.ucas.ac.cn yxinwang@stanford.edu",
39
+ "bbox": [
40
+ 192,
41
+ 276,
42
+ 764,
43
+ 292
44
+ ],
45
+ "page_idx": 0
46
+ },
47
+ {
48
+ "type": "text",
49
+ "text": "{zhangyang20,tianjiang1,shizc2,jfan1,hezq}@lenovo.com",
50
+ "bbox": [
51
+ 246,
52
+ 294,
53
+ 718,
54
+ 309
55
+ ],
56
+ "page_idx": 0
57
+ },
58
+ {
59
+ "type": "text",
60
+ "text": "Abstract",
61
+ "text_level": 1,
62
+ "bbox": [
63
+ 233,
64
+ 344,
65
+ 313,
66
+ 359
67
+ ],
68
+ "page_idx": 0
69
+ },
70
+ {
71
+ "type": "text",
72
+ "text": "Recently, the dominant DETR-based approaches apply central-concept spatial prior to accelerating Transformer detector convergence. These methods gradually refine the reference points to the center of target objects and imbue object queries with the updated central reference information for spatially conditional attention. However, centralizing reference points may severely deteriorate queries' saliency and confuse detectors due to the indiscriminative spatial prior. To bridge the gap between the reference points of salient queries and Transformer detectors, we propose SAlient Point-based DETR (SAP-DETR) by treating object detection as a transformation from salient points to instance objects. Concretely, we explicitly initialize a query-specific reference point for each object query, gradually aggregate them into an instance object, and then predict the distance from each side of the bounding box to these points. By rapidly attending to query-specific reference regions and the conditional box edges, SAP-DETR can effectively bridge the gap between the salient point and the query-based Transformer detector with a significant convergence speed. Experimentally, SAP-DETR achieves $1.4 \\times$ convergence speed with competitive performance and stably promotes the SoTA approaches by $\\sim 1.0$ AP. Based on ResNet-DC-101, SAP-DETR achieves 46.9 AP. The code will be released at https://github.com/liuyang-ict/SAP-DETR.",
73
+ "bbox": [
74
+ 73,
75
+ 376,
76
+ 473,
77
+ 753
78
+ ],
79
+ "page_idx": 0
80
+ },
81
+ {
82
+ "type": "text",
83
+ "text": "1. Introduction",
84
+ "text_level": 1,
85
+ "bbox": [
86
+ 76,
87
+ 782,
88
+ 209,
89
+ 797
90
+ ],
91
+ "page_idx": 0
92
+ },
93
+ {
94
+ "type": "text",
95
+ "text": "Object detection is a fundamental task in computer vision, whose target is to recognize and localize each object from input images. In the last decade, various detec",
96
+ "bbox": [
97
+ 75,
98
+ 808,
99
+ 470,
100
+ 853
101
+ ],
102
+ "page_idx": 0
103
+ },
104
+ {
105
+ "type": "image",
106
+ "img_path": "images/e0e7be2e248a69d58cb6ea61a192c0e5d95c0e8c9cff5ff9a8341202b046f68c.jpg",
107
+ "image_caption": [
108
+ "Figure 1. Comparison of SAP-DETR and DAB-DETR under 3-layer decoder model and 36-epoch training scheme. (a) Statistics of the query count in different classification score intervals. (b) and (c) Visualize the distribution of all reference points (pink) and highlight the top-20 classification score queries with their bounding boxes (blue) and reference points (red) in different decoder layers. (d) Visualize the outputs of positive queries and ground truth (red) during the training process, each query has a representative color for its reference point and bounding box. Best viewed in color."
109
+ ],
110
+ "image_footnote": [],
111
+ "bbox": [
112
+ 506,
113
+ 345,
114
+ 888,
115
+ 529
116
+ ],
117
+ "page_idx": 0
118
+ },
119
+ {
120
+ "type": "text",
121
+ "text": "tors [6,11,14,18,20,22] based on Convolutional Neural Networks (CNNs), have received widespread attention and made significant progress. Recently, Carion et al. [2] proposed a new end-to-end paradigm for object detection based on the Transformer [24], called DEtction TRansformer (DETR), which treats object detection as a problem of set prediction. In DETR, a set of learnable positional encodings, namely object queries, are employed to aggregate instance features from the context image in Transformer Decoder. The predictions of queries are finally assigned to the ground truth via bipartite matching to achieve end-to-end detection.",
122
+ "bbox": [
123
+ 496,
124
+ 686,
125
+ 893,
126
+ 853
127
+ ],
128
+ "page_idx": 0
129
+ },
130
+ {
131
+ "type": "text",
132
+ "text": "Despite the promising results of DETR, its application is largely limited by considerably longer training time compared to conventional CNNs. To address this problem, many",
133
+ "bbox": [
134
+ 496,
135
+ 854,
136
+ 893,
137
+ 902
138
+ ],
139
+ "page_idx": 0
140
+ },
141
+ {
142
+ "type": "header",
143
+ "text": "CVF",
144
+ "bbox": [
145
+ 106,
146
+ 2,
147
+ 181,
148
+ 42
149
+ ],
150
+ "page_idx": 0
151
+ },
152
+ {
153
+ "type": "header",
154
+ "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.",
155
+ "bbox": [
156
+ 236,
157
+ 0,
158
+ 810,
159
+ 46
160
+ ],
161
+ "page_idx": 0
162
+ },
163
+ {
164
+ "type": "page_footnote",
165
+ "text": "*This work was done when working as an intern at AI Lab, Lenovo Research, Beijing, China.",
166
+ "bbox": [
167
+ 75,
168
+ 862,
169
+ 468,
170
+ 887
171
+ ],
172
+ "page_idx": 0
173
+ },
174
+ {
175
+ "type": "page_footnote",
176
+ "text": "† Corresponding author.",
177
+ "bbox": [
178
+ 96,
179
+ 887,
180
+ 220,
181
+ 898
182
+ ],
183
+ "page_idx": 0
184
+ },
185
+ {
186
+ "type": "page_number",
187
+ "text": "15539",
188
+ "bbox": [
189
+ 480,
190
+ 944,
191
+ 517,
192
+ 955
193
+ ],
194
+ "page_idx": 0
195
+ },
196
+ {
197
+ "type": "text",
198
+ "text": "variants attempted to take a close look at query paradigm and introduced various spatial priors for model convergence and efficacy. According to the type of spatial prior, they can be categorized into implicit and explicit methods. The implicit ones [5, 16, 31] attempt to decouple a reference point from the object query and make use of this spatial prior to attend to the image context features efficiently. The current state-of-the-arts (SoTAs) are dominated by the explicit ones [13, 25], which suggest to instantiate a position with spatial prior for each query, i.e., explicit reference coordinates with a center point or an anchor box. These reference coordinates serve as helpful priors and enable the queries to focus on their expected regions easily. For instance, Anchor DETR [25] introduced an anchor concept (center point with different box size patterns) to formulate the query position and directly regressed the central offsets of the bounding boxes. DABDETR [13] further stretched the center point to a 4D anchor box concept $[cx, cy, w, h]$ to refine proposal bonding boxes in a cascaded manner. However, instantiating the query location as a target center may severely degrade the classification accuracy and convergence speed. As illustrated in Fig. 1, there exist many plausible queries [19] with high-quality classification scores (Fig. 1(a) within red box) and box Intersection over Union (IoU, see the redundant blue boxes in Fig. 1(b) and (c)), which only brings a slight improvement on precision rate but inevitably confuses the detector on the positive query assignments when training with bipartite matching strategy. This is because the plausible predictions are considered in negative classification loss, which severely decelerates the model convergence. As shown in Fig. 1(b) and (c), the predefined reference point of the positive query may not be the nearest one to the center of the ground truth bounding box, and the reference points tend to be centralized or marginalized (cyan arrows in Fig. 1(b)), hence losing the spatial specificity. With further insight into the one-to-one label assignment during the training process, we find that the query, whose reference point is closest to the center point, also has a high-quality IoU, but it still exists a disparity with the positive query in the classification confidence. Therefore, we argue that such a centralized spatial prior may cause degeneration of target consistency in both classification and localization tasks, which leads to inconsistent predictions.",
199
+ "bbox": [
200
+ 76,
201
+ 90,
202
+ 472,
203
+ 726
204
+ ],
205
+ "page_idx": 1
206
+ },
207
+ {
208
+ "type": "text",
209
+ "text": "Furthermore, the mentioned central point-based variants also have difficulties in detecting occluded objects. For example, Fig. 1(d) shows that DAB-DETR detects the left baseman twice, while the query point in SAP-DETR is not necessarily the center point, so the query point of the bounding box for the occluded baseman is from a pixel on the occluded baseman on the top right area instead of from the left baseman. One solution for center-based method [25] is to predefine different receptive fields (similar to the scaling anchor box in YOLO [17]) for the position of each query. However, increasing the diversity of the receptive fields for",
210
+ "bbox": [
211
+ 76,
212
+ 734,
213
+ 472,
214
+ 901
215
+ ],
216
+ "page_idx": 1
217
+ },
218
+ {
219
+ "type": "text",
220
+ "text": "each position query is unsuitable for non-overlapped targets, as it still generates massive indistinguishable predictions for one position as same as other center-based models.",
221
+ "bbox": [
222
+ 496,
223
+ 90,
224
+ 893,
225
+ 136
226
+ ],
227
+ "page_idx": 1
228
+ },
229
+ {
230
+ "type": "text",
231
+ "text": "To bridge these gaps, in this paper, we present a novel framework for Transformer detector, called SAclient Point-based DETR (SAP-DETR), which treats object detection as a transformation from salient points to instance objects. Instead of regressing the reference point to the target center, we define the reference point belonging to one positive query as a salient point, keep this query-specific spatial prior with a scaling amplitude, and then gradually update them to an instance object by predicting the distance from each side of the bounding box. Specifically, we tile the mesh-grid referenced points and initialize their center/corner as the query-specific reference point. To disentangle the reference sparsity as well as stabilize the training process, a movable strategy with scaling amplitude is applied for reference point adjustment, which prompts queries to consider their reference grid as the salient region to perform image context attention. By localizing each side of the bounding box layer by layer, such query-specific spatial prior enables compensation for the over-smooth/inadequacy problem during center-based detection, thereby vastly promoting model convergence speed. Inspired by [5, 13, 16], we also take advantage of both Gaussian spatial prior and conditional cross-attention mechanism, and then a salient point enhanced cross-attention mechanism is developed to distinguish the salient region and other conditional extreme regions from the context image features.",
232
+ "bbox": [
233
+ 496,
234
+ 138,
235
+ 895,
236
+ 516
237
+ ],
238
+ "page_idx": 1
239
+ },
240
+ {
241
+ "type": "text",
242
+ "text": "We bridge the gap between salient points and query-based Transformer detector by speedily attending to the query-specific region and other conditional regions. The extensive experiments have shown that SAP-DETR achieves superior convergence speed and performance. To the best of our knowledge, this is the first work to introduce the salient point based regression into end-to-end query-based Transformer detectors. Our contributions can be summarized as follows.",
243
+ "bbox": [
244
+ 496,
245
+ 517,
246
+ 893,
247
+ 638
248
+ ],
249
+ "page_idx": 1
250
+ },
251
+ {
252
+ "type": "list",
253
+ "sub_type": "text",
254
+ "list_items": [
255
+ "1) We introduce the salient point concept into query-based Transformer detectors by assigning query-specific reference points to object queries. Unlike center-based methods, we restrict the reference location and define the point of the positive query as the salient one, hence enlarging the discrepancy of query as well as reducing the redundant predictions (see Fig. 1). Thanks to the efficacy of the query-specific prior, our SAP-DETR accelerates the convergence speed greatly, achieving competitive performance with $30\\%$ fewer training epochs. The proposed movable strategy further boosts SAP-DETR to a new SoTA performance.",
256
+ "2) We devise a point-enhanced cross-attention mechanism to imbue query with spatial prior based on both reference point and box sides for final specific region attention.",
257
+ "3) Evaluation over COCO dataset has demonstrated that SAP-DETR achieves superior convergence speed and detection accuracy. Under the same training settings, SAP-DETR"
258
+ ],
259
+ "bbox": [
260
+ 496,
261
+ 640,
262
+ 895,
263
+ 900
264
+ ],
265
+ "page_idx": 1
266
+ },
267
+ {
268
+ "type": "page_number",
269
+ "text": "15540",
270
+ "bbox": [
271
+ 480,
272
+ 944,
273
+ 519,
274
+ 955
275
+ ],
276
+ "page_idx": 1
277
+ },
278
+ {
279
+ "type": "text",
280
+ "text": "outperforms the SoTA approaches with a large margin.",
281
+ "bbox": [
282
+ 76,
283
+ 90,
284
+ 439,
285
+ 107
286
+ ],
287
+ "page_idx": 2
288
+ },
289
+ {
290
+ "type": "text",
291
+ "text": "2. Related Work",
292
+ "text_level": 1,
293
+ "bbox": [
294
+ 76,
295
+ 119,
296
+ 218,
297
+ 135
298
+ ],
299
+ "page_idx": 2
300
+ },
301
+ {
302
+ "type": "text",
303
+ "text": "Anchor-Free Object Detectors. Classical anchor-free object detectors can be grouped into center-based and keypoint-based approaches. The center-based approaches aim to localize the target objects based on the central locations [11] or predefined ROI [22]. For example, FCOS [22] treated all points within the bounding box as positive ones to predict their distances from each side $([\\ell, t, r, b])$ , and a centerness score was then considered to prohibit the low-quality prediction whose point is located near the border. Compared with FCOS, we also restrict the candidate queries within the bounding box but treat only one as positive to perform end-to-end object detection via an inner matching cost.",
304
+ "bbox": [
305
+ 75,
306
+ 145,
307
+ 470,
308
+ 325
309
+ ],
310
+ "page_idx": 2
311
+ },
312
+ {
313
+ "type": "text",
314
+ "text": "The target of keypoint-based approaches is to localize the specific object locations and assign them to the predefined keypoints of the object for box localized training. For instance, diagonal corner points were considered in CornerNet [8], center point was further grouped into CenterNet [29], and ExtremeNet [30] added some conjectural extreme points for object localization. These works showed an impressive performance, but the complicated keypoint matching may limit their upper bound. Our SAP-DETR takes the advantage of salient point regression to focus on the distinct regions without complicated point-based supervision.",
315
+ "bbox": [
316
+ 75,
317
+ 327,
318
+ 468,
319
+ 492
320
+ ],
321
+ "page_idx": 2
322
+ },
323
+ {
324
+ "type": "text",
325
+ "text": "Query-Based Transformer Detectors. DETR [2] pioneered a new paradigm of Transformer detector for end-to-end object detection without any post-processing [1]. In DETR, a new representation, namely object query, aggregates the instance features and then yields a detection result for each instance object [15]. Following DETR, many voterists put efforts on the optimization of convergence and accuracy.",
326
+ "bbox": [
327
+ 75,
328
+ 492,
329
+ 468,
330
+ 598
331
+ ],
332
+ "page_idx": 2
333
+ },
334
+ {
335
+ "type": "text",
336
+ "text": "Sun et al. [21] revealed that the main reason for slow convergence of DETR is attributed to the Transformer decoder, and they considered an encoder-only structure to alleviate such a problem. For in-depth understanding of the object query, one way is to generate a series of implicit spatial priors from queries to guide feature aggregation in cross-attention layers. SMCA [5] applied a Gaussian-like attention map to augment the query-concerned features spatially. The reference point concept was first introduced by Deformable DETR [31], where the sampling offsets are predicted by each reference point to perform deformable cross-attention. Following such a concept, Conditional DETR [16] reformulated the attention operation and rebuilt positional queries based on the reference points to facilitate extreme region discrimination. Another way is towards position-instantiation explicitly, where this position information enables to directly conduct positional query generation. Anchor DETR [25] utilized a predefined 2D anchor point $[cx, cy]$ to explicitly capitalize on the spatial prior during cross-attention and box regression. DAB-DETR [13] extended such a 2D concept to",
337
+ "bbox": [
338
+ 75,
339
+ 599,
340
+ 470,
341
+ 901
342
+ ],
343
+ "page_idx": 2
344
+ },
345
+ {
346
+ "type": "image",
347
+ "img_path": "images/610de358f6711972cd0b18bdd81eb531a04351f72a5b205d48cb34f2440f488e.jpg",
348
+ "image_caption": [
349
+ "Figure 2. Illustration of SAP-DETR. Each object query in SAP-DETR is assigned to a specific grid region and initialized by the corner/center of the grid as its reference point. A learnable 4D coordinate represents the distance from the four sides of the box to the reference point. Both reference points and box sides are served as positional encodings added/concatenated to content embeddings. All embeddings are refined to predict target objects gradually."
350
+ ],
351
+ "image_footnote": [],
352
+ "bbox": [
353
+ 531,
354
+ 87,
355
+ 864,
356
+ 369
357
+ ],
358
+ "page_idx": 2
359
+ },
360
+ {
361
+ "type": "text",
362
+ "text": "a 4D anchor box $[cx, cy, w, h]$ and refined it layer-by-layer.",
363
+ "bbox": [
364
+ 500,
365
+ 494,
366
+ 888,
367
+ 511
368
+ ],
369
+ "page_idx": 2
370
+ },
371
+ {
372
+ "type": "text",
373
+ "text": "The recent accelerating convergence methods are based on auxiliary queries for facilitating detector discrimination. DN-DETR [9] demonstrates the slow model convergence is mainly caused by the instability of bipartite matching, thus providing a denoising training to eliminate this issue. DINO [28] inherits this advance and further introduces negative queries to perform contrastive denoising. Group-DETR [4] proposes a group-wise one-to-many label assignment to match multiple positive object queries with more gradients for fast DETR convergence.",
374
+ "bbox": [
375
+ 496,
376
+ 511,
377
+ 893,
378
+ 662
379
+ ],
380
+ "page_idx": 2
381
+ },
382
+ {
383
+ "type": "text",
384
+ "text": "The most relevant approaches to ours are Point-DETR [3] and SAM-DETR [26, 27]. The former applied a point encoder for annotated point label infusion in teacher model, and the latter directly updated content embeddings by extracting salient points from image features for query-image semantic alignment. Unlike these concepts, we redefine the salient point from the perspective of the positive query's position and replace the center-concept prior with the query-specific position, thereby attending extreme regions, differentiating queries' saliency, and alleviating redundant predictions.",
385
+ "bbox": [
386
+ 496,
387
+ 662,
388
+ 893,
389
+ 814
390
+ ],
391
+ "page_idx": 2
392
+ },
393
+ {
394
+ "type": "text",
395
+ "text": "3. Method",
396
+ "text_level": 1,
397
+ "bbox": [
398
+ 500,
399
+ 829,
400
+ 589,
401
+ 844
402
+ ],
403
+ "page_idx": 2
404
+ },
405
+ {
406
+ "type": "text",
407
+ "text": "We propose SAP-DETR to bridge the gap between salient points and query-based detectors. Following DETR, the extracted image features are fed into Transformer encoder",
408
+ "bbox": [
409
+ 498,
410
+ 854,
411
+ 893,
412
+ 900
413
+ ],
414
+ "page_idx": 2
415
+ },
416
+ {
417
+ "type": "page_number",
418
+ "text": "15541",
419
+ "bbox": [
420
+ 480,
421
+ 944,
422
+ 517,
423
+ 955
424
+ ],
425
+ "page_idx": 2
426
+ },
427
+ {
428
+ "type": "text",
429
+ "text": "after adding positional encodings, and then re-aggregated by object queries in Transformer decoder. In pursuit of the query-specific prior, we dispense a movable strategy for each query based on a fixed grid region. The query, whose reference region overlaps with ground truth objects, is allowed to predict the relative offsets from four sides of the bounding box to the points. Given the query-specific reference point and the proposal box sides, we propose salient point enhanced cross-attention mechanism to imbue query with spatial prior, thereby attending to extreme regions effectively. Additionally, we discuss two common issues in DETR-like models and address them for further improvements.",
430
+ "bbox": [
431
+ 75,
432
+ 90,
433
+ 472,
434
+ 272
435
+ ],
436
+ "page_idx": 3
437
+ },
438
+ {
439
+ "type": "text",
440
+ "text": "3.1. Salient Points-Based Object Detection",
441
+ "text_level": 1,
442
+ "bbox": [
443
+ 76,
444
+ 281,
445
+ 405,
446
+ 296
447
+ ],
448
+ "page_idx": 3
449
+ },
450
+ {
451
+ "type": "text",
452
+ "text": "Overview. Previous methods [13, 16, 25] normally decompose the object query into both content and position embeddings (queries), and form a center-based anchor point/box prior on the position ones. Unlike the central concept, we tile a fixed mesh-grid region, initialize their left-top corner as the reference points with 2D coordinate $\\boldsymbol{r} = \\{x, y\\} \\in [0,1]^2$ , and instantiate a learnable 4D offset distance $s = \\{\\ell, t, r, b\\} \\in [0,1]^4$ from the reference point to the sides of proposal bounding box for each object query. The object query can be referred as $q = \\{\\mathbf{e}; r, s\\}$ , where $\\mathbf{e} \\in \\mathbb{R}^d$ is the content embedding with $d$ dimension. Instead of regressing the center, width, and height of a bounding box, we follow FCOS [22] and directly supervise the 4D offset from the four sides of a bounding box to the reference point. The final box prediction is formulated as $\\hat{\\boldsymbol{b}} = \\{\\hat{x} - \\hat{\\ell}, \\hat{y} - \\hat{\\bar{t}}, \\hat{x} + \\hat{r}, \\hat{y} + \\hat{b}\\}$ . Ideally, we here fix the reference point $\\{\\hat{x}, \\hat{y}\\} = \\{x, y\\}$ (the movable update strategy is introduced in the next subsection) and only update the 4D box side prediction layer by layer. The prediction for each decoder layer can be calculated by",
453
+ "bbox": [
454
+ 75,
455
+ 303,
456
+ 473,
457
+ 592
458
+ ],
459
+ "page_idx": 3
460
+ },
461
+ {
462
+ "type": "equation",
463
+ "text": "\n$$\n\\Delta \\boldsymbol {s} _ {l} = \\operatorname {B o x H e a d} _ {l} (\\boldsymbol {s} _ {l - 1}, \\boldsymbol {e} _ {l - 1}, \\boldsymbol {r} _ {l - 1}),\n$$\n",
464
+ "text_format": "latex",
465
+ "bbox": [
466
+ 153,
467
+ 602,
468
+ 393,
469
+ 618
470
+ ],
471
+ "page_idx": 3
472
+ },
473
+ {
474
+ "type": "equation",
475
+ "text": "\n$$\n\\hat {\\boldsymbol {s}} _ {l} = \\sigma \\left(\\sigma^ {- 1} \\left(\\boldsymbol {s} _ {l - 1}\\right) + \\Delta \\boldsymbol {s} _ {l}\\right), \\quad \\boldsymbol {s} _ {l} = \\operatorname {D e t a c h} (\\hat {\\boldsymbol {s}} _ {l}),\n$$\n",
476
+ "text_format": "latex",
477
+ "bbox": [
478
+ 117,
479
+ 621,
480
+ 429,
481
+ 638
482
+ ],
483
+ "page_idx": 3
484
+ },
485
+ {
486
+ "type": "equation",
487
+ "text": "\n$$\n\\boldsymbol {r} _ {l} = \\hat {\\boldsymbol {r}} _ {l} = \\boldsymbol {r} _ {l - 1}, \\quad \\hat {\\boldsymbol {b}} _ {l} = \\left\\{\\hat {\\boldsymbol {r}} _ {l} - \\hat {\\boldsymbol {s}} _ {l} [: 2 ], \\hat {\\boldsymbol {r}} _ {l} + \\hat {\\boldsymbol {s}} _ {l} [2: ] \\right\\}, \\tag {1}\n$$\n",
488
+ "text_format": "latex",
489
+ "bbox": [
490
+ 102,
491
+ 643,
492
+ 468,
493
+ 672
494
+ ],
495
+ "page_idx": 3
496
+ },
497
+ {
498
+ "type": "text",
499
+ "text": "where $\\sigma$ and $\\sigma^{-1}$ are the sigmoid and inverse sigmoid operation, respectively. $\\Delta s_{l}$ denotes the side offset prediction. $\\hat{s}_l$ , $\\hat{r}_l$ , and $\\hat{\\pmb{b}}_l$ are the predicted side distance, reference points, and box location from the $l$ decoder layer, respectively. The BoxHead is the prediction head following the layer- $l$ decoder, which is independent between different decoder layers in our settings. Detach operation follows DAB-DETR [13].",
500
+ "bbox": [
501
+ 75,
502
+ 674,
503
+ 470,
504
+ 779
505
+ ],
506
+ "page_idx": 3
507
+ },
508
+ {
509
+ "type": "text",
510
+ "text": "During the training process, each query is only allowed to predict the bounding boxes that overlap its reference region. We adapt this rule into the one-to-one bipartite matching process via an inner matching cost $\\mathcal{L}_{\\mathrm{inner}}$ . Given $N$ queries $\\mathbf{Q} = \\{\\pmb{q}_1,\\pmb{q}_2,\\dots ,\\pmb{q}_N\\}$ and $M$ ground truth objects $\\mathbf{G} = \\{\\pmb{g}_1,\\pmb{g}_2,\\dots ,\\pmb{g}_M\\}$ , the $\\mathcal{L}_{\\mathrm{inner}}(\\pmb{g}_i,\\pmb{q}_j)$ of each query-box pair is a step function to penalize the reference point $\\pmb{r}_j$ of $\\pmb{q}_j$ with value of $k$ when $\\pmb{r}_j$ is outside the bounding box of",
511
+ "bbox": [
512
+ 75,
513
+ 780,
514
+ 472,
515
+ 901
516
+ ],
517
+ "page_idx": 3
518
+ },
519
+ {
520
+ "type": "text",
521
+ "text": "$g_{i}$ . We denote $i \\in [1, M]$ and $j \\in [1, N]$ as the index of query and ground truth, respectively. $k$ can be viewed as a penalty cost, and default to $10^{5}$ . The final permutation of the one-to-one label assignment is formulated as",
522
+ "bbox": [
523
+ 496,
524
+ 90,
525
+ 893,
526
+ 151
527
+ ],
528
+ "page_idx": 3
529
+ },
530
+ {
531
+ "type": "equation",
532
+ "text": "\n$$\n\\mathcal {L} _ {\\text {i n n e r}} \\left(\\boldsymbol {g} _ {i}, \\boldsymbol {q} _ {j}\\right) := \\mathbf {k} _ {\\boldsymbol {r} _ {j} \\notin \\boldsymbol {g} _ {i}},\n$$\n",
533
+ "text_format": "latex",
534
+ "bbox": [
535
+ 616,
536
+ 162,
537
+ 779,
538
+ 179
539
+ ],
540
+ "page_idx": 3
541
+ },
542
+ {
543
+ "type": "equation",
544
+ "text": "\n$$\n\\hat {\\eta} = \\underset {\\eta \\in \\mathfrak {Y} _ {N}} {\\operatorname {a r g m i n}} \\sum_ {i} ^ {N} \\mathcal {L} _ {\\text {m a t c h}} + \\mathcal {L} _ {\\text {i n n e r}}, \\tag {2}\n$$\n",
545
+ "text_format": "latex",
546
+ "bbox": [
547
+ 594,
548
+ 181,
549
+ 890,
550
+ 222
551
+ ],
552
+ "page_idx": 3
553
+ },
554
+ {
555
+ "type": "text",
556
+ "text": "where $\\mathcal{L}_{\\mathrm{match}}$ is the original pair-wise matching cost consist of both classification and localization costs [2]. $\\eta \\in \\mathfrak{V}_N$ is a permutation of $N$ elements for bipartite matching.",
557
+ "bbox": [
558
+ 496,
559
+ 231,
560
+ 890,
561
+ 276
562
+ ],
563
+ "page_idx": 3
564
+ },
565
+ {
566
+ "type": "text",
567
+ "text": "Movable Reference Point. Due to the sparseness of the fixed reference point, some small and slender objects may be indistinguishable when there is no reference point inside these objects. Despite the bipartite matching forcing each object to be assigned to one object query, the positive query, whose reference point is outside the assigned bounding box, is unable to accurately regress the distance from each side by a value between 0 and 1. One straightforward solution is to adjust the locations of reference points inside the ground truth bounding boxes to ensure that each object can be detected by an inner reference point. Similar to the aforementioned box refinement, we first perform a movable reference point design to dynamically update the reference points of each query layer by layer. However, such a full-image point regression inevitably expands the search space as vast variable determinations, causing the final reference point to be trapped in an unexpected corner of the bounding box. To reduce the training search spaces, we scale the offset amplitude of points within their specific grid regions, as illustrated in Fig. 3. Such an operation limits the range of offset values, and hence prevents a large searching space. It is implemented by applying the sigmoid activation $\\sigma$ and multiplying a scale factor $s_{\\mathrm{grid}}$ whose value equals to the height and width of one grid. The update process of the reference points is formulated as",
568
+ "bbox": [
569
+ 496,
570
+ 277,
571
+ 893,
572
+ 654
573
+ ],
574
+ "page_idx": 3
575
+ },
576
+ {
577
+ "type": "equation",
578
+ "text": "\n$$\n\\Delta \\boldsymbol {r} _ {l} ^ {\\prime} = \\operatorname {P o i n t H e a d} _ {l} \\left(\\boldsymbol {s} _ {l - 1}, \\boldsymbol {e} _ {l - 1}, \\boldsymbol {r} _ {l - 1}\\right),\n$$\n",
579
+ "text_format": "latex",
580
+ "bbox": [
581
+ 571,
582
+ 665,
583
+ 821,
584
+ 681
585
+ ],
586
+ "page_idx": 3
587
+ },
588
+ {
589
+ "type": "equation",
590
+ "text": "\n$$\n\\Delta \\boldsymbol {r} _ {l} = \\sigma \\left(\\sigma^ {- 1} \\left(\\boldsymbol {r} _ {l - 1} - \\boldsymbol {r} _ {0}\\right) + \\Delta \\boldsymbol {r} _ {l} ^ {\\prime}\\right), \\tag {3}\n$$\n",
591
+ "text_format": "latex",
592
+ "bbox": [
593
+ 580,
594
+ 685,
595
+ 890,
596
+ 702
597
+ ],
598
+ "page_idx": 3
599
+ },
600
+ {
601
+ "type": "equation",
602
+ "text": "\n$$\n\\hat {\\boldsymbol {r}} _ {l} = \\boldsymbol {r} _ {0} + \\Delta \\boldsymbol {r} _ {l} \\cdot \\boldsymbol {s} _ {\\text {g r i d}}, \\quad \\boldsymbol {r} _ {l} = \\operatorname {D e t a c h} (\\hat {\\boldsymbol {r}} _ {l}),\n$$\n",
603
+ "text_format": "latex",
604
+ "bbox": [
605
+ 557,
606
+ 705,
607
+ 834,
608
+ 722
609
+ ],
610
+ "page_idx": 3
611
+ },
612
+ {
613
+ "type": "text",
614
+ "text": "where $\\Delta \\boldsymbol{r}_l'$ and $\\Delta \\boldsymbol{r}_l$ are the predicted offsets from $\\boldsymbol{r}_l$ to both $\\boldsymbol{r}_{l-1}$ and $\\boldsymbol{r}_0$ before the sigmoid activation $\\sigma$ , respectively.",
615
+ "bbox": [
616
+ 500,
617
+ 732,
618
+ 890,
619
+ 763
620
+ ],
621
+ "page_idx": 3
622
+ },
623
+ {
624
+ "type": "text",
625
+ "text": "3.2. Salient Point Enhanced Cross-Attention",
626
+ "text_level": 1,
627
+ "bbox": [
628
+ 500,
629
+ 771,
630
+ 841,
631
+ 786
632
+ ],
633
+ "page_idx": 3
634
+ },
635
+ {
636
+ "type": "text",
637
+ "text": "In cross-attention layers, existing center-based methods are limited to the attention on both center and sides of the ground truth bounding box, causing detector confusion among the queries with the same center and side attention. To this end, we expect the queries to focus on their specific regions based on the reference points, four box sides, and other conditional regions in different heads. Accordingly,",
638
+ "bbox": [
639
+ 496,
640
+ 795,
641
+ 893,
642
+ 900
643
+ ],
644
+ "page_idx": 3
645
+ },
646
+ {
647
+ "type": "page_number",
648
+ "text": "15542",
649
+ "bbox": [
650
+ 480,
651
+ 944,
652
+ 517,
653
+ 955
654
+ ],
655
+ "page_idx": 3
656
+ },
657
+ {
658
+ "type": "image",
659
+ "img_path": "images/db807192e02719a82d755a7878de64a4f620ef8a0f63351e89cf8572332e9f7c.jpg",
660
+ "image_caption": [
661
+ "Figure 3. Movable reference point. The reference points are initialized by the center/corner points of the mesh-grid. Based on the inner loss, only the green dashed box can be predicted by the inner points when reference points are fixed. By moving the reference points within their grid, the blue dashed boxes can be detected accurately without extended searching space."
662
+ ],
663
+ "image_footnote": [],
664
+ "bbox": [
665
+ 81,
666
+ 88,
667
+ 472,
668
+ 210
669
+ ],
670
+ "page_idx": 4
671
+ },
672
+ {
673
+ "type": "text",
674
+ "text": "we consider an improved Gaussian [5] $\\mathbf{G}$ and conditional attention [16] $\\mathbf{A}_{\\mathrm{peca}}$ to enhance query specificity and spatially extreme region discrimination. The final attention map $\\mathbf{A}_{\\mathrm{cross}}$ is the sum of the two attentions $\\mathbf{A}_{\\mathrm{cross}} = \\mathbf{G} + \\mathbf{A}_{\\mathrm{peca}}$ .",
675
+ "bbox": [
676
+ 75,
677
+ 318,
678
+ 468,
679
+ 380
680
+ ],
681
+ "page_idx": 4
682
+ },
683
+ {
684
+ "type": "text",
685
+ "text": "Side Directed Gaussian (SDG). Similar to the movable strategy, we enforce the predicted Gaussian attention to be inside the proposal bounding box to reduce the searching space. Given a reference point $r$ , the offset scales $\\mathbf{o} \\in [-1,1]^2$ for $H$ heads are produced by a simple MLP with a tanh activation, and then multiply to the two sides of the proposal bounding box for head-specific point offset generation, where the direction is guided by the sign of the offset scales. The head-specific points are generated by Algorithm 1. For each head, the Gaussian-like spatial weight map $G_{i}$ effecting on each pixel $(x,y)$ of context features is then formulated as",
686
+ "bbox": [
687
+ 75,
688
+ 380,
689
+ 473,
690
+ 546
691
+ ],
692
+ "page_idx": 4
693
+ },
694
+ {
695
+ "type": "equation",
696
+ "text": "\n$$\nG _ {i} (x, y) = \\exp \\left(- \\frac {\\left(x - c _ {w , i}\\right) ^ {2}}{v _ {w , i} ^ {2}} - \\frac {\\left(y - c _ {h , i}\\right) ^ {2}}{v _ {h , i} ^ {2}}\\right). \\tag {4}\n$$\n",
697
+ "text_format": "latex",
698
+ "bbox": [
699
+ 117,
700
+ 556,
701
+ 468,
702
+ 597
703
+ ],
704
+ "page_idx": 4
705
+ },
706
+ {
707
+ "type": "text",
708
+ "text": "Point Enhanced Cross-Attention (PECA). As aforementioned, the semantic class for the query is closely related to its referenced location in our SAP-DETR. To enhance the correlation between queries and their references, we concatenate the locations to the content queries after the sinusoidal positional encoding (PE). Take a close look at the conditional attention [16], we find that the linear positional embedding mostly focuses on one box side in each attention head. So we introduce a more straightforward attention mechanism, where the four side coordinates are concatenated and assigned to the corresponding head for side attention. The process of PECA is formulated as",
709
+ "bbox": [
710
+ 75,
711
+ 604,
712
+ 472,
713
+ 787
714
+ ],
715
+ "page_idx": 4
716
+ },
717
+ {
718
+ "type": "equation",
719
+ "text": "\n$$\n\\begin{array}{r l} \\mathbf {A} _ {\\text {p e c a}} & = \\mathbf {e} _ {q} \\mathbf {e} _ {k} ^ {\\top} + \\mathbf {T P E} (\\boldsymbol {r} _ {q}) \\mathbf {P E} (\\boldsymbol {r} _ {k}) ^ {\\top} \\\\ & + \\mathbf {T} g (\\mathbf {P E} (\\boldsymbol {r} _ {q} - \\{\\ell , t \\}, \\boldsymbol {r} _ {q} + \\{r, b \\})) \\mathbf {P E} (\\boldsymbol {r} _ {k}) ^ {\\top}, \\end{array} \\tag {5}\n$$\n",
720
+ "text_format": "latex",
721
+ "bbox": [
722
+ 84,
723
+ 792,
724
+ 468,
725
+ 835
726
+ ],
727
+ "page_idx": 4
728
+ },
729
+ {
730
+ "type": "text",
731
+ "text": "where $g$ is a linear layer mapping 4D-PE into 2D to keep channel dimension consistency. T [16] is a scaling transformation matrix. More details of T and comparison with conditional cross-attention are available in Appendix C.",
732
+ "bbox": [
733
+ 75,
734
+ 839,
735
+ 472,
736
+ 901
737
+ ],
738
+ "page_idx": 4
739
+ },
740
+ {
741
+ "type": "text",
742
+ "text": "Algorithm 1 Side Directed Gaussian",
743
+ "text_level": 1,
744
+ "bbox": [
745
+ 500,
746
+ 90,
747
+ 746,
748
+ 104
749
+ ],
750
+ "page_idx": 4
751
+ },
752
+ {
753
+ "type": "text",
754
+ "text": "Input: Content embedding e, reference point $r$ and box s.",
755
+ "bbox": [
756
+ 500,
757
+ 109,
758
+ 852,
759
+ 123
760
+ ],
761
+ "page_idx": 4
762
+ },
763
+ {
764
+ "type": "text",
765
+ "text": "Output: Head-specific points $\\mathbf{c} = \\{(c_{w,i}, c_{h,i}) | i \\in H\\}$ and head-specific attention $\\mathbf{v} = \\{(v_{w,i}, v_{h,i}) | i \\in H\\}$ .",
766
+ "bbox": [
767
+ 500,
768
+ 123,
769
+ 893,
770
+ 151
771
+ ],
772
+ "page_idx": 4
773
+ },
774
+ {
775
+ "type": "list",
776
+ "sub_type": "text",
777
+ "list_items": [
778
+ "1: Predict offset scale and attention scale based on content embedding, $\\mathbf{o} = \\tanh (\\mathrm{MLP}(\\mathbf{e}))$ , $\\mathbf{v} = \\mathrm{MLP}(\\mathbf{e})$",
779
+ "2: for $h\\gets 1\\in H$ do",
780
+ "3: Select the index of direction guided by the sign of offset scale, $\\{a,b\\} = \\mathrm{sgn}(\\mathbf{o}_i) + \\{1,2\\}$ , $a,b\\in \\{0,1,2,3\\}$ ;",
781
+ "4: According to the index of direction, predict head-specific point, $\\mathbf{c}_i = \\mathbf{o}_i\\cdot \\mathbf{s}[a,b] + \\mathbf{r}$",
782
+ "5: end for",
783
+ "6: return $\\mathbf{c}_i, \\mathbf{v}_i, \\forall i = 1, \\dots, H$"
784
+ ],
785
+ "bbox": [
786
+ 508,
787
+ 151,
788
+ 893,
789
+ 275
790
+ ],
791
+ "page_idx": 4
792
+ },
793
+ {
794
+ "type": "text",
795
+ "text": "3.3. SAP-DETR with Denoising Strategy",
796
+ "text_level": 1,
797
+ "bbox": [
798
+ 500,
799
+ 308,
800
+ 813,
801
+ 325
802
+ ],
803
+ "page_idx": 4
804
+ },
805
+ {
806
+ "type": "text",
807
+ "text": "To further explore the capability of our proposed SAP-DETR, we develop SAP-DN-DETR and SAP-DINO-DETR by adding the denoising auxiliary loss [9,28] into the training process. In the denoised SAP-DETR, the main difference from both DN-DETR and DINO lies in the noise design. Instead of the center point, we perform the box jittering and randomly sample a point from the intersection region between the original bounding box and the jittering one as the reference point. As the denoising strategy only serves as an auxiliary training loss increasing the training cost, the variants of denoising models are test-free whose Params and GFLOPs are the same as SAP-DETR models.",
808
+ "bbox": [
809
+ 496,
810
+ 332,
811
+ 893,
812
+ 513
813
+ ],
814
+ "page_idx": 4
815
+ },
816
+ {
817
+ "type": "text",
818
+ "text": "4. Experimental Results",
819
+ "text_level": 1,
820
+ "bbox": [
821
+ 500,
822
+ 531,
823
+ 705,
824
+ 547
825
+ ],
826
+ "page_idx": 4
827
+ },
828
+ {
829
+ "type": "text",
830
+ "text": "4.1. Implementation Details",
831
+ "text_level": 1,
832
+ "bbox": [
833
+ 500,
834
+ 556,
835
+ 717,
836
+ 573
837
+ ],
838
+ "page_idx": 4
839
+ },
840
+ {
841
+ "type": "text",
842
+ "text": "We conduct the experiments on the COCO 2017 [12] object detection dataset, containing about 118K training images and 5K validation images. All models are evaluated by the standard COCO evaluation metrics. We follow the vanilla DETR [2] structure that consists of a CNN backbone, a stack of Transformer encoder-decoder layers, and two prediction heads for class label and bounding box prediction. We use ImageNet-pretrained ResNet [7] as our backbone, and report results based on the ResNet and its $\\times 1/16$ -resolution extension ResNet-DC. Unlike DAB-DETR [13] sharing box and label head for each layer, we share the class head except the first layer and use an independent box head for the box regression of each layer (for more details please refer to Appendix D). As the mesh-grid initialization for reference points in SAP-DETR, we consider the number of queries $N$ as a perfect square for uniform distribution. Unless otherwise specified, we use $N = 400$ queries in the experiments. Precisely, we also provide a comparison under $N = 300$ in Tab. 2, the standard setting in DETR-like models.",
843
+ "bbox": [
844
+ 496,
845
+ 580,
846
+ 893,
847
+ 868
848
+ ],
849
+ "page_idx": 4
850
+ },
851
+ {
852
+ "type": "text",
853
+ "text": "We adopt two different Transformer structures for experiments where a 3-layer encoder-decoder stack is evaluated to",
854
+ "bbox": [
855
+ 500,
856
+ 869,
857
+ 893,
858
+ 900
859
+ ],
860
+ "page_idx": 4
861
+ },
862
+ {
863
+ "type": "page_number",
864
+ "text": "15543",
865
+ "bbox": [
866
+ 480,
867
+ 944,
868
+ 517,
869
+ 955
870
+ ],
871
+ "page_idx": 4
872
+ },
873
+ {
874
+ "type": "image",
875
+ "img_path": "images/e67c17b52b62bc18f59407f49da518ec9d46208e46c8ed65384428ba5d8ecd32.jpg",
876
+ "image_caption": [
877
+ "Figure 4. Comparison of performance and training losses curves."
