README.md

---
license: cc-by-nc-sa-4.0
task_categories:
- depth-estimation
- image-to-3d
- image-segmentation
tags:
- single-photon-lidar
- spad
- transient-imaging
- time-of-flight
- 3d-vision
- multi-view-reconstruction
- semantic-segmentation
pretty_name: SP-TransientBench
size_categories:
- 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

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 |
| --- | --- |
| SPL device | Adaps ADS6311 Hawk |
| Acquisition mode | Solid-state flash SPL |
| Raw SPAD resolution | `768 x 576` |
| Released output resolution | `256 x 192` after `3 x 3` on-chip binning |
| Histogram bins | `672` |
| Bin width | `750 ps` |
| Field of view | `128 deg x 96 deg` |
| Frame rate | `10-20 Hz` |
| Detection range | Up to `30 m` |
| Range accuracy | `< 5 cm` |

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

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
|-- Histogram/              semantic-track raw histograms and depth maps
|-- codes/                  annotation, point-cloud conversion, simulation, and reconstruction tools
`-- config/                 SPL calibration, extrinsics, parsing settings, and IRF data
```

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/
|-- Histogram/
|   `-- 1.txt ... 10.txt
`-- gt/
    `-- 1.csv ... 10.csv
```

`DepthEstimate/Histogram/{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

Histogram/
|-- 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.

## Code

The `codes/` folder contains utility code released with STB:

- `codes/annotation tool/labelLidarwave.py`: a single-photon LiDAR waveform annotation tool.
- `codes/pointcloud_extract/raw2pc.py`: converts raw histogram `.txt` files into point clouds using the maximum-bin histogram method.
- `codes/pointcloud_extract/ann2pc.py`: converts annotated `.npy` semantic data into corresponding semantic point clouds.
- `codes/simulator/`: simulation tools used for STB experiments.
- `codes/reconstruction/`: STB reconstruction code based on the TransientNeRF method.

## 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}
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 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 |
| `Histogram/` | Raw histogram packages and sequence-level depth maps for the semantic track |
| `codes/` | Annotation, point-cloud conversion, simulation, and TransientNeRF-based reconstruction code |
| `config/config.yaml` | SPL intrinsics, distortion coefficients, temporal bin width, image size, and parsing settings |
| `config/IRF_global.csv` | Global instrument response function (IRF) measurements for transient calibration and analysis |
| `config/final_extrinsic_sp_to_livox.txt` | SPL-to-Livox extrinsic calibration used for cross-sensor alignment |
| `config/offset.txt` | Additional extrinsic offset parameters used for SPL/Livox alignment |
| `fig/` | Dataset-card figures, qualitative examples, annotation diagrams, and statistics |

## Statistics

![Semantic category distribution](fig/semantic_category_distribution.png)

![Light intensity distribution](fig/Light%20Intensity%20Distribution.png)

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}
}
```