SkelMo — Skeletal Motion Release v1

BVH motion companion for SkelMo: Universal Skeletal Motion Generation for 3D Rigged Shapes (ECCV 2026).

Project · Paper · Code

This release contains 20,316 skeletal motion clips, 12,031 rig IDs, and 2,033,903 frames at 20 FPS. It is the non-Truebones subset of our screened mixed-v7 collection. Meshes, textures, skinning weights, source asset binaries, rendered videos, and model training tensors are not included. The BVHs retain the skeleton hierarchy and joint offsets required to interpret the motion.

This is a motion companion release, not the complete set of paper training inputs. Its counts refer to clips and rig IDs in this release and should not be substituted for the paper's model/video counts.

Contents

File	Contents
archives/skelmo-v1-sketchfab.tar.zst	2,578 BVHs from the Sketchfab source group
archives/skelmo-v1-github.tar.zst	17,738 BVHs with GitHub asset provenance recovered through Articulation-XL 2.0
metadata/motions.jsonl	Complete per-motion provenance, hashes, license observations, and preprocessing records
metadata/{train,val,test}.parquet	The same records, split for the Hugging Face dataset viewer
metadata/attribution.jsonl	Per-asset source and available creator/license metadata
metadata/statistics.json	Counts and historical license distribution
metadata/source_revisions.json	Pinned upstream metadata revisions
splits/*.txt	Frozen sample IDs, one per line
SHA256SUMS	Release file checksums
tools/verify_release.py	File and archive integrity verification
tools/fetch_source.py	Retrieve a selected source asset from its recorded upstream location

Download and verify

pip install -U huggingface_hub
hf download momoya2000/SkelMo --repo-type dataset --local-dir SkelMo
python SkelMo/tools/verify_release.py SkelMo
# Install zstd, then verify every BVH inside both archives:
python SkelMo/tools/verify_release.py SkelMo --archives
mkdir -p SkelMo/extracted
tar --zstd -xf SkelMo/archives/skelmo-v1-sketchfab.tar.zst -C SkelMo/extracted
tar --zstd -xf SkelMo/archives/skelmo-v1-github.tar.zst -C SkelMo/extracted

The extraction command assumes GNU tar and zstd. On systems without tar --zstd, use zstd -dc ARCHIVE | tar -xf - -C DESTINATION. Files extract under normalized_bvhs/<rig_id>/. For a clip, join that extraction directory with its bvh_path.

from datasets import load_dataset
metadata = load_dataset("momoya2000/SkelMo", "motions")
sample = metadata["train"][0]
print(sample["sample_id"], sample["bvh_path"], sample["source_url"])

This loads the metadata table; download the archives separately for BVH payloads. For a reproducible experiment, pin the Hugging Face commit via revision= or hf download --revision and retain SHA256SUMS.

Frozen splits

Partition	Clips
Train	19,761
Validation	282
Test	273

Validation comprises 72 original val_seen and 210 val_unseen clips. Test comprises 71 test_seen and 202 test_unseen clips. The lists preserve the existing split600_clean assignments after removing 1,005 Truebones clips; no new random split was generated. The names seen and unseen are inherited labels, not a new claim about asset-level or topology-level independence. Exact correspondence to the paper's reported experimental splits has not been established for this release.

Processing and reproducibility

The payloads are the existing normalized BVHs from the frozen mixed-v7 collection. Processing records include source and target frame counts, scale normalization, root alignment, rest-facing alignment where recorded, and forward-kinematics consistency metrics. The collection uses 20 FPS target sampling, scale normalization, initial root alignment, and the existing v7 rest-facing/FK corrections. Per-motion records describe the selected alignment; they should take precedence over assuming one transform for every source.

We distribute the processed BVHs without re-exporting them. Every payload is matched to its previously recorded SHA-256. Stable sample IDs and split files reproduce this exact released subset. IDs are frozen identifiers and should not be recomputed from a new machine's paths.

Source asset URLs/IDs are provided for retrieving meshes and original rigging separately. GitHub URLs are pinned to the recorded commit. The helper below prints the source and download method by default; add --download --output PATH to fetch a GitHub source file and verify its recorded hash:

python tools/fetch_source.py metadata/motions.jsonl SAMPLE_ID

Sketchfab access follows its official download flow and the source author's current availability and terms. Source assets may become unavailable. Downloading an original asset does not itself reproduce all intermediate alignment/conversion steps. An end-to-end raw-asset-to-identical-BVH rebuild is not included in v1.

Provenance and licensing

Source assets retain their individual licenses and restrictions. This dataset does not apply a single permissive license to all BVHs. See LICENSE.md, NOTICE.md, and SCHEMA.md.

We preserve historical Objaverse-XL license annotations separately from partial live Sketchfab observations. Missing metadata is represented as null; a repository-level license is not proof that every embedded asset is covered. Some clips have noncommercial/share-alike annotations, unknown licenses, or provenance signals recorded in the manifest. These records document available evidence, not a completed rights clearance. Users should review the source terms for their intended use, especially commercial redistribution.

Original creators are credited through metadata/attribution.jsonl and the source links. Creator names are missing where upstream evidence did not supply them; the attribution index is not a claim of complete authorship verification. The Articulation-XL 2.0 UUID mapping and Objaverse-XL metadata are credited in NOTICE.md.

For a correction or removal request, open a dataset discussion with the sample ID/source URL and supporting details. Please avoid posting private personal information.

Citation

@inproceedings{tao2026skelmo,
  title={SkelMo: Universal Skeletal Motion Generation for 3D Rigged Shapes},
  author={Tao, Ye and Yao, Yuxin and Liu, Kendong and Wu, Dapeng and Hou, Junhui},
  booktitle={European Conference on Computer Vision (ECCV)},
  year={2026}
}

Please also acknowledge the original asset creators and the upstream collections listed in NOTICE.md when using their contributions.