Datasets:

shuinb
/

SP-TransientBench

@@ -17,23 +17,23 @@ size_categories:
 - 10K<n<100K
 ---
-# SP-TransientBench
-**SP-TransientBench (STB)** is a real-captured single-photon LiDAR benchmark for photon-starved 3D perception. It provides full per-pixel time-of-flight histograms, calibrated metadata, and task-specific supervision for depth estimation, multi-view 3D reconstruction, and 3D semantic segmentation.
 ![SP-TransientBench overview](fig/overview.png)
-## Highlights
-- Real-captured SPAD transient benchmark collected with an Adaps ADS6311 solid-state single-photon LiDAR.
-- Full waveform data for each view: `256 x 192 x 672` time-resolved photon histograms.
-- Three benchmark tracks: depth estimation, multi-view 3D reconstruction, and 3D semantic segmentation.
-- Task-specific calibration, reference geometry, poses, semantic labels, illumination metadata, and pile-up metadata.
-- Full release size: approximately `168.7 GB`.
-## Sensor Setup
-STB is captured with a flash single-photon LiDAR system based on Direct Time of Flight (DToF) and Time-Correlated Single Photon Counting (TCSPC). The transmitter uses a 940 nm VCSEL array, and the receiver records photon arrival timestamps with a SPAD array.
 | Item | Value |
 | --- | --- |
@@ -50,190 +50,142 @@ STB is captured with a flash single-photon LiDAR system based on Direct Time of
 An auxiliary Livox Avia LiDAR is mounted with the SPL device during collection. It is used for pose estimation, SPL-Livox calibration, and depth-reference generation where required by the benchmark track.
-## Dataset Overview
-The current release contains three complementary subsets:
-| Track | Scale | Main purpose |
-| --- | ---: | --- |
-| Depth Estimation | `10` samples | Single-view depth recovery from raw transient histograms |
-| Multi-view 3D Reconstruction | `10` scenes, `20-40` transient views per scene | Scene reconstruction and novel-view evaluation from calibrated SPL views |
-| 3D Semantic Segmentation | `27` sequences, `10,297` frames | Semantic understanding from SPAD-derived geometry |
-Every released view contains the same core sensing data:
-- Raw SPAD histogram: full waveform tensor with shape `256 x 192 x 672`.
-- Standard metadata: timestamps and sensor/capture metadata.
-- Calibrated SPL intrinsics where geometric back-projection is required.
-Additional metadata includes ambient illumination measurements for sensing-condition analysis and per-frame pile-up metadata for studying transient distortion.
-## Task-wise Data Composition
-### Task 1: Depth Estimation
-![Depth estimation qualitative results](fig/Depth%20Estimation.png)
-This track evaluates depth recovery directly from raw photon time-of-flight histograms.
-| Component | Included | Description |
-| --- | --- | --- |
-| Raw SPAD histograms | Yes | `256 x 192 x 672` full time-of-flight waveform per sample |
-| SPL intrinsics | Yes | Used to back-project predicted depth maps into 3D |
-| Livox reference data | Yes | Auxiliary LiDAR point clouds/depth references for evaluation |
-| SPL-Livox extrinsics | Yes | Used to align SPL predictions with the Livox reference frame |
-| Camera poses | No | Not required for the single-view depth track |
-| Semantic labels | No | Not used in this track |
-| Light intensity metadata | No | Not used as input or supervision |
-Evaluation follows point-cloud geometry metrics after back-projecting predicted depth maps:
-- Chamfer Distance (CD, meters).
-- Recall under temporal-bin tolerances of `1`, `3`, and `5` bins.
-#### Repository data format
 ```text
-DepthEstimate/
-  Histgram/
-    1.txt ... 10.txt
-  gt/
-    1.csv ... 10.csv
 ```
-- `DepthEstimate/Histgram/{id}.txt` stores the SPAD transient input for one depth-estimation sample. Each non-empty line is a whitespace-separated photon-count vector with `672` temporal bins. The valid pixel rows correspond to the flattened `256 x 192` SPL image grid.
-- `DepthEstimate/gt/{id}.csv` stores the corresponding Livox reference point cloud. The CSV header includes `Timestamp`, metric coordinates `X,Y,Z`, `Reflectivity`, and original Livox fields such as `Ori_x,Ori_y,Ori_z`.
-- Matching sample ids are used across `Histgram/` and `gt/`, e.g. `Histgram/1.txt` pairs with `gt/1.csv`.