878
+ ],
879
+ "image_footnote": [],
880
+ "bbox": [
881
+ 187,
882
+ 87,
883
+ 379,
884
+ 210
885
+ ],
886
+ "page_idx": 5
887
+ },
888
+ {
889
+ "type": "image",
890
+ "img_path": "images/c4e4471d6732af893721aadaca57cebf551f6985172900d817e0947edfa5e3c0.jpg",
891
+ "image_caption": [],
892
+ "image_footnote": [],
893
+ "bbox": [
894
+ 385,
895
+ 87,
896
+ 580,
897
+ 210
898
+ ],
899
+ "page_idx": 5
900
+ },
901
+ {
902
+ "type": "image",
903
+ "img_path": "images/a2fba28d8751276d87d046916bae807f8b5b51e25a59642266ec0fb9122b2506.jpg",
904
+ "image_caption": [],
905
+ "image_footnote": [],
906
+ "bbox": [
907
+ 583,
908
+ 85,
909
+ 781,
910
+ 210
911
+ ],
912
+ "page_idx": 5
913
+ },
914
+ {
915
+ "type": "text",
916
+ "text": "demonstrate our lightweight model efficacy compared with the traditional CNN detectors, and a 6-layer encoder-decoder stack is aligned with previous DETR variants to investigate the performance of large model. Both are trained on two training schemes: the 12-epoch and 36-epoch schemes with a learning rate drop after 11 and 30 epochs, respectively. All models are trained on the Nvidia A100 GPUs with batch size of 16 and 8 for ResNet and ResNet-DC, respectively. For more training details, please refer to Appendix F.",
917
+ "bbox": [
918
+ 75,
919
+ 250,
920
+ 472,
921
+ 387
922
+ ],
923
+ "page_idx": 5
924
+ },
925
+ {
926
+ "type": "text",
927
+ "text": "4.2. Main Results",
928
+ "text_level": 1,
929
+ "bbox": [
930
+ 76,
931
+ 393,
932
+ 217,
933
+ 409
934
+ ],
935
+ "page_idx": 5
936
+ },
937
+ {
938
+ "type": "text",
939
+ "text": "As shown in Tab. 1 and Tab. 2, we comprehensively compare our proposed SAP-DETR with the traditional CNN detectors [18], the original DETR [2], and other DETR-like detectors [5, 13, 16, 25, 26, 31] on COCO 2017 validation dataset. For in-depth analysis, we conduct the comparison in two aspects: model convergence and efficacy.",
940
+ "bbox": [
941
+ 75,
942
+ 417,
943
+ 470,
944
+ 508
945
+ ],
946
+ "page_idx": 5
947
+ },
948
+ {
949
+ "type": "text",
950
+ "text": "Model Convergence. Compared with traditional CNN detectors, Transformer detectors are always subject to laborious training time. For example, under the same 12-epoch training scheme, Faster RCNN [18] still achieves good performance, but the mainstream DETR-like models may suffer from inadequate training and perform poorly without the help of auxiliary losses [9]. Under the 12-epoch training scheme, our proposed SAP-DETR can accelerate model convergence significantly, boosting DAB-DETR [13] by 3.9 AP and 2.7 AP on 3-layer and 6-layer encoder-decoder structures, respectively. Compared with the current SoTA, our SAP-DETR also outperforms SAM-DETR [26] by $\\sim 1.3$ AP, with reducing $\\sim 17\\%$ parameters and $\\sim 10\\%$ GFLOPs. Take a close look at the training process, as illustrated in Fig. 4, SAP-DETR conducts with rapid descent curves in both classification and box regression losses. Notably, there is a large margin in classification loss between ours and SoTA methods, which is benefited from the query-specific reference point, hence boosting model performance in early epochs.",
951
+ "bbox": [
952
+ 75,
953
+ 507,
954
+ 470,
955
+ 794
956
+ ],
957
+ "page_idx": 5
958
+ },
959
+ {
960
+ "type": "text",
961
+ "text": "Model Efficacy. To analyze model efficacy, we report results on long training epochs and high-resolution features in Tab. 1. Under the 36-epoch training scheme, SAP-DETR achieves superior performance among all single-scale Transformer detectors, especially on middle and large targets. For example, SAP-DETR boosts DAB-DETR by $2.0\\mathrm{AP}_{\\mathrm{M}}$ and $4.1\\mathrm{AP}_{\\mathrm{L}}$ with 3-layer models, $1.0\\mathrm{AP}_{\\mathrm{M}}$ and $1.9\\mathrm{AP}_{\\mathrm{L}}$ with",
962
+ "bbox": [
963
+ 75,
964
+ 795,
965
+ 470,
966
+ 901
967
+ ],
968
+ "page_idx": 5
969
+ },
970
+ {
971
+ "type": "text",
972
+ "text": "6-layer models, which further verifies the effectiveness of our proposed salient point concept for overlapping object detection. Along with layer increase, a deficient upper-bound of SAM-DETR is exposed, with obviously lower 0.5 AP promotion compared to our 1.0 AP improvement. Persuasively, we also report the 50-epoch training results based on the 300-query setting. To align with our mesh-grid initialization strategy, we tile a $17 \\times 18$ mesh-grid (306 queries) for each reference point initialization. Tab. 2 shows our main results and the most representative approaches with their original reported performance. Notably, SAP-DETR outperforms the current SoTAs with comparable costs based on all backbones. With low-resolution features ( $\\times 1/32$ ), it significantly boosts both middle and large object detection accuracy.",
973
+ "bbox": [
974
+ 496,
975
+ 250,
976
+ 893,
977
+ 462
978
+ ],
979
+ "page_idx": 5
980
+ },
981
+ {
982
+ "type": "text",
983
+ "text": "Combine with Other Fast Convergence Methods. As shown in Tab. 3, we compare our SAP-DETR variants with the current fast convergence methods [4, 9, 28]. With such a subtle modification, our SAP-DETR (grey rows) results in a significant performance improvement compared with the original methods (white rows). Under the 12-epoch training scheme, there exist 0.5-1.9 AP improvements on DN-DETR [9] and 0.7-1.6 AP improvements on Group-DETR [4], but the promotions are slightly reduced when implemented on DINO [28]. We hypothesise that there exists the same effect between the negative query of contrastive denoising [28] and our query-salient reference point. Moreover, we observe that the performance improvements largely originate from the large object detection, especially based on ResNet-DC5 family backbones. We speculate that DETR may prefer the high-resolution features ( $\\times 1/16$ ) rather than the low-resolution ones ( $\\times 1/32$ ), and our SAP-DETR can distinguish the salient points accurately on the high-resolution, thereby taking full advantage of the large object detection.",
984
+ "bbox": [
985
+ 496,
986
+ 462,
987
+ 895,
988
+ 750
989
+ ],
990
+ "page_idx": 5
991
+ },
992
+ {
993
+ "type": "text",
994
+ "text": "4.3. Ablation Study",
995
+ "text_level": 1,
996
+ "bbox": [
997
+ 500,
998
+ 760,
999
+ 653,
1000
+ 776
1001
+ ],
1002
+ "page_idx": 5
1003
+ },
1004
+ {
1005
+ "type": "text",
1006
+ "text": "Effectiveness of Each Component. To offer an intuitionistic comparison of model convergence for each component, Tab. 4 reports the effectiveness of them based on the 3-layer encoder-decoder structure and 12-epoch training scheme. I). The proposed salient point concept based on content embeddings improves the performance from 32.3 AP to 33.5",
1007
+ "bbox": [
1008
+ 496,
1009
+ 784,
1010
+ 893,
1011
+ 876
1012
+ ],
1013
+ "page_idx": 5
1014
+ },
1015
+ {
1016
+ "type": "page_footnote",
1017
+ "text": "<sup>1</sup>https://github.com/facebookresearch/detectron2",
1018
+ "bbox": [
1019
+ 517,
1020
+ 886,
1021
+ 781,
1022
+ 900
1023
+ ],
1024
+ "page_idx": 5
1025
+ },
1026
+ {
1027
+ "type": "page_number",
1028
+ "text": "15544",
1029
+ "bbox": [
1030
+ 480,
1031
+ 944,
1032
+ 519,
1033
+ 955
1034
+ ],
1035
+ "page_idx": 5
1036
+ },
1037
+ {
1038
+ "type": "table",
1039
+ "img_path": "images/f0e495db48712226ee91ed149febaa6e161ba66f364fb9954946855d4923b05a.jpg",
1040
+ "table_caption": [],
1041
+ "table_footnote": [],
1042
+ "table_body": "<table><tr><td>Method</td><td>#Epochs</td><td>#Params(M)</td><td>GFLOPs</td><td>AP</td><td>\\( AP_{50} \\)</td><td>\\( AP_{75} \\)</td><td>\\( AP_S \\)</td><td>\\( AP_M \\)</td><td>\\( AP_L \\)</td></tr><tr><td colspan=\"10\">3-Layer Encoder-Decoder Transformer Neck with ResNet-50 Backbone</td></tr><tr><td>DETR-R50 [2]</td><td>36</td><td>33</td><td>82</td><td>15.8</td><td>28.0</td><td>15.4</td><td>5.3</td><td>16.7</td><td>24.6</td></tr><tr><td>Deformable DETR-R50 [31]</td><td>36</td><td>30</td><td>77</td><td>37.1</td><td>57.6</td><td>39.4</td><td>18.3</td><td>40.8</td><td>51.6</td></tr><tr><td>SMCA-DETR-R50 [5]</td><td>12 / 36</td><td>-</td><td>-</td><td>28.8 / 37.7</td><td>48.1 / 58.7</td><td>29.9 / 40.1</td><td>13.8 / 19.4</td><td>31.3 / 40.5</td><td>41.3 / 54.8</td></tr><tr><td>Conditional DETR-R50 [16]</td><td>12 / 36</td><td>40</td><td>82</td><td>29.6 / 37.1</td><td>48.7 / 57.9</td><td>30.7 / 39.0</td><td>13.0 / 17.6</td><td>32.3 / 40.3</td><td>43.1 / 55.0</td></tr><tr><td>Anchor DETR-R50 [25]</td><td>12 / 36</td><td>31</td><td>79</td><td>30.8 / 37.6</td><td>51.1 / 58.7</td><td>31.8 / 39.7</td><td>14.3 / 18.8</td><td>34.1 / 41.5</td><td>44.3 / 53.5</td></tr><tr><td>DAB-DETR-R50 [13]</td><td>12 / 36</td><td>34</td><td>83</td><td>32.3 / 39.0</td><td>51.3 / 58.6</td><td>34.0 / 41.8</td><td>15.7 / 20.0</td><td>35.2 / 42.5</td><td>45.7 / 56.0</td></tr><tr><td>SAM-DETR-w/SMCA-R50 [26]</td><td>12 / 36</td><td>41</td><td>89</td><td>35.1 / 40.4</td><td>54.7 / 60.7</td><td>36.7 / 42.7</td><td>16.0 / 20.2</td><td>38.4 / 44.4</td><td>52.1 / 58.3</td></tr><tr><td>SAP-DETR-R50 (Ours)</td><td>12 / 36</td><td>36</td><td>84</td><td>36.2 / 41.2</td><td>56.2 / 61.6</td><td>37.9 / 43.4</td><td>16.4 / 21.0</td><td>39.5 / 44.5</td><td>53.8 / 60.1</td></tr><tr><td colspan=\"10\">6-Layer Encoder-Decoder Transformer Neck with ResNet-50 Backbone</td></tr><tr><td>DETR-R50 [2]</td><td>36</td><td>42</td><td>89</td><td>14.0</td><td>24.4</td><td>14.0</td><td>4.2</td><td>13.7</td><td>22.5</td></tr><tr><td>Deformable DETR-R50 [31]</td><td>36</td><td>34</td><td>81</td><td>38.0</td><td>58.2</td><td>40.4</td><td>18.5</td><td>41.7</td><td>54.2</td></tr><tr><td>SMCA-DETR-R50 [5]</td><td>12 / 36</td><td>-</td><td>-</td><td>32.4 / 40.1</td><td>52.3 / 61.4</td><td>34.0 / 42.8</td><td>15.5 / 20.3</td><td>34.9 / 43.3</td><td>47.7 / 57.1</td></tr><tr><td>Conditional DETR-R50 [16]</td><td>12 / 36</td><td>44</td><td>90</td><td>33.1 / 40.2</td><td>53.0 / 61.0</td><td>34.8 / 42.4</td><td>14.5 / 19.9</td><td>35.9 / 43.5</td><td>49.2 / 58.8</td></tr><tr><td>Anchor DETR-R50 [25]</td><td>12 / 36</td><td>37</td><td>85</td><td>33.7 / 39.7</td><td>54.5 / 60.5</td><td>35.1 / 41.9</td><td>15.6 / 19.9</td><td>37.3 / 43.5</td><td>49.8 / 57.3</td></tr><tr><td>DAB-DETR-R50 [13]</td><td>12 / 36</td><td>44</td><td>92</td><td>34.9 / 41.0</td><td>55.5 / 61.7</td><td>36.4 / 43.4</td><td>16.2 / 21.3</td><td>38.4 / 44.7</td><td>51.5 / 58.9</td></tr><tr><td>SAM-DETR-w/SMCA-R50 [26]</td><td>12 / 36</td><td>59</td><td>105</td><td>36.2 / 40.9</td><td>57.2 / 62.2</td><td>37.4 / 43.1</td><td>16.1 / 20.1</td><td>39.8 / 44.7</td><td>55.3 / 60.7</td></tr><tr><td>SAP-DETR-R50 (Ours)</td><td>12 / 36</td><td>47</td><td>94</td><td>37.5 / 42.2</td><td>58.5 / 62.7</td><td>39.2 / 44.6</td><td>17.3 / 22.6</td><td>40.6 / 45.7</td><td>55.4 / 60.8</td></tr></table>",
1043
+ "bbox": [
1044
+ 99,
1045
+ 88,
1046
+ 870,
1047
+ 308
1048
+ ],
1049
+ "page_idx": 6
1050
+ },
1051
+ {
1052
+ "type": "table",
1053
+ "img_path": "images/af9d626399bf3a1424d9d21efba19a0704c74c33aff6a018ede81bfd149db747.jpg",
1054
+ "table_caption": [
1055
+ "Table 1. Comparison between Transformer necks. Based on ResNet-50 backbone, all models are trained by the official source codes with their original settings and evaluated on COCO val2017. All models uses 400 queries except Anchor DETR, while Anchor DETR uses 200 queries with 2 pattern embeddings. GFLOPs and Params are measured by $\\text{Detectron}^2$ ."
1056
+ ],
1057
+ "table_footnote": [],
1058
+ "table_body": "<table><tr><td>Method</td><td>#Epochs</td><td>#Params(M)</td><td>GFLOPs</td><td>AP</td><td>\\( AP_{50} \\)</td><td>\\( AP_{75} \\)</td><td>\\( AP_S \\)</td><td>\\( AP_M \\)</td><td>\\( AP_L \\)</td><td>Infer. \\( Time(s / img)^{\\dagger} \\)</td></tr><tr><td colspan=\"11\">ResNet-50 Backbone</td></tr><tr><td>Faster RCNN-FPN-R50 [10, 18]</td><td>108</td><td>42</td><td>180</td><td>42.0</td><td>62.1</td><td>45.5</td><td>26.6</td><td>45.5</td><td>53.4</td><td>0.039</td></tr><tr><td>DETR-R50 [2]</td><td>500</td><td>41</td><td>86</td><td>42.0</td><td>62.4</td><td>44.2</td><td>20.5</td><td>45.8</td><td>61.1</td><td>0.040</td></tr><tr><td>Deformable DETR-R50 [31]</td><td>50</td><td>34</td><td>78</td><td>39.4</td><td>59.6</td><td>42.3</td><td>20.6</td><td>43.0</td><td>55.5</td><td>0.043</td></tr><tr><td>SMCA-DETR-R50 [5]</td><td>50</td><td>42</td><td>86</td><td>41.0</td><td>-</td><td>-</td><td>21.9</td><td>44.3</td><td>59.1</td><td>0.045</td></tr><tr><td>Conditional DETR-R50 [16]</td><td>50</td><td>44</td><td>90</td><td>40.9</td><td>61.8</td><td>43.3</td><td>20.8</td><td>44.6</td><td>59.2</td><td>0.057</td></tr><tr><td>Anchor DETR-R50 [25]</td><td>50</td><td>39</td><td>85</td><td>42.1</td><td>63.1</td><td>44.9</td><td>22.3</td><td>46.2</td><td>60.0</td><td>0.050</td></tr><tr><td>DAB-DETR-R50 [13]</td><td>50</td><td>44</td><td>\\( 90^† \\)</td><td>42.2</td><td>63.1</td><td>44.7</td><td>21.5</td><td>45.7</td><td>60.3</td><td>0.059</td></tr><tr><td>SAM-DETR-w/SMCA-R50 [26]</td><td>50</td><td>58</td><td>100</td><td>41.8</td><td>63.2</td><td>43.9</td><td>22.1</td><td>45.9</td><td>60.9</td><td>0.065</td></tr><tr><td>SAP-DETR-R50 (Ours)</td><td>50</td><td>47</td><td>92</td><td>43.1</td><td>63.8</td><td>45.4</td><td>22.9</td><td>47.1</td><td>62.1</td><td>0.063</td></tr><tr><td colspan=\"11\">ResNet-101 Backbone</td></tr><tr><td>Faster RCNN-FPN-R101 [10, 18]</td><td>108</td><td>60</td><td>246</td><td>44.0</td><td>63.9</td><td>47.8</td><td>27.2</td><td>48.1</td><td>56.0</td><td>0.050</td></tr><tr><td>DETR-R101 [2]</td><td>500</td><td>60</td><td>152</td><td>43.5</td><td>63.8</td><td>46.4</td><td>21.9</td><td>48.0</td><td>61.8</td><td>0.066</td></tr><tr><td>Conditional DETR-R101 [16]</td><td>50</td><td>63</td><td>156</td><td>42.8</td><td>63.7</td><td>46.0</td><td>21.7</td><td>46.6</td><td>60.9</td><td>0.070</td></tr><tr><td>Anchor DETR-R101 [25]</td><td>50</td><td>58</td><td>150</td><td>43.5</td><td>64.3</td><td>46.6</td><td>23.2</td><td>47.7</td><td>61.4</td><td>0.068</td></tr><tr><td>DAB-DETR-R101 [13]</td><td>50</td><td>63</td><td>\\( 157^† \\)</td><td>43.5</td><td>63.9</td><td>46.6</td><td>23.6</td><td>47.3</td><td>61.5</td><td>0.072</td></tr><tr><td>SAP-DETR-R101 (Ours)</td><td>50</td><td>67</td><td>158</td><td>44.4</td><td>64.9</td><td>47.1</td><td>24.1</td><td>48.7</td><td>63.1</td><td>0.078</td></tr><tr><td colspan=\"11\">DC5-ResNet-50 Backbone</td></tr><tr><td>DETR-DC5-R50 [2]</td><td>500</td><td>41</td><td>187</td><td>43.3</td><td>63.1</td><td>45.9</td><td>22.5</td><td>47.3</td><td>61.1</td><td>0.087</td></tr><tr><td>Conditional DETR-DC5-R50 [16]</td><td>50</td><td>44</td><td>195</td><td>43.8</td><td>64.4</td><td>46.7</td><td>24.0</td><td>47.6</td><td>60.7</td><td>0.093</td></tr><tr><td>Anchor DETR-DC5-R50 [25]</td><td>50</td><td>39</td><td>151</td><td>44.2</td><td>64.7</td><td>47.5</td><td>24.7</td><td>48.2</td><td>60.6</td><td>0.069</td></tr><tr><td>DAB-DETR-DC5-R50 [13]</td><td>50</td><td>44</td><td>\\( 194^† \\)</td><td>44.5</td><td>65.1</td><td>47.7</td><td>25.3</td><td>48.2</td><td>62.3</td><td>0.094</td></tr><tr><td>SAM-DETR-w/SMCA-DC5-R50 [26]</td><td>50</td><td>58</td><td>210</td><td>45.0</td><td>65.4</td><td>47.9</td><td>26.2</td><td>49.0</td><td>63.3</td><td>0.126</td></tr><tr><td>SAP-DETR-DC5-R50 (Ours)</td><td>50</td><td>47</td><td>197</td><td>46.0</td><td>65.5</td><td>48.9</td><td>26.4</td><td>50.2</td><td>62.6</td><td>0.116</td></tr><tr><td colspan=\"11\">DC5-ResNet-101 Backbone</td></tr><tr><td>DETR-DC5-R101 [2]</td><td>500</td><td>60</td><td>253</td><td>44.9</td><td>64.7</td><td>47.7</td><td>23.7</td><td>49.5</td><td>62.3</td><td>0.101</td></tr><tr><td>Conditional DETR-DC5-R101 [16]</td><td>50</td><td>63</td><td>262</td><td>45.0</td><td>65.6</td><td>48.4</td><td>26.1</td><td>48.9</td><td>62.8</td><td>0.105</td></tr><tr><td>Anchor DETR-DC5-R101 [25]</td><td>50</td><td>58</td><td>227</td><td>45.1</td><td>65.7</td><td>48.8</td><td>25.8</td><td>49.4</td><td>61.6</td><td>0.083</td></tr><tr><td>DAB-DETR-DC5-R101 [13]</td><td>50</td><td>63</td><td>\\( 263^† \\)</td><td>45.8</td><td>65.9</td><td>49.3</td><td>27.0</td><td>49.8</td><td>63.8</td><td>0.110</td></tr><tr><td>SAP-DETR-DC5-R101 (Ours)</td><td>50</td><td>67</td><td>266</td><td>46.9</td><td>66.7</td><td>50.5</td><td>27.9</td><td>51.3</td><td>64.3</td><td>0.130</td></tr></table>",
1059
+ "bbox": [
1060
+ 98,
1061
+ 362,
1062
+ 870,
1063
+ 705
1064
+ ],
1065
+ "page_idx": 6
1066
+ },
1067
+ {
1068
+ "type": "text",
1069
+ "text": "Table 2. Comparison of Transformer necks with 300 queries on COCO val2017. All results are reported from their original paper. All models uses 300 queries except Anchor DETR, while Anchor DETR uses 100 queries with 3 pattern embeddings. All inference speeds are measured by a single Nvidia A100 GPU. † denotes the results are measured by ourselves.",
1070
+ "bbox": [
1071
+ 75,
1072
+ 710,
1073
+ 893,
1074
+ 755
1075
+ ],
1076
+ "page_idx": 6
1077
+ },
1078
+ {
1079
+ "type": "text",
1080
+ "text": "AP compared to baseline DAB-DETR (row 8-9). Such a query-specific spatial prior enables queries to attend to their expected region from content features (see Figures in Appendix G) and reduces the false detection rate on occluded and partial objects (see Fig. 1(d)), hence boosting detection performance on middle and large objects. However, there exists a drop in small object detection (15.7 APs vs. 14.3 APs), for which we consider that the failure is mainly caused",
1081
+ "bbox": [
1082
+ 75,
1083
+ 765,
1084
+ 472,
1085
+ 887
1086
+ ],
1087
+ "page_idx": 6
1088
+ },
1089
+ {
1090
+ "type": "text",
1091
+ "text": "by the query sparsity. 2). Therefore, the movable strategy is applied to alleviate the constraint, improving the final model by $1.0\\mathrm{AP}$ and $0.6\\mathrm{AP}_{\\mathrm{S}}$ (row 1 and 6). 3). Compared with the final model, the inner loss greatly improves the performance on high-quality $\\mathrm{AP}_{75}$ and large objects $\\mathrm{AP}_{\\mathrm{L}}$ detection (row 1 and 7), with just a slight drop on low-quality object $\\mathrm{AP}_{50}$ . That is reasonable because the outside reference points are unable to localize objects accurately, and this phenomenon",
1092
+ "bbox": [
1093
+ 496,
1094
+ 765,
1095
+ 893,
1096
+ 887
1097
+ ],
1098
+ "page_idx": 6
1099
+ },
1100
+ {
1101
+ "type": "page_number",
1102
+ "text": "15545",
1103
+ "bbox": [
1104
+ 480,
1105
+ 944,
1106
+ 519,
1107
+ 957
1108
+ ],
1109
+ "page_idx": 6
1110
+ },
1111
+ {
1112
+ "type": "table",
1113
+ "img_path": "images/20d6ddd9750f06eeb1fa780dc674d1a5a429f28f28f646e61a1b317c601627f8.jpg",
1114
+ "table_caption": [],
1115
+ "table_footnote": [],
1116
+ "table_body": "<table><tr><td rowspan=\"2\">Backbone</td><td rowspan=\"2\">Epoch</td><td rowspan=\"2\">w/ SAP</td><td colspan=\"2\">DN-DETR [9]</td><td colspan=\"2\">DINO (Single-Scale) [28]</td><td colspan=\"2\">Group DETR [4]</td></tr><tr><td>AP / AP50 / AP75</td><td>APS / APM / APL</td><td>AP / AP50 / AP75</td><td>APS / APM / APL</td><td>AP / AP50 / AP75</td><td>APS / APM / APL</td></tr><tr><td rowspan=\"2\">R50</td><td rowspan=\"2\">12</td><td rowspan=\"2\">✓(Ours)</td><td>38.3 / 58.6 / 40.5</td><td>18.4 / 41.6 / 57.1</td><td>39.7 / 58.3 / 42.4</td><td>19.1 / 43.7 / 57.1</td><td>39.1 / - / -</td><td>19.7 / 42.5 / 56.8</td></tr><tr><td>39.5 / 59.7 / 41.5</td><td>18.7 / 42.8 / 59.0</td><td>40.0 / 60.1 / 42.1</td><td>20.2 / 43.4 / 58.5</td><td>39.8 / 60.2 / 42.0</td><td>20.2 / 43.5 / 58.6</td></tr><tr><td rowspan=\"2\">R101</td><td rowspan=\"2\">12</td><td rowspan=\"2\">✓(Ours)</td><td>40.5 / 60.8 / 43.0</td><td>19.3 / 44.3 / 59.6</td><td>41.9 / 60.8 / 44.4</td><td>22.5 / 46.3 / 59.5</td><td>- / - / -</td><td>- / - / -</td></tr><tr><td>41.0 / 61.2 / 43.4</td><td>19.8 / 45.3 / 60.0</td><td>41.5 / 61.4 / 43.6</td><td>20.3 / 45.2 / 60.0</td><td>41.1 / 61.5 / 43.4</td><td>20.5 / 45.5 / 59.4</td></tr><tr><td rowspan=\"2\">R50-DC</td><td rowspan=\"2\">12</td><td rowspan=\"2\">✓(Ours)</td><td>41.7 / 61.4 / 44.1</td><td>21.2 / 45.0 / 60.2</td><td>43.6 / 61.4 / 47.0</td><td>24.8 / 47.3 / 59.5</td><td>41.9 / - / -</td><td>23.3 / 45.6 / 58.4</td></tr><tr><td>43.6 / 62.5 / 46.2</td><td>23.3 / 47.3 / 61.0</td><td>44.0 / 63.1 / 46.5</td><td>24.8 / 47.3 / 61.1</td><td>43.9 / 63.2 / 46.8</td><td>24.5 / 47.6 / 61.3</td></tr><tr><td rowspan=\"2\">R101-DC</td><td rowspan=\"2\">12</td><td rowspan=\"2\">✓(Ours)</td><td>43.4 / 61.9 / 47.2</td><td>24.8 / 46.8 / 59.4</td><td>45.4 / 63.5 / 49.2</td><td>26.4 / 49.5 / 61.1</td><td>- / - / -</td><td>- / - / -</td></tr><tr><td>44.6 / 63.9 / 48.0</td><td>25.5 / 48.9 / 62.5</td><td>45.6 / 64.5 / 48.7</td><td>25.0 / 49.7 / 62.5</td><td>44.4 / 63.9 / 47.4</td><td>25.9 / 48.5 / 61.4</td></tr></table>",
1117
+ "bbox": [
1118
+ 119,
1119
+ 88,
1120
+ 854,
1121
+ 200
1122
+ ],
1123
+ "page_idx": 7
1124
+ },
1125
+ {
1126
+ "type": "table",
1127
+ "img_path": "images/32be50eeddec82a689fc94ef01756152c3c617029bb5f1e4522e5711f0a553dd.jpg",
1128
+ "table_caption": [
1129
+ "Table 3. Comparison with denoised methods on COCO dataset based on the 12-epoch training schedule and 300 object queries."
1130
+ ],
1131
+ "table_footnote": [],
1132
+ "table_body": "<table><tr><td>Comment</td><td>Movable</td><td>Inner Loss</td><td>PECA</td><td>SDG</td><td>AP</td><td>\\( AP_{50} \\)</td><td>\\( AP_{75} \\)</td><td>\\( AP_S \\)</td><td>\\( AP_M \\)</td><td>\\( AP_L \\)</td></tr><tr><td>SAP-DETR (Ours)</td><td>✓</td><td>✓</td><td>✓</td><td>✓</td><td>36.2</td><td>56.2</td><td>37.9</td><td>16.4</td><td>39.5</td><td>53.8</td></tr><tr><td>-SDG</td><td>✓</td><td>✓</td><td>✓</td><td></td><td>35.6</td><td>56.2</td><td>36.9</td><td>16.3</td><td>38.9</td><td>52.7</td></tr><tr><td>-PECA</td><td>✓</td><td>✓</td><td></td><td>✓</td><td>34.8</td><td>55.5</td><td>36.0</td><td>15.7</td><td>37.3</td><td>52.0</td></tr><tr><td>-PECA &amp; SDG</td><td>✓</td><td>✓</td><td></td><td></td><td>34.0</td><td>54.9</td><td>35.3</td><td>15.0</td><td>36.7</td><td>51.5</td></tr><tr><td>-Movable</td><td></td><td>✓</td><td>✓</td><td>✓</td><td>35.2</td><td>55.4</td><td>36.8</td><td>15.8</td><td>38.5</td><td>53.8</td></tr><tr><td>-Inner Loss</td><td>✓</td><td></td><td>✓</td><td>✓</td><td>35.9</td><td>56.3</td><td>37.4</td><td>16.2</td><td>39.3</td><td>52.5</td></tr><tr><td>DAB-DETR (Baseline)</td><td>-</td><td>-</td><td>-</td><td>-</td><td>32.3</td><td>51.3</td><td>34.0</td><td>15.7</td><td>35.2</td><td>45.7</td></tr><tr><td>+Salient Point Concept</td><td>-</td><td>-</td><td>-</td><td>-</td><td>33.5</td><td>54.3</td><td>35.1</td><td>14.3</td><td>36.5</td><td>51.0</td></tr></table>",
1133
+ "bbox": [
1134
+ 119,
1135
+ 234,
1136
+ 849,
1137
+ 330
1138
+ ],
1139
+ "page_idx": 7
1140
+ },
1141
+ {
1142
+ "type": "table",
1143
+ "img_path": "images/bfe2dfcf6438eb46ef01ad8830cf28ffb81daf459a4571ee5e1b0268a7e20227.jpg",
1144
+ "table_caption": [
1145
+ "Table 4. Ablation on each components"
1146
+ ],
1147
+ "table_footnote": [],
1148
+ "table_body": "<table><tr><td>Inner Cost (Linner)</td><td>Movable within Grid (sgrid)</td><td>AP</td><td>APS</td><td>APM</td><td>APL</td></tr><tr><td></td><td></td><td>35.9</td><td>17.0</td><td>38.8</td><td>52.7</td></tr><tr><td>✓</td><td></td><td>26.3</td><td>11.3</td><td>28.0</td><td>39.5</td></tr><tr><td>✓</td><td>✓</td><td>36.2</td><td>16.4</td><td>39.5</td><td>53.8</td></tr></table>",
1149
+ "bbox": [
1150
+ 78,
1151
+ 364,
1152
+ 468,
1153
+ 417
1154
+ ],
1155
+ "page_idx": 7
1156
+ },
1157
+ {
1158
+ "type": "table",
1159
+ "img_path": "images/3854a99cdefcd33b32e9cc587580d384c38233fb149d2d205dd74e8137192ae2.jpg",
1160
+ "table_caption": [
1161
+ "Table 5. Ablation on scaling factor of grid"
1162
+ ],
1163
+ "table_footnote": [],
1164
+ "table_body": "<table><tr><td>PECA</td><td>Scaling Factor of SDG</td><td>AP</td><td>APS</td><td>APM</td><td>APL</td></tr><tr><td></td><td></td><td>33.6</td><td>14.7</td><td>36.0</td><td>50.7</td></tr><tr><td>✓</td><td></td><td>35.7</td><td>17.5</td><td>38.8</td><td>52.6</td></tr><tr><td>✓</td><td>✓</td><td>36.2</td><td>16.4</td><td>39.5</td><td>53.8</td></tr></table>",
1165
+ "bbox": [
1166
+ 78,
1167
+ 446,
1168
+ 468,
1169
+ 500
1170
+ ],
1171
+ "page_idx": 7
1172
+ },
1173
+ {
1174
+ "type": "text",
1175
+ "text": "Table 6. Ablation on scaling factor of SDG",
1176
+ "bbox": [
1177
+ 143,
1178
+ 510,
1179
+ 401,
1180
+ 523
1181
+ ],
1182
+ "page_idx": 7
1183
+ },
1184
+ {
1185
+ "type": "text",
1186
+ "text": "always exists in large objects. 4). For salient point enhanced cross-attention, both SDG and PECA serve as the essential components, independently emerging 0.8 AP and 1.6 AP improvements compared to the standard model (rows 2-4). Interestingly, there exists an effectiveness overlap on small objects, with only $0.1\\mathrm{AP}_{\\mathrm{S}}$ improvement when adding SDG to the equipped PECA model (row 1 and 2). We argue that the Gaussian-like map of SDG might be easily overlapped with PECA on small objects.",
1187
+ "bbox": [
1188
+ 75,
1189
+ 535,
1190
+ 468,
1191
+ 671
1192
+ ],
1193
+ "page_idx": 7
1194
+ },
1195
+ {
1196
+ "type": "text",
1197
+ "text": "Scaling Factor of Movable Strategy. We perform an ablation study on the scaling factor of the movable strategy and further investigate the effectiveness of the inner cost in Tab. 5. Notably, it is observed that there exists a conflict between the inner loss and the global search strategy, behaving a sharp drop when only reserving the inner loss. Furthermore, searching within the grid enables the detector to more attend to small objects and avoid a drastic deterioration in normal object detection. See Appendix E for more detailed analyses. Scaling Factor of SDG. We also compare our side-directed manner with the standard offset prediction method in Tab. 6. Based on PECA, the side-directed scaling factor may limit the detector on small object detection but significantly promote the performance on other objects. This phenomenon would be broken without the help of PECA in which a preci",
1198
+ "bbox": [
1199
+ 75,
1200
+ 672,
1201
+ 470,
1202
+ 902
1203
+ ],
1204
+ "page_idx": 7
1205
+ },
1206
+ {
1207
+ "type": "text",
1208
+ "text": "itous decline is emerged on all-scale object detection (row 5 in Tab. 4 vs. row 1 in Tab. 6). We hypothesise that it because the predicted reference points may be outside of the proposal boxes, or even the region of the image.",
1209
+ "bbox": [
1210
+ 496,
1211
+ 366,
1212
+ 890,
1213
+ 428
1214
+ ],
1215
+ "page_idx": 7
1216
+ },
1217
+ {
1218
+ "type": "text",
1219
+ "text": "5. Conclusion",
1220
+ "text_level": 1,
1221
+ "bbox": [
1222
+ 500,
1223
+ 441,
1224
+ 617,
1225
+ 458
1226
+ ],
1227
+ "page_idx": 7
1228
+ },
1229
+ {
1230
+ "type": "text",
1231
+ "text": "In this paper, we propose SAP-DETR for promoting model convergence by treating object detection as a transformation from the salient points to the instance objects. Our SAP-DETR explicitly initializes a query-specific reference point for each object query, gradually aggregates them into an instance object, and predicts the distance from each side of the bounding box. By speedily attending to the query-specific region and other extreme regions from contextual image features, it thus can effectively bridge the gap between the salient points and the query-based Transformer detector. Our extensive experiments have demonstrated that SAP-DETR achieves superior model convergence speed. With the same training settings, our proposed SAP-DETR outperforms SoTA approaches with large margins.",
1232
+ "bbox": [
1233
+ 496,
1234
+ 469,
1235
+ 893,
1236
+ 680
1237
+ ],
1238
+ "page_idx": 7
1239
+ },
1240
+ {
1241
+ "type": "text",
1242
+ "text": "6. Future Work",
1243
+ "text_level": 1,
1244
+ "bbox": [
1245
+ 500,
1246
+ 696,
1247
+ 635,
1248
+ 712
1249
+ ],
1250
+ "page_idx": 7
1251
+ },
1252
+ {
1253
+ "type": "text",
1254
+ "text": "This point-based design for DETR-like models also comes with remaining issues, in particular regarding training with deformable attention, multi-scale features, and negative query design. Following current center-based methods working for similar issues, we expect future work to successfully address them for point-based design of SAP-DETR.",
1255
+ "bbox": [
1256
+ 496,
1257
+ 722,
1258
+ 893,
1259
+ 813
1260
+ ],
1261
+ "page_idx": 7
1262
+ },
1263
+ {
1264
+ "type": "text",
1265
+ "text": "Acknowledgements",
1266
+ "text_level": 1,
1267
+ "bbox": [
1268
+ 500,
1269
+ 829,
1270
+ 666,
1271
+ 845
1272
+ ],
1273
+ "page_idx": 7
1274
+ },
1275
+ {
1276
+ "type": "text",
1277
+ "text": "We thank the Lenovo Research AI Master platform for computing GPU supports without which this work would not be possible.",
1278
+ "bbox": [
1279
+ 496,
1280
+ 854,
1281
+ 890,
1282
+ 900
1283
+ ],
1284
+ "page_idx": 7
1285
+ },
1286
+ {
1287
+ "type": "page_number",
1288
+ "text": "15546",
1289
+ "bbox": [
1290
+ 480,
1291
+ 944,
1292
+ 519,
1293
+ 955
1294
+ ],
1295
+ "page_idx": 7
1296
+ },
1297
+ {
1298
+ "type": "text",
1299
+ "text": "References",
1300
+ "text_level": 1,
1301
+ "bbox": [
1302
+ 78,
1303
+ 89,
1304
+ 173,
1305
+ 104
1306
+ ],
1307
+ "page_idx": 8
1308
+ },
1309
+ {
1310
+ "type": "list",
1311
+ "sub_type": "ref_text",
1312
+ "list_items": [
1313
+ "[1] Navaneeth Bodla, Bharat Singh, Rama Chellappa, and Larry S Davis. Soft-nms-improving object detection with one line of code. In ICCV, pages 5561-5569, 2017. 3",
1314
+ "[2] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In ECCV, pages 213-229. Springer, 2020. 1, 3, 4, 5, 6, 7, 10",
1315
+ "[3] Liangyu Chen, Tong Yang, Xiangyu Zhang, Wei Zhang, and Jian Sun. Points as queries: Weakly semi-supervised object detection by points. In CVPR, pages 8823-8832, 2021. 3",
1316
+ "[4] Qiang Chen, Xiaokang Chen, Gang Zeng, and Jingdong Wang. Group detr:fast training convergence with deco-upplied one-to-many label assignment. arXiv preprint arXiv:2207.13085, 2022. 3, 6, 8",
1317
+ "[5] Peng Gao, Minghang Zheng, Xiaogang Wang, Jifeng Dai, and Hongsheng Li. Fast convergence of detr with spatially modulated co-attention. In ICCV, pages 3621–3630, 2021. 2, 3, 5, 6, 7, 10",
1318
+ "[6] Zheng Ge, Songtao Liu, Feng Wang, Zeming Li, and Jian Sun. Yolox: Exceeding yolo series in 2021. arXiv preprint arXiv:2107.08430, 2021. 1",
1319
+ "[7] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 5",
1320
+ "[8] Hei Law and Jia Deng. Cornernet: Detecting objects as paired keypoints. In ECCV, pages 734-750, 2018. 3",
1321
+ "[9] Feng Li, Hao Zhang, Shilong Liu, Jian Guo, Lionel M Ni, and Lei Zhang. Dn-detr: Accelerate detr training by introducing query denoising. In CVPR, pages 13619-13627, 2022. 3, 5, 6, 8, 13",
1322
+ "[10] Tsung-Yi Lin, Piotr Dollar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In CVPR, pages 2117-2125, 2017. 7",
1323
+ "[11] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólar. Focal loss for dense object detection. In ICCV, pages 2980-2988, 2017. 1, 3",
1324
+ "[12] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, pages 740-755, 2014. 5",
1325
+ "[13] Shilong Liu, Feng Li, Hao Zhang, Xiao Yang, Xianbiao Qi, Hang Su, Jun Zhu, and Lei Zhang. Dab-detr: Dynamic anchor boxes are better queries for detr. ICLR, 2022. 2, 3, 4, 5, 6, 7, 10",
1326
+ "[14] Wei Liu, Dragomir Anguelov, Dumitru Erhan, Christian Szegedy, Scott Reed, Cheng-Yang Fu, and Alexander C Berg. Ssd: Single shot multibox detector. In ECCV, pages 21-37. Springer, 2016. 1",
1327
+ "[15] Yang Liu, Yao Zhang, Yixin Wang, Feng Hou, Jin Yuan, Jiang Tian, Yang Zhang, Zhongchao Shi, Jianping Fan, and Zhiqiang He. A survey of visual transformers. arXiv preprint arXiv:2111.06091, 2021. 3",
1328
+ "[16] Depu Meng, Xiaokang Chen, Zejia Fan, Gang Zeng, Houqiang Li, Yuhui Yuan, Lei Sun, and Jingdong Wang. Con"
1329
+ ],
1330
+ "bbox": [
1331
+ 78,
1332
+ 114,
1333
+ 470,
1334
+ 900
1335
+ ],
1336
+ "page_idx": 8
1337
+ },
1338
+ {
1339
+ "type": "list",
1340
+ "sub_type": "ref_text",
1341
+ "list_items": [
1342
+ "ditional detr for fast training convergence. In ICCV, pages 3651-3660, 2021. 2, 3, 4, 5, 6, 7, 10, 11",
1343
+ "[17] Joseph Redmon and Ali Farhadi. Yolov3: An incremental improvement. arXiv preprint arXiv:1804.02767, 2018. 2",
1344
+ "[18] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. NeurIPS, 28, 2015. 1, 6, 7",
1345
+ "[19] Peize Sun, Yi Jiang, Enze Xie, Wenqi Shao, Zehuan Yuan, Changhu Wang, and Ping Luo. What makes for end-to-end object detection? In ICML, pages 9934-9944, 2021. 2",
1346
+ "[20] Peize Sun, Rufeng Zhang, Yi Jiang, Tao Kong, Chenfeng Xu, Wei Zhan, Masayoshi Tomizuka, Lei Li, et al. Sparse r-cnn: End-to-end object detection with learnable proposals. arXiv preprint arXiv:2011.12450, 2020. 1",
1347
+ "[21] Zhiqing Sun, Shengcao Cao, Yiming Yang, and Kris M Kitani. Rethinking transformer-based set prediction for object detection. In ICCV, pages 3611-3620, 2021. 3",
1348
+ "[22] Zhi Tian, Chunhua Shen, Hao Chen, and Tong He. Fcos: Fully convolutional one-stage object detection. In ICCV, pages 9627-9636, 2019. 1, 3, 4",
1349
+ "[23] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of machine learning research, 9(11), 2008. 12",
1350
+ "[24] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. NeurIPS, 30, 2017. 1, 10",
1351
+ "[25] Yingming Wang, Xiangyu Zhang, Tong Yang, and Jian Sun. Anchor detr: Query design for transformer-based detector. In AAAI, volume 36, pages 2567-2575, 2022. 2, 3, 4, 6, 7, 10, 12",
1352
+ "[26] Gongjie Zhang, Zhipeng Luo, Yingchen Yu, Kaiwen Cui, and Shijian Lu. Accelerating detr convergence via semantic-aligned matching. In CVPR, pages 949-958, 2022. 3, 6, 7, 10",
1353
+ "[27] Gongjie Zhang, Zhipeng Luo, Yingchen Yu, Jiaxing Huang, Kaiwen Cui, Shijian Lu, and Eric P Xing. Semantic-aligned matching for enhanced detr convergence and multi-scale feature fusion. arXiv preprint arXiv:2207.14172, 2022. 3",
1354
+ "[28] Hao Zhang, Feng Li, Shilong Liu, Lei Zhang, Hang Su, Jun Zhu, Lionel M Ni, and Heung-Yeung Shum. Dino: Detr with improved denoising anchor boxes for end-to-end object detection. ICLR, 2022. 3, 5, 6, 8",
1355
+ "[29] Xingyi Zhou, Dequan Wang, and Philipp Krahenbuhl. Objects as points. In arXiv preprint arXiv:1904.07850, 2019. 3",
1356
+ "[30] Xingyi Zhou, Jiacheng Zhuo, and Philipp Krahenbuhl. Bottom-up object detection by grouping extreme and center points. In CVPR, pages 850-859, 2019. 3",
1357
+ "[31] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. In ICLR, 2020. 2, 3, 6, 7, 10"
1358
+ ],
1359
+ "bbox": [
1360
+ 503,
1361
+ 92,
1362
+ 890,
1363
+ 816
1364
+ ],
1365
+ "page_idx": 8
1366
+ },
1367
+ {
1368
+ "type": "page_number",
1369
+ "text": "15547",
1370
+ "bbox": [
1371
+ 480,
1372
+ 944,
1373
+ 517,
1374
+ 955
1375
+ ],
1376
+ "page_idx": 8
1377
+ }
1378
+ ]
2023/SAP-DETR_ Bridging the Gap Between Salient Points and Queries-Based Transformer Detector for Fast Model Convergency/909bbbe4-4e26-472f-ab37-1d60d8bab4b6_model.json ADDED
@@ -0,0 +1,1824 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ [
2
+ [
3
+ {
4
+ "type": "header",
5
+ "bbox": [
6
+ 0.107,
7
+ 0.003,
8
+ 0.182,
9
+ 0.043
10
+ ],
11
+ "angle": 0,
12
+ "content": "CVF"
13
+ },
14
+ {
15
+ "type": "header",
16
+ "bbox": [
17
+ 0.237,
18
+ 0.001,
19
+ 0.812,
20
+ 0.047
21
+ ],
22
+ "angle": 0,
23
+ "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore."
