README.md

---
language:
- en
pretty_name: "Light-Stage OLAT Subsurface-Scattering Dataset"
tags:
- computer-vision
- 3d-reconstruction
- subsurface-scattering
- gaussian-splatting
- inverse-rendering
- photometric-stereo
- light-stage
- olat
- multi-view
- multi-light
- image
license: "other"
task_categories:
- image-to-3d
- other
size_categories:
- 10K<n<100K
dataset_info:
  features:
  - name: image
    dtype: image
  - name: camera_pose
    dtype: json
  - name: light_pose
    dtype: json
  - name: mask
    dtype: image
  splits:
  - name: train
    num_bytes: 30000000000
    num_examples: 30000
  - name: test
    num_bytes: 7000000000
    num_examples: 7000
  download_size: 30000000000
  dataset_size: 37000000000
configs:
- config_name: real_world
  data_files:
  - split: train
    path: real_world/*/transforms_train.json
  - split: test
    path: real_world/*/transforms_test.json
- config_name: synthetic
  data_files:
  - split: train
    path: synthetic/*_full/transforms_train.json
  - split: test
    path: synthetic/*_full/transforms_test.json
- config_name: synthetic_small
  data_files:
  - split: train
    path: synthetic/*_small/transforms_train.json
  - split: test
    path: synthetic/*_small/transforms_test.json
  - split: eval
    path: synthetic/*_small/transforms_eval.json
---

# 🕯️ Light-Stage OLAT Subsurface-Scattering Dataset

*Companion data for the paper **"Subsurface Scattering for 3D Gaussian Splatting"***

> **This README documents *only the dataset*.**  
> A separate repo covers the training / rendering **code**: <https://github.com/cgtuebingen/SSS-GS>

<p align="center">
  <img src="other/dataset.png" width="80%" alt="Dataset overview"/>
</p>

## Overview

Subsurface scattering (SSS) gives translucent materials (wax, soap, jade, skin) their distinctive soft glow. Our paper introduces **SSS-GS**, the first 3D Gaussian-Splatting framework that *jointly* reconstructs shape, BRDF and volumetric SSS while running at real-time framerates. Training such a model requires dense **multi-view ⇄ multi-light OLAT** data.

This dataset delivers exactly that:

* **25 objects** – 20 captured on a physical light-stage, 5 rendered in a synthetic stage
* **> 37k images** (≈ 1 TB raw / ≈ 30 GB processed) with **known camera & light poses**
* Ready-to-use JSON transform files compatible with NeRF & 3D GS toolchains
* Processed to 800 px images + masks; **raw 16 MP capture** available on request

### Applications

* Research on SSS, inverse-rendering, radiance-field relighting, differentiable shading
* Benchmarking OLAT pipelines or light-stage calibration
* Teaching datasets for photometric 3D reconstruction

## Quick Start

```bash
# Download and extract one real-world object
curl -L https://…/real_world/candle.tar | tar -x
```

## Directory Layout
```
dataset_root/
├── real_world/          # Captured objects (processed, ready to train)
│   └── <object>.tar     # Each tar = one object (≈ 4–8 GB)
└── synthetic/           # Procedurally rendered objects
    ├── <object>_full/   # full-resolution (800 px)
    └── <object>_small/  # 256 px "quick-train" version
```

### Inside a **real-world** tar
```
<object>/
├── resized/                 # θ_φ_board_i.png  (≈ 800 × 650 px)
├── transforms_train.json    # (train-set only) ⇄  camera / light metadata
├── transforms_test.json     # (test-set only) ⇄  camera / light metadata
├── light_positions.json     # all θ_φ_board_i → (x,y,z)
├── exclude_list.json        # bad views (lens flare, matting error, …)
└── cam_lights_aligned.png   # sanity-check visualisation
```
*Raw capture* Full-resolution, unprocessed RGB-bayer images (~ 1 TB per object) are kept offline—contact us to arrange transfer.

