---
license: other
pretty_name: "D-RE10K: Dynamic Real-Estate 10K"
language:
  - en
tags:
  - video
  - computer-vision
  - novel-view-synthesis
  - 3d-reconstruction
  - dynamic-scenes
  - research
  - gated

# ---- Gating UI text ----
extra_gated_heading: "Request access to D-RE10K (Research-Only)"
extra_gated_description: >
  D-RE10K contains processed real-estate walkthrough video clips derived from third-party sources.
  Access is granted for non-commercial research only.
  We do not grant rights to any underlying third-party content. You are responsible for ensuring you have the necessary rights to use the media.
  By requesting access, you agree to use this dataset for non-commercial research purposes only.
extra_gated_button_content: "Agree & Request Access"
---

# D-RE10K: Dynamic Real-Estate 10K Dataset

## Overview

This dataset contains the **DRE10K training** split (15,467 clips, 147,422 frames) and the **DRE10K mask** test split (76 clips, 1,541 frames), released on Hugging Face for research on self-supervised large view synthesis in dynamic environments. The data is collected from real-estate walkthrough videos and curated specifically for training and evaluating novel view synthesis models in scenes with dynamic objects.

Our dataset builds on the Real-Estate 10K collection and extends it with per-frame binary masks, masked videos, COLMAP reconstructions, and DPVO camera trajectories for the test split. Each clip is accompanied by JSON metadata containing camera intrinsics and world-to-camera poses, making it a versatile resource for tasks such as novel view synthesis, camera pose estimation, and dynamic scene understanding.

For more details, please refer to our paper [WildRayZer: Self-supervised Large View Synthesis in Dynamic Environments](https://arxiv.org/abs/2601.10716).

| Split | Clips | Extracted Frames | Metadata (JSON) | Binary Masks | Masked Videos | COLMAP | DPVO |
|-------|------:|------------------:|------------------:|--------------:|--------------:|-------:|-----:|
| Train | 15,467 | 147,422 | 15,467 | — | — | — | — |
| Test  | 76     | 1,541             | 76               | 1,540         | 76            | 76 scenes | 76 |

## Key Features

- **Size**: 15,467 training clips with 147,422 extracted frames; 76 test clips with 1,541 frames.
- **Representation**: Extracted PNG frames from real-estate walkthrough videos, with per-clip JSON metadata (camera intrinsics, world-to-camera poses, frame paths).
- **Train split** includes:
  - Video clips (`.mp4`)
  - Extracted frames (`.png`)
  - Per-clip JSON metadata
- **Test split** additionally includes:
  - Per-frame binary masks (`.png`) for dynamic objects
  - Masked videos with dynamic objects removed (`.mp4`)
  - COLMAP reconstructions (sparse models in binary & text, masks, database)
  - DPVO estimated camera trajectories (`.txt`)

## Dataset Format

The dataset is provided in a format ready for view-synthesis and 3D-reconstruction research:

- **Videos**: Stored as `.mp4` files under `videos/`.
- **Frames**: Stored as `.png` files under `images/<clip_id>/`.
- **Metadata**: Stored as `.json` files under `metadata/`. Each JSON file contains camera intrinsics (`fxfycxcy`), 4×4 world-to-camera matrices (`w2c`), and frame paths.
- **Binary Masks** (test only): Stored as `.png` files under `binary_masks/<clip_id>/`.
- **COLMAP** (test only): Full sparse reconstructions under `colmap/<clip_id>/` (includes `sparse/`, `masks/`, `database.db`).
- **DPVO** (test only): Camera trajectory files under `dpvo/<clip_id>.txt`.

The dataset is distributed as multi-part zip archives. After downloading, unzip them as follows:

```bash
# Unzip training data (8 parts)
mkdir -p train
for f in train_zip/train_*.zip; do
  unzip -o "$f" -d ./train
done

# Unzip test data (3 parts)
mkdir -p test
for f in test_zip/test_*.zip; do
  unzip -o "$f" -d ./test
done
```

After unzipping, you should see the `train/` and `test/` directories with the structure described above.

## License

This dataset is released for **non-commercial research use only**. 
The video clips and frames are derived from third-party sources. We do not hold the copyright to the underlying audio-visual content.
Users must agree to the terms outlined in the [LICENSE](LICENSE.md) file, which include:
- Use for non-commercial research only.
- No redistribution of the dataset.
- Acknowledgment of third-party rights.

## Takedown Policy

The video clips in this dataset are derived from third-party sources. If any clips need to be taken down (e.g., due to privacy concerns or copyright requests), we will promptly delete them from this dataset. Please contact us at `xuweic@virginia.edu` for such requests.

## Citation

If you find this dataset useful in your research, please cite our work:

```bibtex
@article{chen2026wildrayzerselfsupervisedlargeview,
      title={WildRayZer: Self-supervised Large View Synthesis in Dynamic Environments}, 
      author={Xuweiyi Chen and Wentao Zhou and Zezhou Cheng},
      year={2026},
      eprint={2601.10716},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2601.10716}, 
}
```