README.md

---
license: other
task_categories:
- image-to-3d
tags:
- volumetric-video
- multi-view-video
- dynamic-scene-reconstruction
pretty_name: SiggraphAsia2026VVC
---

# SiggraphAsia2026VVC

This public release contains the Siggraph Asia 2026 Volumetric Video Challenge data for `siga` and `selfcap`.


## Structure

```text
SiggraphAisa2026VVC/
  siga/
    compression/{val,test}/<case>.tar
    sparse/{val,test}/<case>.tar
  selfcap/
    compression/test/<case>.tar
    sparse/test/<case>.tar
  release_manifest.json
  archive_manifest.json
  checksums.sha256
  license.txt
```

Each extracted case keeps the camera/data layout:

```text
<case>/
  images/<view>/<frame>.jpg or .png
  extri.yml
  intri.yml
  train_extri.yml
  train_intri.yml
  t_offsets.json
  pcds/                 # when available
```

For public test cases, held-out test images and held-out test camera files are intentionally omitted. Public `extri.yml`, `intri.yml`, and `t_offsets.json` are filtered to released views only.

All `selfcap` cases are provided as public test cases only. The released selfcap frames are indexed `000900.jpg` through `001799.jpg`.

## Held-out Test Cameras

The main public archives omit held-out test-view images and held-out test camera files. When the challenge phase requires camera release for hidden test views, use the separate camera-only package:

```text
heldout_test_cameras/SiggraphAisa2026VVC_heldout_test_cameras.tar
```

This package contains no RGB images, masks, point clouds, checkpoints, or rendered results. Each extracted case mirrors the main dataset path and contains only camera metadata for held-out test views:

```text
intri.yml
extri.yml
test_intri.yml
test_extri.yml
t_offsets.json
README.txt
```

`heldout_test_cameras/heldout_test_camera_manifest.json` lists the included cases and held-out view names. `heldout_test_cameras/checksums.sha256` gives the archive checksum.

## Baseline Code

The repository also includes a baseline implementation under `baseline_code/`.
This code release is the cleaned FTGSPP baseline used for the challenge data.
It is provided in the same dataset repository for convenience and does not
change the dataset layout under `siga/` or `selfcap/`.

```text
baseline_code/
  README.md
  LICENSE
  THIRD_PARTY.md
  pyproject.toml
  uv.lock
  Dockerfile
  configs/
  ftgspp/
  scripts/
  CODE_RELEASE_MANIFEST.json
  checksums.sha256
```

Start from the baseline README:

```bash
cd baseline_code
bash scripts/install_env.sh
.venv/bin/python -m ftgspp.run data=<your_config> pipeline=check
.venv/bin/python -m ftgspp.run data=<your_config> run.gpu=0
```

`baseline_code/CODE_RELEASE_MANIFEST.json` lists the files included in the
code release. Large training outputs, checkpoints, videos, caches, and datasets
are intentionally excluded.

## Manifest

`release_manifest.json` lists the released public view IDs and counts for every dataset / track / split / case. It does not expose held-out test-view IDs.

`archive_manifest.json` and `checksums.sha256` describe the upload-ready tar files.

## License

See `license.txt`.