README.md

---
license: cc-by-4.0
---

# BeyondMasks

Dataset repository for **BeyondMasks: Evaluating Causal and Physical Consistency in Video Object Removal**, accepted to **ECCV 2026**.

Authors: [Yiğit Ekin](https://yigitekin.github.io/), [Enes Sanli](https://openreview.net/profile?id=~Enes_Sanli1), [Aykut Erdem](https://aykuterdem.github.io/), [Erkut Erdem](https://web.cs.hacettepe.edu.tr/~erkut/), [Aysegul Dundar](https://www.cs.bilkent.edu.tr/~adundar/)

BeyondMasks evaluates whether video object removal methods remove not only the target object, but also its causal and physical aftereffects such as shadows, reflections, and light effects.

## Repository Contents

```text
BeyondMasks/
├── masks/                # Binary mask videos indicating the object to remove
├── object_not_present/   # Ground-truth videos where the object and aftereffects are absent
├── objects_added/        # Input videos where the object and aftereffects are present
├── data_example.json     # Example metadata format
└── eval.py               # Proposed evaluation metric code
```

The three video folders are aligned by sample id. For example, `masks/1.mp4`, `object_not_present/1.mp4`, and `objects_added/1.mp4` correspond to the same sample.

## Dataset Fields

- `objects_added/`: input videos containing the target object and its physical aftereffects.
- `object_not_present/`: reference videos where the target object and corresponding aftereffects are not present.
- `masks/`: object masks for the region to remove.
- `data_example.json`: example annotation entries with the sample id, foreground object, and aftereffect type.

## Evaluation

`eval.py` contains the proposed LLM-based evaluation metric. It scores object removal quality and aftereffect removal quality using the input video, ground-truth video, and a method output video. Adjust the folder paths according to your filesystem

The script expects an evaluation root with this structure:

```text
eval_root/
├── fg/       # Input videos with object and aftereffects present
├── bg/       # Ground-truth videos with object and aftereffects absent
├── result/   # Method output videos to evaluate
└── data.json # Metadata annotations
```

Run:

```bash
export GEMINI_API_KEY=your_api_key
python eval.py --root /path/to/eval_root
```

Results are written to `core_evaluation_results.json`.

## License

This dataset is released under the CC BY 4.0 license.

## Citation

Citation information will be added when the paper metadata is available.