24
+ },
25
+ {
26
+ "type": "title",
27
+ "bbox": [
28
+ 0.116,
29
+ 0.13,
30
+ 0.853,
31
+ 0.177
32
+ ],
33
+ "angle": 0,
34
+ "content": "SAP-DETR: Bridging the Gap between Salient Points and Queries-Based Transformer Detector for Fast Model Convergence"
35
+ },
36
+ {
37
+ "type": "text",
38
+ "bbox": [
39
+ 0.186,
40
+ 0.202,
41
+ 0.783,
42
+ 0.24
43
+ ],
44
+ "angle": 0,
45
+ "content": "Yang Liu\\(^{1,3}\\) * Yao Zhang\\(^{1,3}\\) * Yixin Wang\\(^{2}\\) * Yang Zhang\\(^{4}\\) Jiang Tian\\(^{4}\\) Zhongchao Shi\\(^{4}\\) Jianping Fan\\(^{4}\\) Zhiqiang He\\(^{1,3,5\\dagger}\\)"
46
+ },
47
+ {
48
+ "type": "text",
49
+ "bbox": [
50
+ 0.1,
51
+ 0.239,
52
+ 0.87,
53
+ 0.275
54
+ ],
55
+ "angle": 0,
56
+ "content": "\\(^{1}\\)Institute of Computing Technology (ICT), Chinese Academy of Sciences \\(^{2}\\)Stanford University \\(^{3}\\)University of Chinese Academy of Sciences \\(^{4}\\)AI Lab, Lenovo Research \\(^{5}\\)Lenovo Ltd."
57
+ },
58
+ {
59
+ "type": "text",
60
+ "bbox": [
61
+ 0.194,
62
+ 0.277,
63
+ 0.766,
64
+ 0.293
65
+ ],
66
+ "angle": 0,
67
+ "content": "{liuyang20c,zhangyao215}@mails.ucas.ac.cn yxinwang@stanford.edu"
68
+ },
69
+ {
70
+ "type": "text",
71
+ "bbox": [
72
+ 0.248,
73
+ 0.295,
74
+ 0.719,
75
+ 0.31
76
+ ],
77
+ "angle": 0,
78
+ "content": "{zhangyang20,tianjiang1,shizc2,jfan1,hezq}@lenovo.com"
79
+ },
80
+ {
81
+ "type": "title",
82
+ "bbox": [
83
+ 0.235,
84
+ 0.345,
85
+ 0.314,
86
+ 0.361
87
+ ],
88
+ "angle": 0,
89
+ "content": "Abstract"
90
+ },
91
+ {
92
+ "type": "text",
93
+ "bbox": [
94
+ 0.075,
95
+ 0.377,
96
+ 0.474,
97
+ 0.754
98
+ ],
99
+ "angle": 0,
100
+ "content": "Recently, the dominant DETR-based approaches apply central-concept spatial prior to accelerating Transformer detector convergence. These methods gradually refine the reference points to the center of target objects and imbue object queries with the updated central reference information for spatially conditional attention. However, centralizing reference points may severely deteriorate queries' saliency and confuse detectors due to the indiscriminative spatial prior. To bridge the gap between the reference points of salient queries and Transformer detectors, we propose SAlient Point-based DETR (SAP-DETR) by treating object detection as a transformation from salient points to instance objects. Concretely, we explicitly initialize a query-specific reference point for each object query, gradually aggregate them into an instance object, and then predict the distance from each side of the bounding box to these points. By rapidly attending to query-specific reference regions and the conditional box edges, SAP-DETR can effectively bridge the gap between the salient point and the query-based Transformer detector with a significant convergence speed. Experimentally, SAP-DETR achieves \\(1.4 \\times\\) convergence speed with competitive performance and stably promotes the SoTA approaches by \\(\\sim 1.0\\) AP. Based on ResNet-DC-101, SAP-DETR achieves 46.9 AP. The code will be released at https://github.com/liuyang-ict/SAP-DETR."
101
+ },
102
+ {
103
+ "type": "title",
104
+ "bbox": [
105
+ 0.078,
106
+ 0.783,
107
+ 0.21,
108
+ 0.799
109
+ ],
110
+ "angle": 0,
111
+ "content": "1. Introduction"
112
+ },
113
+ {
114
+ "type": "text",
115
+ "bbox": [
116
+ 0.076,
117
+ 0.809,
118
+ 0.472,
119
+ 0.854
120
+ ],
121
+ "angle": 0,
122
+ "content": "Object detection is a fundamental task in computer vision, whose target is to recognize and localize each object from input images. In the last decade, various detec"
123
+ },
124
+ {
125
+ "type": "image",
126
+ "bbox": [
127
+ 0.508,
128
+ 0.346,
129
+ 0.89,
130
+ 0.53
131
+ ],
132
+ "angle": 0,
133
+ "content": null
134
+ },
135
+ {
136
+ "type": "image_caption",
137
+ "bbox": [
138
+ 0.498,
139
+ 0.544,
140
+ 0.895,
141
+ 0.67
142
+ ],
143
+ "angle": 0,
144
+ "content": "Figure 1. Comparison of SAP-DETR and DAB-DETR under 3-layer decoder model and 36-epoch training scheme. (a) Statistics of the query count in different classification score intervals. (b) and (c) Visualize the distribution of all reference points (pink) and highlight the top-20 classification score queries with their bounding boxes (blue) and reference points (red) in different decoder layers. (d) Visualize the outputs of positive queries and ground truth (red) during the training process, each query has a representative color for its reference point and bounding box. Best viewed in color."
145
+ },
146
+ {
147
+ "type": "text",
148
+ "bbox": [
149
+ 0.498,
150
+ 0.688,
151
+ 0.895,
152
+ 0.854
153
+ ],
154
+ "angle": 0,
155
+ "content": "tors [6,11,14,18,20,22] based on Convolutional Neural Networks (CNNs), have received widespread attention and made significant progress. Recently, Carion et al. [2] proposed a new end-to-end paradigm for object detection based on the Transformer [24], called DEtction TRansformer (DETR), which treats object detection as a problem of set prediction. In DETR, a set of learnable positional encodings, namely object queries, are employed to aggregate instance features from the context image in Transformer Decoder. The predictions of queries are finally assigned to the ground truth via bipartite matching to achieve end-to-end detection."
156
+ },
157
+ {
158
+ "type": "text",
159
+ "bbox": [
160
+ 0.498,
161
+ 0.856,
162
+ 0.895,
163
+ 0.903
164
+ ],
165
+ "angle": 0,
166
+ "content": "Despite the promising results of DETR, its application is largely limited by considerably longer training time compared to conventional CNNs. To address this problem, many"
167
+ },
168
+ {
169
+ "type": "page_footnote",
170
+ "bbox": [
171
+ 0.076,
172
+ 0.863,
173
+ 0.47,
174
+ 0.888
175
+ ],
176
+ "angle": 0,
177
+ "content": "*This work was done when working as an intern at AI Lab, Lenovo Research, Beijing, China."
178
+ },
179
+ {
180
+ "type": "page_footnote",
181
+ "bbox": [
182
+ 0.098,
183
+ 0.888,
184
+ 0.222,
185
+ 0.9
186
+ ],
187
+ "angle": 0,
188
+ "content": "† Corresponding author."
189
+ },
190
+ {
191
+ "type": "list",
192
+ "bbox": [
193
+ 0.076,
194
+ 0.863,
195
+ 0.47,
196
+ 0.9
197
+ ],
198
+ "angle": 0,
199
+ "content": null
200
+ },
201
+ {
202
+ "type": "page_number",
203
+ "bbox": [
204
+ 0.481,
205
+ 0.945,
206
+ 0.519,
207
+ 0.957
208
+ ],
209
+ "angle": 0,
210
+ "content": "15539"
211
+ }
212
+ ],
213
+ [
214
+ {
215
+ "type": "text",
216
+ "bbox": [
217
+ 0.077,
218
+ 0.092,
219
+ 0.473,
220
+ 0.727
221
+ ],
222
+ "angle": 0,
223
+ "content": "variants attempted to take a close look at query paradigm and introduced various spatial priors for model convergence and efficacy. According to the type of spatial prior, they can be categorized into implicit and explicit methods. The implicit ones [5, 16, 31] attempt to decouple a reference point from the object query and make use of this spatial prior to attend to the image context features efficiently. The current state-of-the-arts (SoTAs) are dominated by the explicit ones [13, 25], which suggest to instantiate a position with spatial prior for each query, i.e., explicit reference coordinates with a center point or an anchor box. These reference coordinates serve as helpful priors and enable the queries to focus on their expected regions easily. For instance, Anchor DETR [25] introduced an anchor concept (center point with different box size patterns) to formulate the query position and directly regressed the central offsets of the bounding boxes. DABDETR [13] further stretched the center point to a 4D anchor box concept \\([cx, cy, w, h]\\) to refine proposal bonding boxes in a cascaded manner. However, instantiating the query location as a target center may severely degrade the classification accuracy and convergence speed. As illustrated in Fig. 1, there exist many plausible queries [19] with high-quality classification scores (Fig. 1(a) within red box) and box Intersection over Union (IoU, see the redundant blue boxes in Fig. 1(b) and (c)), which only brings a slight improvement on precision rate but inevitably confuses the detector on the positive query assignments when training with bipartite matching strategy. This is because the plausible predictions are considered in negative classification loss, which severely decelerates the model convergence. As shown in Fig. 1(b) and (c), the predefined reference point of the positive query may not be the nearest one to the center of the ground truth bounding box, and the reference points tend to be centralized or marginalized (cyan arrows in Fig. 1(b)), hence losing the spatial specificity. With further insight into the one-to-one label assignment during the training process, we find that the query, whose reference point is closest to the center point, also has a high-quality IoU, but it still exists a disparity with the positive query in the classification confidence. Therefore, we argue that such a centralized spatial prior may cause degeneration of target consistency in both classification and localization tasks, which leads to inconsistent predictions."
224
+ },
225
+ {
226
+ "type": "text",
227
+ "bbox": [
228
+ 0.077,
229
+ 0.736,
230
+ 0.473,
231
+ 0.902
232
+ ],
233
+ "angle": 0,
234
+ "content": "Furthermore, the mentioned central point-based variants also have difficulties in detecting occluded objects. For example, Fig. 1(d) shows that DAB-DETR detects the left baseman twice, while the query point in SAP-DETR is not necessarily the center point, so the query point of the bounding box for the occluded baseman is from a pixel on the occluded baseman on the top right area instead of from the left baseman. One solution for center-based method [25] is to predefine different receptive fields (similar to the scaling anchor box in YOLO [17]) for the position of each query. However, increasing the diversity of the receptive fields for"
235
+ },
236
+ {
237
+ "type": "text",
238
+ "bbox": [
239
+ 0.498,
240
+ 0.092,
241
+ 0.895,
242
+ 0.137
243
+ ],
244
+ "angle": 0,
245
+ "content": "each position query is unsuitable for non-overlapped targets, as it still generates massive indistinguishable predictions for one position as same as other center-based models."
246
+ },
247
+ {
248
+ "type": "text",
249
+ "bbox": [
250
+ 0.498,
251
+ 0.139,
252
+ 0.897,
253
+ 0.517
254
+ ],
255
+ "angle": 0,
256
+ "content": "To bridge these gaps, in this paper, we present a novel framework for Transformer detector, called SAclient Point-based DETR (SAP-DETR), which treats object detection as a transformation from salient points to instance objects. Instead of regressing the reference point to the target center, we define the reference point belonging to one positive query as a salient point, keep this query-specific spatial prior with a scaling amplitude, and then gradually update them to an instance object by predicting the distance from each side of the bounding box. Specifically, we tile the mesh-grid referenced points and initialize their center/corner as the query-specific reference point. To disentangle the reference sparsity as well as stabilize the training process, a movable strategy with scaling amplitude is applied for reference point adjustment, which prompts queries to consider their reference grid as the salient region to perform image context attention. By localizing each side of the bounding box layer by layer, such query-specific spatial prior enables compensation for the over-smooth/inadequacy problem during center-based detection, thereby vastly promoting model convergence speed. Inspired by [5, 13, 16], we also take advantage of both Gaussian spatial prior and conditional cross-attention mechanism, and then a salient point enhanced cross-attention mechanism is developed to distinguish the salient region and other conditional extreme regions from the context image features."
257
+ },
258
+ {
259
+ "type": "text",
260
+ "bbox": [
261
+ 0.498,
262
+ 0.518,
263
+ 0.895,
264
+ 0.639
265
+ ],
266
+ "angle": 0,
267
+ "content": "We bridge the gap between salient points and query-based Transformer detector by speedily attending to the query-specific region and other conditional regions. The extensive experiments have shown that SAP-DETR achieves superior convergence speed and performance. To the best of our knowledge, this is the first work to introduce the salient point based regression into end-to-end query-based Transformer detectors. Our contributions can be summarized as follows."
268
+ },
269
+ {
270
+ "type": "text",
271
+ "bbox": [
272
+ 0.498,
273
+ 0.641,
274
+ 0.897,
275
+ 0.807
276
+ ],
277
+ "angle": 0,
278
+ "content": "1) We introduce the salient point concept into query-based Transformer detectors by assigning query-specific reference points to object queries. Unlike center-based methods, we restrict the reference location and define the point of the positive query as the salient one, hence enlarging the discrepancy of query as well as reducing the redundant predictions (see Fig. 1). Thanks to the efficacy of the query-specific prior, our SAP-DETR accelerates the convergence speed greatly, achieving competitive performance with \\(30\\%\\) fewer training epochs. The proposed movable strategy further boosts SAP-DETR to a new SoTA performance."
279
+ },
280
+ {
281
+ "type": "text",
282
+ "bbox": [
283
+ 0.498,
284
+ 0.809,
285
+ 0.895,
286
+ 0.854
287
+ ],
288
+ "angle": 0,
289
+ "content": "2) We devise a point-enhanced cross-attention mechanism to imbue query with spatial prior based on both reference point and box sides for final specific region attention."
290
+ },
291
+ {
292
+ "type": "text",
293
+ "bbox": [
294
+ 0.498,
295
+ 0.856,
296
+ 0.895,
297
+ 0.901
298
+ ],
299
+ "angle": 0,
300
+ "content": "3) Evaluation over COCO dataset has demonstrated that SAP-DETR achieves superior convergence speed and detection accuracy. Under the same training settings, SAP-DETR"
301
+ },
302
+ {
303
+ "type": "list",
304
+ "bbox": [
305
+ 0.498,
306
+ 0.641,
307
+ 0.897,
308
+ 0.901
309
+ ],
310
+ "angle": 0,
311
+ "content": null
312
+ },
313
+ {
314
+ "type": "page_number",
315
+ "bbox": [
316
+ 0.481,
317
+ 0.945,
318
+ 0.521,
319
+ 0.957
320
+ ],
321
+ "angle": 0,
322
+ "content": "15540"
323
+ }
324
+ ],
325
+ [
326
+ {
327
+ "type": "text",
328
+ "bbox": [
329
+ 0.078,
330
+ 0.092,
331
+ 0.44,
332
+ 0.108
333
+ ],
334
+ "angle": 0,
335
+ "content": "outperforms the SoTA approaches with a large margin."
336
+ },
337
+ {
338
+ "type": "title",
339
+ "bbox": [
340
+ 0.078,
341
+ 0.121,
342
+ 0.219,
343
+ 0.136
344
+ ],
345
+ "angle": 0,
346
+ "content": "2. Related Work"
347
+ },
348
+ {
349
+ "type": "text",
350
+ "bbox": [
351
+ 0.076,
352
+ 0.146,
353
+ 0.471,
354
+ 0.327
355
+ ],
356
+ "angle": 0,
357
+ "content": "Anchor-Free Object Detectors. Classical anchor-free object detectors can be grouped into center-based and keypoint-based approaches. The center-based approaches aim to localize the target objects based on the central locations [11] or predefined ROI [22]. For example, FCOS [22] treated all points within the bounding box as positive ones to predict their distances from each side \\(([\\ell, t, r, b])\\), and a centerness score was then considered to prohibit the low-quality prediction whose point is located near the border. Compared with FCOS, we also restrict the candidate queries within the bounding box but treat only one as positive to perform end-to-end object detection via an inner matching cost."
358
+ },
359
+ {
360
+ "type": "text",
361
+ "bbox": [
362
+ 0.076,
363
+ 0.328,
364
+ 0.47,
365
+ 0.493
366
+ ],
367
+ "angle": 0,
368
+ "content": "The target of keypoint-based approaches is to localize the specific object locations and assign them to the predefined keypoints of the object for box localized training. For instance, diagonal corner points were considered in CornerNet [8], center point was further grouped into CenterNet [29], and ExtremeNet [30] added some conjectural extreme points for object localization. These works showed an impressive performance, but the complicated keypoint matching may limit their upper bound. Our SAP-DETR takes the advantage of salient point regression to focus on the distinct regions without complicated point-based supervision."
369
+ },
370
+ {
371
+ "type": "text",
372
+ "bbox": [
373
+ 0.076,
374
+ 0.493,
375
+ 0.47,
376
+ 0.599
377
+ ],
378
+ "angle": 0,
379
+ "content": "Query-Based Transformer Detectors. DETR [2] pioneered a new paradigm of Transformer detector for end-to-end object detection without any post-processing [1]. In DETR, a new representation, namely object query, aggregates the instance features and then yields a detection result for each instance object [15]. Following DETR, many voterists put efforts on the optimization of convergence and accuracy."
380
+ },
381
+ {
382
+ "type": "text",
383
+ "bbox": [
384
+ 0.076,
385
+ 0.6,
386
+ 0.471,
387
+ 0.902
388
+ ],
389
+ "angle": 0,
390
+ "content": "Sun et al. [21] revealed that the main reason for slow convergence of DETR is attributed to the Transformer decoder, and they considered an encoder-only structure to alleviate such a problem. For in-depth understanding of the object query, one way is to generate a series of implicit spatial priors from queries to guide feature aggregation in cross-attention layers. SMCA [5] applied a Gaussian-like attention map to augment the query-concerned features spatially. The reference point concept was first introduced by Deformable DETR [31], where the sampling offsets are predicted by each reference point to perform deformable cross-attention. Following such a concept, Conditional DETR [16] reformulated the attention operation and rebuilt positional queries based on the reference points to facilitate extreme region discrimination. Another way is towards position-instantiation explicitly, where this position information enables to directly conduct positional query generation. Anchor DETR [25] utilized a predefined 2D anchor point \\([cx, cy]\\) to explicitly capitalize on the spatial prior during cross-attention and box regression. DAB-DETR [13] extended such a 2D concept to"
391
+ },
392
+ {
393
+ "type": "image",
394
+ "bbox": [
395
+ 0.532,
396
+ 0.088,
397
+ 0.865,
398
+ 0.37
399
+ ],
400
+ "angle": 0,
401
+ "content": null
402
+ },
403
+ {
404
+ "type": "image_caption",
405
+ "bbox": [
406
+ 0.499,
407
+ 0.379,
408
+ 0.895,
409
+ 0.478
410
+ ],
411
+ "angle": 0,
412
+ "content": "Figure 2. Illustration of SAP-DETR. Each object query in SAP-DETR is assigned to a specific grid region and initialized by the corner/center of the grid as its reference point. A learnable 4D coordinate represents the distance from the four sides of the box to the reference point. Both reference points and box sides are served as positional encodings added/concatenated to content embeddings. All embeddings are refined to predict target objects gradually."
413
+ },
414
+ {
415
+ "type": "text",
416
+ "bbox": [
417
+ 0.5,
418
+ 0.496,
419
+ 0.89,
420
+ 0.512
421
+ ],
422
+ "angle": 0,
423
+ "content": "a 4D anchor box \\([cx, cy, w, h]\\) and refined it layer-by-layer."
424
+ },
425
+ {
426
+ "type": "text",
427
+ "bbox": [
428
+ 0.498,
429
+ 0.512,
430
+ 0.895,
431
+ 0.663
432
+ ],
433
+ "angle": 0,
434
+ "content": "The recent accelerating convergence methods are based on auxiliary queries for facilitating detector discrimination. DN-DETR [9] demonstrates the slow model convergence is mainly caused by the instability of bipartite matching, thus providing a denoising training to eliminate this issue. DINO [28] inherits this advance and further introduces negative queries to perform contrastive denoising. Group-DETR [4] proposes a group-wise one-to-many label assignment to match multiple positive object queries with more gradients for fast DETR convergence."
435
+ },
436
+ {
437
+ "type": "text",
438
+ "bbox": [
439
+ 0.498,
440
+ 0.663,
441
+ 0.895,
442
+ 0.815
443
+ ],
444
+ "angle": 0,
445
+ "content": "The most relevant approaches to ours are Point-DETR [3] and SAM-DETR [26, 27]. The former applied a point encoder for annotated point label infusion in teacher model, and the latter directly updated content embeddings by extracting salient points from image features for query-image semantic alignment. Unlike these concepts, we redefine the salient point from the perspective of the positive query's position and replace the center-concept prior with the query-specific position, thereby attending extreme regions, differentiating queries' saliency, and alleviating redundant predictions."
446
+ },
447
+ {
448
+ "type": "title",
449
+ "bbox": [
450
+ 0.5,
451
+ 0.83,
452
+ 0.591,
453
+ 0.845
454
+ ],
455
+ "angle": 0,
456
+ "content": "3. Method"
457
+ },
458
+ {
459
+ "type": "text",
460
+ "bbox": [
461
+ 0.499,
462
+ 0.856,
463
+ 0.894,
464
+ 0.901
465
+ ],
466
+ "angle": 0,
467
+ "content": "We propose SAP-DETR to bridge the gap between salient points and query-based detectors. Following DETR, the extracted image features are fed into Transformer encoder"
468
+ },
469
+ {
470
+ "type": "page_number",
471
+ "bbox": [
472
+ 0.481,
473
+ 0.945,
474
+ 0.518,
475
+ 0.957
476
+ ],
477
+ "angle": 0,
478
+ "content": "15541"
479
+ }
480
+ ],
481
+ [
482
+ {
483
+ "type": "text",
484
+ "bbox": [
485
+ 0.076,
486
+ 0.092,
487
+ 0.473,
488
+ 0.273
489
+ ],
490
+ "angle": 0,
491
+ "content": "after adding positional encodings, and then re-aggregated by object queries in Transformer decoder. In pursuit of the query-specific prior, we dispense a movable strategy for each query based on a fixed grid region. The query, whose reference region overlaps with ground truth objects, is allowed to predict the relative offsets from four sides of the bounding box to the points. Given the query-specific reference point and the proposal box sides, we propose salient point enhanced cross-attention mechanism to imbue query with spatial prior, thereby attending to extreme regions effectively. Additionally, we discuss two common issues in DETR-like models and address them for further improvements."
492
+ },
493
+ {
494
+ "type": "title",
495
+ "bbox": [
496
+ 0.077,
497
+ 0.282,
498
+ 0.406,
499
+ 0.297
500
+ ],
501
+ "angle": 0,
502
+ "content": "3.1. Salient Points-Based Object Detection"
503
+ },
504
+ {
505
+ "type": "text",
506
+ "bbox": [
507
+ 0.076,
508
+ 0.304,
509
+ 0.474,
510
+ 0.593
511
+ ],
512
+ "angle": 0,
513
+ "content": "Overview. Previous methods [13, 16, 25] normally decompose the object query into both content and position embeddings (queries), and form a center-based anchor point/box prior on the position ones. Unlike the central concept, we tile a fixed mesh-grid region, initialize their left-top corner as the reference points with 2D coordinate \\( \\boldsymbol{r} = \\{x, y\\} \\in [0,1]^2 \\), and instantiate a learnable 4D offset distance \\( s = \\{\\ell, t, r, b\\} \\in [0,1]^4 \\) from the reference point to the sides of proposal bounding box for each object query. The object query can be referred as \\( q = \\{\\mathbf{e}; r, s\\} \\), where \\( \\mathbf{e} \\in \\mathbb{R}^d \\) is the content embedding with \\( d \\) dimension. Instead of regressing the center, width, and height of a bounding box, we follow FCOS [22] and directly supervise the 4D offset from the four sides of a bounding box to the reference point. The final box prediction is formulated as \\( \\hat{\\boldsymbol{b}} = \\{\\hat{x} - \\hat{\\ell}, \\hat{y} - \\hat{\\bar{t}}, \\hat{x} + \\hat{r}, \\hat{y} + \\hat{b}\\} \\). Ideally, we here fix the reference point \\( \\{\\hat{x}, \\hat{y}\\} = \\{x, y\\} \\) (the movable update strategy is introduced in the next subsection) and only update the 4D box side prediction layer by layer. The prediction for each decoder layer can be calculated by"
514
+ },
515
+ {
516
+ "type": "equation",
517
+ "bbox": [
518
+ 0.155,
519
+ 0.603,
520
+ 0.395,
521
+ 0.619
522
+ ],
523
+ "angle": 0,
524
+ "content": "\\[\n\\Delta \\boldsymbol {s} _ {l} = \\operatorname {B o x H e a d} _ {l} (\\boldsymbol {s} _ {l - 1}, \\boldsymbol {e} _ {l - 1}, \\boldsymbol {r} _ {l - 1}),\n\\]"
525
+ },
526
+ {
527
+ "type": "equation",
528
+ "bbox": [
529
+ 0.118,
530
+ 0.622,
531
+ 0.431,
532
+ 0.639
533
+ ],
534
+ "angle": 0,
535
+ "content": "\\[\n\\hat {\\boldsymbol {s}} _ {l} = \\sigma \\left(\\sigma^ {- 1} \\left(\\boldsymbol {s} _ {l - 1}\\right) + \\Delta \\boldsymbol {s} _ {l}\\right), \\quad \\boldsymbol {s} _ {l} = \\operatorname {D e t a c h} (\\hat {\\boldsymbol {s}} _ {l}),\n\\]"
536
+ },
537
+ {
538
+ "type": "equation",
539
+ "bbox": [
540
+ 0.104,
541
+ 0.644,
542
+ 0.47,
543
+ 0.674
544
+ ],
545
+ "angle": 0,
546
+ "content": "\\[\n\\boldsymbol {r} _ {l} = \\hat {\\boldsymbol {r}} _ {l} = \\boldsymbol {r} _ {l - 1}, \\quad \\hat {\\boldsymbol {b}} _ {l} = \\left\\{\\hat {\\boldsymbol {r}} _ {l} - \\hat {\\boldsymbol {s}} _ {l} [: 2 ], \\hat {\\boldsymbol {r}} _ {l} + \\hat {\\boldsymbol {s}} _ {l} [2: ] \\right\\}, \\tag {1}\n\\]"
547
+ },
548
+ {
549
+ "type": "text",
550
+ "bbox": [
551
+ 0.076,
552
+ 0.675,
553
+ 0.472,
554
+ 0.78
555
+ ],
556
+ "angle": 0,
557
+ "content": "where \\(\\sigma\\) and \\(\\sigma^{-1}\\) are the sigmoid and inverse sigmoid operation, respectively. \\(\\Delta s_{l}\\) denotes the side offset prediction. \\(\\hat{s}_l\\), \\(\\hat{r}_l\\), and \\(\\hat{\\pmb{b}}_l\\) are the predicted side distance, reference points, and box location from the \\(l\\) decoder layer, respectively. The BoxHead is the prediction head following the layer-\\(l\\) decoder, which is independent between different decoder layers in our settings. Detach operation follows DAB-DETR [13]."
558
+ },
559
+ {
560
+ "type": "text",
561
+ "bbox": [
562
+ 0.076,
563
+ 0.781,
564
+ 0.473,
565
+ 0.902
566
+ ],
567
+ "angle": 0,
568
+ "content": "During the training process, each query is only allowed to predict the bounding boxes that overlap its reference region. We adapt this rule into the one-to-one bipartite matching process via an inner matching cost \\(\\mathcal{L}_{\\mathrm{inner}}\\). Given \\(N\\) queries \\(\\mathbf{Q} = \\{\\pmb{q}_1,\\pmb{q}_2,\\dots ,\\pmb{q}_N\\}\\) and \\(M\\) ground truth objects \\(\\mathbf{G} = \\{\\pmb{g}_1,\\pmb{g}_2,\\dots ,\\pmb{g}_M\\}\\), the \\(\\mathcal{L}_{\\mathrm{inner}}(\\pmb{g}_i,\\pmb{q}_j)\\) of each query-box pair is a step function to penalize the reference point \\(\\pmb{r}_j\\) of \\(\\pmb{q}_j\\) with value of \\(k\\) when \\(\\pmb{r}_j\\) is outside the bounding box of"
569
+ },
570
+ {
571
+ "type": "text",
572
+ "bbox": [
573
+ 0.498,
574
+ 0.092,
575
+ 0.895,
576
+ 0.152
577
+ ],
578
+ "angle": 0,
579
+ "content": "\\(g_{i}\\). We denote \\(i \\in [1, M]\\) and \\(j \\in [1, N]\\) as the index of query and ground truth, respectively. \\(k\\) can be viewed as a penalty cost, and default to \\(10^{5}\\). The final permutation of the one-to-one label assignment is formulated as"
580
+ },
581
+ {
582
+ "type": "equation",
583
+ "bbox": [
584
+ 0.617,
585
+ 0.163,
586
+ 0.781,
587
+ 0.18
588
+ ],
589
+ "angle": 0,
590
+ "content": "\\[\n\\mathcal {L} _ {\\text {i n n e r}} \\left(\\boldsymbol {g} _ {i}, \\boldsymbol {q} _ {j}\\right) := \\mathbf {k} _ {\\boldsymbol {r} _ {j} \\notin \\boldsymbol {g} _ {i}},\n\\]"
591
+ },
592
+ {
593
+ "type": "equation",
594
+ "bbox": [
595
+ 0.596,
596
+ 0.183,
597
+ 0.892,
598
+ 0.223
599
+ ],
600
+ "angle": 0,
601
+ "content": "\\[\n\\hat {\\eta} = \\underset {\\eta \\in \\mathfrak {Y} _ {N}} {\\operatorname {a r g m i n}} \\sum_ {i} ^ {N} \\mathcal {L} _ {\\text {m a t c h}} + \\mathcal {L} _ {\\text {i n n e r}}, \\tag {2}\n\\]"
602
+ },
603
+ {
604
+ "type": "text",
605
+ "bbox": [
606
+ 0.498,
607
+ 0.232,
608
+ 0.892,
609
+ 0.277
610
+ ],
611
+ "angle": 0,
612
+ "content": "where \\(\\mathcal{L}_{\\mathrm{match}}\\) is the original pair-wise matching cost consist of both classification and localization costs [2]. \\(\\eta \\in \\mathfrak{V}_N\\) is a permutation of \\(N\\) elements for bipartite matching."
613
+ },
614
+ {
615
+ "type": "text",
616
+ "bbox": [
617
+ 0.498,
618
+ 0.278,
619
+ 0.895,
620
+ 0.655
621
+ ],
622
+ "angle": 0,
623
+ "content": "Movable Reference Point. Due to the sparseness of the fixed reference point, some small and slender objects may be indistinguishable when there is no reference point inside these objects. Despite the bipartite matching forcing each object to be assigned to one object query, the positive query, whose reference point is outside the assigned bounding box, is unable to accurately regress the distance from each side by a value between 0 and 1. One straightforward solution is to adjust the locations of reference points inside the ground truth bounding boxes to ensure that each object can be detected by an inner reference point. Similar to the aforementioned box refinement, we first perform a movable reference point design to dynamically update the reference points of each query layer by layer. However, such a full-image point regression inevitably expands the search space as vast variable determinations, causing the final reference point to be trapped in an unexpected corner of the bounding box. To reduce the training search spaces, we scale the offset amplitude of points within their specific grid regions, as illustrated in Fig. 3. Such an operation limits the range of offset values, and hence prevents a large searching space. It is implemented by applying the sigmoid activation \\(\\sigma\\) and multiplying a scale factor \\(s_{\\mathrm{grid}}\\) whose value equals to the height and width of one grid. The update process of the reference points is formulated as"
624
+ },
625
+ {
626
+ "type": "equation",
627
+ "bbox": [
628
+ 0.573,
629
+ 0.666,
630
+ 0.822,
631
+ 0.682
632
+ ],
633
+ "angle": 0,
634
+ "content": "\\[\n\\Delta \\boldsymbol {r} _ {l} ^ {\\prime} = \\operatorname {P o i n t H e a d} _ {l} \\left(\\boldsymbol {s} _ {l - 1}, \\boldsymbol {e} _ {l - 1}, \\boldsymbol {r} _ {l - 1}\\right),\n\\]"
635
+ },
636
+ {
637
+ "type": "equation",
638
+ "bbox": [
639
+ 0.581,
640
+ 0.686,
641
+ 0.892,
642
+ 0.703
643
+ ],
644
+ "angle": 0,
645
+ "content": "\\[\n\\Delta \\boldsymbol {r} _ {l} = \\sigma \\left(\\sigma^ {- 1} \\left(\\boldsymbol {r} _ {l - 1} - \\boldsymbol {r} _ {0}\\right) + \\Delta \\boldsymbol {r} _ {l} ^ {\\prime}\\right), \\tag {3}\n\\]"
646
+ },
647
+ {
648
+ "type": "equation",
649
+ "bbox": [
650
+ 0.558,
651
+ 0.707,
652
+ 0.836,
653
+ 0.723
654
+ ],
655
+ "angle": 0,
656
+ "content": "\\[\n\\hat {\\boldsymbol {r}} _ {l} = \\boldsymbol {r} _ {0} + \\Delta \\boldsymbol {r} _ {l} \\cdot \\boldsymbol {s} _ {\\text {g r i d}}, \\quad \\boldsymbol {r} _ {l} = \\operatorname {D e t a c h} (\\hat {\\boldsymbol {r}} _ {l}),\n\\]"
657
+ },
658
+ {
659
+ "type": "text",
660
+ "bbox": [
661
+ 0.5,
662
+ 0.733,
663
+ 0.892,
664
+ 0.764
665
+ ],
666
+ "angle": 0,
667
+ "content": "where \\(\\Delta \\boldsymbol{r}_l'\\) and \\(\\Delta \\boldsymbol{r}_l\\) are the predicted offsets from \\(\\boldsymbol{r}_l\\) to both \\(\\boldsymbol{r}_{l-1}\\) and \\(\\boldsymbol{r}_0\\) before the sigmoid activation \\(\\sigma\\), respectively."
668
+ },
669
+ {
670
+ "type": "title",
671
+ "bbox": [
672
+ 0.5,
673
+ 0.772,
674
+ 0.843,
675
+ 0.787
676
+ ],
677
+ "angle": 0,
678
+ "content": "3.2. Salient Point Enhanced Cross-Attention"
679
+ },
680
+ {
681
+ "type": "text",
682
+ "bbox": [
683
+ 0.498,
684
+ 0.796,
685
+ 0.895,
686
+ 0.901
687
+ ],
688
+ "angle": 0,
689
+ "content": "In cross-attention layers, existing center-based methods are limited to the attention on both center and sides of the ground truth bounding box, causing detector confusion among the queries with the same center and side attention. To this end, we expect the queries to focus on their specific regions based on the reference points, four box sides, and other conditional regions in different heads. Accordingly,"
690
+ },
691
+ {
692
+ "type": "page_number",
693
+ "bbox": [
694
+ 0.481,
695
+ 0.945,
696
+ 0.519,
697
+ 0.957
698
+ ],
699
+ "angle": 0,
700
+ "content": "15542"
701
+ }
702
+ ],
703
+ [
704
+ {
705
+ "type": "image",
706
+ "bbox": [
707
+ 0.082,
708
+ 0.089,
709
+ 0.473,
710
+ 0.212
711
+ ],
712
+ "angle": 0,
713
+ "content": null
714
+ },
715
+ {
716
+ "type": "image_caption",
717
+ "bbox": [
718
+ 0.076,
719
+ 0.221,
720
+ 0.473,
721
+ 0.306
722
+ ],
723
+ "angle": 0,
724
+ "content": "Figure 3. Movable reference point. The reference points are initialized by the center/corner points of the mesh-grid. Based on the inner loss, only the green dashed box can be predicted by the inner points when reference points are fixed. By moving the reference points within their grid, the blue dashed boxes can be detected accurately without extended searching space."
725
+ },
726
+ {
727
+ "type": "text",
728
+ "bbox": [
729
+ 0.076,
730
+ 0.319,
731
+ 0.47,
732
+ 0.381
733
+ ],
734
+ "angle": 0,
735
+ "content": "we consider an improved Gaussian [5] \\(\\mathbf{G}\\) and conditional attention [16] \\(\\mathbf{A}_{\\mathrm{peca}}\\) to enhance query specificity and spatially extreme region discrimination. The final attention map \\(\\mathbf{A}_{\\mathrm{cross}}\\) is the sum of the two attentions \\(\\mathbf{A}_{\\mathrm{cross}} = \\mathbf{G} + \\mathbf{A}_{\\mathrm{peca}}\\)."
736
+ },
737
+ {
738
+ "type": "text",
739
+ "bbox": [
740
+ 0.076,
741
+ 0.381,
742
+ 0.474,
743
+ 0.547
744
+ ],
745
+ "angle": 0,
746
+ "content": "Side Directed Gaussian (SDG). Similar to the movable strategy, we enforce the predicted Gaussian attention to be inside the proposal bounding box to reduce the searching space. Given a reference point \\( r \\), the offset scales \\( \\mathbf{o} \\in [-1,1]^2 \\) for \\( H \\) heads are produced by a simple MLP with a tanh activation, and then multiply to the two sides of the proposal bounding box for head-specific point offset generation, where the direction is guided by the sign of the offset scales. The head-specific points are generated by Algorithm 1. For each head, the Gaussian-like spatial weight map \\( G_{i} \\) effecting on each pixel \\( (x,y) \\) of context features is then formulated as"
747
+ },
748
+ {
749
+ "type": "equation",
750
+ "bbox": [
751
+ 0.118,
752
+ 0.557,
753
+ 0.47,
754
+ 0.598
755
+ ],
756
+ "angle": 0,
757
+ "content": "\\[\nG _ {i} (x, y) = \\exp \\left(- \\frac {\\left(x - c _ {w , i}\\right) ^ {2}}{v _ {w , i} ^ {2}} - \\frac {\\left(y - c _ {h , i}\\right) ^ {2}}{v _ {h , i} ^ {2}}\\right). \\tag {4}\n\\]"
758
+ },
759
+ {
760
+ "type": "text",
761
+ "bbox": [
762
+ 0.076,
763
+ 0.606,
764
+ 0.473,
765
+ 0.788
766
+ ],
767
+ "angle": 0,
768
+ "content": "Point Enhanced Cross-Attention (PECA). As aforementioned, the semantic class for the query is closely related to its referenced location in our SAP-DETR. To enhance the correlation between queries and their references, we concatenate the locations to the content queries after the sinusoidal positional encoding (PE). Take a close look at the conditional attention [16], we find that the linear positional embedding mostly focuses on one box side in each attention head. So we introduce a more straightforward attention mechanism, where the four side coordinates are concatenated and assigned to the corresponding head for side attention. The process of PECA is formulated as"
769
+ },
770
+ {
771
+ "type": "equation",
772
+ "bbox": [
773
+ 0.085,
774
+ 0.794,
775
+ 0.47,
776
+ 0.836
777
+ ],
778
+ "angle": 0,
779
+ "content": "\\[\n\\begin{array}{r l} \\mathbf {A} _ {\\text {p e c a}} & = \\mathbf {e} _ {q} \\mathbf {e} _ {k} ^ {\\top} + \\mathbf {T P E} (\\boldsymbol {r} _ {q}) \\mathbf {P E} (\\boldsymbol {r} _ {k}) ^ {\\top} \\\\ & + \\mathbf {T} g (\\mathbf {P E} (\\boldsymbol {r} _ {q} - \\{\\ell , t \\}, \\boldsymbol {r} _ {q} + \\{r, b \\})) \\mathbf {P E} (\\boldsymbol {r} _ {k}) ^ {\\top}, \\end{array} \\tag {5}\n\\]"
780
+ },
781
+ {
782
+ "type": "text",
783
+ "bbox": [
784
+ 0.076,
785
+ 0.84,
786
+ 0.473,
787
+ 0.902
788
+ ],
789
+ "angle": 0,
790
+ "content": "where \\( g \\) is a linear layer mapping 4D-PE into 2D to keep channel dimension consistency. T [16] is a scaling transformation matrix. More details of T and comparison with conditional cross-attention are available in Appendix C."
