Linzhan's picture
Add dataset card, copyright notice, indexes and tooling
539a502 verified
|
Raw
History Blame Contribute Delete
1.24 kB
# Scripts
Derives both CSV indexes from the assets. Pure Python 3, no dependencies: GLB metadata is read
from each file's glTF JSON chunk and the binary payload is never touched, so probing the whole
23 GB collection takes about ten seconds.
| script | purpose |
|---|---|
| `glb_probe.py` | Library + CLI. Parses one `.glb` and returns rig, animation and geometry stats. |
| `build_dataset.py` | Walks every asset once and writes `../metadata.csv` and `../animations.csv`. `--link` also hardlinks a source directory into `../glb/`. |
```bash
python3 build_dataset.py # rebuild both tables in place
python3 build_dataset.py --src /path/raw --link # ingest a fresh download
python3 glb_probe.py ../glb/<file>.glb # inspect a single asset
```
## Notes on the probe
- Kinematic-tree roots are counted as joints whose parent is not itself a joint, so
`num_skeleton_roots > 1` marks a disconnected or multi-tree skeleton.
- Clip duration comes from the `max` of each animation sampler's input accessor, which glTF
stores in the JSON chunk.
- `animated_joints` intersects a clip's target nodes with the skin's joint set, separating clips
that drive a rig from clips that only move object-level nodes.