### Inside a **synthetic** object folder
```
<object>_full/
├── <object>.blend         # Blender scene with 112 HDR stage lights
├── train/                 # r_<cam>_l_<light>.png (= 800 × 800 px)
├── test/                  # r_<cam>_l_<light>.png (= 800 × 800 px)
├── eval/                  # only in "_small" subsets
├── transforms_train.json  # (train-set only) ⇄  camera / light metadata
└── transforms_test.json   # (test-set only) ⇄  camera / light metadata
```
The *small* variant differs only in image resolution & optional `eval/`.

## Data Collection

### Real-World Subset

**Capture Setup:**
- **Stage**: 4 m diameter light-stage with 167 individually addressable LEDs
- **Camera**: FLIR Oryx 12 MP with 35 mm F-mount, motorized turntable & vertical rail
- **Processing**: COLMAP SfM, automatic masking (SAM + ViTMatte), resize → PNG

| Objects | Avg. Views | Lights/View | Resolution | Masks |
|---------|------------|-------------|------------|-------|
| 20      | 158        | 167         | 800×650 px | α-mattes |

<p align="center">
  <img src="other/preprocessing.png" width="60%" alt="Preprocessing pipeline"/>
</p>

### Synthetic Subset

**Rendering Setup:**
- **Models**: Stanford 3D Scans and BlenderKit
- **Renderer**: Blender Cycles with spectral SSS (Principled BSDF)
- **Lights**: 112 positions (7 rings × 16), 200 test cameras on NeRF spiral path

| Variant | Images | Views × Lights | Resolution | Notes |
|---------|--------|----------------|------------|-------|
| _full   | 11,200 | 100 × 112      | 800²       | Filmic tonemapping |
| _small  | 1,500  | 15 × 100       | 256²       | Quick prototyping |

## File & Naming Conventions
* **Real images** `theta_<θ>_phi_<φ>_board_<id>.png`  
  *θ, φ* in degrees; *board* 0-195 indexes the LED PCBs.  
* **Synthetic images** `r_<camera>_l_<light>.png`  
* **JSON schema**  
  ```jsonc
  {
    "camera_angle_x": 0.3558,
    "frames": [{
      "file_paths": ["resized/theta_10.0_phi_0.0_board_1", …],
      "light_positions": [[x,y,z], …],   // metres, stage origin
      "transform_matrix": [[...], ...],  // 4×4 extrinsic
      "width": 800, "height": 650, "cx": 400.0, "cy": 324.5
    }]
  }
  ```
  For synthetic files: identical structure, naming `r_<cam>_l_<light>`.

## Licensing & Third-Party Assets
| Asset | Source | License / Note |
|-------|--------|----------------|
| Synthetic models | [Stanford 3-D Scans](https://graphics.stanford.edu/data/3Dscanrep/) | Varies (non-commercial / research) |
|                  | [BlenderKit](https://www.blenderkit.com/) | CC-0, CC-BY or Royalty-Free (check per-asset page) |
| HDR env-maps     | [Poly Haven](https://polyhaven.com/) | CC-0 |
| Code             | MIT (see repo) |

The dataset is released **for non-commercial research and educational use**.  
If you plan to redistribute or use individual synthetic assets commercially, verify the upstream license first.

## Citation
If you use this dataset, please cite the paper:

```bibtex
@inproceeding{sss_gs,
 author = {Dihlmann, Jan-Niklas and Majumdar, Arjun and Engelhardt, Andreas and Braun, Raphael and Lensch, Hendrik P.A.},
 booktitle = {Advances in Neural Information Processing Systems},
 editor = {A. Globerson and L. Mackey and D. Belgrave and A. Fan and U. Paquet and J. Tomczak and C. Zhang},
 pages = {121765--121789},
 publisher = {Curran Associates, Inc.},
 title = {Subsurface Scattering for Gaussian Splatting},
 url = {https://proceedings.neurips.cc/paper_files/paper/2024/file/dc72529d604962a86b7730806b6113fa-Paper-Conference.pdf},
 volume = {37},
 year = {2024}
}

```

## Contact & Acknowledgements
Questions, raw-capture requests, or pull-requests?  
📧 `jan-niklas.dihlmann (at) uni-tuebingen.de`

This work was funded by DFG (EXC 2064/1, SFB 1233) and the Tübingen AI Center.