791
+ },
792
+ {
793
+ "type": "title",
794
+ "bbox": [
795
+ 0.501,
796
+ 0.091,
797
+ 0.747,
798
+ 0.106
799
+ ],
800
+ "angle": 0,
801
+ "content": "Algorithm 1 Side Directed Gaussian"
802
+ },
803
+ {
804
+ "type": "text",
805
+ "bbox": [
806
+ 0.501,
807
+ 0.111,
808
+ 0.854,
809
+ 0.124
810
+ ],
811
+ "angle": 0,
812
+ "content": "Input: Content embedding e, reference point \\( r \\) and box s."
813
+ },
814
+ {
815
+ "type": "text",
816
+ "bbox": [
817
+ 0.501,
818
+ 0.124,
819
+ 0.894,
820
+ 0.152
821
+ ],
822
+ "angle": 0,
823
+ "content": "Output: Head-specific points \\(\\mathbf{c} = \\{(c_{w,i}, c_{h,i}) | i \\in H\\}\\) and head-specific attention \\(\\mathbf{v} = \\{(v_{w,i}, v_{h,i}) | i \\in H\\}\\)."
824
+ },
825
+ {
826
+ "type": "text",
827
+ "bbox": [
828
+ 0.509,
829
+ 0.152,
830
+ 0.894,
831
+ 0.18
832
+ ],
833
+ "angle": 0,
834
+ "content": "1: Predict offset scale and attention scale based on content embedding, \\(\\mathbf{o} = \\tanh (\\mathrm{MLP}(\\mathbf{e}))\\), \\(\\mathbf{v} = \\mathrm{MLP}(\\mathbf{e})\\)"
835
+ },
836
+ {
837
+ "type": "text",
838
+ "bbox": [
839
+ 0.509,
840
+ 0.18,
841
+ 0.645,
842
+ 0.193
843
+ ],
844
+ "angle": 0,
845
+ "content": "2: for \\(h\\gets 1\\in H\\) do"
846
+ },
847
+ {
848
+ "type": "text",
849
+ "bbox": [
850
+ 0.509,
851
+ 0.193,
852
+ 0.892,
853
+ 0.222
854
+ ],
855
+ "angle": 0,
856
+ "content": "3: Select the index of direction guided by the sign of offset scale, \\(\\{a,b\\} = \\mathrm{sgn}(\\mathbf{o}_i) + \\{1,2\\}\\), \\(a,b\\in \\{0,1,2,3\\}\\);"
857
+ },
858
+ {
859
+ "type": "text",
860
+ "bbox": [
861
+ 0.509,
862
+ 0.222,
863
+ 0.892,
864
+ 0.249
865
+ ],
866
+ "angle": 0,
867
+ "content": "4: According to the index of direction, predict head-specific point, \\(\\mathbf{c}_i = \\mathbf{o}_i\\cdot \\mathbf{s}[a,b] + \\mathbf{r}\\)"
868
+ },
869
+ {
870
+ "type": "text",
871
+ "bbox": [
872
+ 0.509,
873
+ 0.25,
874
+ 0.576,
875
+ 0.261
876
+ ],
877
+ "angle": 0,
878
+ "content": "5: end for"
879
+ },
880
+ {
881
+ "type": "text",
882
+ "bbox": [
883
+ 0.509,
884
+ 0.262,
885
+ 0.704,
886
+ 0.276
887
+ ],
888
+ "angle": 0,
889
+ "content": "6: return \\(\\mathbf{c}_i, \\mathbf{v}_i, \\forall i = 1, \\dots, H\\)"
890
+ },
891
+ {
892
+ "type": "list",
893
+ "bbox": [
894
+ 0.509,
895
+ 0.152,
896
+ 0.894,
897
+ 0.276
898
+ ],
899
+ "angle": 0,
900
+ "content": null
901
+ },
902
+ {
903
+ "type": "title",
904
+ "bbox": [
905
+ 0.5,
906
+ 0.309,
907
+ 0.814,
908
+ 0.326
909
+ ],
910
+ "angle": 0,
911
+ "content": "3.3. SAP-DETR with Denoising Strategy"
912
+ },
913
+ {
914
+ "type": "text",
915
+ "bbox": [
916
+ 0.498,
917
+ 0.333,
918
+ 0.895,
919
+ 0.514
920
+ ],
921
+ "angle": 0,
922
+ "content": "To further explore the capability of our proposed SAP-DETR, we develop SAP-DN-DETR and SAP-DINO-DETR by adding the denoising auxiliary loss [9,28] into the training process. In the denoised SAP-DETR, the main difference from both DN-DETR and DINO lies in the noise design. Instead of the center point, we perform the box jittering and randomly sample a point from the intersection region between the original bounding box and the jittering one as the reference point. As the denoising strategy only serves as an auxiliary training loss increasing the training cost, the variants of denoising models are test-free whose Params and GFLOPs are the same as SAP-DETR models."
923
+ },
924
+ {
925
+ "type": "title",
926
+ "bbox": [
927
+ 0.5,
928
+ 0.532,
929
+ 0.706,
930
+ 0.549
931
+ ],
932
+ "angle": 0,
933
+ "content": "4. Experimental Results"
934
+ },
935
+ {
936
+ "type": "title",
937
+ "bbox": [
938
+ 0.5,
939
+ 0.558,
940
+ 0.718,
941
+ 0.574
942
+ ],
943
+ "angle": 0,
944
+ "content": "4.1. Implementation Details"
945
+ },
946
+ {
947
+ "type": "text",
948
+ "bbox": [
949
+ 0.498,
950
+ 0.582,
951
+ 0.895,
952
+ 0.869
953
+ ],
954
+ "angle": 0,
955
+ "content": "We conduct the experiments on the COCO 2017 [12] object detection dataset, containing about 118K training images and 5K validation images. All models are evaluated by the standard COCO evaluation metrics. We follow the vanilla DETR [2] structure that consists of a CNN backbone, a stack of Transformer encoder-decoder layers, and two prediction heads for class label and bounding box prediction. We use ImageNet-pretrained ResNet [7] as our backbone, and report results based on the ResNet and its \\(\\times 1/16\\)-resolution extension ResNet-DC. Unlike DAB-DETR [13] sharing box and label head for each layer, we share the class head except the first layer and use an independent box head for the box regression of each layer (for more details please refer to Appendix D). As the mesh-grid initialization for reference points in SAP-DETR, we consider the number of queries \\(N\\) as a perfect square for uniform distribution. Unless otherwise specified, we use \\(N = 400\\) queries in the experiments. Precisely, we also provide a comparison under \\(N = 300\\) in Tab. 2, the standard setting in DETR-like models."
956
+ },
957
+ {
958
+ "type": "text",
959
+ "bbox": [
960
+ 0.5,
961
+ 0.871,
962
+ 0.894,
963
+ 0.901
964
+ ],
965
+ "angle": 0,
966
+ "content": "We adopt two different Transformer structures for experiments where a 3-layer encoder-decoder stack is evaluated to"
967
+ },
968
+ {
969
+ "type": "page_number",
970
+ "bbox": [
971
+ 0.481,
972
+ 0.945,
973
+ 0.519,
974
+ 0.957
975
+ ],
976
+ "angle": 0,
977
+ "content": "15543"
978
+ }
979
+ ],
980
+ [
981
+ {
982
+ "type": "image",
983
+ "bbox": [
984
+ 0.188,
985
+ 0.088,
986
+ 0.38,
987
+ 0.212
988
+ ],
989
+ "angle": 0,
990
+ "content": null
991
+ },
992
+ {
993
+ "type": "image",
994
+ "bbox": [
995
+ 0.387,
996
+ 0.088,
997
+ 0.581,
998
+ 0.212
999
+ ],
1000
+ "angle": 0,
1001
+ "content": null
1002
+ },
1003
+ {
1004
+ "type": "image",
1005
+ "bbox": [
1006
+ 0.584,
1007
+ 0.087,
1008
+ 0.782,
1009
+ 0.212
1010
+ ],
1011
+ "angle": 0,
1012
+ "content": null
1013
+ },
1014
+ {
1015
+ "type": "image_caption",
1016
+ "bbox": [
1017
+ 0.291,
1018
+ 0.22,
1019
+ 0.678,
1020
+ 0.234
1021
+ ],
1022
+ "angle": 0,
1023
+ "content": "Figure 4. Comparison of performance and training losses curves."
1024
+ },
1025
+ {
1026
+ "type": "text",
1027
+ "bbox": [
1028
+ 0.076,
1029
+ 0.251,
1030
+ 0.473,
1031
+ 0.388
1032
+ ],
1033
+ "angle": 0,
1034
+ "content": "demonstrate our lightweight model efficacy compared with the traditional CNN detectors, and a 6-layer encoder-decoder stack is aligned with previous DETR variants to investigate the performance of large model. Both are trained on two training schemes: the 12-epoch and 36-epoch schemes with a learning rate drop after 11 and 30 epochs, respectively. All models are trained on the Nvidia A100 GPUs with batch size of 16 and 8 for ResNet and ResNet-DC, respectively. For more training details, please refer to Appendix F."
1035
+ },
1036
+ {
1037
+ "type": "title",
1038
+ "bbox": [
1039
+ 0.077,
1040
+ 0.395,
1041
+ 0.218,
1042
+ 0.41
1043
+ ],
1044
+ "angle": 0,
1045
+ "content": "4.2. Main Results"
1046
+ },
1047
+ {
1048
+ "type": "text",
1049
+ "bbox": [
1050
+ 0.076,
1051
+ 0.418,
1052
+ 0.471,
1053
+ 0.509
1054
+ ],
1055
+ "angle": 0,
1056
+ "content": "As shown in Tab. 1 and Tab. 2, we comprehensively compare our proposed SAP-DETR with the traditional CNN detectors [18], the original DETR [2], and other DETR-like detectors [5, 13, 16, 25, 26, 31] on COCO 2017 validation dataset. For in-depth analysis, we conduct the comparison in two aspects: model convergence and efficacy."
1057
+ },
1058
+ {
1059
+ "type": "text",
1060
+ "bbox": [
1061
+ 0.076,
1062
+ 0.508,
1063
+ 0.472,
1064
+ 0.795
1065
+ ],
1066
+ "angle": 0,
1067
+ "content": "Model Convergence. Compared with traditional CNN detectors, Transformer detectors are always subject to laborious training time. For example, under the same 12-epoch training scheme, Faster RCNN [18] still achieves good performance, but the mainstream DETR-like models may suffer from inadequate training and perform poorly without the help of auxiliary losses [9]. Under the 12-epoch training scheme, our proposed SAP-DETR can accelerate model convergence significantly, boosting DAB-DETR [13] by 3.9 AP and 2.7 AP on 3-layer and 6-layer encoder-decoder structures, respectively. Compared with the current SoTA, our SAP-DETR also outperforms SAM-DETR [26] by \\(\\sim 1.3\\) AP, with reducing \\(\\sim 17\\%\\) parameters and \\(\\sim 10\\%\\) GFLOPs. Take a close look at the training process, as illustrated in Fig. 4, SAP-DETR conducts with rapid descent curves in both classification and box regression losses. Notably, there is a large margin in classification loss between ours and SoTA methods, which is benefited from the query-specific reference point, hence boosting model performance in early epochs."
1068
+ },
1069
+ {
1070
+ "type": "text",
1071
+ "bbox": [
1072
+ 0.076,
1073
+ 0.796,
1074
+ 0.472,
1075
+ 0.902
1076
+ ],
1077
+ "angle": 0,
1078
+ "content": "Model Efficacy. To analyze model efficacy, we report results on long training epochs and high-resolution features in Tab. 1. Under the 36-epoch training scheme, SAP-DETR achieves superior performance among all single-scale Transformer detectors, especially on middle and large targets. For example, SAP-DETR boosts DAB-DETR by \\(2.0\\mathrm{AP}_{\\mathrm{M}}\\) and \\(4.1\\mathrm{AP}_{\\mathrm{L}}\\) with 3-layer models, \\(1.0\\mathrm{AP}_{\\mathrm{M}}\\) and \\(1.9\\mathrm{AP}_{\\mathrm{L}}\\) with"
1079
+ },
1080
+ {
1081
+ "type": "text",
1082
+ "bbox": [
1083
+ 0.498,
1084
+ 0.251,
1085
+ 0.895,
1086
+ 0.463
1087
+ ],
1088
+ "angle": 0,
1089
+ "content": "6-layer models, which further verifies the effectiveness of our proposed salient point concept for overlapping object detection. Along with layer increase, a deficient upper-bound of SAM-DETR is exposed, with obviously lower 0.5 AP promotion compared to our 1.0 AP improvement. Persuasively, we also report the 50-epoch training results based on the 300-query setting. To align with our mesh-grid initialization strategy, we tile a \\(17 \\times 18\\) mesh-grid (306 queries) for each reference point initialization. Tab. 2 shows our main results and the most representative approaches with their original reported performance. Notably, SAP-DETR outperforms the current SoTAs with comparable costs based on all backbones. With low-resolution features (\\(\\times 1/32\\)), it significantly boosts both middle and large object detection accuracy."
1090
+ },
1091
+ {
1092
+ "type": "text",
1093
+ "bbox": [
1094
+ 0.498,
1095
+ 0.463,
1096
+ 0.897,
1097
+ 0.751
1098
+ ],
1099
+ "angle": 0,
1100
+ "content": "Combine with Other Fast Convergence Methods. As shown in Tab. 3, we compare our SAP-DETR variants with the current fast convergence methods [4, 9, 28]. With such a subtle modification, our SAP-DETR (grey rows) results in a significant performance improvement compared with the original methods (white rows). Under the 12-epoch training scheme, there exist 0.5-1.9 AP improvements on DN-DETR [9] and 0.7-1.6 AP improvements on Group-DETR [4], but the promotions are slightly reduced when implemented on DINO [28]. We hypothesise that there exists the same effect between the negative query of contrastive denoising [28] and our query-salient reference point. Moreover, we observe that the performance improvements largely originate from the large object detection, especially based on ResNet-DC5 family backbones. We speculate that DETR may prefer the high-resolution features (\\( \\times 1/16 \\)) rather than the low-resolution ones (\\( \\times 1/32 \\)), and our SAP-DETR can distinguish the salient points accurately on the high-resolution, thereby taking full advantage of the large object detection."
1101
+ },
1102
+ {
1103
+ "type": "title",
1104
+ "bbox": [
1105
+ 0.5,
1106
+ 0.761,
1107
+ 0.655,
1108
+ 0.777
1109
+ ],
1110
+ "angle": 0,
1111
+ "content": "4.3. Ablation Study"
1112
+ },
1113
+ {
1114
+ "type": "text",
1115
+ "bbox": [
1116
+ 0.498,
1117
+ 0.785,
1118
+ 0.895,
1119
+ 0.877
1120
+ ],
1121
+ "angle": 0,
1122
+ "content": "Effectiveness of Each Component. To offer an intuitionistic comparison of model convergence for each component, Tab. 4 reports the effectiveness of them based on the 3-layer encoder-decoder structure and 12-epoch training scheme. I). The proposed salient point concept based on content embeddings improves the performance from 32.3 AP to 33.5"
1123
+ },
1124
+ {
1125
+ "type": "page_footnote",
1126
+ "bbox": [
1127
+ 0.518,
1128
+ 0.887,
1129
+ 0.782,
1130
+ 0.901
1131
+ ],
1132
+ "angle": 0,
1133
+ "content": "<sup>1</sup>https://github.com/facebookresearch/detectron2"
1134
+ },
1135
+ {
1136
+ "type": "page_number",
1137
+ "bbox": [
1138
+ 0.481,
1139
+ 0.945,
1140
+ 0.52,
1141
+ 0.957
1142
+ ],
1143
+ "angle": 0,
1144
+ "content": "15544"
1145
+ }
1146
+ ],
1147
+ [
1148
+ {
1149
+ "type": "table",
1150
+ "bbox": [
1151
+ 0.1,
1152
+ 0.089,
1153
+ 0.872,
1154
+ 0.309
1155
+ ],
1156
+ "angle": 0,
1157
+ "content": "<table><tr><td>Method</td><td>#Epochs</td><td>#Params(M)</td><td>GFLOPs</td><td>AP</td><td>\\( AP_{50} \\)</td><td>\\( AP_{75} \\)</td><td>\\( AP_S \\)</td><td>\\( AP_M \\)</td><td>\\( AP_L \\)</td></tr><tr><td colspan=\"10\">3-Layer Encoder-Decoder Transformer Neck with ResNet-50 Backbone</td></tr><tr><td>DETR-R50 [2]</td><td>36</td><td>33</td><td>82</td><td>15.8</td><td>28.0</td><td>15.4</td><td>5.3</td><td>16.7</td><td>24.6</td></tr><tr><td>Deformable DETR-R50 [31]</td><td>36</td><td>30</td><td>77</td><td>37.1</td><td>57.6</td><td>39.4</td><td>18.3</td><td>40.8</td><td>51.6</td></tr><tr><td>SMCA-DETR-R50 [5]</td><td>12 / 36</td><td>-</td><td>-</td><td>28.8 / 37.7</td><td>48.1 / 58.7</td><td>29.9 / 40.1</td><td>13.8 / 19.4</td><td>31.3 / 40.5</td><td>41.3 / 54.8</td></tr><tr><td>Conditional DETR-R50 [16]</td><td>12 / 36</td><td>40</td><td>82</td><td>29.6 / 37.1</td><td>48.7 / 57.9</td><td>30.7 / 39.0</td><td>13.0 / 17.6</td><td>32.3 / 40.3</td><td>43.1 / 55.0</td></tr><tr><td>Anchor DETR-R50 [25]</td><td>12 / 36</td><td>31</td><td>79</td><td>30.8 / 37.6</td><td>51.1 / 58.7</td><td>31.8 / 39.7</td><td>14.3 / 18.8</td><td>34.1 / 41.5</td><td>44.3 / 53.5</td></tr><tr><td>DAB-DETR-R50 [13]</td><td>12 / 36</td><td>34</td><td>83</td><td>32.3 / 39.0</td><td>51.3 / 58.6</td><td>34.0 / 41.8</td><td>15.7 / 20.0</td><td>35.2 / 42.5</td><td>45.7 / 56.0</td></tr><tr><td>SAM-DETR-w/SMCA-R50 [26]</td><td>12 / 36</td><td>41</td><td>89</td><td>35.1 / 40.4</td><td>54.7 / 60.7</td><td>36.7 / 42.7</td><td>16.0 / 20.2</td><td>38.4 / 44.4</td><td>52.1 / 58.3</td></tr><tr><td>SAP-DETR-R50 (Ours)</td><td>12 / 36</td><td>36</td><td>84</td><td>36.2 / 41.2</td><td>56.2 / 61.6</td><td>37.9 / 43.4</td><td>16.4 / 21.0</td><td>39.5 / 44.5</td><td>53.8 / 60.1</td></tr><tr><td colspan=\"10\">6-Layer Encoder-Decoder Transformer Neck with ResNet-50 Backbone</td></tr><tr><td>DETR-R50 [2]</td><td>36</td><td>42</td><td>89</td><td>14.0</td><td>24.4</td><td>14.0</td><td>4.2</td><td>13.7</td><td>22.5</td></tr><tr><td>Deformable DETR-R50 [31]</td><td>36</td><td>34</td><td>81</td><td>38.0</td><td>58.2</td><td>40.4</td><td>18.5</td><td>41.7</td><td>54.2</td></tr><tr><td>SMCA-DETR-R50 [5]</td><td>12 / 36</td><td>-</td><td>-</td><td>32.4 / 40.1</td><td>52.3 / 61.4</td><td>34.0 / 42.8</td><td>15.5 / 20.3</td><td>34.9 / 43.3</td><td>47.7 / 57.1</td></tr><tr><td>Conditional DETR-R50 [16]</td><td>12 / 36</td><td>44</td><td>90</td><td>33.1 / 40.2</td><td>53.0 / 61.0</td><td>34.8 / 42.4</td><td>14.5 / 19.9</td><td>35.9 / 43.5</td><td>49.2 / 58.8</td></tr><tr><td>Anchor DETR-R50 [25]</td><td>12 / 36</td><td>37</td><td>85</td><td>33.7 / 39.7</td><td>54.5 / 60.5</td><td>35.1 / 41.9</td><td>15.6 / 19.9</td><td>37.3 / 43.5</td><td>49.8 / 57.3</td></tr><tr><td>DAB-DETR-R50 [13]</td><td>12 / 36</td><td>44</td><td>92</td><td>34.9 / 41.0</td><td>55.5 / 61.7</td><td>36.4 / 43.4</td><td>16.2 / 21.3</td><td>38.4 / 44.7</td><td>51.5 / 58.9</td></tr><tr><td>SAM-DETR-w/SMCA-R50 [26]</td><td>12 / 36</td><td>59</td><td>105</td><td>36.2 / 40.9</td><td>57.2 / 62.2</td><td>37.4 / 43.1</td><td>16.1 / 20.1</td><td>39.8 / 44.7</td><td>55.3 / 60.7</td></tr><tr><td>SAP-DETR-R50 (Ours)</td><td>12 / 36</td><td>47</td><td>94</td><td>37.5 / 42.2</td><td>58.5 / 62.7</td><td>39.2 / 44.6</td><td>17.3 / 22.6</td><td>40.6 / 45.7</td><td>55.4 / 60.8</td></tr></table>"
1158
+ },
1159
+ {
1160
+ "type": "table_caption",
1161
+ "bbox": [
1162
+ 0.076,
1163
+ 0.317,
1164
+ 0.895,
1165
+ 0.36
1166
+ ],
1167
+ "angle": 0,
1168
+ "content": "Table 1. Comparison between Transformer necks. Based on ResNet-50 backbone, all models are trained by the official source codes with their original settings and evaluated on COCO val2017. All models uses 400 queries except Anchor DETR, while Anchor DETR uses 200 queries with 2 pattern embeddings. GFLOPs and Params are measured by \\( \\text{Detectron}^2 \\)."
1169
+ },
1170
+ {
1171
+ "type": "table",
1172
+ "bbox": [
1173
+ 0.099,
1174
+ 0.363,
1175
+ 0.872,
1176
+ 0.707
1177
+ ],
1178
+ "angle": 0,
1179
+ "content": "<table><tr><td>Method</td><td>#Epochs</td><td>#Params(M)</td><td>GFLOPs</td><td>AP</td><td>\\( AP_{50} \\)</td><td>\\( AP_{75} \\)</td><td>\\( AP_S \\)</td><td>\\( AP_M \\)</td><td>\\( AP_L \\)</td><td>Infer. \\( Time(s / img)^{\\dagger} \\)</td></tr><tr><td colspan=\"11\">ResNet-50 Backbone</td></tr><tr><td>Faster RCNN-FPN-R50 [10, 18]</td><td>108</td><td>42</td><td>180</td><td>42.0</td><td>62.1</td><td>45.5</td><td>26.6</td><td>45.5</td><td>53.4</td><td>0.039</td></tr><tr><td>DETR-R50 [2]</td><td>500</td><td>41</td><td>86</td><td>42.0</td><td>62.4</td><td>44.2</td><td>20.5</td><td>45.8</td><td>61.1</td><td>0.040</td></tr><tr><td>Deformable DETR-R50 [31]</td><td>50</td><td>34</td><td>78</td><td>39.4</td><td>59.6</td><td>42.3</td><td>20.6</td><td>43.0</td><td>55.5</td><td>0.043</td></tr><tr><td>SMCA-DETR-R50 [5]</td><td>50</td><td>42</td><td>86</td><td>41.0</td><td>-</td><td>-</td><td>21.9</td><td>44.3</td><td>59.1</td><td>0.045</td></tr><tr><td>Conditional DETR-R50 [16]</td><td>50</td><td>44</td><td>90</td><td>40.9</td><td>61.8</td><td>43.3</td><td>20.8</td><td>44.6</td><td>59.2</td><td>0.057</td></tr><tr><td>Anchor DETR-R50 [25]</td><td>50</td><td>39</td><td>85</td><td>42.1</td><td>63.1</td><td>44.9</td><td>22.3</td><td>46.2</td><td>60.0</td><td>0.050</td></tr><tr><td>DAB-DETR-R50 [13]</td><td>50</td><td>44</td><td>\\( 90^† \\)</td><td>42.2</td><td>63.1</td><td>44.7</td><td>21.5</td><td>45.7</td><td>60.3</td><td>0.059</td></tr><tr><td>SAM-DETR-w/SMCA-R50 [26]</td><td>50</td><td>58</td><td>100</td><td>41.8</td><td>63.2</td><td>43.9</td><td>22.1</td><td>45.9</td><td>60.9</td><td>0.065</td></tr><tr><td>SAP-DETR-R50 (Ours)</td><td>50</td><td>47</td><td>92</td><td>43.1</td><td>63.8</td><td>45.4</td><td>22.9</td><td>47.1</td><td>62.1</td><td>0.063</td></tr><tr><td colspan=\"11\">ResNet-101 Backbone</td></tr><tr><td>Faster RCNN-FPN-R101 [10, 18]</td><td>108</td><td>60</td><td>246</td><td>44.0</td><td>63.9</td><td>47.8</td><td>27.2</td><td>48.1</td><td>56.0</td><td>0.050</td></tr><tr><td>DETR-R101 [2]</td><td>500</td><td>60</td><td>152</td><td>43.5</td><td>63.8</td><td>46.4</td><td>21.9</td><td>48.0</td><td>61.8</td><td>0.066</td></tr><tr><td>Conditional DETR-R101 [16]</td><td>50</td><td>63</td><td>156</td><td>42.8</td><td>63.7</td><td>46.0</td><td>21.7</td><td>46.6</td><td>60.9</td><td>0.070</td></tr><tr><td>Anchor DETR-R101 [25]</td><td>50</td><td>58</td><td>150</td><td>43.5</td><td>64.3</td><td>46.6</td><td>23.2</td><td>47.7</td><td>61.4</td><td>0.068</td></tr><tr><td>DAB-DETR-R101 [13]</td><td>50</td><td>63</td><td>\\( 157^† \\)</td><td>43.5</td><td>63.9</td><td>46.6</td><td>23.6</td><td>47.3</td><td>61.5</td><td>0.072</td></tr><tr><td>SAP-DETR-R101 (Ours)</td><td>50</td><td>67</td><td>158</td><td>44.4</td><td>64.9</td><td>47.1</td><td>24.1</td><td>48.7</td><td>63.1</td><td>0.078</td></tr><tr><td colspan=\"11\">DC5-ResNet-50 Backbone</td></tr><tr><td>DETR-DC5-R50 [2]</td><td>500</td><td>41</td><td>187</td><td>43.3</td><td>63.1</td><td>45.9</td><td>22.5</td><td>47.3</td><td>61.1</td><td>0.087</td></tr><tr><td>Conditional DETR-DC5-R50 [16]</td><td>50</td><td>44</td><td>195</td><td>43.8</td><td>64.4</td><td>46.7</td><td>24.0</td><td>47.6</td><td>60.7</td><td>0.093</td></tr><tr><td>Anchor DETR-DC5-R50 [25]</td><td>50</td><td>39</td><td>151</td><td>44.2</td><td>64.7</td><td>47.5</td><td>24.7</td><td>48.2</td><td>60.6</td><td>0.069</td></tr><tr><td>DAB-DETR-DC5-R50 [13]</td><td>50</td><td>44</td><td>\\( 194^† \\)</td><td>44.5</td><td>65.1</td><td>47.7</td><td>25.3</td><td>48.2</td><td>62.3</td><td>0.094</td></tr><tr><td>SAM-DETR-w/SMCA-DC5-R50 [26]</td><td>50</td><td>58</td><td>210</td><td>45.0</td><td>65.4</td><td>47.9</td><td>26.2</td><td>49.0</td><td>63.3</td><td>0.126</td></tr><tr><td>SAP-DETR-DC5-R50 (Ours)</td><td>50</td><td>47</td><td>197</td><td>46.0</td><td>65.5</td><td>48.9</td><td>26.4</td><td>50.2</td><td>62.6</td><td>0.116</td></tr><tr><td colspan=\"11\">DC5-ResNet-101 Backbone</td></tr><tr><td>DETR-DC5-R101 [2]</td><td>500</td><td>60</td><td>253</td><td>44.9</td><td>64.7</td><td>47.7</td><td>23.7</td><td>49.5</td><td>62.3</td><td>0.101</td></tr><tr><td>Conditional DETR-DC5-R101 [16]</td><td>50</td><td>63</td><td>262</td><td>45.0</td><td>65.6</td><td>48.4</td><td>26.1</td><td>48.9</td><td>62.8</td><td>0.105</td></tr><tr><td>Anchor DETR-DC5-R101 [25]</td><td>50</td><td>58</td><td>227</td><td>45.1</td><td>65.7</td><td>48.8</td><td>25.8</td><td>49.4</td><td>61.6</td><td>0.083</td></tr><tr><td>DAB-DETR-DC5-R101 [13]</td><td>50</td><td>63</td><td>\\( 263^† \\)</td><td>45.8</td><td>65.9</td><td>49.3</td><td>27.0</td><td>49.8</td><td>63.8</td><td>0.110</td></tr><tr><td>SAP-DETR-DC5-R101 (Ours)</td><td>50</td><td>67</td><td>266</td><td>46.9</td><td>66.7</td><td>50.5</td><td>27.9</td><td>51.3</td><td>64.3</td><td>0.130</td></tr></table>"
1180
+ },
1181
+ {
1182
+ "type": "table_caption",
1183
+ "bbox": [
1184
+ 0.076,
1185
+ 0.712,
1186
+ 0.895,
1187
+ 0.756
1188
+ ],
1189
+ "angle": 0,
1190
+ "content": "Table 2. Comparison of Transformer necks with 300 queries on COCO val2017. All results are reported from their original paper. All models uses 300 queries except Anchor DETR, while Anchor DETR uses 100 queries with 3 pattern embeddings. All inference speeds are measured by a single Nvidia A100 GPU. † denotes the results are measured by ourselves."
1191
+ },
1192
+ {
1193
+ "type": "text",
1194
+ "bbox": [
1195
+ 0.076,
1196
+ 0.766,
1197
+ 0.473,
1198
+ 0.888
1199
+ ],
1200
+ "angle": 0,
1201
+ "content": "AP compared to baseline DAB-DETR (row 8-9). Such a query-specific spatial prior enables queries to attend to their expected region from content features (see Figures in Appendix G) and reduces the false detection rate on occluded and partial objects (see Fig. 1(d)), hence boosting detection performance on middle and large objects. However, there exists a drop in small object detection (15.7 APs vs. 14.3 APs), for which we consider that the failure is mainly caused"
1202
+ },
1203
+ {
1204
+ "type": "text",
1205
+ "bbox": [
1206
+ 0.498,
1207
+ 0.766,
1208
+ 0.895,
1209
+ 0.888
1210
+ ],
1211
+ "angle": 0,
1212
+ "content": "by the query sparsity. 2). Therefore, the movable strategy is applied to alleviate the constraint, improving the final model by \\(1.0\\mathrm{AP}\\) and \\(0.6\\mathrm{AP}_{\\mathrm{S}}\\) (row 1 and 6). 3). Compared with the final model, the inner loss greatly improves the performance on high-quality \\(\\mathrm{AP}_{75}\\) and large objects \\(\\mathrm{AP}_{\\mathrm{L}}\\) detection (row 1 and 7), with just a slight drop on low-quality object \\(\\mathrm{AP}_{50}\\). That is reasonable because the outside reference points are unable to localize objects accurately, and this phenomenon"
1213
+ },
1214
+ {
1215
+ "type": "page_number",
1216
+ "bbox": [
1217
+ 0.481,
1218
+ 0.945,
1219
+ 0.52,
1220
+ 0.958
1221
+ ],
1222
+ "angle": 0,
1223
+ "content": "15545"
1224
+ }
1225
+ ],
1226
+ [
1227
+ {
1228
+ "type": "table",
1229
+ "bbox": [
1230
+ 0.12,
1231
+ 0.089,
1232
+ 0.855,
1233
+ 0.201
1234
+ ],
1235
+ "angle": 0,
1236
+ "content": "<table><tr><td rowspan=\"2\">Backbone</td><td rowspan=\"2\">Epoch</td><td rowspan=\"2\">w/ SAP</td><td colspan=\"2\">DN-DETR [9]</td><td colspan=\"2\">DINO (Single-Scale) [28]</td><td colspan=\"2\">Group DETR [4]</td></tr><tr><td>AP / AP50 / AP75</td><td>APS / APM / APL</td><td>AP / AP50 / AP75</td><td>APS / APM / APL</td><td>AP / AP50 / AP75</td><td>APS / APM / APL</td></tr><tr><td rowspan=\"2\">R50</td><td rowspan=\"2\">12</td><td rowspan=\"2\">✓(Ours)</td><td>38.3 / 58.6 / 40.5</td><td>18.4 / 41.6 / 57.1</td><td>39.7 / 58.3 / 42.4</td><td>19.1 / 43.7 / 57.1</td><td>39.1 / - / -</td><td>19.7 / 42.5 / 56.8</td></tr><tr><td>39.5 / 59.7 / 41.5</td><td>18.7 / 42.8 / 59.0</td><td>40.0 / 60.1 / 42.1</td><td>20.2 / 43.4 / 58.5</td><td>39.8 / 60.2 / 42.0</td><td>20.2 / 43.5 / 58.6</td></tr><tr><td rowspan=\"2\">R101</td><td rowspan=\"2\">12</td><td rowspan=\"2\">✓(Ours)</td><td>40.5 / 60.8 / 43.0</td><td>19.3 / 44.3 / 59.6</td><td>41.9 / 60.8 / 44.4</td><td>22.5 / 46.3 / 59.5</td><td>- / - / -</td><td>- / - / -</td></tr><tr><td>41.0 / 61.2 / 43.4</td><td>19.8 / 45.3 / 60.0</td><td>41.5 / 61.4 / 43.6</td><td>20.3 / 45.2 / 60.0</td><td>41.1 / 61.5 / 43.4</td><td>20.5 / 45.5 / 59.4</td></tr><tr><td rowspan=\"2\">R50-DC</td><td rowspan=\"2\">12</td><td rowspan=\"2\">✓(Ours)</td><td>41.7 / 61.4 / 44.1</td><td>21.2 / 45.0 / 60.2</td><td>43.6 / 61.4 / 47.0</td><td>24.8 / 47.3 / 59.5</td><td>41.9 / - / -</td><td>23.3 / 45.6 / 58.4</td></tr><tr><td>43.6 / 62.5 / 46.2</td><td>23.3 / 47.3 / 61.0</td><td>44.0 / 63.1 / 46.5</td><td>24.8 / 47.3 / 61.1</td><td>43.9 / 63.2 / 46.8</td><td>24.5 / 47.6 / 61.3</td></tr><tr><td rowspan=\"2\">R101-DC</td><td rowspan=\"2\">12</td><td rowspan=\"2\">✓(Ours)</td><td>43.4 / 61.9 / 47.2</td><td>24.8 / 46.8 / 59.4</td><td>45.4 / 63.5 / 49.2</td><td>26.4 / 49.5 / 61.1</td><td>- / - / -</td><td>- / - / -</td></tr><tr><td>44.6 / 63.9 / 48.0</td><td>25.5 / 48.9 / 62.5</td><td>45.6 / 64.5 / 48.7</td><td>25.0 / 49.7 / 62.5</td><td>44.4 / 63.9 / 47.4</td><td>25.9 / 48.5 / 61.4</td></tr></table>"
1237
+ },
1238
+ {
1239
+ "type": "table_caption",
1240
+ "bbox": [
1241
+ 0.108,
1242
+ 0.206,
1243
+ 0.86,
1244
+ 0.221
1245
+ ],
1246
+ "angle": 0,
1247
+ "content": "Table 3. Comparison with denoised methods on COCO dataset based on the 12-epoch training schedule and 300 object queries."
1248
+ },
1249
+ {
1250
+ "type": "table",
1251
+ "bbox": [
1252
+ 0.12,
1253
+ 0.235,
1254
+ 0.851,
1255
+ 0.331
1256
+ ],
1257
+ "angle": 0,
1258
+ "content": "<table><tr><td>Comment</td><td>Movable</td><td>Inner Loss</td><td>PECA</td><td>SDG</td><td>AP</td><td>\\( AP_{50} \\)</td><td>\\( AP_{75} \\)</td><td>\\( AP_S \\)</td><td>\\( AP_M \\)</td><td>\\( AP_L \\)</td></tr><tr><td>SAP-DETR (Ours)</td><td>✓</td><td>✓</td><td>✓</td><td>✓</td><td>36.2</td><td>56.2</td><td>37.9</td><td>16.4</td><td>39.5</td><td>53.8</td></tr><tr><td>-SDG</td><td>✓</td><td>✓</td><td>✓</td><td></td><td>35.6</td><td>56.2</td><td>36.9</td><td>16.3</td><td>38.9</td><td>52.7</td></tr><tr><td>-PECA</td><td>✓</td><td>✓</td><td></td><td>✓</td><td>34.8</td><td>55.5</td><td>36.0</td><td>15.7</td><td>37.3</td><td>52.0</td></tr><tr><td>-PECA &amp; SDG</td><td>✓</td><td>✓</td><td></td><td></td><td>34.0</td><td>54.9</td><td>35.3</td><td>15.0</td><td>36.7</td><td>51.5</td></tr><tr><td>-Movable</td><td></td><td>✓</td><td>✓</td><td>✓</td><td>35.2</td><td>55.4</td><td>36.8</td><td>15.8</td><td>38.5</td><td>53.8</td></tr><tr><td>-Inner Loss</td><td>✓</td><td></td><td>✓</td><td>✓</td><td>35.9</td><td>56.3</td><td>37.4</td><td>16.2</td><td>39.3</td><td>52.5</td></tr><tr><td>DAB-DETR (Baseline)</td><td>-</td><td>-</td><td>-</td><td>-</td><td>32.3</td><td>51.3</td><td>34.0</td><td>15.7</td><td>35.2</td><td>45.7</td></tr><tr><td>+Salient Point Concept</td><td>-</td><td>-</td><td>-</td><td>-</td><td>33.5</td><td>54.3</td><td>35.1</td><td>14.3</td><td>36.5</td><td>51.0</td></tr></table>"
1259
+ },
1260
+ {
1261
+ "type": "table_caption",
1262
+ "bbox": [
1263
+ 0.369,
1264
+ 0.335,
1265
+ 0.6,
1266
+ 0.35
1267
+ ],
1268
+ "angle": 0,
1269
+ "content": "Table 4. Ablation on each components"
1270
+ },
1271
+ {
1272
+ "type": "table",
1273
+ "bbox": [
1274
+ 0.079,
1275
+ 0.365,
1276
+ 0.47,
1277
+ 0.418
1278
+ ],
1279
+ "angle": 0,
1280
+ "content": "<table><tr><td>Inner Cost (Linner)</td><td>Movable within Grid (sgrid)</td><td>AP</td><td>APS</td><td>APM</td><td>APL</td></tr><tr><td></td><td></td><td>35.9</td><td>17.0</td><td>38.8</td><td>52.7</td></tr><tr><td>✓</td><td></td><td>26.3</td><td>11.3</td><td>28.0</td><td>39.5</td></tr><tr><td>✓</td><td>✓</td><td>36.2</td><td>16.4</td><td>39.5</td><td>53.8</td></tr></table>"
1281
+ },
1282
+ {
1283
+ "type": "table_caption",
1284
+ "bbox": [
1285
+ 0.147,
1286
+ 0.427,
1287
+ 0.4,
1288
+ 0.44
1289
+ ],
1290
+ "angle": 0,
1291
+ "content": "Table 5. Ablation on scaling factor of grid"
1292
+ },
1293
+ {
1294
+ "type": "table",
1295
+ "bbox": [
1296
+ 0.079,
1297
+ 0.447,
1298
+ 0.469,
1299
+ 0.501
1300
+ ],
1301
+ "angle": 0,
1302
+ "content": "<table><tr><td>PECA</td><td>Scaling Factor of SDG</td><td>AP</td><td>APS</td><td>APM</td><td>APL</td></tr><tr><td></td><td></td><td>33.6</td><td>14.7</td><td>36.0</td><td>50.7</td></tr><tr><td>✓</td><td></td><td>35.7</td><td>17.5</td><td>38.8</td><td>52.6</td></tr><tr><td>✓</td><td>✓</td><td>36.2</td><td>16.4</td><td>39.5</td><td>53.8</td></tr></table>"
1303
+ },
1304
+ {
1305
+ "type": "table_caption",
1306
+ "bbox": [
1307
+ 0.144,
1308
+ 0.511,
1309
+ 0.402,
1310
+ 0.525
1311
+ ],
1312
+ "angle": 0,
1313
+ "content": "Table 6. Ablation on scaling factor of SDG"
1314
+ },
1315
+ {
1316
+ "type": "text",
1317
+ "bbox": [
1318
+ 0.076,
1319
+ 0.536,
1320
+ 0.47,
1321
+ 0.672
1322
+ ],
1323
+ "angle": 0,
1324
+ "content": "always exists in large objects. 4). For salient point enhanced cross-attention, both SDG and PECA serve as the essential components, independently emerging 0.8 AP and 1.6 AP improvements compared to the standard model (rows 2-4). Interestingly, there exists an effectiveness overlap on small objects, with only \\(0.1\\mathrm{AP}_{\\mathrm{S}}\\) improvement when adding SDG to the equipped PECA model (row 1 and 2). We argue that the Gaussian-like map of SDG might be easily overlapped with PECA on small objects."
