README.md

---
pretty_name: "OmniLife360"
license: cc-by-nc-4.0
tags:
  - 3d
  - video
  - computer-vision
  - benchmark
  - panoramic
  - 360-panorama
  - reconstruction
  - gaussian-splatting
---

# OmniLife360

OmniLife360 is a benchmark for 3D reconstruction from in-the-wild 360-degree captures. This repository provides source links, benchmark metadata, COLMAP/SfM reconstruction results, and helper scripts for frame preparation and panorama SfM reproduction.

The benchmark contains 2,407 panoramic sequences, totaling about 66 hours, with 1,937 training sequences and 470 testing sequences across 64 scene categories and 31 action classes. The metadata includes 2,177 publicly linked web-video sequences and 230 author-captured sequences that are not publicly linked.

## Repository Contents

```text
metadata/train_test_split_with_links.json
sfm/omnilife360_colmap_sfm_results.tar.zst
scripts/prepare_frames_from_metadata.py
scripts/run_panorama_sfm_from_frames.sh
```

`metadata/train_test_split_with_links.json` is the main benchmark index. It stores the official split, source links where available, temporal segments, scene/action labels, captions, and basic media metadata.

`sfm/omnilife360_colmap_sfm_results.tar.zst` contains COLMAP/SfM reconstruction outputs organized by the same split/scene/sequence structure. The archive includes reconstruction files under `colmap_nonoverlap/sparse/`.

The `scripts/` directory contains helpers for preparing frames from public source links and reproducing the panorama SfM directory layout.

## Metadata Fields

Each sequence is identified by the full path `<split>/<scene>/<clip_id>`. A small number of `clip_id` values appear in more than one scene or split, so the full path should be used as the unique key.

Important fields include:

- `source_video_id`: normalized source identifier.
- `source_url`: public source URL for web videos; `null` for author-captured sequences.
- `source_url_status`: `inferred_from_video_id` or `author_team_capture_not_publicly_linked`.
- `start_sec`, `end_sec`, `duration`: temporal segment definition, where `duration = end_sec - start_sec`.
- `scene`, `action`, `area`: benchmark labels.
- `resolution`, `frame_rate`: source media properties.
- `caption`: sequence-level caption.

## Frame Preparation

For public web videos, users can download the source video from `source_url` and extract the benchmark segment defined by `start_sec` and `end_sec`. Our preprocessing samples panoramic frames at 1 FPS, resizes them to `1920x960`, and stores them as:

```text
<data_root>/<split>/<scene>/<clip_id>/images/%04d.jpg
```

To prepare frames for all publicly linked sequences:

```bash
python scripts/prepare_frames_from_metadata.py \
  --metadata metadata/train_test_split_with_links.json \
  --output-root <data_root> \
  --download-root <download_cache> \
  --split all
```

Author-captured sequences with `source_url: null` are skipped by this script.

## Panorama SfM

The released SfM models were produced from the extracted panoramic frames using COLMAP's panorama pipeline with non-overlapping perspective rendering. To reproduce the same directory layout, set `PANO_SCRIPT` to COLMAP's `python/examples/panorama_sfm.py` and run:

```bash
DATA_ROOT=<data_root> \
PANO_SCRIPT=<path_to_colmap>/python/examples/panorama_sfm.py \
GPU_IDS="0" \
bash scripts/run_panorama_sfm_from_frames.sh
```

This writes results under:

```text
<data_root>/<split>/<scene>/<clip_id>/colmap_nonoverlap/
```

Image names inside the COLMAP sparse models follow the rendered layout `colmap_nonoverlap/images/pano_camera*/%04d.jpg`. Public web videos may become unavailable or may be transcoded by hosting platforms over time, so exact frame-level reproduction is best-effort.

## Responsible Use And Takedown

OmniLife360 was constructed from publicly accessible 360-degree videos from the web, together with a small number of panoramic sequences captured by the author team. Candidate videos were manually reviewed and screened for privacy risks. Videos containing sensitive, private, harmful, inappropriate, or otherwise privacy-risky content were excluded. Author-captured sequences were reviewed under the same criteria before inclusion.

We do not annotate personal identities, usernames, demographic attributes, or other personally identifiable information. The benchmark is intended for academic evaluation of 360-degree 3D reconstruction methods, not for identifying, analyzing, or profiling individuals.

Please use the dataset responsibly and avoid privacy-invasive applications such as person re-identification, face recognition, biometric identification, surveillance, profiling, individual tracking, targeting individuals, or attempts to recover uploader/platform identities.

For dataset questions or requests to remove links or associated metadata, contact:

`zonglinzhao@hust.edu.cn`

Please include, when available, your name, contact email, relationship to the content, relevant source URL or sequence ID, and the reason for the request.

## Citation

If you use OmniLife360, please cite:

```bibtex
@inproceedings{omnilife360_2026,
  title     = {OmniLife360: A Benchmark for 3D Reconstruction from In-the-Wild 360-Degree Captures},
  author    = {OmniLife360 Authors},
  booktitle = {European Conference on Computer Vision (ECCV)},
  year      = {2026}
}
```

## License

OmniLife360 is released under the Creative Commons Attribution-NonCommercial 4.0 International license (`CC BY-NC 4.0`).