README.md

---
license: cc-by-nc-4.0
task_categories:
- video-to-video
language:
- en
tags:
- computer-vision
- video-restoration
- video-dereflection
- reflection-removal
- benchmark
- 4k-video
size_categories:
- n<1K
pretty_name: S2R-Bench
---

# S2R-Bench

S2R-Bench is a real-world video reflection removal benchmark introduced in **[From Synthesis to Removal:
Physics-Grounded Reflection Simulation and Diffusion-Based Video Dereflection](https://arxiv.org/abs/2608.11562)**. It evaluates whether a video dereflection method can recover a temporally consistent transmission layer from videos recorded through reflective glass.

The benchmark contains two complementary test sets:

- **S2R-Ref**: 60 real captured video pairs constructed from the [DRR dataset](https://github.com/Abuuu122/Dereflection-Any-Image), each containing a reflection-contaminated input and its reflection-free transmission-layer ground truth.
- **S2R-Real**: 50 in-the-wild reflection-contaminated videos without ground truth, intended for qualitative and no-reference evaluation.

> **Attention:** S2R-Ref is constructed from the **training split** of the DRR dataset.
> Therefore, it may not provide an accurate measure of generalization for methods trained on DRR,
> as those methods may have seen the underlying training data during development.
> Results on S2R-Ref should be interpreted with this potential data overlap in mind.


## Dataset Details

Video reflection removal separates a video recorded through glass into a desired transmission layer and an unwanted reflection layer. In addition to the challenges of single-image reflection removal, a video method must preserve temporal consistency under camera motion, scene motion, spatially varying reflections, blur, ghosting, and different reflection strengths.

| Subset | Input videos | Ground truth | Primary use |
|---|---:|---:|---|
| S2R-Ref | 60 | 60 paired transmission videos | Full-reference quantitative and qualitative evaluation |
| S2R-Real | 50 | Not available | Real-world qualitative and no-reference evaluation |

## Qualitative Examples

### S2R-Ref

<p align="center">
  <img src="https://codingwzp.github.io/VideoDereflection_S2R/static/images/fig7_s2r_ref.jpg" alt="Qualitative comparison on the S2R-Ref benchmark" width="100%">
</p>

<p align="center"><em>Qualitative comparison on the S2R-Ref benchmark. Image from the project homepage.</em></p>

### S2R-Real

<p align="center">
  <img src="https://codingwzp.github.io/VideoDereflection_S2R/static/images/fig8_s2r_real.jpg" alt="Qualitative comparison on the S2R-Real benchmark" width="100%">
</p>

<p align="center"><em>Qualitative comparison on the S2R-Real benchmark. Image from the project homepage.</em></p>

- **Homepage:** https://codingwzp.github.io/VideoDereflection_S2R/
- **Paper:** https://arxiv.org/abs/2608.11562
- **Task:** Video reflection removal / video dereflection
- **License:** Creative Commons Attribution-NonCommercial 4.0 International (CC BY-NC 4.0)

## Dataset Structure

```text
S2R-Bench/
|-- S2R-Ref/
|   |-- blended/                 # Reflection-contaminated input videos
|   `-- transmission_layer/      # Paired reflection-free ground truth
`-- S2R-Real/
    `-- blended/                 # In-the-wild input videos without ground truth
```

For **S2R-Ref**, an input video and its ground-truth transmission video share the same filename:

```text
S2R-Ref/blended/DSC_0006.mp4
S2R-Ref/transmission_layer/DSC_0006.mp4
```

Use the filename stem as the sample identifier when matching predictions to ground truth. Each file in `S2R-Real/blended/` is an independent test sequence.


## Citation

If you use S2R-Bench, please cite the S2R paper:

```bibtex
@article{wang2026s2r,
  title   = {From Synthesis to Removal: Physics-Grounded Reflection Simulation and Diffusion-Based Video Dereflection},
  author  = {Wang, Zepeng and Hu, Jiagao and Li, Fuhao and Chen, Yuxuan and Wang, Fei and Zhou, Daiguo},
  journal = {arXiv preprint arXiv:2608.11562},
  year    = {2026}
}
```