1325
+ },
1326
+ {
1327
+ "type": "text",
1328
+ "bbox": [
1329
+ 0.076,
1330
+ 0.673,
1331
+ 0.472,
1332
+ 0.904
1333
+ ],
1334
+ "angle": 0,
1335
+ "content": "Scaling Factor of Movable Strategy. We perform an ablation study on the scaling factor of the movable strategy and further investigate the effectiveness of the inner cost in Tab. 5. Notably, it is observed that there exists a conflict between the inner loss and the global search strategy, behaving a sharp drop when only reserving the inner loss. Furthermore, searching within the grid enables the detector to more attend to small objects and avoid a drastic deterioration in normal object detection. See Appendix E for more detailed analyses. Scaling Factor of SDG. We also compare our side-directed manner with the standard offset prediction method in Tab. 6. Based on PECA, the side-directed scaling factor may limit the detector on small object detection but significantly promote the performance on other objects. This phenomenon would be broken without the help of PECA in which a preci"
1336
+ },
1337
+ {
1338
+ "type": "text",
1339
+ "bbox": [
1340
+ 0.498,
1341
+ 0.367,
1342
+ 0.892,
1343
+ 0.429
1344
+ ],
1345
+ "angle": 0,
1346
+ "content": "itous decline is emerged on all-scale object detection (row 5 in Tab. 4 vs. row 1 in Tab. 6). We hypothesise that it because the predicted reference points may be outside of the proposal boxes, or even the region of the image."
1347
+ },
1348
+ {
1349
+ "type": "title",
1350
+ "bbox": [
1351
+ 0.5,
1352
+ 0.443,
1353
+ 0.619,
1354
+ 0.459
1355
+ ],
1356
+ "angle": 0,
1357
+ "content": "5. Conclusion"
1358
+ },
1359
+ {
1360
+ "type": "text",
1361
+ "bbox": [
1362
+ 0.498,
1363
+ 0.47,
1364
+ 0.895,
1365
+ 0.681
1366
+ ],
1367
+ "angle": 0,
1368
+ "content": "In this paper, we propose SAP-DETR for promoting model convergence by treating object detection as a transformation from the salient points to the instance objects. Our SAP-DETR explicitly initializes a query-specific reference point for each object query, gradually aggregates them into an instance object, and predicts the distance from each side of the bounding box. By speedily attending to the query-specific region and other extreme regions from contextual image features, it thus can effectively bridge the gap between the salient points and the query-based Transformer detector. Our extensive experiments have demonstrated that SAP-DETR achieves superior model convergence speed. With the same training settings, our proposed SAP-DETR outperforms SoTA approaches with large margins."
1369
+ },
1370
+ {
1371
+ "type": "title",
1372
+ "bbox": [
1373
+ 0.5,
1374
+ 0.697,
1375
+ 0.636,
1376
+ 0.713
1377
+ ],
1378
+ "angle": 0,
1379
+ "content": "6. Future Work"
1380
+ },
1381
+ {
1382
+ "type": "text",
1383
+ "bbox": [
1384
+ 0.498,
1385
+ 0.723,
1386
+ 0.894,
1387
+ 0.814
1388
+ ],
1389
+ "angle": 0,
1390
+ "content": "This point-based design for DETR-like models also comes with remaining issues, in particular regarding training with deformable attention, multi-scale features, and negative query design. Following current center-based methods working for similar issues, we expect future work to successfully address them for point-based design of SAP-DETR."
1391
+ },
1392
+ {
1393
+ "type": "title",
1394
+ "bbox": [
1395
+ 0.5,
1396
+ 0.83,
1397
+ 0.667,
1398
+ 0.846
1399
+ ],
1400
+ "angle": 0,
1401
+ "content": "Acknowledgements"
1402
+ },
1403
+ {
1404
+ "type": "text",
1405
+ "bbox": [
1406
+ 0.498,
1407
+ 0.856,
1408
+ 0.892,
1409
+ 0.901
1410
+ ],
1411
+ "angle": 0,
1412
+ "content": "We thank the Lenovo Research AI Master platform for computing GPU supports without which this work would not be possible."
1413
+ },
1414
+ {
1415
+ "type": "page_number",
1416
+ "bbox": [
1417
+ 0.481,
1418
+ 0.945,
1419
+ 0.52,
1420
+ 0.957
1421
+ ],
1422
+ "angle": 0,
1423
+ "content": "15546"
1424
+ }
1425
+ ],
1426
+ [
1427
+ {
1428
+ "type": "title",
1429
+ "bbox": [
1430
+ 0.08,
1431
+ 0.09,
1432
+ 0.174,
1433
+ 0.106
1434
+ ],
1435
+ "angle": 0,
1436
+ "content": "References"
1437
+ },
1438
+ {
1439
+ "type": "ref_text",
1440
+ "bbox": [
1441
+ 0.087,
1442
+ 0.116,
1443
+ 0.47,
1444
+ 0.157
1445
+ ],
1446
+ "angle": 0,
1447
+ "content": "[1] Navaneeth Bodla, Bharat Singh, Rama Chellappa, and Larry S Davis. Soft-nms-improving object detection with one line of code. In ICCV, pages 5561-5569, 2017. 3"
1448
+ },
1449
+ {
1450
+ "type": "ref_text",
1451
+ "bbox": [
1452
+ 0.087,
1453
+ 0.159,
1454
+ 0.472,
1455
+ 0.214
1456
+ ],
1457
+ "angle": 0,
1458
+ "content": "[2] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In ECCV, pages 213-229. Springer, 2020. 1, 3, 4, 5, 6, 7, 10"
1459
+ },
1460
+ {
1461
+ "type": "ref_text",
1462
+ "bbox": [
1463
+ 0.087,
1464
+ 0.216,
1465
+ 0.47,
1466
+ 0.257
1467
+ ],
1468
+ "angle": 0,
1469
+ "content": "[3] Liangyu Chen, Tong Yang, Xiangyu Zhang, Wei Zhang, and Jian Sun. Points as queries: Weakly semi-supervised object detection by points. In CVPR, pages 8823-8832, 2021. 3"
1470
+ },
1471
+ {
1472
+ "type": "ref_text",
1473
+ "bbox": [
1474
+ 0.087,
1475
+ 0.259,
1476
+ 0.472,
1477
+ 0.313
1478
+ ],
1479
+ "angle": 0,
1480
+ "content": "[4] Qiang Chen, Xiaokang Chen, Gang Zeng, and Jingdong Wang. Group detr:fast training convergence with deco-upplied one-to-many label assignment. arXiv preprint arXiv:2207.13085, 2022. 3, 6, 8"
1481
+ },
1482
+ {
1483
+ "type": "ref_text",
1484
+ "bbox": [
1485
+ 0.087,
1486
+ 0.316,
1487
+ 0.472,
1488
+ 0.37
1489
+ ],
1490
+ "angle": 0,
1491
+ "content": "[5] Peng Gao, Minghang Zheng, Xiaogang Wang, Jifeng Dai, and Hongsheng Li. Fast convergence of detr with spatially modulated co-attention. In ICCV, pages 3621–3630, 2021. 2, 3, 5, 6, 7, 10"
1492
+ },
1493
+ {
1494
+ "type": "ref_text",
1495
+ "bbox": [
1496
+ 0.087,
1497
+ 0.373,
1498
+ 0.47,
1499
+ 0.413
1500
+ ],
1501
+ "angle": 0,
1502
+ "content": "[6] Zheng Ge, Songtao Liu, Feng Wang, Zeming Li, and Jian Sun. Yolox: Exceeding yolo series in 2021. arXiv preprint arXiv:2107.08430, 2021. 1"
1503
+ },
1504
+ {
1505
+ "type": "ref_text",
1506
+ "bbox": [
1507
+ 0.087,
1508
+ 0.416,
1509
+ 0.47,
1510
+ 0.457
1511
+ ],
1512
+ "angle": 0,
1513
+ "content": "[7] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 5"
1514
+ },
1515
+ {
1516
+ "type": "ref_text",
1517
+ "bbox": [
1518
+ 0.087,
1519
+ 0.459,
1520
+ 0.47,
1521
+ 0.487
1522
+ ],
1523
+ "angle": 0,
1524
+ "content": "[8] Hei Law and Jia Deng. Cornernet: Detecting objects as paired keypoints. In ECCV, pages 734-750, 2018. 3"
1525
+ },
1526
+ {
1527
+ "type": "ref_text",
1528
+ "bbox": [
1529
+ 0.087,
1530
+ 0.489,
1531
+ 0.47,
1532
+ 0.543
1533
+ ],
1534
+ "angle": 0,
1535
+ "content": "[9] Feng Li, Hao Zhang, Shilong Liu, Jian Guo, Lionel M Ni, and Lei Zhang. Dn-detr: Accelerate detr training by introducing query denoising. In CVPR, pages 13619-13627, 2022. 3, 5, 6, 8, 13"
1536
+ },
1537
+ {
1538
+ "type": "ref_text",
1539
+ "bbox": [
1540
+ 0.08,
1541
+ 0.546,
1542
+ 0.47,
1543
+ 0.599
1544
+ ],
1545
+ "angle": 0,
1546
+ "content": "[10] Tsung-Yi Lin, Piotr Dollar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In CVPR, pages 2117-2125, 2017. 7"
1547
+ },
1548
+ {
1549
+ "type": "ref_text",
1550
+ "bbox": [
1551
+ 0.08,
1552
+ 0.603,
1553
+ 0.47,
1554
+ 0.643
1555
+ ],
1556
+ "angle": 0,
1557
+ "content": "[11] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólar. Focal loss for dense object detection. In ICCV, pages 2980-2988, 2017. 1, 3"
1558
+ },
1559
+ {
1560
+ "type": "ref_text",
1561
+ "bbox": [
1562
+ 0.08,
1563
+ 0.646,
1564
+ 0.47,
1565
+ 0.7
1566
+ ],
1567
+ "angle": 0,
1568
+ "content": "[12] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, pages 740-755, 2014. 5"
1569
+ },
1570
+ {
1571
+ "type": "ref_text",
1572
+ "bbox": [
1573
+ 0.08,
1574
+ 0.703,
1575
+ 0.47,
1576
+ 0.756
1577
+ ],
1578
+ "angle": 0,
1579
+ "content": "[13] Shilong Liu, Feng Li, Hao Zhang, Xiao Yang, Xianbiao Qi, Hang Su, Jun Zhu, and Lei Zhang. Dab-detr: Dynamic anchor boxes are better queries for detr. ICLR, 2022. 2, 3, 4, 5, 6, 7, 10"
1580
+ },
1581
+ {
1582
+ "type": "ref_text",
1583
+ "bbox": [
1584
+ 0.08,
1585
+ 0.76,
1586
+ 0.47,
1587
+ 0.814
1588
+ ],
1589
+ "angle": 0,
1590
+ "content": "[14] Wei Liu, Dragomir Anguelov, Dumitru Erhan, Christian Szegedy, Scott Reed, Cheng-Yang Fu, and Alexander C Berg. Ssd: Single shot multibox detector. In ECCV, pages 21-37. Springer, 2016. 1"
1591
+ },
1592
+ {
1593
+ "type": "ref_text",
1594
+ "bbox": [
1595
+ 0.08,
1596
+ 0.817,
1597
+ 0.47,
1598
+ 0.87
1599
+ ],
1600
+ "angle": 0,
1601
+ "content": "[15] Yang Liu, Yao Zhang, Yixin Wang, Feng Hou, Jin Yuan, Jiang Tian, Yang Zhang, Zhongchao Shi, Jianping Fan, and Zhiqiang He. A survey of visual transformers. arXiv preprint arXiv:2111.06091, 2021. 3"
1602
+ },
1603
+ {
1604
+ "type": "ref_text",
1605
+ "bbox": [
1606
+ 0.08,
1607
+ 0.874,
1608
+ 0.47,
1609
+ 0.901
1610
+ ],
1611
+ "angle": 0,
1612
+ "content": "[16] Depu Meng, Xiaokang Chen, Zejia Fan, Gang Zeng, Houqiang Li, Yuhui Yuan, Lei Sun, and Jingdong Wang. Con"
1613
+ },
1614
+ {
1615
+ "type": "list",
1616
+ "bbox": [
1617
+ 0.08,
1618
+ 0.116,
1619
+ 0.472,
1620
+ 0.901
1621
+ ],
1622
+ "angle": 0,
1623
+ "content": null
1624
+ },
1625
+ {
1626
+ "type": "ref_text",
1627
+ "bbox": [
1628
+ 0.533,
1629
+ 0.093,
1630
+ 0.892,
1631
+ 0.12
1632
+ ],
1633
+ "angle": 0,
1634
+ "content": "ditional detr for fast training convergence. In ICCV, pages 3651-3660, 2021. 2, 3, 4, 5, 6, 7, 10, 11"
1635
+ },
1636
+ {
1637
+ "type": "ref_text",
1638
+ "bbox": [
1639
+ 0.504,
1640
+ 0.122,
1641
+ 0.892,
1642
+ 0.149
1643
+ ],
1644
+ "angle": 0,
1645
+ "content": "[17] Joseph Redmon and Ali Farhadi. Yolov3: An incremental improvement. arXiv preprint arXiv:1804.02767, 2018. 2"
1646
+ },
1647
+ {
1648
+ "type": "ref_text",
1649
+ "bbox": [
1650
+ 0.504,
1651
+ 0.151,
1652
+ 0.892,
1653
+ 0.192
1654
+ ],
1655
+ "angle": 0,
1656
+ "content": "[18] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. NeurIPS, 28, 2015. 1, 6, 7"
1657
+ },
1658
+ {
1659
+ "type": "ref_text",
1660
+ "bbox": [
1661
+ 0.504,
1662
+ 0.194,
1663
+ 0.892,
1664
+ 0.235
1665
+ ],
1666
+ "angle": 0,
1667
+ "content": "[19] Peize Sun, Yi Jiang, Enze Xie, Wenqi Shao, Zehuan Yuan, Changhu Wang, and Ping Luo. What makes for end-to-end object detection? In ICML, pages 9934-9944, 2021. 2"
1668
+ },
1669
+ {
1670
+ "type": "ref_text",
1671
+ "bbox": [
1672
+ 0.504,
1673
+ 0.237,
1674
+ 0.892,
1675
+ 0.291
1676
+ ],
1677
+ "angle": 0,
1678
+ "content": "[20] Peize Sun, Rufeng Zhang, Yi Jiang, Tao Kong, Chenfeng Xu, Wei Zhan, Masayoshi Tomizuka, Lei Li, et al. Sparse r-cnn: End-to-end object detection with learnable proposals. arXiv preprint arXiv:2011.12450, 2020. 1"
1679
+ },
1680
+ {
1681
+ "type": "ref_text",
1682
+ "bbox": [
1683
+ 0.504,
1684
+ 0.294,
1685
+ 0.892,
1686
+ 0.334
1687
+ ],
1688
+ "angle": 0,
1689
+ "content": "[21] Zhiqing Sun, Shengcao Cao, Yiming Yang, and Kris M Kitani. Rethinking transformer-based set prediction for object detection. In ICCV, pages 3611-3620, 2021. 3"
1690
+ },
1691
+ {
1692
+ "type": "ref_text",
1693
+ "bbox": [
1694
+ 0.504,
1695
+ 0.336,
1696
+ 0.892,
1697
+ 0.377
1698
+ ],
1699
+ "angle": 0,
1700
+ "content": "[22] Zhi Tian, Chunhua Shen, Hao Chen, and Tong He. Fcos: Fully convolutional one-stage object detection. In ICCV, pages 9627-9636, 2019. 1, 3, 4"
1701
+ },
1702
+ {
1703
+ "type": "ref_text",
1704
+ "bbox": [
1705
+ 0.504,
1706
+ 0.379,
1707
+ 0.892,
1708
+ 0.418
1709
+ ],
1710
+ "angle": 0,
1711
+ "content": "[23] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of machine learning research, 9(11), 2008. 12"
1712
+ },
1713
+ {
1714
+ "type": "ref_text",
1715
+ "bbox": [
1716
+ 0.504,
1717
+ 0.421,
1718
+ 0.892,
1719
+ 0.475
1720
+ ],
1721
+ "angle": 0,
1722
+ "content": "[24] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. NeurIPS, 30, 2017. 1, 10"
1723
+ },
1724
+ {
1725
+ "type": "ref_text",
1726
+ "bbox": [
1727
+ 0.504,
1728
+ 0.478,
1729
+ 0.892,
1730
+ 0.531
1731
+ ],
1732
+ "angle": 0,
1733
+ "content": "[25] Yingming Wang, Xiangyu Zhang, Tong Yang, and Jian Sun. Anchor detr: Query design for transformer-based detector. In AAAI, volume 36, pages 2567-2575, 2022. 2, 3, 4, 6, 7, 10, 12"
1734
+ },
1735
+ {
1736
+ "type": "ref_text",
1737
+ "bbox": [
1738
+ 0.504,
1739
+ 0.534,
1740
+ 0.892,
1741
+ 0.587
1742
+ ],
1743
+ "angle": 0,
1744
+ "content": "[26] Gongjie Zhang, Zhipeng Luo, Yingchen Yu, Kaiwen Cui, and Shijian Lu. Accelerating detr convergence via semantic-aligned matching. In CVPR, pages 949-958, 2022. 3, 6, 7, 10"
1745
+ },
1746
+ {
1747
+ "type": "ref_text",
1748
+ "bbox": [
1749
+ 0.504,
1750
+ 0.591,
1751
+ 0.892,
1752
+ 0.646
1753
+ ],
1754
+ "angle": 0,
1755
+ "content": "[27] Gongjie Zhang, Zhipeng Luo, Yingchen Yu, Jiaxing Huang, Kaiwen Cui, Shijian Lu, and Eric P Xing. Semantic-aligned matching for enhanced detr convergence and multi-scale feature fusion. arXiv preprint arXiv:2207.14172, 2022. 3"
1756
+ },
1757
+ {
1758
+ "type": "ref_text",
1759
+ "bbox": [
1760
+ 0.504,
1761
+ 0.648,
1762
+ 0.892,
1763
+ 0.702
1764
+ ],
1765
+ "angle": 0,
1766
+ "content": "[28] Hao Zhang, Feng Li, Shilong Liu, Lei Zhang, Hang Su, Jun Zhu, Lionel M Ni, and Heung-Yeung Shum. Dino: Detr with improved denoising anchor boxes for end-to-end object detection. ICLR, 2022. 3, 5, 6, 8"
1767
+ },
1768
+ {
1769
+ "type": "ref_text",
1770
+ "bbox": [
1771
+ 0.504,
1772
+ 0.705,
1773
+ 0.892,
1774
+ 0.731
1775
+ ],
1776
+ "angle": 0,
1777
+ "content": "[29] Xingyi Zhou, Dequan Wang, and Philipp Krahenbuhl. Objects as points. In arXiv preprint arXiv:1904.07850, 2019. 3"
1778
+ },
1779
+ {
1780
+ "type": "ref_text",
1781
+ "bbox": [
1782
+ 0.504,
1783
+ 0.733,
1784
+ 0.892,
1785
+ 0.774
1786
+ ],
1787
+ "angle": 0,
1788
+ "content": "[30] Xingyi Zhou, Jiacheng Zhuo, and Philipp Krahenbuhl. Bottom-up object detection by grouping extreme and center points. In CVPR, pages 850-859, 2019. 3"
1789
+ },
1790
+ {
1791
+ "type": "ref_text",
1792
+ "bbox": [
1793
+ 0.504,
1794
+ 0.776,
1795
+ 0.892,
1796
+ 0.817
1797
+ ],
1798
+ "angle": 0,
1799
+ "content": "[31] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. In ICLR, 2020. 2, 3, 6, 7, 10"
1800
+ },
1801
+ {
1802
+ "type": "list",
1803
+ "bbox": [
1804
+ 0.504,
1805
+ 0.093,
1806
+ 0.892,
1807
+ 0.817
1808
+ ],
1809
+ "angle": 0,
1810
+ "content": null
1811
+ },
1812
+ {
1813
+ "type": "page_number",
1814
+ "bbox": [
1815
+ 0.481,
1816
+ 0.945,
1817
+ 0.519,
1818
+ 0.957
1819
+ ],
1820
+ "angle": 0,
1821
+ "content": "15547"
1822
+ }
1823
+ ]
1824
+ ]
2023/SAP-DETR_ Bridging the Gap Between Salient Points and Queries-Based Transformer Detector for Fast Model Convergency/909bbbe4-4e26-472f-ab37-1d60d8bab4b6_origin.pdf ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:4041f74fc5012d002d394546532efcbb5d9ebaf2fbf62bae196979cbfcf9dba5
3
+ size 3355680
2023/SAP-DETR_ Bridging the Gap Between Salient Points and Queries-Based Transformer Detector for Fast Model Convergency/full.md ADDED
@@ -0,0 +1,268 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # SAP-DETR: Bridging the Gap between Salient Points and Queries-Based Transformer Detector for Fast Model Convergence
2
+
3
+ Yang Liu $^{1,3}$ * Yao Zhang $^{1,3}$ * Yixin Wang $^{2}$ * Yang Zhang $^{4}$ Jiang Tian $^{4}$ Zhongchao Shi $^{4}$ Jianping Fan $^{4}$ Zhiqiang He $^{1,3,5\dagger}$
4
+
5
+ $^{1}$ Institute of Computing Technology (ICT), Chinese Academy of Sciences $^{2}$ Stanford University $^{3}$ University of Chinese Academy of Sciences $^{4}$ AI Lab, Lenovo Research $^{5}$ Lenovo Ltd.
6
+
7
+ {liuyang20c,zhangyao215}@mails.ucas.ac.cn yxinwang@stanford.edu
8
+
9
+ {zhangyang20,tianjiang1,shizc2,jfan1,hezq}@lenovo.com
10
+
11
+ # Abstract
12
+
13
+ Recently, the dominant DETR-based approaches apply central-concept spatial prior to accelerating Transformer detector convergence. These methods gradually refine the reference points to the center of target objects and imbue object queries with the updated central reference information for spatially conditional attention. However, centralizing reference points may severely deteriorate queries' saliency and confuse detectors due to the indiscriminative spatial prior. To bridge the gap between the reference points of salient queries and Transformer detectors, we propose SAlient Point-based DETR (SAP-DETR) by treating object detection as a transformation from salient points to instance objects. Concretely, we explicitly initialize a query-specific reference point for each object query, gradually aggregate them into an instance object, and then predict the distance from each side of the bounding box to these points. By rapidly attending to query-specific reference regions and the conditional box edges, SAP-DETR can effectively bridge the gap between the salient point and the query-based Transformer detector with a significant convergence speed. Experimentally, SAP-DETR achieves $1.4 \times$ convergence speed with competitive performance and stably promotes the SoTA approaches by $\sim 1.0$ AP. Based on ResNet-DC-101, SAP-DETR achieves 46.9 AP. The code will be released at https://github.com/liuyang-ict/SAP-DETR.
14
+
15
+ # 1. Introduction
16
+
17
+ Object detection is a fundamental task in computer vision, whose target is to recognize and localize each object from input images. In the last decade, various detec
18
+
19
+ ![](images/e0e7be2e248a69d58cb6ea61a192c0e5d95c0e8c9cff5ff9a8341202b046f68c.jpg)
20
+ Figure 1. Comparison of SAP-DETR and DAB-DETR under 3-layer decoder model and 36-epoch training scheme. (a) Statistics of the query count in different classification score intervals. (b) and (c) Visualize the distribution of all reference points (pink) and highlight the top-20 classification score queries with their bounding boxes (blue) and reference points (red) in different decoder layers. (d) Visualize the outputs of positive queries and ground truth (red) during the training process, each query has a representative color for its reference point and bounding box. Best viewed in color.
21
+
22
+ tors [6,11,14,18,20,22] based on Convolutional Neural Networks (CNNs), have received widespread attention and made significant progress. Recently, Carion et al. [2] proposed a new end-to-end paradigm for object detection based on the Transformer [24], called DEtction TRansformer (DETR), which treats object detection as a problem of set prediction. In DETR, a set of learnable positional encodings, namely object queries, are employed to aggregate instance features from the context image in Transformer Decoder. The predictions of queries are finally assigned to the ground truth via bipartite matching to achieve end-to-end detection.
23
+
24
+ Despite the promising results of DETR, its application is largely limited by considerably longer training time compared to conventional CNNs. To address this problem, many
25
+
26
+ variants attempted to take a close look at query paradigm and introduced various spatial priors for model convergence and efficacy. According to the type of spatial prior, they can be categorized into implicit and explicit methods. The implicit ones [5, 16, 31] attempt to decouple a reference point from the object query and make use of this spatial prior to attend to the image context features efficiently. The current state-of-the-arts (SoTAs) are dominated by the explicit ones [13, 25], which suggest to instantiate a position with spatial prior for each query, i.e., explicit reference coordinates with a center point or an anchor box. These reference coordinates serve as helpful priors and enable the queries to focus on their expected regions easily. For instance, Anchor DETR [25] introduced an anchor concept (center point with different box size patterns) to formulate the query position and directly regressed the central offsets of the bounding boxes. DABDETR [13] further stretched the center point to a 4D anchor box concept $[cx, cy, w, h]$ to refine proposal bonding boxes in a cascaded manner. However, instantiating the query location as a target center may severely degrade the classification accuracy and convergence speed. As illustrated in Fig. 1, there exist many plausible queries [19] with high-quality classification scores (Fig. 1(a) within red box) and box Intersection over Union (IoU, see the redundant blue boxes in Fig. 1(b) and (c)), which only brings a slight improvement on precision rate but inevitably confuses the detector on the positive query assignments when training with bipartite matching strategy. This is because the plausible predictions are considered in negative classification loss, which severely decelerates the model convergence. As shown in Fig. 1(b) and (c), the predefined reference point of the positive query may not be the nearest one to the center of the ground truth bounding box, and the reference points tend to be centralized or marginalized (cyan arrows in Fig. 1(b)), hence losing the spatial specificity. With further insight into the one-to-one label assignment during the training process, we find that the query, whose reference point is closest to the center point, also has a high-quality IoU, but it still exists a disparity with the positive query in the classification confidence. Therefore, we argue that such a centralized spatial prior may cause degeneration of target consistency in both classification and localization tasks, which leads to inconsistent predictions.
27
+
28
+ Furthermore, the mentioned central point-based variants also have difficulties in detecting occluded objects. For example, Fig. 1(d) shows that DAB-DETR detects the left baseman twice, while the query point in SAP-DETR is not necessarily the center point, so the query point of the bounding box for the occluded baseman is from a pixel on the occluded baseman on the top right area instead of from the left baseman. One solution for center-based method [25] is to predefine different receptive fields (similar to the scaling anchor box in YOLO [17]) for the position of each query. However, increasing the diversity of the receptive fields for
29
+
30
+ each position query is unsuitable for non-overlapped targets, as it still generates massive indistinguishable predictions for one position as same as other center-based models.
31
+
32
+ To bridge these gaps, in this paper, we present a novel framework for Transformer detector, called SAclient Point-based DETR (SAP-DETR), which treats object detection as a transformation from salient points to instance objects. Instead of regressing the reference point to the target center, we define the reference point belonging to one positive query as a salient point, keep this query-specific spatial prior with a scaling amplitude, and then gradually update them to an instance object by predicting the distance from each side of the bounding box. Specifically, we tile the mesh-grid referenced points and initialize their center/corner as the query-specific reference point. To disentangle the reference sparsity as well as stabilize the training process, a movable strategy with scaling amplitude is applied for reference point adjustment, which prompts queries to consider their reference grid as the salient region to perform image context attention. By localizing each side of the bounding box layer by layer, such query-specific spatial prior enables compensation for the over-smooth/inadequacy problem during center-based detection, thereby vastly promoting model convergence speed. Inspired by [5, 13, 16], we also take advantage of both Gaussian spatial prior and conditional cross-attention mechanism, and then a salient point enhanced cross-attention mechanism is developed to distinguish the salient region and other conditional extreme regions from the context image features.
33
+
34
+ We bridge the gap between salient points and query-based Transformer detector by speedily attending to the query-specific region and other conditional regions. The extensive experiments have shown that SAP-DETR achieves superior convergence speed and performance. To the best of our knowledge, this is the first work to introduce the salient point based regression into end-to-end query-based Transformer detectors. Our contributions can be summarized as follows.
35
+
36
+ 1) We introduce the salient point concept into query-based Transformer detectors by assigning query-specific reference points to object queries. Unlike center-based methods, we restrict the reference location and define the point of the positive query as the salient one, hence enlarging the discrepancy of query as well as reducing the redundant predictions (see Fig. 1). Thanks to the efficacy of the query-specific prior, our SAP-DETR accelerates the convergence speed greatly, achieving competitive performance with $30\%$ fewer training epochs. The proposed movable strategy further boosts SAP-DETR to a new SoTA performance.
37
+ 2) We devise a point-enhanced cross-attention mechanism to imbue query with spatial prior based on both reference point and box sides for final specific region attention.
38
+ 3) Evaluation over COCO dataset has demonstrated that SAP-DETR achieves superior convergence speed and detection accuracy. Under the same training settings, SAP-DETR
39
+
40
+ outperforms the SoTA approaches with a large margin.
41
+
42
+ # 2. Related Work
43
+
44
+ Anchor-Free Object Detectors. Classical anchor-free object detectors can be grouped into center-based and keypoint-based approaches. The center-based approaches aim to localize the target objects based on the central locations [11] or predefined ROI [22]. For example, FCOS [22] treated all points within the bounding box as positive ones to predict their distances from each side $([\ell, t, r, b])$ , and a centerness score was then considered to prohibit the low-quality prediction whose point is located near the border. Compared with FCOS, we also restrict the candidate queries within the bounding box but treat only one as positive to perform end-to-end object detection via an inner matching cost.
45
+
46
+ The target of keypoint-based approaches is to localize the specific object locations and assign them to the predefined keypoints of the object for box localized training. For instance, diagonal corner points were considered in CornerNet [8], center point was further grouped into CenterNet [29], and ExtremeNet [30] added some conjectural extreme points for object localization. These works showed an impressive performance, but the complicated keypoint matching may limit their upper bound. Our SAP-DETR takes the advantage of salient point regression to focus on the distinct regions without complicated point-based supervision.
47
+
48
+ Query-Based Transformer Detectors. DETR [2] pioneered a new paradigm of Transformer detector for end-to-end object detection without any post-processing [1]. In DETR, a new representation, namely object query, aggregates the instance features and then yields a detection result for each instance object [15]. Following DETR, many voterists put efforts on the optimization of convergence and accuracy.
49
+
50
+ Sun et al. [21] revealed that the main reason for slow convergence of DETR is attributed to the Transformer decoder, and they considered an encoder-only structure to alleviate such a problem. For in-depth understanding of the object query, one way is to generate a series of implicit spatial priors from queries to guide feature aggregation in cross-attention layers. SMCA [5] applied a Gaussian-like attention map to augment the query-concerned features spatially. The reference point concept was first introduced by Deformable DETR [31], where the sampling offsets are predicted by each reference point to perform deformable cross-attention. Following such a concept, Conditional DETR [16] reformulated the attention operation and rebuilt positional queries based on the reference points to facilitate extreme region discrimination. Another way is towards position-instantiation explicitly, where this position information enables to directly conduct positional query generation. Anchor DETR [25] utilized a predefined 2D anchor point $[cx, cy]$ to explicitly capitalize on the spatial prior during cross-attention and box regression. DAB-DETR [13] extended such a 2D concept to
51
+
52
+ ![](images/610de358f6711972cd0b18bdd81eb531a04351f72a5b205d48cb34f2440f488e.jpg)
53
+ Figure 2. Illustration of SAP-DETR. Each object query in SAP-DETR is assigned to a specific grid region and initialized by the corner/center of the grid as its reference point. A learnable 4D coordinate represents the distance from the four sides of the box to the reference point. Both reference points and box sides are served as positional encodings added/concatenated to content embeddings. All embeddings are refined to predict target objects gradually.
54
+
55
+ a 4D anchor box $[cx, cy, w, h]$ and refined it layer-by-layer.
56
+
57
+ The recent accelerating convergence methods are based on auxiliary queries for facilitating detector discrimination. DN-DETR [9] demonstrates the slow model convergence is mainly caused by the instability of bipartite matching, thus providing a denoising training to eliminate this issue. DINO [28] inherits this advance and further introduces negative queries to perform contrastive denoising. Group-DETR [4] proposes a group-wise one-to-many label assignment to match multiple positive object queries with more gradients for fast DETR convergence.
58
+
59
+ The most relevant approaches to ours are Point-DETR [3] and SAM-DETR [26, 27]. The former applied a point encoder for annotated point label infusion in teacher model, and the latter directly updated content embeddings by extracting salient points from image features for query-image semantic alignment. Unlike these concepts, we redefine the salient point from the perspective of the positive query's position and replace the center-concept prior with the query-specific position, thereby attending extreme regions, differentiating queries' saliency, and alleviating redundant predictions.
60
+
61
+ # 3. Method
62
+
63
+ We propose SAP-DETR to bridge the gap between salient points and query-based detectors. Following DETR, the extracted image features are fed into Transformer encoder
64
+
65
+ after adding positional encodings, and then re-aggregated by object queries in Transformer decoder. In pursuit of the query-specific prior, we dispense a movable strategy for each query based on a fixed grid region. The query, whose reference region overlaps with ground truth objects, is allowed to predict the relative offsets from four sides of the bounding box to the points. Given the query-specific reference point and the proposal box sides, we propose salient point enhanced cross-attention mechanism to imbue query with spatial prior, thereby attending to extreme regions effectively. Additionally, we discuss two common issues in DETR-like models and address them for further improvements.
66
+
67
+ # 3.1. Salient Points-Based Object Detection
68
+
69
+ Overview. Previous methods [13, 16, 25] normally decompose the object query into both content and position embeddings (queries), and form a center-based anchor point/box prior on the position ones. Unlike the central concept, we tile a fixed mesh-grid region, initialize their left-top corner as the reference points with 2D coordinate $\boldsymbol{r} = \{x, y\} \in [0,1]^2$ , and instantiate a learnable 4D offset distance $s = \{\ell, t, r, b\} \in [0,1]^4$ from the reference point to the sides of proposal bounding box for each object query. The object query can be referred as $q = \{\mathbf{e}; r, s\}$ , where $\mathbf{e} \in \mathbb{R}^d$ is the content embedding with $d$ dimension. Instead of regressing the center, width, and height of a bounding box, we follow FCOS [22] and directly supervise the 4D offset from the four sides of a bounding box to the reference point. The final box prediction is formulated as $\hat{\boldsymbol{b}} = \{\hat{x} - \hat{\ell}, \hat{y} - \hat{\bar{t}}, \hat{x} + \hat{r}, \hat{y} + \hat{b}\}$ . Ideally, we here fix the reference point $\{\hat{x}, \hat{y}\} = \{x, y\}$ (the movable update strategy is introduced in the next subsection) and only update the 4D box side prediction layer by layer. The prediction for each decoder layer can be calculated by
70
+
71
+ $$
72
+ \Delta \boldsymbol {s} _ {l} = \operatorname {B o x H e a d} _ {l} (\boldsymbol {s} _ {l - 1}, \boldsymbol {e} _ {l - 1}, \boldsymbol {r} _ {l - 1}),
73
+ $$
74
+
75
+ $$
76
+ \hat {\boldsymbol {s}} _ {l} = \sigma \left(\sigma^ {- 1} \left(\boldsymbol {s} _ {l - 1}\right) + \Delta \boldsymbol {s} _ {l}\right), \quad \boldsymbol {s} _ {l} = \operatorname {D e t a c h} (\hat {\boldsymbol {s}} _ {l}),
77
+ $$
78
+
79
+ $$
80
+ \boldsymbol {r} _ {l} = \hat {\boldsymbol {r}} _ {l} = \boldsymbol {r} _ {l - 1}, \quad \hat {\boldsymbol {b}} _ {l} = \left\{\hat {\boldsymbol {r}} _ {l} - \hat {\boldsymbol {s}} _ {l} [: 2 ], \hat {\boldsymbol {r}} _ {l} + \hat {\boldsymbol {s}} _ {l} [2: ] \right\}, \tag {1}
81
+ $$
82
+
83
+ where $\sigma$ and $\sigma^{-1}$ are the sigmoid and inverse sigmoid operation, respectively. $\Delta s_{l}$ denotes the side offset prediction. $\hat{s}_l$ , $\hat{r}_l$ , and $\hat{\pmb{b}}_l$ are the predicted side distance, reference points, and box location from the $l$ decoder layer, respectively. The BoxHead is the prediction head following the layer- $l$ decoder, which is independent between different decoder layers in our settings. Detach operation follows DAB-DETR [13].
84
+
85
+ During the training process, each query is only allowed to predict the bounding boxes that overlap its reference region. We adapt this rule into the one-to-one bipartite matching process via an inner matching cost $\mathcal{L}_{\mathrm{inner}}$ . Given $N$ queries $\mathbf{Q} = \{\pmb{q}_1,\pmb{q}_2,\dots ,\pmb{q}_N\}$ and $M$ ground truth objects $\mathbf{G} = \{\pmb{g}_1,\pmb{g}_2,\dots ,\pmb{g}_M\}$ , the $\mathcal{L}_{\mathrm{inner}}(\pmb{g}_i,\pmb{q}_j)$ of each query-box pair is a step function to penalize the reference point $\pmb{r}_j$ of $\pmb{q}_j$ with value of $k$ when $\pmb{r}_j$ is outside the bounding box of
86
+
87
+ $g_{i}$ . We denote $i \in [1, M]$ and $j \in [1, N]$ as the index of query and ground truth, respectively. $k$ can be viewed as a penalty cost, and default to $10^{5}$ . The final permutation of the one-to-one label assignment is formulated as
88
+
89
+ $$
90
+ \mathcal {L} _ {\text {i n n e r}} \left(\boldsymbol {g} _ {i}, \boldsymbol {q} _ {j}\right) := \mathbf {k} _ {\boldsymbol {r} _ {j} \notin \boldsymbol {g} _ {i}},
91
+ $$
92
+
93
+ $$
94
+ \hat {\eta} = \underset {\eta \in \mathfrak {Y} _ {N}} {\operatorname {a r g m i n}} \sum_ {i} ^ {N} \mathcal {L} _ {\text {m a t c h}} + \mathcal {L} _ {\text {i n n e r}}, \tag {2}
95
+ $$
96
+
97
+ where $\mathcal{L}_{\mathrm{match}}$ is the original pair-wise matching cost consist of both classification and localization costs [2]. $\eta \in \mathfrak{V}_N$ is a permutation of $N$ elements for bipartite matching.
98
+
99
+ Movable Reference Point. Due to the sparseness of the fixed reference point, some small and slender objects may be indistinguishable when there is no reference point inside these objects. Despite the bipartite matching forcing each object to be assigned to one object query, the positive query, whose reference point is outside the assigned bounding box, is unable to accurately regress the distance from each side by a value between 0 and 1. One straightforward solution is to adjust the locations of reference points inside the ground truth bounding boxes to ensure that each object can be detected by an inner reference point. Similar to the aforementioned box refinement, we first perform a movable reference point design to dynamically update the reference points of each query layer by layer. However, such a full-image point regression inevitably expands the search space as vast variable determinations, causing the final reference point to be trapped in an unexpected corner of the bounding box. To reduce the training search spaces, we scale the offset amplitude of points within their specific grid regions, as illustrated in Fig. 3. Such an operation limits the range of offset values, and hence prevents a large searching space. It is implemented by applying the sigmoid activation $\sigma$ and multiplying a scale factor $s_{\mathrm{grid}}$ whose value equals to the height and width of one grid. The update process of the reference points is formulated as
100
+
101
+ $$
102
+ \Delta \boldsymbol {r} _ {l} ^ {\prime} = \operatorname {P o i n t H e a d} _ {l} \left(\boldsymbol {s} _ {l - 1}, \boldsymbol {e} _ {l - 1}, \boldsymbol {r} _ {l - 1}\right),
103
+ $$
104
+
105
+ $$
106
+ \Delta \boldsymbol {r} _ {l} = \sigma \left(\sigma^ {- 1} \left(\boldsymbol {r} _ {l - 1} - \boldsymbol {r} _ {0}\right) + \Delta \boldsymbol {r} _ {l} ^ {\prime}\right), \tag {3}
107
+ $$
108
+
109
+ $$
110
+ \hat {\boldsymbol {r}} _ {l} = \boldsymbol {r} _ {0} + \Delta \boldsymbol {r} _ {l} \cdot \boldsymbol {s} _ {\text {g r i d}}, \quad \boldsymbol {r} _ {l} = \operatorname {D e t a c h} (\hat {\boldsymbol {r}} _ {l}),
111
+ $$
112
+
113
+ where $\Delta \boldsymbol{r}_l'$ and $\Delta \boldsymbol{r}_l$ are the predicted offsets from $\boldsymbol{r}_l$ to both $\boldsymbol{r}_{l-1}$ and $\boldsymbol{r}_0$ before the sigmoid activation $\sigma$ , respectively.
114
+
115
+ # 3.2. Salient Point Enhanced Cross-Attention
116
+
117
+ In cross-attention layers, existing center-based methods are limited to the attention on both center and sides of the ground truth bounding box, causing detector confusion among the queries with the same center and side attention. To this end, we expect the queries to focus on their specific regions based on the reference points, four box sides, and other conditional regions in different heads. Accordingly,
118
+
119
+ ![](images/db807192e02719a82d755a7878de64a4f620ef8a0f63351e89cf8572332e9f7c.jpg)
120
+ Figure 3. Movable reference point. The reference points are initialized by the center/corner points of the mesh-grid. Based on the inner loss, only the green dashed box can be predicted by the inner points when reference points are fixed. By moving the reference points within their grid, the blue dashed boxes can be detected accurately without extended searching space.
121
+
122
+ we consider an improved Gaussian [5] $\mathbf{G}$ and conditional attention [16] $\mathbf{A}_{\mathrm{peca}}$ to enhance query specificity and spatially extreme region discrimination. The final attention map $\mathbf{A}_{\mathrm{cross}}$ is the sum of the two attentions $\mathbf{A}_{\mathrm{cross}} = \mathbf{G} + \mathbf{A}_{\mathrm{peca}}$ .