-### Task 2: Multi-view 3D Reconstruction
-![Multi-view reconstruction qualitative results](fig/Multi-view%20reconstruction.png)
-This track evaluates reconstruction of scene geometry and novel-view rendering from multiple calibrated SPL views.
-| Component | Included | Description |
-| --- | --- | --- |
-| Raw SPAD histograms | Yes | Full transient waveform for each view |
-| SPL intrinsics | Yes | Used for geometric projection and view alignment |
-| Livox data | Yes | Supports pose estimation and geometric reference generation |
-| SPL-Livox extrinsics | Yes | Re-calibrated for reconstruction sequences |
-| Camera poses | Yes | Livox-SLAM poses transformed into the SPL frame |
-| Semantic labels | No | Not used in this track |
-| Light intensity metadata | No | Not used as input or supervision |
-The reconstruction subset contains `10` scenes, each with `20-40` transient views. For each scene, benchmark settings use `3`, `5`, or `10` input views for training and reserve the remaining views for novel-view rendering and geometry evaluation.
-Reported metrics cover three output domains:
-- Intensity rendering: SSIM and LPIPS.
-- Depth rendering: per-pixel `L1` error over valid pixels.
-- Histogram rendering: PSNR for methods that explicitly render transient histograms.
-#### Repository data format
-```text
-Reconstruction/
-  AI_floor2.zip
-  artbuilding_floor2.zip
-  c4floor2.zip
-  design_floor1.zip
-  library_floor2.zip
-  material_building.zip
-  parking.zip
-  physics_building2.zip
-  physics_building3.zip
-config/
-  config.yaml
-```
-- Each file under `Reconstruction/` is one compressed scene package for multi-view reconstruction.
-- After decompression, a scene package follows the same organization as the `AI_floor2` example:
-```text
-AI_floor2/
-  RawDataHistogramMap_frame_0_<timestamp>.txt
-  1.csv ... 26.csv
-  sp_pose_results.csv
-  sp_merged_map.ply
-  json/
-    three_views/
-      train.json
-      test.json
-    five_views/
-      train.json
-      test.json
-    ten_views/
-      train.json
-      test.json
-```
-- `RawDataHistogramMap_frame_0_<timestamp>.txt` stores one SPL view. Each valid row is a whitespace-separated `672`-bin photon-count histogram, and the valid rows correspond to the flattened `256 x 192` image grid (`49152` pixels).
-- `{view_id}.csv` stores the Livox point cloud associated with a captured view. The CSV columns include `Timestamp`, metric coordinates `X,Y,Z`, `Reflectivity`, and original Livox fields such as `Ori_x,Ori_y,Ori_z`.
-- `sp_pose_results.csv` maps each SPL histogram file to the matched Livox CSV/pose and records both `livox_pose_*` and `sp_pose_*` as flattened `4 x 4` transformation matrices. It also includes bookkeeping fields such as `used_in_final_merge`, `matched_livox_index`, and `match_mode`.
-- `sp_merged_map.ply` is the scene-level merged reference map generated from the registered captures. In the provided example it is an Open3D binary little-endian PLY with `x,y,z` and RGB vertex fields.
-- `json/{three_views,five_views,ten_views}/train.json` and `test.json` define the official sparse-view reconstruction splits. They follow a NeRF-style format with `camera_angle_x`, a list of `frames`, each frame's `file_path`, and a `transform_matrix`.
-- `config/config.yaml` provides common sensor settings and SPL calibration parameters used when parsing the released data, including `dt_ps: 750`, `image_sizes: [256, 192]`, and calibrated intrinsics/distortion parameters for `p1` and `p2`.
-- Benchmark experiments select `3`, `5`, or `10` views from a scene package as input views and reserve the remaining views for novel-view evaluation.
-### Task 3: 3D Semantic Segmentation
-![Semantic segmentation qualitative results](fig/Semantic%20Segmentation.png)
-This track evaluates semantic understanding from SPAD time-resolved measurements. Histograms are preprocessed, converted into single-photon point clouds through histogram-to-range projection, and then segmented with point-cloud backbones.
