ageppert/single_photon_challenge_full_preprocessed

---
license: cc-by-4.0
task_categories:
- image-to-image
tags:
- single-photon
- denoising
- computational-imaging
- diffusion
pretty_name: Single Photon Challenge - Full Preprocessed
---

# Single Photon Challenge — Full Preprocessed Dataset

Preprocessed measurement/target PNG pairs derived from the
[Single Photon Challenge](https://singlephotonchallenge.com/) reconstruction dataset.

## Source

The raw dataset (~425GB training, ~42GB test) is hosted by the
[WISION Lab](https://wisionlab.com/) at UW-Madison. Photoncubes contain 1024
binary frames from a simulated single-photon camera, paired with ground-truth
RGB reconstructions.

- **Challenge website:** <https://singlephotonchallenge.com/>
- **Download page:** <https://singlephotonchallenge.com/download>
- **VisionSIM toolkit:** <https://visionsim.readthedocs.io/>

## Preprocessing pipeline

Each photoncube was preprocessed using the same approach as the
[challenge FAQ naive sum](https://singlephotonchallenge.com/faq):

1. **Average** the last 16 binary frames → detection probability in [0, 1]
2. **Invert SPC response** (`invert_response=True`, `factor=0.5`)
→ linear RGB flux via `flux = -log(1 - p) / factor`
3. **sRGB tonemap** (`tonemap=True`) → standard gamma curve
4. **Save** as uint8 PNG

Measurements and targets are stored as 800×800 RGB PNGs.

## Dataset statistics

| Split | Measurements | Targets | Paired |
|-------|-------------|---------|--------|
| train | 1850 | 1850 | yes |
| test | 185 | 0 | no (test set has no ground truth) |
| **total** | **2035** | **1850** | |

## Directory structure

```
single_photon_challenge_full_preprocessed/
metadata.json
train/
<scene>/<frame>_measurement.png
<scene>/<frame>_target.png
test/
<scene>/<frame>_measurement.png
```

## Usage

```python
from huggingface_hub import snapshot_download

# Download the full preprocessed dataset
root = snapshot_download(
repo_id="ageppert/single_photon_challenge_full_preprocessed",
repo_type="dataset",
)

# Or use with the diffusion training codebase:
# Set in config.py:
# PREPROCESSED_DATA_CONFIG["dataset_source"] = "hf"
# PREPROCESSED_DATA_CONFIG["dataset_hf_repo"] = "ageppert/single_photon_challenge_full_preprocessed"
```

## Preprocessing parameters

```json
{
"source": "Single Photon Challenge reconstruction dataset",
"source_url": "https://singlephotonchallenge.com/download",
"num_frames": 16,
"invert_response": true,
"invert_factor": 0.5,
"tonemap": true,
"split": "all",
"notes": "Measurements are preprocessed from raw photoncubes using: naive sum averaging, SPC response inversion, and sRGB tonemapping. Saved as uint8 PNGs. Targets are copied from original ground-truth PNGs."
}
```

## Citation

If you use this dataset, please cite the Single Photon Challenge:

```
@misc{singlephotonchallenge,
title={The Single Photon Challenge},
author={Jungerman, Sacha and Ingle, Atul and Nousias, Sotiris and Wei, Mian and White, Mel and Gupta, Mohit},
year={2025},
url={https://singlephotonchallenge.com/}
}
```

---
许可证：CC BY 4.0
任务类别：
- 图像到图像（image-to-image）
标签：
- 单光子（single-photon）
- 去噪（denoising）
- 计算成像（computational-imaging）
- 扩散模型（diffusion）
美观名称：单光子挑战赛——完整预处理数据集
---

# 单光子挑战赛——完整预处理数据集

本数据集为源自[单光子挑战赛](https://singlephotonchallenge.com/)重建数据集的预处理后测量值与目标值PNG配对样本。

## 数据集来源

原始数据集（训练集约425GB，测试集约42GB）由威斯康星大学麦迪逊分校的WISION实验室（WISION Lab）托管。每个光子立方体（Photoncube）包含来自模拟单光子相机（single-photon camera）的1024帧二进制帧，并与真实RGB重建结果配对。
- **挑战赛官网**：<https://singlephotonchallenge.com/>
- **下载页面**：<https://singlephotonchallenge.com/download>
- **VisionSIM工具包**：<https://visionsim.readthedocs.io/>

## 预处理流程

每个光子立方体均按照挑战赛常见问题解答中的"朴素求和法"（https://singlephotonchallenge.com/faq）进行预处理，步骤如下：
1. 对最后16帧二进制帧取平均，得到范围在[0, 1]的检测概率
2. 反转单光子相机响应（`invert_response=True`，`factor=0.5`），通过公式`flux = -log(1 - p) / factor`计算得到线性RGB光通量
3. 应用sRGB色调映射（`tonemap=True`），采用标准伽马曲线
4. 保存为uint8格式的PNG文件

测量值与目标值均存储为800×800分辨率的RGB PNG图像。

## 数据集统计信息

| 数据集划分 | 测量样本数 | 目标样本数 | 配对情况 |
|-------|-------------|---------|--------|
| 训练集 | 1850 | 1850 | 是 |
| 测试集 | 185 | 0 | 否（测试集无真实标签） |
| **总计** | **2035** | **1850** | |

## 目录结构

single_photon_challenge_full_preprocessed/
metadata.json
train/
<场景>/<帧编号>_measurement.png
<场景>/<帧编号>_target.png
test/
<场景>/<帧编号>_measurement.png

## 使用方法

python
from huggingface_hub import snapshot_download

# 下载完整预处理数据集
root = snapshot_download(
repo_id="ageppert/single_photon_challenge_full_preprocessed",
repo_type="dataset",
)

# 或配合扩散模型训练代码库使用：
# 在config.py中设置：
# PREPROCESSED_DATA_CONFIG["dataset_source"] = "hf"
# PREPROCESSED_DATA_CONFIG["dataset_hf_repo"] = "ageppert/single_photon_challenge_full_preprocessed"

## 预处理参数

json
{
"source": "单光子挑战赛重建数据集",
"source_url": "https://singlephotonchallenge.com/download",
"num_frames": 16,
"invert_response": true,
"invert_factor": 0.5,
"tonemap": true,
"split": "all",
"notes": "测量值由原始光子立方体经以下步骤预处理得到：朴素求和平均、单光子相机响应反转、sRGB色调映射。保存为uint8格式PNG文件。目标值直接复制自原始真实标签PNG文件。"
}

## 引用说明

若使用本数据集，请引用单光子挑战赛相关文献：

@misc{singlephotonchallenge,
title={The Single Photon Challenge},
author={Jungerman, Sacha and Ingle, Atul and Nousias, Sotiris and Wei, Mian and White, Mel and Gupta, Mohit},
year={2025},
url={https://singlephotonchallenge.com/}
}