---
pretty_name: Real4D
license: other
task_categories:
  - depth-estimation
tags:
  - 4d
  - dynamic-scene
  - multi-view
  - novel-view-synthesis
  - camera-pose
  - gaussian-splatting
size_categories:
  - 100K<n<1M
---

# Real4D

Real4D is a real-domain dynamic-scene dataset generated from synchronized
multi-camera captures through geometry-regularized 4D Gaussian Splatting and
spacetime camera interpolation. It contains synchronized RGB, dense rendered
depth, and analytically prescribed camera metadata.

## Dataset summary

| Source subset | Scenes | Reference cameras | Trajectories | Frames per trajectory | Total frames | Resolution |
|---|---:|---:|---:|---:|---:|---:|
| N3DV | 6 | 18--21 | 2,408 | 300 | 722,400 | 1352 x 1014 |
| MeetRoom | 4 | 13 | 676 | 300 | 202,800 | 1280 x 720 |
| **Total** | **10** | -- | **3,084** | **300** | **925,200** | -- |

Each trajectory is named `cam{a}_{b}`. Camera pose and scene time advance
together from reference camera `a` to reference camera `b`. Identity pairs such
as `cam00_00` keep the viewpoint fixed while time advances.

## Released modalities

- RGB: 8-bit JPEG images.
- Depth: opacity-normalized rendered depth in 16-bit grayscale PNG. Values are
  millimeters; zero denotes an invalid or negligible-opacity pixel.
- Camera metadata: one JSON manifest per trajectory, with one entry per frame.
- Previews: colorized depth images plus 30 fps RGB and depth MP4 files.

Depth is reconstruction-rendered supervision, not sensor ground truth. Camera
poses are exact with respect to the prescribed virtual trajectories.

## Organization

```text
Real4D/
├── dynerf/
│   └── <scene>/
│       ├── images/<trajectory>/<trajectory>_<frame>.jpg
│       ├── depths/<trajectory>/<trajectory>_<frame>.jpg.geometric.png
│       ├── depth_vis/<trajectory>/<trajectory>_<frame>.jpg
│       ├── camera_params/<trajectory>.json
│       └── video/<trajectory>_{rgb,depth}.mp4
├── meetroom/
│   └── <scene>/
│       └── ...
├── dataset_manifest.json
├── docs/DATA_FORMAT.md
└── scripts/
    ├── load_frame.py
    ├── validate_dataset.py
    └── upload_to_hf.sh
```

Frame files are numbered from `000001` to `000300`, while JSON list indices are
zero-based. The JSON field `image_name` identifies the source reference camera;
it is not the generated frame filename. Match frames by trajectory and list
index as shown in `scripts/load_frame.py`.

The local generation directories `_scene_cache`, `plane_depths`,
`plane_depth_vis`, and `*_plane_depth.mp4` are intermediate artifacts and are
not part of the public Real4D release.

See [docs/DATA_FORMAT.md](docs/DATA_FORMAT.md) for field definitions and loading
examples. Run the integrity check before use:

```bash
python scripts/validate_dataset.py .
python scripts/validate_dataset.py . --full
```

## Source data and license

Real4D contains derived renderings of real scenes from two source datasets.
The N3DV source dataset is distributed under CC BY-NC 4.0. The official
StreamRF materials describe MeetRoom as released for research purposes but do
not state an explicit dataset license. Consequently, this dataset card uses
`license: other`; the applicable source-dataset terms and required citations
must be preserved. Public redistribution of the MeetRoom-derived subset should
be confirmed with the original dataset owners before upload.

This notice does not replace the source-dataset licenses or grant additional
rights to the underlying captures.

## Citation

The Real4D paper citation will be added after publication. When using a source
subset, also cite its original dataset paper:

```bibtex
@inproceedings{li2022neural3dvideo,
  title     = {Neural 3D Video Synthesis from Multi-view Video},
  author    = {Li, Tianye and Slavcheva, Miroslava and Zollhoefer, Michael and
               Green, Simon and Lassner, Christoph and Kim, Changil and
               Schmidt, Tanner and Lovegrove, Steven and Goesele, Michael and
               Newcombe, Richard and Lv, Zhaoyang},
  booktitle = {CVPR},
  year      = {2022}
}

@inproceedings{li2022streaming,
  title     = {Streaming Radiance Fields for 3D Video Synthesis},
  author    = {Li, Lingzhi and Shen, Zhen and Wang, Zhongshu and Shen, Li and
               Tan, Ping},
  booktitle = {NeurIPS},
  year      = {2022}
}
```