-| Component | Included | Description |
-| --- | --- | --- |
-| Raw SPAD histograms | Yes | Full transient waveform for each semantic frame |
-| SPL intrinsics | Yes | Device intrinsics for converting labeled bins to 3D points |
-| Livox data | No | Not used in this track |
-| SPL-Livox extrinsics | No | Not used in this track |
-| Camera poses | No | Not used in this track |
-| Semantic labels | Yes | Histogram-domain semantic annotations in `.npy` format |
-| Light intensity metadata | Yes | Ambient illumination recorded for each capture condition |
-The semantic subset contains `10,297` frames captured across `27` sequences and is split into:
 | Split | Samples |
 | --- | ---: |
 | Train | `8,297` |
 | Test | `2,000` |
-The benchmark uses `13` semantic classes. Evaluation reports Overall Accuracy (OA) and mean Intersection-over-Union (mIoU), averaged over three random seeds in the paper protocol.
-#### Repository data format
-```text
-Annotations/
-  p1/
-    Sequence1.zip ... Sequence20.zip
-  p2/
-    Sequence21.zip ... Sequence27.zip
-Histgram/
-  p1.zip
-  p2.zip
-  depth_maps/
-    P1_Sequence1_depth.png ... P1_Sequence20_depth.png
-    P2_Sequence21_depth.png ... P2_Sequence27_depth.png
-config/
-  config.yaml
-```
-- `Annotations/p1/` contains semantic annotation packages for sequences captured by SPL device `p1`.
-- `Annotations/p2/` contains semantic annotation packages for sequences captured by SPL device `p2`.
-- Inside each sequence zip, semantic labels are stored as `.npy` arrays named like `RawDataHistogramMap_frame_*_semantic.npy`.
-- Each semantic array has shape `(49152, 672)`, where `49152 = 256 x 192` flattened pixels and `672` is the number of temporal bins. Values are `uint8` semantic ids, with `0` used for unlabeled/background bins and `1-13` for semantic classes.
-- `Histgram/p1.zip` and `Histgram/p2.zip` store the corresponding raw histogram data for the semantic sequences.
-- `Histgram/depth_maps/` provides sequence-level depth-map PNG files aligned with the `P1_Sequence*` and `P2_Sequence*` naming convention.
-- `config/config.yaml` contains the calibrated intrinsics for both SPL devices. The semantic track uses both `p1` and `p2`, while the depth and reconstruction tracks use the relevant device calibration required by their data.
-## Semantic Annotation Format
 ![Annotation pipeline](fig/annotation_pipeline.png)
-STB uses histogram-domain semantic annotation to handle multi-return SPL measurements. Instead of assigning only one label to a pixel, the annotation is defined over temporal bins:
 ```text
 S in {0, ..., C}^{N x B}
@@ -241,7 +193,18 @@ N = H x W
 B = number of temporal bins
 ```
-The annotation pipeline identifies dominant peaks, assigns semantic labels to peak-support intervals, peels the labeled signal, and repeats the process to reveal weaker returns. This allows one pixel ray to contain multiple semantic entities at different ranges.
 ## Statistics
@@ -251,20 +214,22 @@ The annotation pipeline identifies dominant peaks, assigns semantic labels to pe
 STB records sensing-condition metadata such as ambient illumination and pile-up indicators. These metadata are intended for dataset analysis and robustness studies rather than model input.
-## Expected Release Contents
-The released package is expected to include:
-- Raw SPAD histogram files.
-- Semantic annotations in `.npy` format.
-- Calibrated SPL intrinsics and distortion coefficients.
-- SPL-Livox extrinsics for tracks that require auxiliary LiDAR alignment.
-- Camera poses for multi-view reconstruction.
-- Timestamps and capture metadata.
-- Ambient illumination metadata.
-- Per-frame pile-up metadata.
-- Benchmark evaluation code, data-loading scripts, and annotation tools.
-## License
-The dataset is released under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International license (CC BY-NC-SA 4.0).

 - 10K<n<100K
 ---
+# SP-TransientBench: A Real-Captured Single Photon Perception Benchmark
 ![SP-TransientBench overview](fig/overview.png)
+[Paper](https://arxiv.org/abs/2606.18952) | [Dataset download](https://huggingface.co/datasets/shuinb/SP-TransientBench)
+> Dong, H., Zhang, Z., Wen, Z., Qiang, Y., Deng, R., Dong, W., Jiang, Z., Li, X., Lu, R., Sun, S., Wang, W., Xia, Z., Zheng, H., Shi, G., & Ren, X. SP-TransientBench: A Real-Captured Single Photon Perception Benchmark. arXiv:2606.18952, 2026.
+**SP-TransientBench (STB)** is a real-captured benchmark for single-photon LiDAR (SPL) perception in photon-starved 3D scenes. It provides full per-pixel time-of-flight histograms, calibrated metadata, task-specific supervision, and official splits for depth estimation, multi-view 3D reconstruction, and 3D semantic segmentation.
