Blender-Moments
Blender-Moments is a multi-view benchmark for dynamic surface reconstruction. It has nine animated Blender scenes, each rendered from 50 training and 10 test cameras at every frame, with ground-truth meshes extracted from the animated scene itself.
With the scripts in the GitHub repository, you can extract ground-truth meshes at additional frames, modify the camera rig, or add new scenes, and generate your own version of the data.
Downloading
Use the download scripts in the Blender-Moments GitHub repository, which verify checksums and unpack into the dataset layout. They need only Python 3 and huggingface_hub.
git clone https://github.com/royorel/Blender-Moments.git
cd Blender-Moments
pip install huggingface_hub
python scripts/download_data.py # everything: renders, meshes, point clouds
python scripts/download_blend.py # source .blend files
Downloading part of the data
Both scripts take the scenes to fetch, and download_data.py also takes the components. Each defaults to all of them, and repeated runs skip files already downloaded and verified.
python scripts/download_data.py --scenes lego_4d book # two scenes, all components
python scripts/download_data.py --components meshes # ground-truth meshes of every scene
python scripts/download_data.py --scenes lego_4d --components images sfm # renders and point clouds of one scene
python scripts/download_data.py --components images --dry-run # list the archives without downloading
python scripts/download_blend.py --scenes lego_4d # one source scene
| Flag | Values | Default |
|---|---|---|
--scenes |
alligator beetle bicycle book elephant ficus_4d lego_4d spiderman toyota |
all nine |
--components (download_data.py only) |
images (training and test renders with their transforms files), meshes (ground-truth meshes and scene_normalization.json), sfm (per-frame COLMAP point clouds) |
all three |
--dataset-root |
where to unpack | the repository's directory |
--dry-run |
list the files instead of downloading them | off |
Single files can also be fetched with huggingface_hub alone, for example hf download royorel/Blender-Moments lego_4d/gt_meshes.zip --repo-type dataset --local-dir ., and unzipped at the dataset root.
Contents
Each scene directory holds four uncompressed zip archives and the source scene. Archive members are relative to the dataset root.
| File | Contents | Size, all scenes |
|---|---|---|
renders_train.zip |
Training images and transforms_train.json |
21.9 GB |
renders_test.zip |
Test images and transforms_test.json |
4.4 GB |
gt_meshes.zip |
10 ground-truth meshes and scene_normalization.json |
0.6 GB |
sfm_init.zip |
One COLMAP point cloud per frame | 0.06 GB |
<scene>.blend |
Source scene | 1.0 GB |
checksums.sha256 lists every file, and file_manifest.json the checksum of every archive member.
| Scene | Frames | Training images | Test images | License |
|---|---|---|---|---|
| alligator | 100 | 5000 | 1000 | CC BY-NC 4.0 |
| beetle | 150 | 7500 | 1500 | CC BY 4.0 |
| bicycle | 100 | 5000 | 1000 | CC BY 4.0 |
| book | 170 | 8500 | 1700 | CC BY 4.0 |
| elephant | 100 | 5000 | 1000 | CC BY 4.0 |
| ficus_4d | 100 | 5000 | 1000 | CC0 |
| lego_4d | 100 | 5000 | 1000 | CC BY-NC |
| spiderman | 120 | 6000 | 1200 | CC BY 4.0 |
| toyota | 150 | 7500 | 1500 | CC BY 4.0 |
Evaluation
To compare with the results reported in the paper, evaluate with the scripts and protocol of the MEMO repository, which cover both image metrics on the test views and geometric metrics against the ground-truth meshes. Each scene's scene_normalization.json holds the diagonal of its bounding box over the ground-truth frames, the constant by which the geometric metrics are normalized.
License
The data is released for non-commercial use under CC BY-NC 4.0, with each scene also subject to the license of its source model, listed with full attribution in the repository README.
Citation
The citation will be added with the arXiv version of the paper.
- Downloads last month
- 20