123
+
124
+ Side Directed Gaussian (SDG). Similar to the movable strategy, we enforce the predicted Gaussian attention to be inside the proposal bounding box to reduce the searching space. Given a reference point $r$ , the offset scales $\mathbf{o} \in [-1,1]^2$ for $H$ heads are produced by a simple MLP with a tanh activation, and then multiply to the two sides of the proposal bounding box for head-specific point offset generation, where the direction is guided by the sign of the offset scales. The head-specific points are generated by Algorithm 1. For each head, the Gaussian-like spatial weight map $G_{i}$ effecting on each pixel $(x,y)$ of context features is then formulated as
125
+
126
+ $$
127
+ G _ {i} (x, y) = \exp \left(- \frac {\left(x - c _ {w , i}\right) ^ {2}}{v _ {w , i} ^ {2}} - \frac {\left(y - c _ {h , i}\right) ^ {2}}{v _ {h , i} ^ {2}}\right). \tag {4}
128
+ $$
129
+
130
+ Point Enhanced Cross-Attention (PECA). As aforementioned, the semantic class for the query is closely related to its referenced location in our SAP-DETR. To enhance the correlation between queries and their references, we concatenate the locations to the content queries after the sinusoidal positional encoding (PE). Take a close look at the conditional attention [16], we find that the linear positional embedding mostly focuses on one box side in each attention head. So we introduce a more straightforward attention mechanism, where the four side coordinates are concatenated and assigned to the corresponding head for side attention. The process of PECA is formulated as
131
+
132
+ $$
133
+ \begin{array}{r l} \mathbf {A} _ {\text {p e c a}} & = \mathbf {e} _ {q} \mathbf {e} _ {k} ^ {\top} + \mathbf {T P E} (\boldsymbol {r} _ {q}) \mathbf {P E} (\boldsymbol {r} _ {k}) ^ {\top} \\ & + \mathbf {T} g (\mathbf {P E} (\boldsymbol {r} _ {q} - \{\ell , t \}, \boldsymbol {r} _ {q} + \{r, b \})) \mathbf {P E} (\boldsymbol {r} _ {k}) ^ {\top}, \end{array} \tag {5}
134
+ $$
135
+
136
+ where $g$ is a linear layer mapping 4D-PE into 2D to keep channel dimension consistency. T [16] is a scaling transformation matrix. More details of T and comparison with conditional cross-attention are available in Appendix C.
137
+
138
+ # Algorithm 1 Side Directed Gaussian
139
+
140
+ Input: Content embedding e, reference point $r$ and box s.
141
+
142
+ Output: Head-specific points $\mathbf{c} = \{(c_{w,i}, c_{h,i}) | i \in H\}$ and head-specific attention $\mathbf{v} = \{(v_{w,i}, v_{h,i}) | i \in H\}$ .
143
+
144
+ 1: Predict offset scale and attention scale based on content embedding, $\mathbf{o} = \tanh (\mathrm{MLP}(\mathbf{e}))$ , $\mathbf{v} = \mathrm{MLP}(\mathbf{e})$
145
+ 2: for $h\gets 1\in H$ do
146
+ 3: Select the index of direction guided by the sign of offset scale, $\{a,b\} = \mathrm{sgn}(\mathbf{o}_i) + \{1,2\}$ , $a,b\in \{0,1,2,3\}$ ;
147
+ 4: According to the index of direction, predict head-specific point, $\mathbf{c}_i = \mathbf{o}_i\cdot \mathbf{s}[a,b] + \mathbf{r}$
148
+ 5: end for
149
+ 6: return $\mathbf{c}_i, \mathbf{v}_i, \forall i = 1, \dots, H$
150
+
151
+ # 3.3. SAP-DETR with Denoising Strategy
152
+
153
+ To further explore the capability of our proposed SAP-DETR, we develop SAP-DN-DETR and SAP-DINO-DETR by adding the denoising auxiliary loss [9,28] into the training process. In the denoised SAP-DETR, the main difference from both DN-DETR and DINO lies in the noise design. Instead of the center point, we perform the box jittering and randomly sample a point from the intersection region between the original bounding box and the jittering one as the reference point. As the denoising strategy only serves as an auxiliary training loss increasing the training cost, the variants of denoising models are test-free whose Params and GFLOPs are the same as SAP-DETR models.
154
+
155
+ # 4. Experimental Results
156
+
157
+ # 4.1. Implementation Details
158
+
159
+ We conduct the experiments on the COCO 2017 [12] object detection dataset, containing about 118K training images and 5K validation images. All models are evaluated by the standard COCO evaluation metrics. We follow the vanilla DETR [2] structure that consists of a CNN backbone, a stack of Transformer encoder-decoder layers, and two prediction heads for class label and bounding box prediction. We use ImageNet-pretrained ResNet [7] as our backbone, and report results based on the ResNet and its $\times 1/16$ -resolution extension ResNet-DC. Unlike DAB-DETR [13] sharing box and label head for each layer, we share the class head except the first layer and use an independent box head for the box regression of each layer (for more details please refer to Appendix D). As the mesh-grid initialization for reference points in SAP-DETR, we consider the number of queries $N$ as a perfect square for uniform distribution. Unless otherwise specified, we use $N = 400$ queries in the experiments. Precisely, we also provide a comparison under $N = 300$ in Tab. 2, the standard setting in DETR-like models.
160
+
161
+ We adopt two different Transformer structures for experiments where a 3-layer encoder-decoder stack is evaluated to
162
+
163
+ ![](images/e67c17b52b62bc18f59407f49da518ec9d46208e46c8ed65384428ba5d8ecd32.jpg)
164
+ Figure 4. Comparison of performance and training losses curves.
165
+
166
+ ![](images/c4e4471d6732af893721aadaca57cebf551f6985172900d817e0947edfa5e3c0.jpg)
167
+
168
+ ![](images/a2fba28d8751276d87d046916bae807f8b5b51e25a59642266ec0fb9122b2506.jpg)
169
+
170
+ demonstrate our lightweight model efficacy compared with the traditional CNN detectors, and a 6-layer encoder-decoder stack is aligned with previous DETR variants to investigate the performance of large model. Both are trained on two training schemes: the 12-epoch and 36-epoch schemes with a learning rate drop after 11 and 30 epochs, respectively. All models are trained on the Nvidia A100 GPUs with batch size of 16 and 8 for ResNet and ResNet-DC, respectively. For more training details, please refer to Appendix F.
171
+
172
+ # 4.2. Main Results
173
+
174
+ As shown in Tab. 1 and Tab. 2, we comprehensively compare our proposed SAP-DETR with the traditional CNN detectors [18], the original DETR [2], and other DETR-like detectors [5, 13, 16, 25, 26, 31] on COCO 2017 validation dataset. For in-depth analysis, we conduct the comparison in two aspects: model convergence and efficacy.
175
+
176
+ Model Convergence. Compared with traditional CNN detectors, Transformer detectors are always subject to laborious training time. For example, under the same 12-epoch training scheme, Faster RCNN [18] still achieves good performance, but the mainstream DETR-like models may suffer from inadequate training and perform poorly without the help of auxiliary losses [9]. Under the 12-epoch training scheme, our proposed SAP-DETR can accelerate model convergence significantly, boosting DAB-DETR [13] by 3.9 AP and 2.7 AP on 3-layer and 6-layer encoder-decoder structures, respectively. Compared with the current SoTA, our SAP-DETR also outperforms SAM-DETR [26] by $\sim 1.3$ AP, with reducing $\sim 17\%$ parameters and $\sim 10\%$ GFLOPs. Take a close look at the training process, as illustrated in Fig. 4, SAP-DETR conducts with rapid descent curves in both classification and box regression losses. Notably, there is a large margin in classification loss between ours and SoTA methods, which is benefited from the query-specific reference point, hence boosting model performance in early epochs.
177
+
178
+ Model Efficacy. To analyze model efficacy, we report results on long training epochs and high-resolution features in Tab. 1. Under the 36-epoch training scheme, SAP-DETR achieves superior performance among all single-scale Transformer detectors, especially on middle and large targets. For example, SAP-DETR boosts DAB-DETR by $2.0\mathrm{AP}_{\mathrm{M}}$ and $4.1\mathrm{AP}_{\mathrm{L}}$ with 3-layer models, $1.0\mathrm{AP}_{\mathrm{M}}$ and $1.9\mathrm{AP}_{\mathrm{L}}$ with
179
+
180
+ 6-layer models, which further verifies the effectiveness of our proposed salient point concept for overlapping object detection. Along with layer increase, a deficient upper-bound of SAM-DETR is exposed, with obviously lower 0.5 AP promotion compared to our 1.0 AP improvement. Persuasively, we also report the 50-epoch training results based on the 300-query setting. To align with our mesh-grid initialization strategy, we tile a $17 \times 18$ mesh-grid (306 queries) for each reference point initialization. Tab. 2 shows our main results and the most representative approaches with their original reported performance. Notably, SAP-DETR outperforms the current SoTAs with comparable costs based on all backbones. With low-resolution features ( $\times 1/32$ ), it significantly boosts both middle and large object detection accuracy.
181
+
182
+ Combine with Other Fast Convergence Methods. As shown in Tab. 3, we compare our SAP-DETR variants with the current fast convergence methods [4, 9, 28]. With such a subtle modification, our SAP-DETR (grey rows) results in a significant performance improvement compared with the original methods (white rows). Under the 12-epoch training scheme, there exist 0.5-1.9 AP improvements on DN-DETR [9] and 0.7-1.6 AP improvements on Group-DETR [4], but the promotions are slightly reduced when implemented on DINO [28]. We hypothesise that there exists the same effect between the negative query of contrastive denoising [28] and our query-salient reference point. Moreover, we observe that the performance improvements largely originate from the large object detection, especially based on ResNet-DC5 family backbones. We speculate that DETR may prefer the high-resolution features ( $\times 1/16$ ) rather than the low-resolution ones ( $\times 1/32$ ), and our SAP-DETR can distinguish the salient points accurately on the high-resolution, thereby taking full advantage of the large object detection.
183
+
184
+ # 4.3. Ablation Study
185
+
186
+ Effectiveness of Each Component. To offer an intuitionistic comparison of model convergence for each component, Tab. 4 reports the effectiveness of them based on the 3-layer encoder-decoder structure and 12-epoch training scheme. I). The proposed salient point concept based on content embeddings improves the performance from 32.3 AP to 33.5
187
+
188
+ <table><tr><td>Method</td><td>#Epochs</td><td>#Params(M)</td><td>GFLOPs</td><td>AP</td><td>\( AP_{50} \)</td><td>\( AP_{75} \)</td><td>\( AP_S \)</td><td>\( AP_M \)</td><td>\( AP_L \)</td></tr><tr><td colspan="10">3-Layer Encoder-Decoder Transformer Neck with ResNet-50 Backbone</td></tr><tr><td>DETR-R50 [2]</td><td>36</td><td>33</td><td>82</td><td>15.8</td><td>28.0</td><td>15.4</td><td>5.3</td><td>16.7</td><td>24.6</td></tr><tr><td>Deformable DETR-R50 [31]</td><td>36</td><td>30</td><td>77</td><td>37.1</td><td>57.6</td><td>39.4</td><td>18.3</td><td>40.8</td><td>51.6</td></tr><tr><td>SMCA-DETR-R50 [5]</td><td>12 / 36</td><td>-</td><td>-</td><td>28.8 / 37.7</td><td>48.1 / 58.7</td><td>29.9 / 40.1</td><td>13.8 / 19.4</td><td>31.3 / 40.5</td><td>41.3 / 54.8</td></tr><tr><td>Conditional DETR-R50 [16]</td><td>12 / 36</td><td>40</td><td>82</td><td>29.6 / 37.1</td><td>48.7 / 57.9</td><td>30.7 / 39.0</td><td>13.0 / 17.6</td><td>32.3 / 40.3</td><td>43.1 / 55.0</td></tr><tr><td>Anchor DETR-R50 [25]</td><td>12 / 36</td><td>31</td><td>79</td><td>30.8 / 37.6</td><td>51.1 / 58.7</td><td>31.8 / 39.7</td><td>14.3 / 18.8</td><td>34.1 / 41.5</td><td>44.3 / 53.5</td></tr><tr><td>DAB-DETR-R50 [13]</td><td>12 / 36</td><td>34</td><td>83</td><td>32.3 / 39.0</td><td>51.3 / 58.6</td><td>34.0 / 41.8</td><td>15.7 / 20.0</td><td>35.2 / 42.5</td><td>45.7 / 56.0</td></tr><tr><td>SAM-DETR-w/SMCA-R50 [26]</td><td>12 / 36</td><td>41</td><td>89</td><td>35.1 / 40.4</td><td>54.7 / 60.7</td><td>36.7 / 42.7</td><td>16.0 / 20.2</td><td>38.4 / 44.4</td><td>52.1 / 58.3</td></tr><tr><td>SAP-DETR-R50 (Ours)</td><td>12 / 36</td><td>36</td><td>84</td><td>36.2 / 41.2</td><td>56.2 / 61.6</td><td>37.9 / 43.4</td><td>16.4 / 21.0</td><td>39.5 / 44.5</td><td>53.8 / 60.1</td></tr><tr><td colspan="10">6-Layer Encoder-Decoder Transformer Neck with ResNet-50 Backbone</td></tr><tr><td>DETR-R50 [2]</td><td>36</td><td>42</td><td>89</td><td>14.0</td><td>24.4</td><td>14.0</td><td>4.2</td><td>13.7</td><td>22.5</td></tr><tr><td>Deformable DETR-R50 [31]</td><td>36</td><td>34</td><td>81</td><td>38.0</td><td>58.2</td><td>40.4</td><td>18.5</td><td>41.7</td><td>54.2</td></tr><tr><td>SMCA-DETR-R50 [5]</td><td>12 / 36</td><td>-</td><td>-</td><td>32.4 / 40.1</td><td>52.3 / 61.4</td><td>34.0 / 42.8</td><td>15.5 / 20.3</td><td>34.9 / 43.3</td><td>47.7 / 57.1</td></tr><tr><td>Conditional DETR-R50 [16]</td><td>12 / 36</td><td>44</td><td>90</td><td>33.1 / 40.2</td><td>53.0 / 61.0</td><td>34.8 / 42.4</td><td>14.5 / 19.9</td><td>35.9 / 43.5</td><td>49.2 / 58.8</td></tr><tr><td>Anchor DETR-R50 [25]</td><td>12 / 36</td><td>37</td><td>85</td><td>33.7 / 39.7</td><td>54.5 / 60.5</td><td>35.1 / 41.9</td><td>15.6 / 19.9</td><td>37.3 / 43.5</td><td>49.8 / 57.3</td></tr><tr><td>DAB-DETR-R50 [13]</td><td>12 / 36</td><td>44</td><td>92</td><td>34.9 / 41.0</td><td>55.5 / 61.7</td><td>36.4 / 43.4</td><td>16.2 / 21.3</td><td>38.4 / 44.7</td><td>51.5 / 58.9</td></tr><tr><td>SAM-DETR-w/SMCA-R50 [26]</td><td>12 / 36</td><td>59</td><td>105</td><td>36.2 / 40.9</td><td>57.2 / 62.2</td><td>37.4 / 43.1</td><td>16.1 / 20.1</td><td>39.8 / 44.7</td><td>55.3 / 60.7</td></tr><tr><td>SAP-DETR-R50 (Ours)</td><td>12 / 36</td><td>47</td><td>94</td><td>37.5 / 42.2</td><td>58.5 / 62.7</td><td>39.2 / 44.6</td><td>17.3 / 22.6</td><td>40.6 / 45.7</td><td>55.4 / 60.8</td></tr></table>
189
+
190
+ Table 1. Comparison between Transformer necks. Based on ResNet-50 backbone, all models are trained by the official source codes with their original settings and evaluated on COCO val2017. All models uses 400 queries except Anchor DETR, while Anchor DETR uses 200 queries with 2 pattern embeddings. GFLOPs and Params are measured by $\text{Detectron}^2$ .
191
+
192
+ <table><tr><td>Method</td><td>#Epochs</td><td>#Params(M)</td><td>GFLOPs</td><td>AP</td><td>\( AP_{50} \)</td><td>\( AP_{75} \)</td><td>\( AP_S \)</td><td>\( AP_M \)</td><td>\( AP_L \)</td><td>Infer. \( Time(s / img)^{\dagger} \)</td></tr><tr><td colspan="11">ResNet-50 Backbone</td></tr><tr><td>Faster RCNN-FPN-R50 [10, 18]</td><td>108</td><td>42</td><td>180</td><td>42.0</td><td>62.1</td><td>45.5</td><td>26.6</td><td>45.5</td><td>53.4</td><td>0.039</td></tr><tr><td>DETR-R50 [2]</td><td>500</td><td>41</td><td>86</td><td>42.0</td><td>62.4</td><td>44.2</td><td>20.5</td><td>45.8</td><td>61.1</td><td>0.040</td></tr><tr><td>Deformable DETR-R50 [31]</td><td>50</td><td>34</td><td>78</td><td>39.4</td><td>59.6</td><td>42.3</td><td>20.6</td><td>43.0</td><td>55.5</td><td>0.043</td></tr><tr><td>SMCA-DETR-R50 [5]</td><td>50</td><td>42</td><td>86</td><td>41.0</td><td>-</td><td>-</td><td>21.9</td><td>44.3</td><td>59.1</td><td>0.045</td></tr><tr><td>Conditional DETR-R50 [16]</td><td>50</td><td>44</td><td>90</td><td>40.9</td><td>61.8</td><td>43.3</td><td>20.8</td><td>44.6</td><td>59.2</td><td>0.057</td></tr><tr><td>Anchor DETR-R50 [25]</td><td>50</td><td>39</td><td>85</td><td>42.1</td><td>63.1</td><td>44.9</td><td>22.3</td><td>46.2</td><td>60.0</td><td>0.050</td></tr><tr><td>DAB-DETR-R50 [13]</td><td>50</td><td>44</td><td>\( 90^† \)</td><td>42.2</td><td>63.1</td><td>44.7</td><td>21.5</td><td>45.7</td><td>60.3</td><td>0.059</td></tr><tr><td>SAM-DETR-w/SMCA-R50 [26]</td><td>50</td><td>58</td><td>100</td><td>41.8</td><td>63.2</td><td>43.9</td><td>22.1</td><td>45.9</td><td>60.9</td><td>0.065</td></tr><tr><td>SAP-DETR-R50 (Ours)</td><td>50</td><td>47</td><td>92</td><td>43.1</td><td>63.8</td><td>45.4</td><td>22.9</td><td>47.1</td><td>62.1</td><td>0.063</td></tr><tr><td colspan="11">ResNet-101 Backbone</td></tr><tr><td>Faster RCNN-FPN-R101 [10, 18]</td><td>108</td><td>60</td><td>246</td><td>44.0</td><td>63.9</td><td>47.8</td><td>27.2</td><td>48.1</td><td>56.0</td><td>0.050</td></tr><tr><td>DETR-R101 [2]</td><td>500</td><td>60</td><td>152</td><td>43.5</td><td>63.8</td><td>46.4</td><td>21.9</td><td>48.0</td><td>61.8</td><td>0.066</td></tr><tr><td>Conditional DETR-R101 [16]</td><td>50</td><td>63</td><td>156</td><td>42.8</td><td>63.7</td><td>46.0</td><td>21.7</td><td>46.6</td><td>60.9</td><td>0.070</td></tr><tr><td>Anchor DETR-R101 [25]</td><td>50</td><td>58</td><td>150</td><td>43.5</td><td>64.3</td><td>46.6</td><td>23.2</td><td>47.7</td><td>61.4</td><td>0.068</td></tr><tr><td>DAB-DETR-R101 [13]</td><td>50</td><td>63</td><td>\( 157^† \)</td><td>43.5</td><td>63.9</td><td>46.6</td><td>23.6</td><td>47.3</td><td>61.5</td><td>0.072</td></tr><tr><td>SAP-DETR-R101 (Ours)</td><td>50</td><td>67</td><td>158</td><td>44.4</td><td>64.9</td><td>47.1</td><td>24.1</td><td>48.7</td><td>63.1</td><td>0.078</td></tr><tr><td colspan="11">DC5-ResNet-50 Backbone</td></tr><tr><td>DETR-DC5-R50 [2]</td><td>500</td><td>41</td><td>187</td><td>43.3</td><td>63.1</td><td>45.9</td><td>22.5</td><td>47.3</td><td>61.1</td><td>0.087</td></tr><tr><td>Conditional DETR-DC5-R50 [16]</td><td>50</td><td>44</td><td>195</td><td>43.8</td><td>64.4</td><td>46.7</td><td>24.0</td><td>47.6</td><td>60.7</td><td>0.093</td></tr><tr><td>Anchor DETR-DC5-R50 [25]</td><td>50</td><td>39</td><td>151</td><td>44.2</td><td>64.7</td><td>47.5</td><td>24.7</td><td>48.2</td><td>60.6</td><td>0.069</td></tr><tr><td>DAB-DETR-DC5-R50 [13]</td><td>50</td><td>44</td><td>\( 194^† \)</td><td>44.5</td><td>65.1</td><td>47.7</td><td>25.3</td><td>48.2</td><td>62.3</td><td>0.094</td></tr><tr><td>SAM-DETR-w/SMCA-DC5-R50 [26]</td><td>50</td><td>58</td><td>210</td><td>45.0</td><td>65.4</td><td>47.9</td><td>26.2</td><td>49.0</td><td>63.3</td><td>0.126</td></tr><tr><td>SAP-DETR-DC5-R50 (Ours)</td><td>50</td><td>47</td><td>197</td><td>46.0</td><td>65.5</td><td>48.9</td><td>26.4</td><td>50.2</td><td>62.6</td><td>0.116</td></tr><tr><td colspan="11">DC5-ResNet-101 Backbone</td></tr><tr><td>DETR-DC5-R101 [2]</td><td>500</td><td>60</td><td>253</td><td>44.9</td><td>64.7</td><td>47.7</td><td>23.7</td><td>49.5</td><td>62.3</td><td>0.101</td></tr><tr><td>Conditional DETR-DC5-R101 [16]</td><td>50</td><td>63</td><td>262</td><td>45.0</td><td>65.6</td><td>48.4</td><td>26.1</td><td>48.9</td><td>62.8</td><td>0.105</td></tr><tr><td>Anchor DETR-DC5-R101 [25]</td><td>50</td><td>58</td><td>227</td><td>45.1</td><td>65.7</td><td>48.8</td><td>25.8</td><td>49.4</td><td>61.6</td><td>0.083</td></tr><tr><td>DAB-DETR-DC5-R101 [13]</td><td>50</td><td>63</td><td>\( 263^† \)</td><td>45.8</td><td>65.9</td><td>49.3</td><td>27.0</td><td>49.8</td><td>63.8</td><td>0.110</td></tr><tr><td>SAP-DETR-DC5-R101 (Ours)</td><td>50</td><td>67</td><td>266</td><td>46.9</td><td>66.7</td><td>50.5</td><td>27.9</td><td>51.3</td><td>64.3</td><td>0.130</td></tr></table>
193
+
194
+ Table 2. Comparison of Transformer necks with 300 queries on COCO val2017. All results are reported from their original paper. All models uses 300 queries except Anchor DETR, while Anchor DETR uses 100 queries with 3 pattern embeddings. All inference speeds are measured by a single Nvidia A100 GPU. † denotes the results are measured by ourselves.
195
+
196
+ AP compared to baseline DAB-DETR (row 8-9). Such a query-specific spatial prior enables queries to attend to their expected region from content features (see Figures in Appendix G) and reduces the false detection rate on occluded and partial objects (see Fig. 1(d)), hence boosting detection performance on middle and large objects. However, there exists a drop in small object detection (15.7 APs vs. 14.3 APs), for which we consider that the failure is mainly caused
197
+
198
+ by the query sparsity. 2). Therefore, the movable strategy is applied to alleviate the constraint, improving the final model by $1.0\mathrm{AP}$ and $0.6\mathrm{AP}_{\mathrm{S}}$ (row 1 and 6). 3). Compared with the final model, the inner loss greatly improves the performance on high-quality $\mathrm{AP}_{75}$ and large objects $\mathrm{AP}_{\mathrm{L}}$ detection (row 1 and 7), with just a slight drop on low-quality object $\mathrm{AP}_{50}$ . That is reasonable because the outside reference points are unable to localize objects accurately, and this phenomenon
199
+
200
+ <table><tr><td rowspan="2">Backbone</td><td rowspan="2">Epoch</td><td rowspan="2">w/ SAP</td><td colspan="2">DN-DETR [9]</td><td colspan="2">DINO (Single-Scale) [28]</td><td colspan="2">Group DETR [4]</td></tr><tr><td>AP / AP50 / AP75</td><td>APS / APM / APL</td><td>AP / AP50 / AP75</td><td>APS / APM / APL</td><td>AP / AP50 / AP75</td><td>APS / APM / APL</td></tr><tr><td rowspan="2">R50</td><td rowspan="2">12</td><td rowspan="2">✓(Ours)</td><td>38.3 / 58.6 / 40.5</td><td>18.4 / 41.6 / 57.1</td><td>39.7 / 58.3 / 42.4</td><td>19.1 / 43.7 / 57.1</td><td>39.1 / - / -</td><td>19.7 / 42.5 / 56.8</td></tr><tr><td>39.5 / 59.7 / 41.5</td><td>18.7 / 42.8 / 59.0</td><td>40.0 / 60.1 / 42.1</td><td>20.2 / 43.4 / 58.5</td><td>39.8 / 60.2 / 42.0</td><td>20.2 / 43.5 / 58.6</td></tr><tr><td rowspan="2">R101</td><td rowspan="2">12</td><td rowspan="2">✓(Ours)</td><td>40.5 / 60.8 / 43.0</td><td>19.3 / 44.3 / 59.6</td><td>41.9 / 60.8 / 44.4</td><td>22.5 / 46.3 / 59.5</td><td>- / - / -</td><td>- / - / -</td></tr><tr><td>41.0 / 61.2 / 43.4</td><td>19.8 / 45.3 / 60.0</td><td>41.5 / 61.4 / 43.6</td><td>20.3 / 45.2 / 60.0</td><td>41.1 / 61.5 / 43.4</td><td>20.5 / 45.5 / 59.4</td></tr><tr><td rowspan="2">R50-DC</td><td rowspan="2">12</td><td rowspan="2">✓(Ours)</td><td>41.7 / 61.4 / 44.1</td><td>21.2 / 45.0 / 60.2</td><td>43.6 / 61.4 / 47.0</td><td>24.8 / 47.3 / 59.5</td><td>41.9 / - / -</td><td>23.3 / 45.6 / 58.4</td></tr><tr><td>43.6 / 62.5 / 46.2</td><td>23.3 / 47.3 / 61.0</td><td>44.0 / 63.1 / 46.5</td><td>24.8 / 47.3 / 61.1</td><td>43.9 / 63.2 / 46.8</td><td>24.5 / 47.6 / 61.3</td></tr><tr><td rowspan="2">R101-DC</td><td rowspan="2">12</td><td rowspan="2">✓(Ours)</td><td>43.4 / 61.9 / 47.2</td><td>24.8 / 46.8 / 59.4</td><td>45.4 / 63.5 / 49.2</td><td>26.4 / 49.5 / 61.1</td><td>- / - / -</td><td>- / - / -</td></tr><tr><td>44.6 / 63.9 / 48.0</td><td>25.5 / 48.9 / 62.5</td><td>45.6 / 64.5 / 48.7</td><td>25.0 / 49.7 / 62.5</td><td>44.4 / 63.9 / 47.4</td><td>25.9 / 48.5 / 61.4</td></tr></table>
201
+
202
+ Table 3. Comparison with denoised methods on COCO dataset based on the 12-epoch training schedule and 300 object queries.
203
+
204
+ <table><tr><td>Comment</td><td>Movable</td><td>Inner Loss</td><td>PECA</td><td>SDG</td><td>AP</td><td>\( AP_{50} \)</td><td>\( AP_{75} \)</td><td>\( AP_S \)</td><td>\( AP_M \)</td><td>\( AP_L \)</td></tr><tr><td>SAP-DETR (Ours)</td><td>✓</td><td>✓</td><td>✓</td><td>✓</td><td>36.2</td><td>56.2</td><td>37.9</td><td>16.4</td><td>39.5</td><td>53.8</td></tr><tr><td>-SDG</td><td>✓</td><td>✓</td><td>✓</td><td></td><td>35.6</td><td>56.2</td><td>36.9</td><td>16.3</td><td>38.9</td><td>52.7</td></tr><tr><td>-PECA</td><td>✓</td><td>✓</td><td></td><td>✓</td><td>34.8</td><td>55.5</td><td>36.0</td><td>15.7</td><td>37.3</td><td>52.0</td></tr><tr><td>-PECA &amp; SDG</td><td>✓</td><td>✓</td><td></td><td></td><td>34.0</td><td>54.9</td><td>35.3</td><td>15.0</td><td>36.7</td><td>51.5</td></tr><tr><td>-Movable</td><td></td><td>✓</td><td>✓</td><td>✓</td><td>35.2</td><td>55.4</td><td>36.8</td><td>15.8</td><td>38.5</td><td>53.8</td></tr><tr><td>-Inner Loss</td><td>✓</td><td></td><td>✓</td><td>✓</td><td>35.9</td><td>56.3</td><td>37.4</td><td>16.2</td><td>39.3</td><td>52.5</td></tr><tr><td>DAB-DETR (Baseline)</td><td>-</td><td>-</td><td>-</td><td>-</td><td>32.3</td><td>51.3</td><td>34.0</td><td>15.7</td><td>35.2</td><td>45.7</td></tr><tr><td>+Salient Point Concept</td><td>-</td><td>-</td><td>-</td><td>-</td><td>33.5</td><td>54.3</td><td>35.1</td><td>14.3</td><td>36.5</td><td>51.0</td></tr></table>
205
+
206
+ Table 4. Ablation on each components
207
+
208
+ <table><tr><td>Inner Cost (Linner)</td><td>Movable within Grid (sgrid)</td><td>AP</td><td>APS</td><td>APM</td><td>APL</td></tr><tr><td></td><td></td><td>35.9</td><td>17.0</td><td>38.8</td><td>52.7</td></tr><tr><td>✓</td><td></td><td>26.3</td><td>11.3</td><td>28.0</td><td>39.5</td></tr><tr><td>✓</td><td>✓</td><td>36.2</td><td>16.4</td><td>39.5</td><td>53.8</td></tr></table>
209
+
210
+ Table 5. Ablation on scaling factor of grid
211
+
212
+ <table><tr><td>PECA</td><td>Scaling Factor of SDG</td><td>AP</td><td>APS</td><td>APM</td><td>APL</td></tr><tr><td></td><td></td><td>33.6</td><td>14.7</td><td>36.0</td><td>50.7</td></tr><tr><td>✓</td><td></td><td>35.7</td><td>17.5</td><td>38.8</td><td>52.6</td></tr><tr><td>✓</td><td>✓</td><td>36.2</td><td>16.4</td><td>39.5</td><td>53.8</td></tr></table>
213
+
214
+ Table 6. Ablation on scaling factor of SDG
215
+
216
+ always exists in large objects. 4). For salient point enhanced cross-attention, both SDG and PECA serve as the essential components, independently emerging 0.8 AP and 1.6 AP improvements compared to the standard model (rows 2-4). Interestingly, there exists an effectiveness overlap on small objects, with only $0.1\mathrm{AP}_{\mathrm{S}}$ improvement when adding SDG to the equipped PECA model (row 1 and 2). We argue that the Gaussian-like map of SDG might be easily overlapped with PECA on small objects.
217
+
218
+ Scaling Factor of Movable Strategy. We perform an ablation study on the scaling factor of the movable strategy and further investigate the effectiveness of the inner cost in Tab. 5. Notably, it is observed that there exists a conflict between the inner loss and the global search strategy, behaving a sharp drop when only reserving the inner loss. Furthermore, searching within the grid enables the detector to more attend to small objects and avoid a drastic deterioration in normal object detection. See Appendix E for more detailed analyses. Scaling Factor of SDG. We also compare our side-directed manner with the standard offset prediction method in Tab. 6. Based on PECA, the side-directed scaling factor may limit the detector on small object detection but significantly promote the performance on other objects. This phenomenon would be broken without the help of PECA in which a preci
219
+
220
+ itous decline is emerged on all-scale object detection (row 5 in Tab. 4 vs. row 1 in Tab. 6). We hypothesise that it because the predicted reference points may be outside of the proposal boxes, or even the region of the image.
221
+
222
+ # 5. Conclusion
223
+
224
+ In this paper, we propose SAP-DETR for promoting model convergence by treating object detection as a transformation from the salient points to the instance objects. Our SAP-DETR explicitly initializes a query-specific reference point for each object query, gradually aggregates them into an instance object, and predicts the distance from each side of the bounding box. By speedily attending to the query-specific region and other extreme regions from contextual image features, it thus can effectively bridge the gap between the salient points and the query-based Transformer detector. Our extensive experiments have demonstrated that SAP-DETR achieves superior model convergence speed. With the same training settings, our proposed SAP-DETR outperforms SoTA approaches with large margins.
225
+
226
+ # 6. Future Work
227
+
228
+ This point-based design for DETR-like models also comes with remaining issues, in particular regarding training with deformable attention, multi-scale features, and negative query design. Following current center-based methods working for similar issues, we expect future work to successfully address them for point-based design of SAP-DETR.
229
+
230
+ # Acknowledgements
231
+
232
+ We thank the Lenovo Research AI Master platform for computing GPU supports without which this work would not be possible.
233
+
234
+ # References
235
+
236
+ [1] Navaneeth Bodla, Bharat Singh, Rama Chellappa, and Larry S Davis. Soft-nms-improving object detection with one line of code. In ICCV, pages 5561-5569, 2017. 3
237
+ [2] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In ECCV, pages 213-229. Springer, 2020. 1, 3, 4, 5, 6, 7, 10
238
+ [3] Liangyu Chen, Tong Yang, Xiangyu Zhang, Wei Zhang, and Jian Sun. Points as queries: Weakly semi-supervised object detection by points. In CVPR, pages 8823-8832, 2021. 3
239
+ [4] Qiang Chen, Xiaokang Chen, Gang Zeng, and Jingdong Wang. Group detr:fast training convergence with deco-upplied one-to-many label assignment. arXiv preprint arXiv:2207.13085, 2022. 3, 6, 8
240
+ [5] Peng Gao, Minghang Zheng, Xiaogang Wang, Jifeng Dai, and Hongsheng Li. Fast convergence of detr with spatially modulated co-attention. In ICCV, pages 3621–3630, 2021. 2, 3, 5, 6, 7, 10
241
+ [6] Zheng Ge, Songtao Liu, Feng Wang, Zeming Li, and Jian Sun. Yolox: Exceeding yolo series in 2021. arXiv preprint arXiv:2107.08430, 2021. 1
242
+ [7] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 5
243
+ [8] Hei Law and Jia Deng. Cornernet: Detecting objects as paired keypoints. In ECCV, pages 734-750, 2018. 3
244
+ [9] Feng Li, Hao Zhang, Shilong Liu, Jian Guo, Lionel M Ni, and Lei Zhang. Dn-detr: Accelerate detr training by introducing query denoising. In CVPR, pages 13619-13627, 2022. 3, 5, 6, 8, 13
245
+ [10] Tsung-Yi Lin, Piotr Dollar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In CVPR, pages 2117-2125, 2017. 7
246
+ [11] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólar. Focal loss for dense object detection. In ICCV, pages 2980-2988, 2017. 1, 3
247
+ [12] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, pages 740-755, 2014. 5
248
+ [13] Shilong Liu, Feng Li, Hao Zhang, Xiao Yang, Xianbiao Qi, Hang Su, Jun Zhu, and Lei Zhang. Dab-detr: Dynamic anchor boxes are better queries for detr. ICLR, 2022. 2, 3, 4, 5, 6, 7, 10
249
+ [14] Wei Liu, Dragomir Anguelov, Dumitru Erhan, Christian Szegedy, Scott Reed, Cheng-Yang Fu, and Alexander C Berg. Ssd: Single shot multibox detector. In ECCV, pages 21-37. Springer, 2016. 1
250
+ [15] Yang Liu, Yao Zhang, Yixin Wang, Feng Hou, Jin Yuan, Jiang Tian, Yang Zhang, Zhongchao Shi, Jianping Fan, and Zhiqiang He. A survey of visual transformers. arXiv preprint arXiv:2111.06091, 2021. 3
251
+ [16] Depu Meng, Xiaokang Chen, Zejia Fan, Gang Zeng, Houqiang Li, Yuhui Yuan, Lei Sun, and Jingdong Wang. Con
252
+
253
+ ditional detr for fast training convergence. In ICCV, pages 3651-3660, 2021. 2, 3, 4, 5, 6, 7, 10, 11
254
+ [17] Joseph Redmon and Ali Farhadi. Yolov3: An incremental improvement. arXiv preprint arXiv:1804.02767, 2018. 2
255
+ [18] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. NeurIPS, 28, 2015. 1, 6, 7
256
+ [19] Peize Sun, Yi Jiang, Enze Xie, Wenqi Shao, Zehuan Yuan, Changhu Wang, and Ping Luo. What makes for end-to-end object detection? In ICML, pages 9934-9944, 2021. 2
257
+ [20] Peize Sun, Rufeng Zhang, Yi Jiang, Tao Kong, Chenfeng Xu, Wei Zhan, Masayoshi Tomizuka, Lei Li, et al. Sparse r-cnn: End-to-end object detection with learnable proposals. arXiv preprint arXiv:2011.12450, 2020. 1
258
+ [21] Zhiqing Sun, Shengcao Cao, Yiming Yang, and Kris M Kitani. Rethinking transformer-based set prediction for object detection. In ICCV, pages 3611-3620, 2021. 3
259
+ [22] Zhi Tian, Chunhua Shen, Hao Chen, and Tong He. Fcos: Fully convolutional one-stage object detection. In ICCV, pages 9627-9636, 2019. 1, 3, 4
260
+ [23] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of machine learning research, 9(11), 2008. 12
261
+ [24] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. NeurIPS, 30, 2017. 1, 10
262
+ [25] Yingming Wang, Xiangyu Zhang, Tong Yang, and Jian Sun. Anchor detr: Query design for transformer-based detector. In AAAI, volume 36, pages 2567-2575, 2022. 2, 3, 4, 6, 7, 10, 12
263
+ [26] Gongjie Zhang, Zhipeng Luo, Yingchen Yu, Kaiwen Cui, and Shijian Lu. Accelerating detr convergence via semantic-aligned matching. In CVPR, pages 949-958, 2022. 3, 6, 7, 10
264
+ [27] Gongjie Zhang, Zhipeng Luo, Yingchen Yu, Jiaxing Huang, Kaiwen Cui, Shijian Lu, and Eric P Xing. Semantic-aligned matching for enhanced detr convergence and multi-scale feature fusion. arXiv preprint arXiv:2207.14172, 2022. 3
265
+ [28] Hao Zhang, Feng Li, Shilong Liu, Lei Zhang, Hang Su, Jun Zhu, Lionel M Ni, and Heung-Yeung Shum. Dino: Detr with improved denoising anchor boxes for end-to-end object detection. ICLR, 2022. 3, 5, 6, 8
266
+ [29] Xingyi Zhou, Dequan Wang, and Philipp Krahenbuhl. Objects as points. In arXiv preprint arXiv:1904.07850, 2019. 3
267
+ [30] Xingyi Zhou, Jiacheng Zhuo, and Philipp Krahenbuhl. Bottom-up object detection by grouping extreme and center points. In CVPR, pages 850-859, 2019. 3
268
+ [31] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. In ICLR, 2020. 2, 3, 6, 7, 10
2023/SAP-DETR_ Bridging the Gap Between Salient Points and Queries-Based Transformer Detector for Fast Model Convergency/images.zip ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:85762c46455be45b3ba870afc73422e15823f2eba2f5f51f8d0e338950da01c1
3
+ size 711193
2023/SAP-DETR_ Bridging the Gap Between Salient Points and Queries-Based Transformer Detector for Fast Model Convergency/layout.json ADDED
The diff for this file is too large to render. See raw diff
 
2023/SCADE_ NeRFs from Space Carving With Ambiguity-Aware Depth Estimates/cb201944-e311-4584-9847-096959c3a596_content_list.json ADDED
@@ -0,0 +1,1770 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ [
2
+ {
3
+ "type": "text",
4
+ "text": "SCADE: NeRFs from Space Carving with Ambiguity-Aware Depth Estimates",
5
+ "text_level": 1,
6
+ "bbox": [
7
+ 94,
8
+ 114,
9
+ 875,
10
+ 138
11
+ ],
12
+ "page_idx": 0
13
+ },
14
+ {
15
+ "type": "text",
16
+ "text": "Mikaela Angelina Uy $^{1,2}$ Ricardo Martin-Brualla $^{2}$ Leonidas Guibas $^{1,2}$ Ke Li $^{2,3}$",
17
+ "bbox": [
18
+ 158,
19
+ 148,
20
+ 803,
21
+ 170
22
+ ],
23
+ "page_idx": 0
24
+ },
25
+ {
26
+ "type": "text",
27
+ "text": "$^{1}$ Stanford University $^{2}$ Google $^{3}$ Simon Fraser University",
28
+ "bbox": [
29
+ 241,
30
+ 174,
31
+ 723,
32
+ 194
33
+ ],
34
+ "page_idx": 0
35
+ },
36
+ {
37
+ "type": "image",
38
+ "img_path": "images/f95b9fa5adb6316ff5da22e4504071a144ea972361b019bf087b8333209ab4ed.jpg",
39
+ "image_caption": [
40
+ "Figure 1. SCADE Overview. We present SCADE, a novel technique for NeRF reconstruction under sparse, unconstrained views for in-the-wild indoor scenes. We leverage on generalizable monocular depth priors and address to represent the inherent ambiguities of monocular depth by exploiting our ambiguity-aware depth estimates (left). Our approach accounts for multimodality of both distributions using our novel space carving loss that seeks to disambiguate and find the common mode to fuse the information between different views (middle). SCADE enables better photometric reconstruction especially in highly ambiguous scenes, e.g. non-opaque surfaces (right)."