+STB contains `10,297` views captured with a solid-state single-photon LiDAR at `256 x 192` spatial resolution. Each view stores a full transient waveform with `672` temporal bins, preserving photon sparsity, background noise, and multi-return structures that are often lost in depth-only releases.
+## Capture Setup
+![Single-photon LiDAR setup](fig/SPL.jpg)
+STB is captured with an Adaps ADS6311 Hawk solid-state SPL device operating under Direct Time-of-Flight (DToF) and Time-Correlated Single Photon Counting (TCSPC). The transmitter uses a `940 nm` VCSEL array, and the receiver records photon arrival timestamps with a SPAD array.
 | Item | Value |
 | --- | --- |
 An auxiliary Livox Avia LiDAR is mounted with the SPL device during collection. It is used for pose estimation, SPL-Livox calibration, and depth-reference generation where required by the benchmark track.
+## Benchmark Tracks
+![Benchmark examples](fig/comparison_figure.png)
+STB is organized around three complementary tasks:
+| Track | Scale | Purpose |
+| --- | ---: | --- |
+| Depth estimation | `10` samples | Recover single-view depth directly from raw transient histograms |
+| Multi-view 3D reconstruction | `10` scenes, `20-40` views per scene | Reconstruct geometry and render novel views from calibrated SPL captures |
+| 3D semantic segmentation | `27` sequences, `10,297` frames | Segment SPAD-derived 3D observations with histogram-domain semantic labels |
+Every track includes raw SPAD histograms. Depending on the task, the release also includes SPL intrinsics, SPL-Livox extrinsics, reference Livox point clouds, camera poses, ambient illumination metadata, pile-up metadata, and semantic annotations.
+## Layout
+```text
+SP-TransientBench/
+|-- README.md
+|-- fig/                    figures used in this dataset card
+|-- DepthEstimate/          depth-estimation histograms and Livox references
+|-- Reconstruction/         multi-view reconstruction scene archives
+|-- Annotations/            semantic annotation archives
+|-- Histgram/               semantic-track raw histograms and depth maps
+`-- config/                 common SPL calibration and parsing settings
+```
+The full release is approximately `168.7 GB`. Large files are stored in the Hugging Face dataset repository and should be downloaded from the dataset files page or with Git LFS/Xet-compatible tooling.
+```bash
+git lfs install
+git clone https://huggingface.co/datasets/shuinb/SP-TransientBench
+```
+## Data
+Depth estimation samples pair raw SPL histograms with Livox reference point clouds:
+```text
+DepthEstimate/
+|-- Histgram/
+|   `-- 1.txt ... 10.txt
+`-- gt/
+    `-- 1.csv ... 10.csv
+```
+`DepthEstimate/Histgram/{id}.txt` stores one flattened `256 x 192` SPL histogram grid, with `672` photon-count bins per valid row. `DepthEstimate/gt/{id}.csv` stores the corresponding Livox reference point cloud with metric `X,Y,Z` coordinates and capture metadata.
+Multi-view reconstruction scenes are released as compressed scene packages:
 ```text
+Reconstruction/
+|-- AI_floor2.zip
+|-- artbuilding_floor2.zip
+|-- c4floor2.zip
+|-- design_floor1.zip
+|-- library_floor2.zip
+|-- material_building.zip
+|-- parking.zip
+|-- physics_building2.zip
+`-- physics_building3.zip
 ```
+After decompression, each scene follows the same structure as `AI_floor2`:
+```text
+AI_floor2/
+|-- RawDataHistogramMap_frame_0_<timestamp>.txt
+|-- 1.csv ... 26.csv
+|-- sp_pose_results.csv
+|-- sp_merged_map.ply
+`-- json/
+    |-- three_views/
+    |   |-- train.json
+    |   `-- test.json
+    |-- five_views/
+    |   |-- train.json
+    |   `-- test.json
+    `-- ten_views/
+        |-- train.json
+        `-- test.json
+```
+The `RawDataHistogramMap_frame_0_<timestamp>.txt` files store SPL views, `{view_id}.csv` files store matched Livox point clouds, `sp_pose_results.csv` records Livox and SPL poses as flattened `4 x 4` transforms, and `sp_merged_map.ply` provides the registered scene-level reference map. The `json/` folders define official `3`, `5`, and `10` input-view reconstruction splits in a NeRF-style format.