41
+ ],
42
+ "image_footnote": [],
43
+ "bbox": [
44
+ 109,
45
+ 202,
46
+ 864,
47
+ 411
48
+ ],
49
+ "page_idx": 0
50
+ },
51
+ {
52
+ "type": "text",
53
+ "text": "Abstract",
54
+ "text_level": 1,
55
+ "bbox": [
56
+ 233,
57
+ 491,
58
+ 313,
59
+ 507
60
+ ],
61
+ "page_idx": 0
62
+ },
63
+ {
64
+ "type": "text",
65
+ "text": "Neural radiance fields (NeRFs) have enabled high fidelity 3D reconstruction from multiple 2D input views. However, a well-known drawback of NeRFs is the less-than-ideal performance under a small number of views, due to insufficient constraints enforced by volumetric rendering. To address this issue, we introduce SCADE, a novel technique that improves NeRF reconstruction quality on sparse, unconstrained input views for in-the-wild indoor scenes. To constrain NeRF reconstruction, we leverage geometric priors in the form of per-view depth estimates produced with state-of-the-art monocular depth estimation models, which can generalize across scenes. A key challenge is that monocular depth estimation is an ill-posed problem, with inherent ambiguities. To handle this issue, we propose a new method that learns to predict, for each view, a continuous, multimodal distribution of depth estimates using conditional Implicit Maximum Likelihood Estimation (cIMLE). In order to disambiguate exploiting multiple views, we introduce an original space carving loss that guides the NeRF representation to fuse multiple hypothesized depth maps from each view and distill from them a common geometry that is consistent with all views. Experiments show that our approach enables higher fidelity novel view synthesis from sparse views. Our project page can be found at scade-spacecarving-nerfs.github.io.",
66
+ "bbox": [
67
+ 73,
68
+ 530,
69
+ 473,
70
+ 883
71
+ ],
72
+ "page_idx": 0
73
+ },
74
+ {
75
+ "type": "text",
76
+ "text": "1. Introduction",
77
+ "text_level": 1,
78
+ "bbox": [
79
+ 501,
80
+ 491,
81
+ 632,
82
+ 507
83
+ ],
84
+ "page_idx": 0
85
+ },
86
+ {
87
+ "type": "text",
88
+ "text": "Neural radiance fields (NeRF) [24] have enabled high fidelity novel view synthesis from dozens of input views. Such number of views are difficult to obtain in practice, however. When given only a small number of sparse views, vanilla NeRF tends to struggle with reconstructing shape accurately, due to inadequate constraints enforced by the volume rendering loss alone.",
89
+ "bbox": [
90
+ 496,
91
+ 511,
92
+ 893,
93
+ 616
94
+ ],
95
+ "page_idx": 0
96
+ },
97
+ {
98
+ "type": "text",
99
+ "text": "Shape priors can help remedy this problem. Various forms of shape priors have been proposed for NeRFs, such as object category-level priors [13], and handcrafted data-independent priors [25]. The former requires knowledge of category labels and is not applicable to scenes, and the latter is agnostic to the specifics of the scene and only encodes low-level regularity (e.g., local smoothness). A form of prior that is both scene-dependent and category-agnostic is per-view monocular depth estimates, which have been explored in prior work [6, 32]. Unfortunately, monocular depth estimates are often inaccurate, due to estimation errors and inherent ambiguities, such as albedo vs. shading (cf. check shadow illusion), concavity vs. convexity (cf. hollow face illusion), distance vs. scale (cf. miniature cinematography), etc. As a result, the incorrect or inconsistent priors imposed by such depth estimates may mislead the NeRF into reconstructing incorrect shape and produce artifacts in the generated views.",
100
+ "bbox": [
101
+ 496,
102
+ 618,
103
+ 895,
104
+ 868
105
+ ],
106
+ "page_idx": 0
107
+ },
108
+ {
109
+ "type": "text",
110
+ "text": "In this paper, we propose a method that embraces the uncertainty and ambiguities present in monocular depth esti",
111
+ "bbox": [
112
+ 500,
113
+ 869,
114
+ 893,
115
+ 901
116
+ ],
117
+ "page_idx": 0
118
+ },
119
+ {
120
+ "type": "header",
121
+ "text": "CVF",
122
+ "bbox": [
123
+ 106,
124
+ 2,
125
+ 181,
126
+ 42
127
+ ],
128
+ "page_idx": 0
129
+ },
130
+ {
131
+ "type": "header",
132
+ "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.",
133
+ "bbox": [
134
+ 236,
135
+ 0,
136
+ 810,
137
+ 46
138
+ ],
139
+ "page_idx": 0
140
+ },
141
+ {
142
+ "type": "page_number",
143
+ "text": "16518",
144
+ "bbox": [
145
+ 480,
146
+ 944,
147
+ 519,
148
+ 957
149
+ ],
150
+ "page_idx": 0
151
+ },
152
+ {
153
+ "type": "text",
154
+ "text": "mates, by modelling a probability distribution over depth estimates. The ambiguities are retained at the stage of monocular depth estimation, and are only resolved once information from multiple views are fused together. We do so with a principled loss defined on probability distributions over depth estimates for different views. This loss selects the subset of modes of probability distributions that are consistent across all views and matches them with the modes of the depth distribution along different rays as modelled by the NeRF. It turns out that this operation can be interpreted as a probabilistic analogue of classical depth-based space carving [17]. For this reason, we dub our method Space Carving with Ambiguity-aware Depth Estimates, or SCADE for short.",
155
+ "bbox": [
156
+ 75,
157
+ 90,
158
+ 472,
159
+ 281
160
+ ],
161
+ "page_idx": 1
162
+ },
163
+ {
164
+ "type": "text",
165
+ "text": "Compared to prior approaches of depth supervision [6,32] that only supervise the moments of NeRF's depth distribution (rather than the modes), our key innovation is that we supervise the modes of NeRF's depth distribution. The supervisory signal provided by the former is weaker than the latter, because the former only constrains the value of an integral aggregated along the ray, whereas the latter constrains the values at different individual points along the ray. Hence, the supervisory signal provided by the former is 2D (because it integrates over a ray), whereas the supervisory signal provided by the our method is 3D (because it is point-wise) and thus can be more fine-grained.",
166
+ "bbox": [
167
+ 75,
168
+ 282,
169
+ 472,
170
+ 458
171
+ ],
172
+ "page_idx": 1
173
+ },
174
+ {
175
+ "type": "text",
176
+ "text": "An important technical challenge is modelling probability distributions over depth estimates. Classical approaches use simple distributions with closed-form probability densities such as Gaussian or Laplace distributions. Unfortunately these distributions are not very expressive, since they only have a single mode (known as \"unimodal\") and have a fixed shape for the tails. Since each interpretation of an ambiguous image should be a distinct mode, these simple unimodal distributions cannot capture the complex ambiguities in depth estimates. Naive extensions like a mixture of Gaussians are also not ideal because some images are more ambiguous than others, and so the number of modes needed may differ for the depth estimates of different images. Moreover, learning such a mixture requires backpropagating through E-M, which is nontrivial. Any attempt at modifying the probability density to make it capable of handling a variable number of modes can easily run into an intractable partition function [4, 10, 31], which makes learning difficult because maximum likelihood requires the ability to evaluate the probability density, which is a function of the partition function.",
177
+ "bbox": [
178
+ 75,
179
+ 459,
180
+ 472,
181
+ 752
182
+ ],
183
+ "page_idx": 1
184
+ },
185
+ {
186
+ "type": "text",
187
+ "text": "To get around this conundrum, we propose representing the probability distribution with a set of samples generated from a neural network. Such a distribution can be learned with a conditional GAN; however, because GANs suffer from mode collapse, they cannot model multiple modes [11] and are therefore unsuited to modelling ambiguity. Instead, we propose leveraging conditional Implicit Maximum Likelihood Estimation (cIMLE) [19, 28] to learn the distribution, which is designed to avoid mode collapse.",
188
+ "bbox": [
189
+ 75,
190
+ 753,
191
+ 472,
192
+ 885
193
+ ],
194
+ "page_idx": 1
195
+ },
196
+ {
197
+ "type": "text",
198
+ "text": "We consider the challenging setting of leveraging out-of",
199
+ "bbox": [
200
+ 96,
201
+ 885,
202
+ 472,
203
+ 901
204
+ ],
205
+ "page_idx": 1
206
+ },
207
+ {
208
+ "type": "image",
209
+ "img_path": "images/89ae8a01be14ce67f619ad8b234580bd782b76021c3c70019a5b9d7fbb59aa90.jpg",
210
+ "image_caption": [],
211
+ "image_footnote": [],
212
+ "bbox": [
213
+ 526,
214
+ 93,
215
+ 607,
216
+ 152
217
+ ],
218
+ "page_idx": 1
219
+ },
220
+ {
221
+ "type": "image",
222
+ "img_path": "images/a06fd61cd86dab5c1198a5da0f1e188d15aa7047d12b01b972958a739eeec331.jpg",
223
+ "image_caption": [],
224
+ "image_footnote": [],
225
+ "bbox": [
226
+ 612,
227
+ 93,
228
+ 694,
229
+ 152
230
+ ],
231
+ "page_idx": 1
232
+ },
233
+ {
234
+ "type": "image",
235
+ "img_path": "images/68d33f6a715203b18ec8f4c3b82c6cb7b9e4b4f0b15c1a3d22ce704b35e778c8.jpg",
236
+ "image_caption": [],
237
+ "image_footnote": [],
238
+ "bbox": [
239
+ 697,
240
+ 93,
241
+ 777,
242
+ 152
243
+ ],
244
+ "page_idx": 1
245
+ },
246
+ {
247
+ "type": "image",
248
+ "img_path": "images/13559aab43aa532959f7217e43d80f5dd5c5d23b0037c75dad3323d3954e0466.jpg",
249
+ "image_caption": [],
250
+ "image_footnote": [],
251
+ "bbox": [
252
+ 779,
253
+ 93,
254
+ 857,
255
+ 152
256
+ ],
257
+ "page_idx": 1
258
+ },
259
+ {
260
+ "type": "image",
261
+ "img_path": "images/22fce397da257929668b856372fb7391c94c99ce422710eedf4bda265c1f0754.jpg",
262
+ "image_caption": [
263
+ "Input Image"
264
+ ],
265
+ "image_footnote": [],
266
+ "bbox": [
267
+ 532,
268
+ 155,
269
+ 602,
270
+ 223
271
+ ],
272
+ "page_idx": 1
273
+ },
274
+ {
275
+ "type": "image",
276
+ "img_path": "images/6b3d450c6b39276ec11d0710dfd745893d459c0730b59fac991eb1a3c2ac5c7f.jpg",
277
+ "image_caption": [
278
+ "Ambiguity-Aware Depth Estimates"
279
+ ],
280
+ "image_footnote": [],
281
+ "bbox": [
282
+ 620,
283
+ 155,
284
+ 691,
285
+ 223
286
+ ],
287
+ "page_idx": 1
288
+ },
289
+ {
290
+ "type": "image",
291
+ "img_path": "images/e81b800d71ea46699ae5cf16dcac3e985927f8f81fd5d4df9b54df2184be02b6.jpg",
292
+ "image_caption": [
293
+ "Figure 2. Ambiguities from a single image. We show samples from our ambiguity-aware depth estimates that is able to handle different types of ambiguities. (Top) An image of a chair taken from the top-view. Without context of the scene, it is unclear that it is an image of a chair. We show different samples from our ambiguity-aware depth estimates that are able to capture different degrees of convexity. (Bottom) An image of a cardboard under bad lighting conditions that capture the albedo vs shading ambiguity that is also represented by our different samples."
294
+ ],
295
+ "image_footnote": [],
296
+ "bbox": [
297
+ 702,
298
+ 155,
299
+ 772,
300
+ 223
301
+ ],
302
+ "page_idx": 1
303
+ },
304
+ {
305
+ "type": "image",
306
+ "img_path": "images/d8d127bc3b443638620fdab80b812ae0ca16954b16deff9a3bd0cf586025dffd.jpg",
307
+ "image_caption": [],
308
+ "image_footnote": [],
309
+ "bbox": [
310
+ 785,
311
+ 155,
312
+ 852,
313
+ 223
314
+ ],
315
+ "page_idx": 1
316
+ },
317
+ {
318
+ "type": "text",
319
+ "text": "domain depth priors to train NeRFs on real-world indoor scenes with sparse views. Under this setting, the depth priors we use are trained on a different dataset (e.g., Taskonomy) from the scenes our NeRFs are trained on (e.g., ScanNet). This setting is more challenging than usual due to domain gap between the dataset the prior is trained on and the scenes NeRF is asked to reconstruct. We demonstrate that our method outperforms vanilla NeRF and NeRFs with supervision from depth-based priors in novel view synthesis.",
320
+ "bbox": [
321
+ 496,
322
+ 366,
323
+ 893,
324
+ 498
325
+ ],
326
+ "page_idx": 1
327
+ },
328
+ {
329
+ "type": "text",
330
+ "text": "In summary, our key contributions include:",
331
+ "bbox": [
332
+ 519,
333
+ 500,
334
+ 803,
335
+ 512
336
+ ],
337
+ "page_idx": 1
338
+ },
339
+ {
340
+ "type": "list",
341
+ "sub_type": "text",
342
+ "list_items": [
343
+ "- An method that allows the creation of NeRFs in unconstrained indoor settings from only a modest number of 2D views by introducing a sophisticated way to exploit ambiguity-aware monocular depth estimation.",
344
+ "- A novel way to sample distributions over image depth estimates based on conditional Implicit Maximum Likelihood Estimation that can represent depth ambiguities and capture a variable number of discrete depth modes.",
345
+ "- A new space-carving loss that can be used in the NeRF formulation to optimize for a mode-seeking 3D density that helps select consistent depth modes across the views and thus compensate for the under-constrained photometric information in the few view regime."
346
+ ],
347
+ "bbox": [
348
+ 517,
349
+ 521,
350
+ 893,
351
+ 732
352
+ ],
353
+ "page_idx": 1
354
+ },
355
+ {
356
+ "type": "text",
357
+ "text": "2. Related Work",
358
+ "text_level": 1,
359
+ "bbox": [
360
+ 500,
361
+ 746,
362
+ 640,
363
+ 762
364
+ ],
365
+ "page_idx": 1
366
+ },
367
+ {
368
+ "type": "text",
369
+ "text": "Novel View Synthesis and 3D Reconstruction. Reconstructing the structure of a scene from a few views is a long standing problem in computer vision with a long literature. Neural Radiance Fields or NeRF [24] revolutionized the field by showing how to use weights of an MLP to represent a scene that is rendered using volume rendering [23]. A key innovation was the use of positional encoding [37, 38] to increase the effective capacity of the MLPs that model a emitted radiance and density as a function of position and",
370
+ "bbox": [
371
+ 496,
372
+ 767,
373
+ 893,
374
+ 901
375
+ ],
376
+ "page_idx": 1
377
+ },
378
+ {
379
+ "type": "page_number",
380
+ "text": "16519",
381
+ "bbox": [
382
+ 480,
383
+ 944,
384
+ 519,
385
+ 955
386
+ ],
387
+ "page_idx": 1
388
+ },
389
+ {
390
+ "type": "text",
391
+ "text": "viewing direction. Extensions of NeRF include works on unconstrained photo collections [22], dynamic scenes [20], deformable scenes [27], and reflective materials [1,39].",
392
+ "bbox": [
393
+ 75,
394
+ 90,
395
+ 470,
396
+ 135
397
+ ],
398
+ "page_idx": 2
399
+ },
400
+ {
401
+ "type": "text",
402
+ "text": "A critical shortcoming of NeRF is its reliance on having many input views of a scene. Several approaches have been proposed, including adding patch likelihood losses [25], data-driven priors [36, 48], semantic consistency prior [12], image features [41], or surface [25, 39], occupancy [26], and depth [15, 32, 43] priors. DS-NeRF [6] uses the sparse point reconstructions recovered during Structure-from-Motion, to supervise the depth of sparse points in the recovered NeRF. In the same spirit, DDP [32] uses a depth completion network, that takes as input sparse point cloud projected to one an input view, and produces a depth estimate. Both works [15, 32] model depth with uncertainty but only supervise with moments of NeRF's depth distribution. In contrast, our work is able to represent multimodal depth estimates, which can handle the inherit ambiguities of depth estimation, and is able to seek the modes for each depth estimate to make a consistent prediction of the scene structure.",
403
+ "bbox": [
404
+ 75,
405
+ 136,
406
+ 472,
407
+ 383
408
+ ],
409
+ "page_idx": 2
410
+ },
411
+ {
412
+ "type": "text",
413
+ "text": "Inspired by traditional multi-view stereo, MVS-NeRF [3] and RC-MVSNet [2] incorporate the use cost volumes [9] to NeRF. To construct the cost volumes, these works look for agreement between features to look for correspondences, which is difficult under the setting of large variations in appearance or viewpoint. In contrast, our approach introduces a novel space carving loss that does not rely on feature correspondences and instead directly defines the loss in 3D.",
414
+ "bbox": [
415
+ 75,
416
+ 385,
417
+ 470,
418
+ 501
419
+ ],
420
+ "page_idx": 2
421
+ },
422
+ {
423
+ "type": "text",
424
+ "text": "Another line of works focuses on geometry reconstruction [40, 45, 49] by using the volumetric rendering scheme to learn a neural SDF representation. These works tackle a different problem on modeling geometry reconstruction, unlike NeRFs whose focus is modeling appearance for novel view synthesis, where our work falls under.",
425
+ "bbox": [
426
+ 75,
427
+ 503,
428
+ 470,
429
+ 589
430
+ ],
431
+ "page_idx": 2
432
+ },
433
+ {
434
+ "type": "text",
435
+ "text": "Depth Estimation from Single View. Depth estimation from a single image is a complex task due multiple ambiguities, including scene scale and shading / albedo ambiguities. Early efforts used MRFs to compute predictions on superpixels [33]. The advent of consumer depth cameras like Kinect [53] enabled the acquisition of larger scale indoor 3D datasets [35], leading to methods that used deep neural networks to predict depth from a single color image [7]. Nonetheless, scaling datasets for depth estimation remained a challenge. Deep3D [44] enabled the creation of stereo views by training on a large dataset of stereo movies, while [8] used the self-supervision of left-right consistency to learn depth from stereo views from driving cars. This approach was extended to videos where the egomotion is also estimated and self-supervision happens across time [54]. MegaDepth [21] uses the depth from SfM reconstructions of internet photos, together with semantic segmentation that conveys ordinal depth supervision cues, i.e. transient objects must be in front of the static scenes. Most recent, multi-task learning has shown promise in training depth estimators that work well in out-of-domain data [30].",
436
+ "bbox": [
437
+ 75,
438
+ 590,
439
+ 472,
440
+ 901
441
+ ],
442
+ "page_idx": 2
443
+ },
444
+ {
445
+ "type": "text",
446
+ "text": "In some cases, surface normal estimation has fewer pitfalls than direct depth estimation, as it suffers somewhat less from scale ambiguities. GeoNet [29] jointly estimates depth and surface normals, then uses the geometric relation between them to refine the estimates. LeRes [46] uses a second geometry reasoning module that refines a focal length estimate and the depth itself. Our depth estimation approach is derived from LeReS without the second stage, as the focal length is already estimated during SfM for NeRF scenes.",
447
+ "bbox": [
448
+ 496,
449
+ 90,
450
+ 893,
451
+ 223
452
+ ],
453
+ "page_idx": 2
454
+ },
455
+ {
456
+ "type": "text",
457
+ "text": "Most relevant is the work of Kendall and Gal [15], that learns depth estimation in a Bayesian setting, where their objective maximizes the likelihood under a predicted distribution. In our work, we use instead a multimodal depth estimation technique built on cIMLE [19], which makes our prior robust to ambiguities in depth prediction.",
458
+ "bbox": [
459
+ 496,
460
+ 223,
461
+ 893,
462
+ 311
463
+ ],
464
+ "page_idx": 2
465
+ },
466
+ {
467
+ "type": "text",
468
+ "text": "3. Background",
469
+ "text_level": 1,
470
+ "bbox": [
471
+ 498,
472
+ 324,
473
+ 627,
474
+ 340
475
+ ],
476
+ "page_idx": 2
477
+ },
478
+ {
479
+ "type": "text",
480
+ "text": "3.1. Neural Radiance Fields (NeRF)",
481
+ "text_level": 1,
482
+ "bbox": [
483
+ 498,
484
+ 348,
485
+ 777,
486
+ 363
487
+ ],
488
+ "page_idx": 2
489
+ },
490
+ {
491
+ "type": "text",
492
+ "text": "A neural radiance field [24], or NeRF for short, represents a field in 3D space, where each point represents an infinitesimal particle with a certain opacity that emits varying amounts of light along different viewing directions. The opacity at a point is represented as a volume density, and the amount of emitted light is represented as a colour. Mathematically, a NeRF is represented as two parameterized functions, a volume density $\\sigma_{\\theta}:\\mathbb{R}^{3}\\to \\mathbb{R}_{\\geq 0}$ and a colour $\\mathbf{c}_{\\psi}:\\mathbb{R}^3\\times S^2\\rightarrow [0,255]^3$ . The former maps a 3D coordinate $\\mathbf{x}\\in \\mathbb{R}^3$ to a volume density $\\sigma \\in \\mathbb{R}_{\\geq 0}$ , and the latter maps a 3D coordinate $\\mathbf{x}\\in \\mathbb{R}^3$ and a viewing direction $\\mathbf{d}\\in S^2$ to a colour $\\mathbf{c}\\in [0,255]^3$ .",
493
+ "bbox": [
494
+ 496,
495
+ 364,
496
+ 893,
497
+ 540
498
+ ],
499
+ "page_idx": 2
500
+ },
501
+ {
502
+ "type": "text",
503
+ "text": "To render a NeRF from a view, we shoot rays from each pixel on the camera sensor and integrate over the product of colour, volume density and visibility along the ray to arrive at each pixel value. Visibility at a point is represented with transmittance, which accumulates the exponentiated negative volume density multiplicatively up to the point. The higher transmittance at a point, the more visible the point is from the camera. Mathematically, if the camera ray is $\\mathbf{r}(t) = \\mathbf{o} + t\\mathbf{d}$ , where $\\mathbf{o}$ is the camera centre and $\\mathbf{d}$ is the ray direction, the pixel value $\\hat{I}_{\\theta,\\psi}(\\mathbf{o},\\mathbf{d})$ we would have is:",
504
+ "bbox": [
505
+ 496,
506
+ 541,
507
+ 893,
508
+ 688
509
+ ],
510
+ "page_idx": 2
511
+ },
512
+ {
513
+ "type": "equation",
514
+ "text": "\n$$\n\\begin{array}{l} \\hat {I} _ {\\theta , \\psi} (\\mathbf {o}, \\mathbf {d}) := \\int_ {t _ {n}} ^ {t _ {f}} T _ {\\theta , \\mathbf {o}, \\mathbf {d}} (t) \\sigma_ {\\theta} (\\mathbf {o} + t \\mathbf {d}) \\mathbf {c} _ {\\psi} (\\mathbf {o} + t \\mathbf {d}, \\mathbf {d}) d t \\\\ \\text {w h e r e} T _ {\\theta , \\mathbf {o}, \\mathbf {d}} (t) = \\exp \\left(- \\int_ {t _ {n}} ^ {t} \\sigma_ {\\theta} (\\mathbf {o} + s \\mathbf {d}) \\mathrm {d} s\\right). \\tag {1} \\\\ \\end{array}\n$$\n",
515
+ "text_format": "latex",
516
+ "bbox": [
517
+ 506,
518
+ 698,
519
+ 893,
520
+ 785
521
+ ],
522
+ "page_idx": 2
523
+ },
524
+ {
525
+ "type": "text",
526
+ "text": "In the expressions above, $t_n$ and $t_f$ denote the points the ray intersects with the near plane and far plane respectively.",
527
+ "bbox": [
528
+ 496,
529
+ 795,
530
+ 890,
531
+ 825
532
+ ],
533
+ "page_idx": 2
534
+ },
535
+ {
536
+ "type": "text",
537
+ "text": "3.2. Inverse Rendering with NeRF",
538
+ "text_level": 1,
539
+ "bbox": [
540
+ 500,
541
+ 833,
542
+ 766,
543
+ 849
544
+ ],
545
+ "page_idx": 2
546
+ },
547
+ {
548
+ "type": "text",
549
+ "text": "Given a set of real-world images from known views, inverse rendering aims to find the scene whose rendered images match the real-world images. If we use $I(\\mathbf{o}, \\mathbf{d})$ to de",
550
+ "bbox": [
551
+ 496,
552
+ 856,
553
+ 893,
554
+ 901
555
+ ],
556
+ "page_idx": 2
557
+ },
558
+ {
559
+ "type": "page_number",
560
+ "text": "16520",
561
+ "bbox": [
562
+ 480,
563
+ 944,
564
+ 519,
565
+ 957
566
+ ],
567
+ "page_idx": 2
568
+ },
569
+ {
570
+ "type": "text",
571
+ "text": "note the pixel value for the ray $\\mathbf{r}(t) = \\mathbf{o} + t\\mathbf{d}$ , this problem can be cast as an optimization problem, namely:",
572
+ "bbox": [
573
+ 76,
574
+ 90,
575
+ 468,
576
+ 121
577
+ ],
578
+ "page_idx": 3
579
+ },
580
+ {
581
+ "type": "equation",
582
+ "text": "\n$$\n\\min _ {\\theta , \\psi} \\sum_ {\\mathbf {o}} \\sum_ {\\mathbf {d}} \\| \\hat {I} _ {\\theta , \\psi} (\\mathbf {o}, \\mathbf {d}) - I (\\mathbf {o}, \\mathbf {d}) \\| _ {2} ^ {2} \\tag {2}\n$$\n",
583
+ "text_format": "latex",
584
+ "bbox": [
585
+ 148,
586
+ 131,
587
+ 468,
588
+ 162
589
+ ],
590
+ "page_idx": 3
591
+ },
592
+ {
593
+ "type": "text",
594
+ "text": "If $\\hat{I}_{\\theta, \\psi}(\\mathbf{o}, \\mathbf{d})$ is differentiable w.r.t. the parameters $\\theta, \\psi$ , this problem can be tackled straightforwardly with gradient-based optimization. To enable this, the volume density function $\\sigma_{\\theta}$ and the colour function $\\mathbf{c}_{\\psi}$ are chosen to be neural networks. As a result, inverse rendering amounts to training the NeRF to reconstruct the scene in a way that is consistent with the images that are given. Once the NeRF is trained, novel view synthesis can be achieved by rendering the NeRF from new, unseen views.",
595
+ "bbox": [
596
+ 75,
597
+ 167,
598
+ 470,
599
+ 299
600
+ ],
601
+ "page_idx": 3
602
+ },
603
+ {
604
+ "type": "text",
605
+ "text": "Since this optimization problem is underconstrained, in general many views are needed to reconstruct the scene accurately. Yet, in practical applications, typically only a few views are available, hence priors are needed to provide sufficient constraints. In this paper, we consider priors in the form of monocular depth estimates from each view. The monocular depth estimates are trained on different datasets from the dataset NeRF is trained on to simulate real-world conditions where there is a domain gap between what the prior was trained on and what the NeRF is trained on.",
606
+ "bbox": [
607
+ 75,
608
+ 300,
609
+ 468,
610
+ 446
611
+ ],
612
+ "page_idx": 3
613
+ },
614
+ {
615
+ "type": "text",
616
+ "text": "3.3. Ray Termination Distance",
617
+ "text_level": 1,
618
+ "bbox": [
619
+ 76,
620
+ 449,
621
+ 313,
622
+ 463
623
+ ],
624
+ "page_idx": 3
625
+ },
626
+ {
627
+ "type": "text",
628
+ "text": "In order to leverage depth priors, a natural way is to use them to constrain the ray termination distance. Because NeRF can represent non-opaque surfaces, there is no single ray termination distance. Instead, there is a distribution of ray termination distances. The cumulative distribution function (CDF) of this distribution represents the probability that the ray terminates before reaching a given point. It turns out that the probability density of this distribution $f_{\\theta ,\\mathbf{o},\\mathbf{d}}(t)$ is given by [6]:",
629
+ "bbox": [
630
+ 75,
631
+ 465,
632
+ 468,
633
+ 597
634
+ ],
635
+ "page_idx": 3
636
+ },
637
+ {
638
+ "type": "equation",
639
+ "text": "\n$$\nf _ {\\theta , \\mathbf {o}, \\mathbf {d}} (t) = T _ {\\theta , \\mathbf {o}, \\mathbf {d}} (t) \\sigma_ {\\theta} (\\mathbf {o} + t \\mathbf {d}) \\tag {3}\n$$\n",
640
+ "text_format": "latex",
641
+ "bbox": [
642
+ 163,
643
+ 609,
644
+ 468,
645
+ 626
646
+ ],
647
+ "page_idx": 3
648
+ },
649
+ {
650
+ "type": "text",
651
+ "text": "4. Method",
652
+ "text_level": 1,
653
+ "bbox": [
654
+ 76,
655
+ 636,
656
+ 166,
657
+ 651
658
+ ],
659
+ "page_idx": 3
660
+ },
661
+ {
662
+ "type": "text",
663
+ "text": "4.1. Supervising Ray Termination Distance",
664
+ "text_level": 1,
665
+ "bbox": [
666
+ 76,
667
+ 660,
668
+ 408,
669
+ 676
670
+ ],
671
+ "page_idx": 3
672
+ },
673
+ {
674
+ "type": "text",
675
+ "text": "If all surfaces were opaque and the ground truth depth were given, supervising NeRF with the ground truth depth would be straightforward. All that is needed is to make the NeRFs distribution of ray termination distances as close as possible to a delta distribution centred at the ground truth depth. However, when the ground truth depth is not known, we have to use monocular depth estimates as supervision. The challenge is that unlike ground truth depth, monocular depth estimates may not be stereo-consistent, due to estimation error and inherent ambiguity. As a result, it may not be possible to find a scene that is consistent with the depth estimates from every view. Therefore, in order to reconstruct the scene, we must model the uncertainty and ambiguity in the depth estimates, which can be mathematically characterized by distributions. Such ambiguities cannot be resolved from",
676
+ "bbox": [
677
+ 75,
678
+ 680,
679
+ 470,
680
+ 900
681
+ ],
682
+ "page_idx": 3
683
+ },
684
+ {
685
+ "type": "image",
686
+ "img_path": "images/d79b1f9b1912712bf049c99bd1ce552f3d28968c73f444a03701056de739514e.jpg",
687
+ "image_caption": [
688
+ "Figure 3. Network Architecture for our Ambiguity-aware Depth Estimates."
689
+ ],
690
+ "image_footnote": [],
691
+ "bbox": [
692
+ 506,
693
+ 87,
694
+ 890,
695
+ 250
696
+ ],
697
+ "page_idx": 3
698
+ },
699
+ {
700
+ "type": "text",
701
+ "text": "a single view and can often be resolved from multiple views. So we need an method that can fuse together uncertain depth estimates from different views.",
702
+ "bbox": [
703
+ 498,
704
+ 296,
705
+ 893,
706
+ 339
707
+ ],
708
+ "page_idx": 3
709
+ },
710
+ {
711
+ "type": "text",
712
+ "text": "A natural way of representing uncertainty is through a probability distribution. When there are ambiguities, the distribution of depth estimates are typically multimodal, i.e., their probability densities have disjoint peaks, where each mode represents one interpretation of the scene structure. For example, albedo vs shading ambiguity as shown in Figure 2 can result in multiple plausible interpretations and make the distribution multimodal.",
713
+ "bbox": [
714
+ 496,
715
+ 340,
716
+ 892,
717
+ 455
718
+ ],
719
+ "page_idx": 3
720
+ },
721
+ {
722
+ "type": "text",
723
+ "text": "An added challenge arises when non-opaque surfaces are present. In this case, there is no single ground truth depth, and so the distribution of ground truth depths is multimodal. So even after training the NeRF, the distribution of ray termination distances induced by NeRF should be multimodal, as illustrated in the glass walls in Figure 1.",
724
+ "bbox": [
725
+ 496,
726
+ 458,
727
+ 893,
728
+ 546
729
+ ],
730
+ "page_idx": 3
731
+ },
732
+ {
733
+ "type": "text",
734
+ "text": "Prior methods compute moments (i.e., mean and/or variance) of the distribution of depth estimates or ray termination distance induced by NeRF. For example, DS-NeRF [6] predicts the mean depth estimate and uses reprojection error as a proxy for the variance. It then minimizes the KL divergence from the distribution of ray termination distances induced by NeRF to a Gaussian whose moments match those of the depth estimates. DDP [32] fits a Gaussian to distribution of ray termination distances and maximizes the likelihood of the depth estimate under the Gaussian.",
735
+ "bbox": [
736
+ 496,
737
+ 547,
738
+ 893,
739
+ 691
740
+ ],
741
+ "page_idx": 3
742
+ },
743
+ {
744
+ "type": "text",
745
+ "text": "Distilling complex distributions to moments has the drawback of ignoring their possible multimodality. Because both the distribution of depth estimates and ray termination distances are potentially multimodal, it is important to handle the multimodality.",
746
+ "bbox": [
747
+ 496,
748
+ 694,
749
+ 893,
750
+ 767
751
+ ],
752
+ "page_idx": 3
753
+ },
754
+ {
755
+ "type": "text",
756
+ "text": "4.2. SCADE",
757
+ "text_level": 1,
758
+ "bbox": [
759
+ 500,
760
+ 773,
761
+ 596,
762
+ 787
763
+ ],
764
+ "page_idx": 3
765
+ },
766
+ {
767
+ "type": "text",
768
+ "text": "We now present our novel method Space Carving with Ambiguity-aware Depth Estimates (SCADE). Our contributions address the existing problems and drawbacks mentioned in the previous section. First, SCADE accounts for the multimodality of both the distributions of monocular depth estimates and ray termination distances that arise due to inherent ambiguities (Sec 4.2.1) and non-opaque surfaces.",
769
+ "bbox": [
770
+ 496,
771
+ 797,
772
+ 893,
773
+ 900
774
+ ],
775
+ "page_idx": 3
776
+ },
777
+ {
778
+ "type": "page_number",
779
+ "text": "16521",
780
+ "bbox": [
781
+ 480,
782
+ 944,
783
+ 517,
784
+ 955
785
+ ],
786
+ "page_idx": 3
787
+ },
788
+ {
789
+ "type": "text",
790
+ "text": "Second, SCADE is also able to resolve these ambiguities by fusing together information from multiple views through our space carving loss (Sec 4.2.3) that uses the reverse cross entropy loss. Because this a loss on the distribution rather than moments of the distribution (which was the paradigm in prior work [6, 32]), it achieves supervision in 3D rather than 2D. Third, our loss formulation is sample-based (Sec 4.2.2) and so is computationally efficient to optimize.",
791
+ "bbox": [
792
+ 75,
793
+ 90,
794
+ 472,
795
+ 209
796
+ ],
797
+ "page_idx": 4
798
+ },
799
+ {
800
+ "type": "text",
801
+ "text": "4.2.1 Our Ambiguity-aware Depth Estimates",
802
+ "text_level": 1,
803
+ "bbox": [
804
+ 76,
805
+ 215,
806
+ 405,
807
+ 229
808
+ ],
809
+ "page_idx": 4
810
+ },
811
+ {
812
+ "type": "text",
813
+ "text": "We handle multimodality of monocular depth distributions by introducing our ambiguity-aware depth estimation module to account for ambiguities in depth estimation from a single view (Figure 1, 2). In contrast to existing monocular depth estimation networks [47] that predict single point estimates for depth, we model the inherent uncertainty in monocular depth estimation by representing the distribution as a set of samples generated by a neural network. We propose leveraging conditional Implicit Maximum Likelihood estimation (cIMLE) [19] to learn a multimodal distribution of depth estimates. We chose cIMLE rather than conditional GANs [14] in order to avoid mode collapse, which can lead to a unimodal distribution.",
814
+ "bbox": [
815
+ 75,
816
+ 231,
817
+ 472,
818
+ 421
819
+ ],
820
+ "page_idx": 4
821
+ },
822
+ {
823
+ "type": "text",
824
+ "text": "Figure 3 shows our network architecture. Concretely, we combine cIMLE with a state-of-the-art monocular depth estimation network, LeReS [47]. Our ambiguity-aware depth estimation module $(G)$ takes an input image $(I)$ and a latent code $z\\sim \\mathcal{N}(0,\\mathbf{I})$ and outputs a conditional depth sample $G(I,z)$ for the input image $I$ . To inject randomness into the network, we follow the technique used in [14] and incorporate AdaIn layers into the network backbone that predicts a scale and shift to the intermediate network features. Specifically, we added four AdaIn layers to the encoder of the depth estimation backbone.",
825
+ "bbox": [
826
+ 75,
827
+ 422,
828
+ 472,
829
+ 583
830
+ ],
831
+ "page_idx": 4
832
+ },
833
+ {
834
+ "type": "text",
835
+ "text": "4.2.2 Sample-based Losses on Distributions",
836
+ "text_level": 1,
837
+ "bbox": [
838
+ 76,
839
+ 590,
840
+ 393,
841
+ 604
842
+ ],
843
+ "page_idx": 4
844
+ },
845
+ {
846
+ "type": "text",
847
+ "text": "We desire to achieve an agreement between the distributions from our ambiguity-aware prior and the ray termination distance from NeRF to constrain the NeRF optimization under the sparse view set-up using our learned prior. As distributions are continuous, taking their analytical integral is computationally intractable. Thus, in order to define a differentiable loss function for our optimization problem, we result in a sample-based loss on these two distributions.",
848
+ "bbox": [
849
+ 75,
850
+ 606,
851
+ 468,
852
+ 724
853
+ ],
854
+ "page_idx": 4
855
+ },
856
+ {
857
+ "type": "text",
858
+ "text": "Samples from depth prior. As previously described, we leverage cIMLE to sample from our ambiguity-aware depth prior. Sampling $z_{1},\\ldots ,z_{M}\\sim \\mathcal{N}(0,\\mathbf{I})$ , we get depth map samples $G(I,z_1),\\dots,G(I,z_M)$ for input image $I$ . Hence we denote corresponding samples to estimate and represent the distribution for ray $\\mathbf{r}$ as $y_{1},y_{2},\\dots,y_{M}\\sim G_{\\mathbf{o,d}}$ .",
859
+ "bbox": [
860
+ 75,
861
+ 731,
862
+ 468,
863
+ 821
864
+ ],
865
+ "page_idx": 4
866
+ },
867
+ {
868
+ "type": "text",
869
+ "text": "Samples from NeRF. For the distribution given by the ray termination distances modelled by NeRF, we sample from its probability density function $f_{\\theta ,\\mathbf{o},\\mathbf{d}}$ (Eq. 3) with inverse transform sampling. Concretely, we define the probability mass function for ray $\\mathbf{r}$ as $f_{\\theta ,\\mathbf{o},\\mathbf{d}}(t_i)$ as given in Eq. 3 with",
870
+ "bbox": [
871
+ 75,
872
+ 825,
873
+ 470,
874
+ 902
875
+ ],
876
+ "page_idx": 4
877
+ },
878
+ {
879
+ "type": "image",
880
+ "img_path": "images/5ce69a59f9693151cf0118715c9aa7a4632ba34e4fa87d3d6c04f8224a7bbbb2.jpg",
881
+ "image_caption": [],
882
+ "image_footnote": [],
883
+ "bbox": [
884
+ 542,
885
+ 92,
886
+ 694,
887
+ 160
888
+ ],
889
+ "page_idx": 4
890
+ },
891
+ {
892
+ "type": "image",
893
+ "img_path": "images/f0f60911c43cb08167196bca975a3842d6efec53ccdedbaafe3d95dc4a38fd5a.jpg",
894
+ "image_caption": [
895
+ "2D Expected Depth"
896
+ ],
897
+ "image_footnote": [],
898
+ "bbox": [
899
+ 542,
900
+ 162,
901
+ 694,
902
+ 228
903
+ ],
904
+ "page_idx": 4
905
+ },
906
+ {
907
+ "type": "image",
908
+ "img_path": "images/9e3fb50f9689e1cee02dd5c1e8baba0ac80730b0def19b79a7f53c0861ad64f9.jpg",
909
+ "image_caption": [],
910
+ "image_footnote": [],
911
+ "bbox": [
912
+ 696,
913
+ 92,
914
+ 846,
915
+ 160
916
+ ],
917
+ "page_idx": 4
918
+ },
919
+ {
920
+ "type": "image",
921
+ "img_path": "images/095433d6285bfd4172af0cf62ac7b64bbf1bcdd6b6a6d1de52d37ce1fe9081d0.jpg",
922
+ "image_caption": [
923
+ "3D Space Carving (Ours)",
924
+ "Figure 4. Our space carving loss supervises the ray termination in 3D as opposed to existing approaches that supervise on 2D expected depth. As shown, this clears out clouds of dust in space. Moreover supervising in 3D allows for wide baseline input views and still finds the common mode that correctly snaps the surfaces."
925
+ ],
926
+ "image_footnote": [],
927
+ "bbox": [
928
+ 696,
929
+ 162,
930
+ 844,
931
+ 228
932
+ ],
933
+ "page_idx": 4
934
+ },
935
+ {
936
+ "type": "text",
937
+ "text": "samples $t_1,\\ldots ,t_K$ , where $t_i\\in [t_n,t_f]\\forall i$ . By inverse transform sampling, we first compute the cumulative distribution function (CDF) of the ray termination distribution at samples $t_i$ , which we denote as $F_{\\theta ,\\mathbf{o},\\mathbf{d}}(t_i) = \\sum_{j < i}f_{\\theta ,\\mathbf{o},\\mathbf{d}}(t_i)$ . Thus our samples $x_{1},x_{2},\\dots,x_{N}$ of ray termination distances from NeRF are then given by:",
938
+ "bbox": [
939
+ 498,
940
+ 313,
941
+ 893,
942
+ 402
943
+ ],
944
+ "page_idx": 4
945
+ },
946
+ {
947
+ "type": "equation",
948
+ "text": "\n$$\nx _ {i} = F _ {\\theta , \\mathbf {o}, \\mathbf {d}} ^ {- 1} (u _ {i}), \\tag {4}\n$$\n",
949
+ "text_format": "latex",
950
+ "bbox": [
951
+ 645,
952
+ 410,
953
+ 890,
954
+ 436
955
+ ],
956
+ "page_idx": 4
957
+ },
958
+ {
959
+ "type": "text",
960
+ "text": "where $u_{i}\\sim \\mathcal{U}(0,1)$",
961
+ "bbox": [
962
+ 630,
963
+ 433,
964
+ 761,
965
+ 448
966
+ ],
967
+ "page_idx": 4
968
+ },
969
+ {
970
+ "type": "text",
971
+ "text": "4.2.3 Our Space Carving Loss",
972
+ "text_level": 1,
973
+ "bbox": [
974
+ 498,
975
+ 460,
976
+ 725,
977
+ 476
978
+ ],
979
+ "page_idx": 4
980
+ },
981
+ {
982
+ "type": "text",
983
+ "text": "We now introduce our novel space carving loss utilizes the samples from both multimodal distributions to constrain the NeRF optimization. The goal is to find a subset of modes captured in the monocular depth distributions from each view that are globally consistent. In doing so, we can fuse together the information from the different views, since the inherent ambiguities are only resolved given information from multiple views. This would allow us to find a common shape that is consistent across all views and \"snap\" objects surfaces together.",
984
+ "bbox": [
985
+ 496,
986
+ 477,
987
+ 890,
988
+ 623
989
+ ],
990
+ "page_idx": 4
991
+ },
992
+ {
993
+ "type": "text",
994
+ "text": "We thus desire a loss that has mode seeking behavior, that is the modes of the distribution being trained should be a subset of the modes of the distribution that provides supervision. One such loss is the reverse cross entropy, which is the cross entropy from the NeRF ray termination distance distribution to the distribution of our ambiguity-aware depth estimates:",
995
+ "bbox": [
996
+ 496,
997
+ 623,
998
+ 893,
999
+ 726
1000
+ ],
1001
+ "page_idx": 4
1002
+ },
1003
+ {
1004
+ "type": "equation",
1005
+ "text": "\n$$\nH \\left(f _ {\\theta , \\mathbf {o}, \\mathbf {d}}, G _ {\\mathbf {o}, \\mathbf {d}}\\right) = - \\mathbb {E} _ {f _ {\\theta , \\mathbf {o}, \\mathbf {d}}} [ \\log G _ {\\mathbf {o}, \\mathbf {d}} ]. \\tag {5}\n$$\n",
1006
+ "text_format": "latex",
1007
+ "bbox": [
1008
+ 563,
1009
+ 734,
1010
+ 890,
1011
+ 753
1012
+ ],
1013
+ "page_idx": 4
1014
+ },
1015
+ {
1016
+ "type": "text",
1017
+ "text": "As shown in the IMLE papers [19, 28], this is equivalent to penalizing the minimum of the L2 norm between the samples from the two distributions, please see [18] for the full proof. Hence, our space carving loss is given by",
1018
+ "bbox": [
1019
+ 496,
1020
+ 760,
1021
+ 890,
1022
+ 821
1023
+ ],
1024
+ "page_idx": 4
1025
+ },
1026
+ {
1027
+ "type": "equation",
1028
+ "text": "\n$$\n\\mathcal {L} _ {\\text {s p a c e _ c a r v i n g}} (\\mathbf {r}) = \\sum_ {i \\in [ N ]} \\min _ {j \\in [ M ]} \\| x _ {i} - y _ {j} \\| _ {2} ^ {2}. \\tag {6}\n$$\n",
1029
+ "text_format": "latex",
1030
+ "bbox": [
1031
+ 555,
1032
+ 828,
1033
+ 890,
1034
+ 863
1035
+ ],
1036
+ "page_idx": 4
1037
+ },
1038
+ {
1039
+ "type": "text",
1040
+ "text": "Because our novel space carving loss operates on distributions rather than moments of distributions, we can have",
1041
+ "bbox": [
1042
+ 498,
1043
+ 869,
1044
+ 893,
1045
+ 900
1046
+ ],
1047
+ "page_idx": 4
1048
+ },
1049
+ {
1050
+ "type": "page_number",
1051
+ "text": "16522",
1052
+ "bbox": [
1053
+ 480,
1054
+ 944,
1055
+ 517,
1056
+ 955
1057
+ ],
1058
+ "page_idx": 4
1059
+ },
1060
+ {
1061
+ "type": "text",
1062
+ "text": "different supervision targets for different points along the ray. In contrast, if the loss were to only to operate on moments, which are integrals along the ray, this only yields the same supervision target for all points along the ray. Therefore, our loss allows us to supervise in 3D, unlike prior methods [6, 32] (which use losses on moments) that supervise in 2D. 3D supervision allows us to clear out dust in space, since in order to move a sample from the ray termination distance distribution farther in depth, the loss would have to decrease the probability density of all the points in front of it. These advantages are highlighted in Figure 4.",
1063
+ "bbox": [
1064
+ 75,
1065
+ 90,
1066
+ 472,
1067
+ 252
1068
+ ],
1069
+ "page_idx": 5
1070
+ },
1071
+ {
1072
+ "type": "text",
1073
+ "text": "Our total loss to optimize and train our NeRF model is given by",
1074
+ "bbox": [
1075
+ 75,
1076
+ 252,
1077
+ 470,
1078
+ 282
1079
+ ],
1080
+ "page_idx": 5
1081
+ },
1082
+ {
1083
+ "type": "equation",
1084
+ "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {\\text {p h o t o m e t r i c}} + \\lambda \\mathcal {L} _ {\\text {s p a c e _ c a r v i n g}}, \\tag {7}\n$$\n",
1085
+ "text_format": "latex",
1086
+ "bbox": [
1087
+ 163,
1088
+ 292,
1089
+ 468,
1090
+ 310
1091
+ ],
1092
+ "page_idx": 5
1093
+ },
1094
+ {
1095
+ "type": "text",
1096
+ "text": "where $\\mathcal{L}_{\\mathrm{photometric}}$ is the standard MSE loss on the predicted and ground truth image. See Fig. 1 for our overall pipeline.",
1097
+ "bbox": [
1098
+ 75,
1099
+ 320,
1100
+ 468,
1101
+ 351
1102
+ ],
1103
+ "page_idx": 5
1104
+ },
1105
+ {
1106
+ "type": "text",
1107
+ "text": "5. Results",
1108
+ "text_level": 1,
1109
+ "bbox": [
1110
+ 76,
1111
+ 363,
1112
+ 163,
1113
+ 378
1114
+ ],
1115
+ "page_idx": 5
1116
+ },
1117
+ {
1118
+ "type": "text",
1119
+ "text": "In this section, we present our experimental evaluation to demonstrate the advantages of SCADE.",
1120
+ "bbox": [
1121
+ 75,
1122
+ 381,
1123
+ 468,
1124
+ 411
1125
+ ],
1126
+ "page_idx": 5
1127
+ },
1128
+ {
1129
+ "type": "text",
1130
+ "text": "5.1. Datasets and Evaluation Metrics",
1131
+ "text_level": 1,
1132
+ "bbox": [
1133
+ 76,
1134
+ 416,
1135
+ 362,
1136
+ 430
1137
+ ],
1138
+ "page_idx": 5
1139
+ },
1140
+ {
1141
+ "type": "text",
1142
+ "text": "We evaluate our method on ScanNet [5] and an in-the-wild dataset we collected<sup>1</sup>. We use the sparse-view ScanNet data used by DDP [32] in their evaluations, which comprises of three sample scenes each with 18 to 20 train images and 8 test images. To further test the robustness of our method, we further evaluated our method on three in-the-wild scenes collected using an iPhoneX. We captured sparse views in three different scenes - basement, kitchen and lounge, where each scene has 18 to 23 train images and 8 test images. Following similar data preprocessing steps as DDP [32], we ran SfM [34] on all images to obtain camera poses for NeRF training. For quantitative comparison, we follow the original NeRF [24] paper and report the PSNR, SSIM [42] and LPIPS [52] on the novel test views.",
1143
+ "bbox": [
1144
+ 75,
1145
+ 431,
1146
+ 472,
1147
+ 636
1148
+ ],
1149
+ "page_idx": 5
1150
+ },
1151
+ {
1152
+ "type": "text",
1153
+ "text": "5.2. Implementation Details",
1154
+ "text_level": 1,
1155
+ "bbox": [
1156
+ 76,
1157
+ 646,
1158
+ 294,
1159
+ 661
1160
+ ],
1161
+ "page_idx": 5
1162
+ },
1163
+ {
1164
+ "type": "text",
1165
+ "text": "We train our ambiguity-aware prior on the Taskonomy [50] dataset and initialize the weights with the pretrained LeReS [47] model. We use $M = 20$ depth estimates in our experiments. See supplement for additional details.",
1166
+ "bbox": [
1167
+ 75,
1168
+ 662,
1169
+ 470,
1170
+ 720
1171
+ ],
1172
+ "page_idx": 5
1173
+ },
1174
+ {
1175
+ "type": "text",
1176
+ "text": "5.3. Experiments on ScanNet",
1177
+ "text_level": 1,
1178
+ "bbox": [
1179
+ 76,
1180
+ 726,
1181
+ 303,
1182
+ 742
1183
+ ],
1184
+ "page_idx": 5
1185
+ },
1186
+ {
1187
+ "type": "text",
1188
+ "text": "We evaluate on ScanNet following the evaluation from DDP [32]. We compare to the original NeRF [24] (Vanilla NeRF) and the recent state-of-the-art NeRF with depth prior-based supervision, Dense Depth Prior [32] (DDP). Table 1 shows SCADE quantitatively outperforming the baselines. Because we are interested in simulating real-world conditions with a domain gap between the prior and the NeRF, our setting requires the use of out-of-domain priors. Since DDP uses a in-domain prior, we retrain DDP's depth completion",
1189
+ "bbox": [
1190
+ 75,
1191
+ 744,
1192
+ 472,
1193
+ 878
1194
+ ],
1195
+ "page_idx": 5
1196
+ },
1197
+ {
1198
+ "type": "image",
1199
+ "img_path": "images/949b1024709f36574ef8fca2f848b5a09b39f2c61a19449dd8a1bfcb1d311674.jpg",
1200
+ "image_caption": [
1201
+ "Figure 5. Depth and Fusion Comparison. Depth map and fusion comparison between SCADE and DDP."
1202
+ ],
1203
+ "image_footnote": [],
1204
+ "bbox": [
1205
+ 503,
1206
+ 82,
1207
+ 888,
1208
+ 190
1209
+ ],
1210
+ "page_idx": 5
1211
+ },
1212
+ {
1213
+ "type": "table",
1214
+ "img_path": "images/9e79add36f31f9f730a8d851dc39149f2ae384c31bbc009d5a948066029dffb3.jpg",
1215
+ "table_caption": [],
1216
+ "table_footnote": [],
1217
+ "table_body": "<table><tr><td></td><td>PSNR ↑</td><td>SSIM ↑</td><td>LPIPS ↓</td></tr><tr><td>Vanilla NeRF [24]</td><td>19.03</td><td>0.670</td><td>0.398</td></tr><tr><td>NerfingMVS [47]</td><td>16.29</td><td>0.626</td><td>0.502</td></tr><tr><td>IBRNet [41]</td><td>13.25</td><td>0.529</td><td>0.673</td></tr><tr><td>MVSNeRF [3]</td><td>15.67</td><td>0.533</td><td>0.635</td></tr><tr><td>DS-NeRF [6]</td><td>20.85</td><td>0.713</td><td>0.344</td></tr><tr><td>DDP [32]</td><td>19.29</td><td>0.695</td><td>0.368</td></tr><tr><td>SCADE (Ours)</td><td>21.54</td><td>0.732</td><td>0.292</td></tr></table>",
1218
+ "bbox": [
1219
+ 501,
1220
+ 226,
1221
+ 890,
1222
+ 366
1223
+ ],
1224
+ "page_idx": 5
1225
+ },
1226
+ {
1227
+ "type": "text",
1228
+ "text": "Table 1. ScanNet Results. Results for DS-NeRF and NerfingMVS follow what was reported in prior literature [32]. Because our setting requires out-of-domain priors, the results for DDP are with out-of-domain priors. The results of DDP with in-domain priors are (20.96, 0.737, 0.236) for PSNR, SSIM and LPIPS, respectively.",
1229
+ "bbox": [
1230
+ 498,
1231
+ 369,
1232
+ 890,
1233
+ 439
1234
+ ],
1235
+ "page_idx": 5
1236
+ },
1237
+ {
1238
+ "type": "text",
1239
+ "text": "network using their official code and hyperparameters on Taskonomy [50], i.e. the same out-of-domain dataset that our prior is trained on.",
1240
+ "bbox": [
1241
+ 498,
1242
+ 446,
1243
+ 890,
1244
+ 491
1245
+ ],
1246
+ "page_idx": 5
1247
+ },
1248
+ {
1249
+ "type": "text",
1250
+ "text": "Figure 6 shows our qualitative results. As shown, compared to the baselines, SCADE is able to avoid producing the clouds of dust that are present in the results of baselines (first, third and last column). Moreover, SCADE is also able to snap and recover objects in the scene such as the details on the blinds (second column), the back of the chair (fourth column) and the legs of the piano stool (fifth column). Moreover, Fig. 5 shows rendered depthmaps and fusion results using [51] on a Scannet scene. Notice that SCADE is able to recover better geometry compared to DDP - see corner of the calendar, cabinets and office chair in the right image.",
1251
+ "bbox": [
1252
+ 496,
1253
+ 491,
1254
+ 893,
1255
+ 654
1256
+ ],
1257
+ "page_idx": 5
1258
+ },
1259
+ {
1260
+ "type": "text",
1261
+ "text": "5.4. Experiments on In-the-Wild Data",
1262
+ "text_level": 1,
1263
+ "bbox": [
1264
+ 500,
1265
+ 660,
1266
+ 794,
1267
+ 675
1268
+ ],
1269
+ "page_idx": 5
1270
+ },
1271
+ {
1272
+ "type": "text",
1273
+ "text": "To further test the robustness of SCADE, we further evaluate on data collected in-the-wild with a standard phone camera (iPhoneX). The intrinsics for different views are different and are not known to the algorithm. As shown in Table 2, SCADE with the same out-of-domain prior (trained on Taskonomy) outperforms the baselines on in-the-wild captured data, which demonstrates its robustness. Figure 7 shows qualitative examples. Interestingly, we are able to recover the room behind the glass better the baselines, whose results are more blurry (first, third and fourth column). We are also able to recover objects behind the glass such as the monitor and chair (first column). As discussed in Sect. 4.1, non-opaque surfaces are the most challenging because both the ground truth distribution of ray termination distances induced by NeRF and the distribution of estimated depths",
1274
+ "bbox": [
1275
+ 496,
1276
+ 679,
1277
+ 893,
1278
+ 900
1279
+ ],
1280
+ "page_idx": 5
1281
+ },
1282
+ {
1283
+ "type": "page_footnote",
1284
+ "text": "<sup>1</sup>Also see supplement for results on the Tanks and Temples dataset [16].",
1285
+ "bbox": [
1286
+ 94,
1287
+ 886,
1288
+ 467,
1289
+ 900
1290
+ ],
1291
+ "page_idx": 5
1292
+ },
1293
+ {
1294
+ "type": "page_number",
1295
+ "text": "16523",
1296
+ "bbox": [
1297
+ 480,
1298
+ 944,
1299
+ 517,
1300
+ 955
1301
+ ],
1302
+ "page_idx": 5
1303
+ },
1304
+ {
1305
+ "type": "image",
1306
+ "img_path": "images/e9189b41264e143ca49e61725e478d56b531c4f517a25de627d60d97f175c7fe.jpg",
1307
+ "image_caption": [
1308
+ "Figure 6. Qualitative results on ScanNet. As shown, SCADE is able to avoid producing the clouds of dust that are present in the results of baselines (1st, 3rd, last columns). Moreover, SCADE is also able to snap and recover objects in the scene such as the details on the blinds (2nd column), the back of the chair (4th column) and the legs of the piano stool (5th column)."
1309
+ ],
1310
+ "image_footnote": [],
1311
+ "bbox": [
1312
+ 99,
1313
+ 90,
1314
+ 869,
1315
+ 343
1316
+ ],
1317
+ "page_idx": 6
1318
+ },
1319
+ {
1320
+ "type": "table",
1321
+ "img_path": "images/fe37034eb23db24e5a1050fedb62f31d15d23bdfad568a658cd709102a20d220.jpg",
1322
+ "table_caption": [],
1323
+ "table_footnote": [],
1324
+ "table_body": "<table><tr><td></td><td>PSNR ↑</td><td>SSIM ↑</td><td>LPIPS ↓</td></tr><tr><td>Vanilla NeRF [24]</td><td>20.46</td><td>0.713</td><td>0.398</td></tr><tr><td>DDP [32]</td><td>21.28</td><td>0.727</td><td>0.366</td></tr><tr><td>SCADE (Ours)</td><td>22.82</td><td>0.743</td><td>0.347</td></tr></table>",
1325
+ "bbox": [
1326
+ 78,
1327
+ 397,
1328
+ 468,
1329
+ 477
1330
+ ],
1331
+ "page_idx": 6
1332
+ },
1333
+ {
1334
+ "type": "text",
1335
+ "text": "from the prior must be multimodal. This validates the ability of our model to capture the multimodal distributions.",
1336
+ "bbox": [
1337
+ 75,
1338
+ 494,
1339
+ 468,
1340
+ 525
1341
+ ],
1342
+ "page_idx": 6
1343
+ },
1344
+ {
1345
+ "type": "text",
1346
+ "text": "Similar to the ScanNet results, we are able to recover crisp shapes of objects such as the thin table leg and the white chair (second and last column), and we are also able to clear up dust near the microwave and printer compared to the baselines (fifth and last column).",
1347
+ "bbox": [
1348
+ 75,
1349
+ 525,
1350
+ 468,
1351
+ 599
1352
+ ],
1353
+ "page_idx": 6
1354
+ },
1355
+ {
1356
+ "type": "text",
1357
+ "text": "5.5. Ablation Study",
1358
+ "text_level": 1,
1359
+ "bbox": [
1360
+ 76,
1361
+ 606,
1362
+ 230,
1363
+ 622
1364
+ ],
1365
+ "page_idx": 6
1366
+ },
1367
+ {
1368
+ "type": "text",
1369
+ "text": "Supervision on full distribution vs moments We first validate the importance of using our novel space carving loss that supervises the full distribution of ray termination distance rather than just its moments. The latter integrates along the ray, and so provides one target value for the entire ray, which effectively makes it 2D supervision. The former provides different target values for different samples along the same ray, which makes it 3D supervision. We adapt our method to use 2D supervision proposed in the recent MonoSDF [49], which computes the expected ray termination distance and aligns the output depth map with a monocular depth estimation prior using the MiDaS [30] loss. We use our learned prior as the monocular depth estimate. Table 3 validates the effectiveness of our space carving loss.",
1370
+ "bbox": [
1371
+ 75,
1372
+ 625,
1373
+ 470,
1374
+ 832
1375
+ ],
1376
+ "page_idx": 6
1377
+ },
1378
+ {
1379
+ "type": "text",
1380
+ "text": "Multimodal prior vs unimodal prior. Our prior models does not constrain the depth distribution to a particular form and allows it to be multimodal. In contrast, DDP's depth completion prior assumes a Gaussian, which is unimodal.",
1381
+ "bbox": [
1382
+ 75,
1383
+ 840,
1384
+ 472,
1385
+ 902
1386
+ ],
1387
+ "page_idx": 6
1388
+ },
1389
+ {
1390
+ "type": "table",
1391
+ "img_path": "images/ffe62631a9c24eccb06a7e079ddd2ff92556d3af67acf0602456f18dfce15e89.jpg",
1392
+ "table_caption": [
1393
+ "Table 2. In-the-wild Results."
1394
+ ],
1395
+ "table_footnote": [],
1396
+ "table_body": "<table><tr><td></td><td>PSNR ↑</td><td>SSIM ↑</td><td>LPIPS ↓</td></tr><tr><td>MonoSDF supervision</td><td>20.13</td><td>0.710</td><td>0.332</td></tr><tr><td>DDP prior - single sample</td><td>20.85</td><td>0.712</td><td>0.320</td></tr><tr><td>DDP prior - multiple samples</td><td>21.00</td><td>0.718</td><td>0.316</td></tr><tr><td>Our prior - single sample</td><td>21.22</td><td>0.714</td><td>0.318</td></tr><tr><td>SCADE (Ours)</td><td>21.54</td><td>0.732</td><td>0.292</td></tr></table>",
1397
+ "bbox": [
1398
+ 501,
1399
+ 397,
1400
+ 890,
1401
+ 507
1402
+ ],
1403
+ "page_idx": 6
1404
+ },
1405
+ {
1406
+ "type": "text",
1407
+ "text": "Table 3. Ablation Study. Results on the Scannet dataset. MonoSDF supervision refers to supervising with MonoSDF loss on expected ray termination distance using our prior.",
1408
+ "bbox": [
1409
+ 498,
1410
+ 511,
1411
+ 893,
1412
+ 553
1413
+ ],
1414
+ "page_idx": 6
1415
+ },
1416
+ {
1417
+ "type": "text",
1418
+ "text": "We ablate our prior against DDP's by sampling from their Gaussian distribution. We train with our space carving loss using both a single sample, as well as $M$ samples drawn from DDP's prior. As shown in Table 3, while both perform better than the 2D supervision provided by MonoSDF, but they are subpar compared to using our multimodal prior.",
1419
+ "bbox": [
1420
+ 496,
1421
+ 561,
1422
+ 893,
1423
+ 651
1424
+ ],
1425
+ "page_idx": 6
1426
+ },
1427
+ {
1428
+ "type": "text",
1429
+ "text": "Multiple vs single sample. Finally, we also ablate on using a single sample vs multiple samples for our space carving loss. For the single sample set-up, we supervise using the sample mean of the depth estimates, which is equivalent to the maximum likelihood estimate of the depth under Gaussian likelihood. Results in Table 3 show the importance of using multiple samples for our space carving loss.",
1430
+ "bbox": [
1431
+ 496,
1432
+ 657,
1433
+ 893,
1434
+ 762
1435
+ ],
1436
+ "page_idx": 6
1437
+ },
1438
+ {
1439
+ "type": "text",
1440
+ "text": "Sparsity. We further show addt'l results under varying number of views in Tab 4. Note that in general sparsity is not fully reflected by the absolute number of views, because all else being equal, a larger scene requires more views to attain complete coverage. Following prior work [32], we also report the average number of views that see the same point as another measure of sparsity. As shown, our setting is at the edge of the theoretical lower limit of 2 for general-purpose reconstruction, which shows view sparsity.",
1441
+ "bbox": [
1442
+ 496,
1443
+ 768,
1444
+ 893,
1445
+ 900
1446
+ ],
1447
+ "page_idx": 6
1448
+ },
1449
+ {
1450
+ "type": "page_number",
1451
+ "text": "16524",
1452
+ "bbox": [
1453
+ 480,
1454
+ 944,
1455
+ 519,
1456
+ 955
1457
+ ],
1458
+ "page_idx": 6
1459
+ },
1460
+ {
1461
+ "type": "image",
1462
+ "img_path": "images/7e2362ab6a15190dbdcc9d678e22a5b9e89861b7021b641f07fec5373cdc845d.jpg",
1463
+ "image_caption": [
1464
+ "Figure 7. Qualitative results on In-the-Wild data. As shown, we are able to recover the room behind the glass better the baselines, whose results are more blurry (1st, 3rd, 4th column). Also, SCADE is able to recover crisp shapes of objects such as the thin table leg and the white chair (2nd, last column), and we are also able to clear up dust near the microwave and printer compared to the baselines (5th, last column)."
1465
+ ],
1466
+ "image_footnote": [],
1467
+ "bbox": [
1468
+ 101,
1469
+ 90,
1470
+ 869,
1471
+ 345
1472
+ ],
1473
+ "page_idx": 7
1474
+ },
1475
+ {
1476
+ "type": "table",
1477
+ "img_path": "images/8a0faede98243af031aa79b018a997b7fb9b6baf0046a053c27e95fac9f5a15c.jpg",
1478
+ "table_caption": [],
1479
+ "table_footnote": [],
1480
+ "table_body": "<table><tr><td>ave. #views visible</td><td>1.87</td><td>1.98</td><td>2.01</td><td>2.2</td></tr><tr><td>absolute #views</td><td>18</td><td>20</td><td>22</td><td>24</td></tr><tr><td>DDP / SCADE</td><td>19.35/ 21.66</td><td>22.4/ 23.67</td><td>23.10/ 24.00</td><td>23.56/ 24.84</td></tr></table>",
1481
+ "bbox": [
1482
+ 78,
1483
+ 398,
1484
+ 470,
1485
+ 448
1486
+ ],
1487
+ "page_idx": 7
1488
+ },
1489
+ {
1490
+ "type": "text",
1491
+ "text": "Table 4. PSNR for DDP [32]/SCADE on scene781 of Scannet.",
1492
+ "bbox": [
1493
+ 86,
1494
+ 450,
1495
+ 459,
1496
+ 464
1497
+ ],
1498
+ "page_idx": 7
1499
+ },
1500
+ {
1501
+ "type": "image",
1502
+ "img_path": "images/29b97ea9fa589445999cb9da35c0638f35b86213f46f987167af3609e7216b4d.jpg",
1503
+ "image_caption": [],
1504
+ "image_footnote": [],
1505
+ "bbox": [
1506
+ 78,
1507
+ 468,
1508
+ 303,
1509
+ 573
1510
+ ],
1511
+ "page_idx": 7
1512
+ },
1513
+ {
1514
+ "type": "image",
1515
+ "img_path": "images/4d6aa5d3f52bf5f17bac43561c98b1ea27cd406ae972d63e2f89a5253283b950.jpg",
1516
+ "image_caption": [
1517
+ "Figure 8. Depth Mode Discussion. a) Train images from Taskonomy [50] and their labels. Notice that non-opaque surfaces are labelled differently. b) Output of our prior on reflective surfaces."
1518
+ ],
1519
+ "image_footnote": [],
1520
+ "bbox": [
1521
+ 308,
1522
+ 468,
1523
+ 467,
1524
+ 575
1525
+ ],
1526
+ "page_idx": 7
1527
+ },
1528
+ {
1529
+ "type": "text",
1530
+ "text": "5.6. Discussion on our Ambiguity-Aware Prior",
1531
+ "text_level": 1,
1532
+ "bbox": [
1533
+ 76,
1534
+ 628,
1535
+ 434,
1536
+ 645
1537
+ ],
1538
+ "page_idx": 7
1539
+ },
1540
+ {
1541
+ "type": "text",
1542
+ "text": "We provide addt'l details on our ambiguity-aware prior on i) how and why it works, ii) its performance on reflective surfaces, which is a common failure case for depth estimate, and iii) provide an intuition on how what our depth modes look like. i) We achieve variable depth modes by exploiting inconsistently labeled training data. As shown in Fig 8-a, in Taskonomy [50], different training images with non-opaque surfaces label depth differently: shooting through the glass (left), on the glass (middle), or a mixture of both (right). Despite multiple possible depth labels, each image only has one ground truth label. Training with cIMLE allows our prior to model these multiple possible (ambiguous) outputs through sampling $^2$ , even when given only one label per image. Interestingly when testing our prior on a test image with a mirror, we find that it is able to capture variable modes",
1543
+ "bbox": [
1544
+ 75,
1545
+ 646,
1546
+ 472,
1547
+ 864
1548
+ ],
1549
+ "page_idx": 7
1550
+ },
1551
+ {
1552
+ "type": "text",
1553
+ "text": "on reflective surfaces, including the \"correct\" flat surface as shown in Fig 8-b. An intuition of what the depth modes look like is for example if a ray intersects $n - 1$ non-opaque surfaces, the $n$ modes are intersections with the non-opaque surfaces and the terminating point on the opaque surface.",
1554
+ "bbox": [
1555
+ 496,
1556
+ 401,
1557
+ 890,
1558
+ 474
1559
+ ],
1560
+ "page_idx": 7
1561
+ },
1562
+ {
1563
+ "type": "text",
1564
+ "text": "6. Conclusion",
1565
+ "text_level": 1,
1566
+ "bbox": [
1567
+ 500,
1568
+ 479,
1569
+ 617,
1570
+ 494
1571
+ ],
1572
+ "page_idx": 7
1573
+ },
1574
+ {
1575
+ "type": "text",
1576
+ "text": "In this paper, we present a new approach towards NeRF reconstruction that can work with a modest number of in-the-wild views of an indoor scene. We address the underconstrained nature of this problem by regularizing the NeRF optimization with additional depth estimates for each view. Our key technical contribution is to model multimodality in the depth estimates, which can capture inherent ambiguities in monocular depth estimation as well as the possible presence of non-opaque surfaces. We resolve ambiguities using a novel space carving loss that fuses the multimodal depth estimates from different views and seeks the modes that are consistent across views so as to arrive at a globally consistent 3D reconstruction. The improved recovery of shape and appearance enables higher fidelity novel view synthesis from sparse views.",
1577
+ "bbox": [
1578
+ 496,
1579
+ 497,
1580
+ 893,
1581
+ 717
1582
+ ],
1583
+ "page_idx": 7
1584
+ },
1585
+ {
1586
+ "type": "text",
1587
+ "text": "Limitations and Future Work. The performance of our method is constrained by the quality of the monocular depth priors. While we found that our prior generalizes well across domains, if the domain gap is too great, the performance of our method will degrade. A future direction would be to detect when this happens and dynamically adjust the strength of depth supervision in response.",
1588
+ "bbox": [
1589
+ 496,
1590
+ 720,
1591
+ 893,
1592
+ 824
1593
+ ],
1594
+ "page_idx": 7
1595
+ },
1596
+ {
1597
+ "type": "text",
1598
+ "text": "Acknowledgements. We sincerely thank the Teleportation and CCI team in Google for all the insightful discussions during the summer. We also thank Mirko Visontai for internship logistics, Guandao Yang for experiment set-up assist, and Weicheng Kuo for looking over the paper.",
1599
+ "bbox": [
1600
+ 496,
1601
+ 825,
1602
+ 893,
1603
+ 902
1604
+ ],
1605
+ "page_idx": 7
1606
+ },
1607
+ {
1608
+ "type": "page_footnote",
1609
+ "text": "2See supplement for more qualitative examples on our ambiguity-aware depth estimates.",
1610
+ "bbox": [
1611
+ 76,
1612
+ 875,
1613
+ 468,
1614
+ 902
1615
+ ],
1616
+ "page_idx": 7
1617
+ },
1618
+ {
1619
+ "type": "page_number",
1620
+ "text": "16525",
1621
+ "bbox": [
1622
+ 480,
1623
+ 944,
1624
+ 517,
1625
+ 955
1626
+ ],
1627
+ "page_idx": 7
1628
+ },
1629
+ {
1630
+ "type": "text",
1631
+ "text": "References",
1632
+ "text_level": 1,
1633
+ "bbox": [
1634
+ 78,
1635
+ 89,
1636
+ 174,
1637
+ 104
1638
+ ],
1639
+ "page_idx": 8
1640
+ },
1641
+ {
1642
+ "type": "list",
1643
+ "sub_type": "ref_text",
1644
+ "list_items": [
1645
+ "[1] Sai Bi, Zexiang Xu, Pratul Srinivasan, Ben Mildenhall, Kalyan Sunkavalli, Miloš Hasan, Yannick Hold-Geoffroy, David Kriegman, and Ravi Ramamoorthi. Neural reflectance fields for appearance acquisition. arXiv preprint arXiv:2008.03824, 2020. 3",
1646
+ "[2] Di Chang, Aljaž Božić, Tong Zhang, Qingsong Yan, Yingcong Chen, Sabine Susstrunk, and Matthias Nießner. Rc-mvsnet: Unsupervised multi-view stereo with neural rendering. ECCV 2022, 2022. 3",
1647
+ "[3] Anpei Chen, Zexiang Xu, Fuqiang Zhao, Xiaoshuai Zhang, Fanbo Xiang, Jingyi Yu, and Hao Su. Mvsnerf: Fast generalizable radiance field reconstruction from multi-view stereo. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14124-14133, 2021. 3, 6",
1648
+ "[4] Kyung Hyun Cho, Tapani Raiko, and Alexander Ilin. Gaussian-bernoulli deep boltzmann machine. In The 2013 International Joint Conference on Neural Networks (IJCNN), pages 1-7. IEEE, 2013. 2",
1649
+ "[5] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5828-5839, 2017. 6",
1650
+ "[6] Kangle Deng, Andrew Liu, Jun-Yan Zhu, and Deva Ramanan. Depth-supervised nerf: Fewer views and faster training for free. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12882-12891, 2022. 1, 2, 3, 4, 5, 6",
1651
+ "[7] David Eigen, Christian Puhrsch, and Rob Fergus. Depth map prediction from a single image using a multi-scale deep network. Advances in neural information processing systems, 27, 2014. 3",
1652
+ "[8] Clément Godard, Oisin Mac Aodha, and Gabriel J Brostow. Unsupervised monocular depth estimation with left-right consistency. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 270-279, 2017. 3",
1653
+ "[9] Xiaodong Gu, Zhiwen Fan, Siyu Zhu, Zuozhuo Dai, Feitong Tan, and Ping Tan. Cascade cost volume for high-resolution multi-view stereo and stereo matching. 2019. 3",
1654
+ "[10] Geoffrey E Hinton. A practical guide to training restricted boltzmann machines. In Neural networks: Tricks of the trade, pages 599-619. Springer, 2012. 2",
1655
+ "[11] Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros. Image-to-image translation with conditional adversarial networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1125-1134, 2017. 2",
1656
+ "[12] Ajay Jain, Matthew Tancik, and Pieter Abbeel. Putting nerf on a diet: Semantically consistent few-shot view synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 5885-5894, October 2021. 3",
1657
+ "[13] Wonbong Jang and Lourdes Agapito. Codenerf: Disentangled neural radiance fields for object categories. In ICCV, 2021. 1",
1658
+ "[14] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 5"
1659
+ ],
1660
+ "bbox": [
1661
+ 78,
1662
+ 114,
1663
+ 470,
1664
+ 900
1665
+ ],
1666
+ "page_idx": 8
1667
+ },
1668
+ {
1669
+ "type": "list",
1670
+ "sub_type": "ref_text",
1671
+ "list_items": [
1672
+ "[15] Alex Kendall and Yarin Gal. What uncertainties do we need in bayesian deep learning for computer vision? Advances in neural information processing systems, 30, 2017. 3",
1673
+ "[16] Arno Knapitsch, Jaesik Park, Qian-Yi Zhou, and Vladlen Koltun. Tanks and temples: Benchmarking large-scale scene reconstruction. ACM Transactions on Graphics, 36(4), 2017. 6",
1674
+ "[17] K.N. Kutulakos and S.M. Seitz. A theory of shape by space carving. In ICCV, 1999. 2",
1675
+ "[18] Ke Li and Jitendra Malik. Implicit maximum likelihood estimation. arXiv preprint arXiv:1809.09087, 2018. 5",
1676
+ "[19] Ke Li*, Shichong Peng*, Tianhao Zhang*, and Jitendra Malik. Multimodal image synthesis with conditional implicit maximum likelihood estimation. International Journal of Computer Vision, May 2020. 2, 3, 5",
1677
+ "[20] Zhengqi Li, Simon Niklaus, Noah Snavely, and Oliver Wang. Neural scene flow fields for space-time view synthesis of dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6498-6508, 2021. 3",
1678
+ "[21] Zhengqi Li and Noah Snavely. Megadepth: Learning single-view depth prediction from internet photos. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2041-2050, 2018. 3",
1679
+ "[22] Ricardo Martin-Brualla, Noha Radwan, Mehdi S. M. Sajjadi, Jonathan T. Barron, Alexey Dosovitskiy, and Daniel Duckworth. NeRF in the Wild: Neural Radiance Fields for Unconstrained Photo Collections. In CVPR, 2021. 3",
1680
+ "[23] Nelson Max. Optical models for direct volume rendering. IEEE Transactions on Visualization and Computer Graphics, 1(2):99-108, 1995. 2",
1681
+ "[24] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part I, pages 405-421, 2020. 1, 2, 3, 6, 7",
1682
+ "[25] Michael Niemeyer, Jonathan T. Barron, Ben Mildenhall, Mehdi S. M. Sajjadi, Andreas Geiger, and Noha Radwan. Regnerf: Regularizing neural radiance fields for view synthesis from sparse inputs. In CVPR, 2022. 1, 3",
1683
+ "[26] Michael Oechsle, Songyou Peng, and Andreas Geiger. Unisurf: Unifying neural implicit surfaces and radiance fields for multi-view reconstruction. In International Conference on Computer Vision (ICCV), 2021. 3",
1684
+ "[27] Keunhong Park, Utkarsh Sinha, Jonathan T. Barron, Sofien Bouaziz, Dan B Goldman, Steven M. Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable neural radiance fields. ICCV, 2021. 3",
1685
+ "[28] Shichong Peng, Seyed Alireza Moazenipourasil, and Ke Li. Chimle: Conditional hierarchical imle. In NeurIPS, 2022. 2, 5",
1686
+ "[29] Xiaojuan Qi, Renjie Liao, Zhengzhe Liu, Raquel Urtasun, and Jiaya Jia. Geonet: Geometric neural network for joint depth and surface normal estimation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 283-291, 2018. 3",
1687
+ "[30] René Ranftl, Katrin Lasinger, David Hafner, Konrad Schindler, and Vladlen Koltun. Towards robust monocular"
1688
+ ],
1689
+ "bbox": [
1690
+ 501,
1691
+ 92,
1692
+ 893,
1693
+ 900
1694
+ ],
1695
+ "page_idx": 8
1696
+ },
1697
+ {
1698
+ "type": "page_number",
1699
+ "text": "16526",
1700
+ "bbox": [
1701
+ 480,
1702
+ 944,
1703
+ 519,
1704
+ 955
1705
+ ],
1706
+ "page_idx": 8
1707
+ },
1708
+ {
1709
+ "type": "list",
1710
+ "sub_type": "ref_text",
1711
+ "list_items": [
1712
+ "depth estimation: Mixing datasets for zero-shot cross-dataset transfer. IEEE transactions on pattern analysis and machine intelligence, 2020. 3, 7",
1713
+ "[31] Marc'Aurelio Ranzato, Y-Lan Boureau, Sumit Chopra, and Yann LeCun. A unified energy-based framework for unsupervised learning. In Artificial Intelligence and Statistics, pages 371-379. PMLR, 2007. 2",
1714
+ "[32] Barbara Roessle, Jonathan T Barron, Ben Mildenhall, Pratul P Srinivasan, and Matthias Nießner. Dense depth priors for neural radiance fields from sparse input views. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12892-12901, 2022. 1, 2, 3, 4, 5, 6, 7, 8",
1715
+ "[33] Ashutosh Saxena, Min Sun, and Andrew Y Ng. Make3d: Learning 3d scene structure from a single still image. IEEE transactions on pattern analysis and machine intelligence, 31(5):824-840, 2008. 3",
1716
+ "[34] Johannes Lutz Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 6",
1717
+ "[35] Nathan Silberman, Derek Hoiem, Pushmeet Kohli, and Rob Fergus. Indoor segmentation and support inference from rgbd images. In European conference on computer vision, pages 746-760. Springer, 2012. 3",
1718
+ "[36] Matthew Tancik, Ben Mildenhall, Terrance Wang, Divi Schmidt, Pratul P. Srinivasan, Jonathan T. Barron, and Ren Ng. Learned initializations for optimizing coordinate-based neural representations. In CVPR, 2021. 3",
1719
+ "[37] Matthew Tancik, Pratul Srinivasan, Ben Mildenhall, Sara Fridovich-Keil, Nithin Raghavan, Utkarsh Singhal, Ravi Ramamoorthi, Jonathan Barron, and Ren Ng. Fourier features let networks learn high frequency functions in low dimensional domains. Advances in Neural Information Processing Systems, 33:7537-7547, 2020. 2",
1720
+ "[38] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 2",
1721
+ "[39] Dor Verbin, Peter Hedman, Ben Mildenhall, Todd Zickler, Jonathan T. Barron, and Pratul P. Srinivasan. Ref-NeRF: Structured view-dependent appearance for neural radiance fields. CVPR, 2022. 3",
1722
+ "[40] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. NeurIPS, 2021. 3",
1723
+ "[41] Qianqian Wang, Zhicheng Wang, Kyle Genova, Pratul Srinivasan, Howard Zhou, Jonathan T. Barron, Ricardo MartinBrualla, Noah Snavely, and Thomas Funkhouser. Ibrnet: Learning multi-view image-based rendering. In CVPR, 2021. 3,6",
1724
+ "[42] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 6",
1725
+ "[43] Yi Wei, Shaohui Liu, Yongming Rao, Wang Zhao, Jiwen Lu, and Jie Zhou. Nerfingmvs: Guided optimization of neural radiance fields for indoor multi-view stereo. In ICCV, 2021. 3"
1726
+ ],
1727
+ "bbox": [
1728
+ 78,
1729
+ 90,
1730
+ 470,
1731
+ 898
1732
+ ],
1733
+ "page_idx": 9
1734
+ },
1735
+ {
1736
+ "type": "list",
1737
+ "sub_type": "ref_text",
1738
+ "list_items": [
1739
+ "[44] Junyuan Xie, Ross Girshick, and Ali Farhadi. Deep3d: Fully automatic 2d-to-3d video conversion with deep convolutional neural networks. In European conference on computer vision, pages 842–857. Springer, 2016. 3",
1740
+ "[45] Lior Yariv, Jiatao Gu, Yoni Kasten, and Yaron Lipman. Volume rendering of neural implicit surfaces. In NeurIPS, 2021. 3",
1741
+ "[46] Wei Yin, Jianming Zhang, Oliver Wang, Simon Niklaus, Simon Chen, Yifan Liu, and Chunhua Shen. Towards accurate reconstruction of 3d scene shape from a single monocular image. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 3",
1742
+ "[47] Wei Yin, Jianming Zhang, Oliver Wang, Simon Niklaus, Long Mai, Simon Chen, and Chunhua Shen. Learning to recover 3d scene shape from a single image. In Proc. IEEE Conf. Comp. Vis. Patt. Recogn. (CVPR), 2021. 5, 6",
1743
+ "[48] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. pixelNeRF: Neural radiance fields from one or few images. In CVPR, 2021. 3",
1744
+ "[49] Zehao Yu, Songyou Peng, Michael Niemeyer, Torsten Sattler, and Andreas Geiger. Monosdf: Exploring monocular geometric cues for neural implicit surface reconstruction. arXiv preprint arXiv:2206.00665, 2022. 3, 7",
1745
+ "[50] Amir R Zamir, Alexander Sax, William Shen, Leonidas J Guibas, Jitendra Malik, and Silvio Savarese. Taskonomy: Disentangling task transfer learning. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3712-3722, 2018. 6, 8",
1746
+ "[51] Andy Zeng, Shuran Song, Matthias Nießner, Matthew Fisher, Jianxiong Xiao, and Thomas Funkhouser. 3dmatch: Learning local geometric descriptors from rgb-d reconstructions. In CVPR, 2017. 6",
1747
+ "[52] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 6",
1748
+ "[53] Zhengyou Zhang. Microsoft Kinect sensor and its effect. IEEE multimedia, 19(2):4-10, 2012. 3",
1749
+ "[54] Tinghui Zhou, Matthew Brown, Noah Snavely, and David G Lowe. Unsupervised learning of depth and ego-motion from video. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1851-1858, 2017. 3"
1750
+ ],
1751
+ "bbox": [
1752
+ 501,
1753
+ 92,
1754
+ 893,
1755
+ 683
1756
+ ],
1757
+ "page_idx": 9
1758
+ },
1759
+ {
1760
+ "type": "page_number",
1761
+ "text": "16527",
1762
+ "bbox": [
1763
+ 480,
1764
+ 944,
1765
+ 517,
1766
+ 955
1767
+ ],
1768
+ "page_idx": 9
1769
+ }
1770
+ ]