+Semantic segmentation data are split by SPL device:
+```text
+Annotations/
+|-- p1/
+|   `-- Sequence1.zip ... Sequence20.zip
+`-- p2/
+    `-- Sequence21.zip ... Sequence27.zip
+Histgram/
+|-- p1.zip
+|-- p2.zip
+`-- depth_maps/
+    |-- P1_Sequence1_depth.png ... P1_Sequence20_depth.png
+    `-- P2_Sequence21_depth.png ... P2_Sequence27_depth.png
+```
+Inside each sequence archive, semantic labels are stored as `.npy` arrays named like `RawDataHistogramMap_frame_*_semantic.npy`. Each array has shape `(49152, 672)`, where `49152 = 256 x 192`, and stores `uint8` semantic ids with `0` for unlabeled/background bins and `1-13` for semantic classes.
+## Task Details
+### Depth Estimation
+![Depth estimation qualitative results](fig/Depth%20Estimation.png)
+This track evaluates depth recovery directly from raw photon time-of-flight histograms. Predictions are back-projected to 3D with calibrated SPL intrinsics and compared against Livox references with Chamfer Distance (CD, meters) and Recall under `1`, `3`, and `5` temporal-bin tolerances.
+### Multi-view 3D Reconstruction
+![Multi-view reconstruction qualitative results](fig/Multi-view%20reconstruction.png)
+This track evaluates scene reconstruction and novel-view rendering from multiple calibrated SPL views. Each scene provides sparse-view settings with `3`, `5`, or `10` input views for training and reserves the remaining views for evaluation. Reported metrics cover intensity rendering (SSIM, LPIPS), depth rendering (`L1` error), and histogram rendering (PSNR).
+### 3D Semantic Segmentation
+![Semantic segmentation qualitative results](fig/Semantic%20Segmentation.png)
+This track evaluates semantic understanding from SPAD time-resolved measurements. Histograms are preprocessed, converted into single-photon point clouds through histogram-to-range projection, and segmented with point-cloud backbones.
 | Split | Samples |
 | --- | ---: |
 | Train | `8,297` |
 | Test | `2,000` |
+The semantic track uses `13` foreground classes. Evaluation reports Overall Accuracy (OA) and mean Intersection-over-Union (mIoU), averaged over three random seeds in the paper protocol.
+## Semantic Labels
 ![Annotation pipeline](fig/annotation_pipeline.png)
+STB uses histogram-domain semantic annotation to handle multi-return SPL measurements. Instead of assigning a single label to each pixel, annotations are defined over temporal bins:
 ```text
 S in {0, ..., C}^{N x B}
 B = number of temporal bins
 ```
+The annotation pipeline identifies dominant peaks, assigns semantic labels to peak-support intervals, peels the labeled signal, and repeats the process to reveal weaker returns. This lets a single pixel ray contain multiple semantic entities at different ranges.
+## Contents
+| File or folder | Purpose |
+| --- | --- |
+| `DepthEstimate/` | Single-view depth-estimation samples with raw histograms and Livox point-cloud references |
+| `Reconstruction/` | Scene packages for sparse-view SPL reconstruction and novel-view evaluation |
+| `Annotations/` | Histogram-domain semantic label packages for `27` sequences |
+| `Histgram/` | Raw histogram packages and sequence-level depth maps for the semantic track |
+| `config/config.yaml` | SPL intrinsics, distortion coefficients, temporal bin width, image size, and parsing settings |
+| `fig/` | Dataset-card figures, qualitative examples, annotation diagrams, and statistics |
 ## Statistics
 STB records sensing-condition metadata such as ambient illumination and pile-up indicators. These metadata are intended for dataset analysis and robustness studies rather than model input.
+## License
+The dataset is released under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International license (CC BY-NC-SA 4.0).
+## Citation
+If you use SP-TransientBench, please cite:
+```bibtex
+@misc{dong2026sptransientbench,
+  title         = {SP-TransientBench: A Real-Captured Single Photon Perception Benchmark},
+  author        = {Dong, Hongzhou and Zhang, Zili and Wen, Ziting and Qiang, Yiheng and Deng, Runrong and Dong, Wenle and Jiang, Ziwen and Li, Xinyang and Lu, Rui and Sun, Shuoyao and Wang, Wenyu and Xia, Ziyi and Zheng, Haitao and Shi, Guodong and Ren, Xiaoqiang},
+  year          = {2026},
+  eprint        = {2606.18952},
+  archivePrefix = {arXiv},
+  primaryClass  = {cs.CV},
+  doi           = {10.48550/arXiv.2606.18952}
+}
+```