Vid-HACO checkpoints

Dense per-vertex hand contact (778 MANO vertices) from video, built on HACO. Code, setup and inference instructions: github.com/Bottle101/Vid-HACO.

file	--model	best at	backbone	training
vith.pt	vith	binary contact detection — is the hand touching something?	V-JEPA 2 ViT-H/256 + DPT	8-frame clips, epoch 1
vitl.pt	vitl	contact map estimation — which of the 778 vertices are in contact?	V-JEPA 2.1 ViT-L/384 + DPT	12 datasets, 8-frame clips, epoch 6

• Use vitl when you need the dense per-vertex contact map.
• Use vith when you need a reliable per-frame contact / no-contact decision.

Validation metrics stored in the checkpoints:

	contact-map F1 (per-vertex)	binary-contact F1	binary-contact accuracy	val samples
vitl	0.791	0.985	0.989	172,896 (12 datasets)
vith	0.637	0.987	0.990	49,664 (9 datasets)

The two were validated on different subsets, so compare within a column with that in mind.

file	sha256
vitl.pt	a663b0f54f7ef7bedf814c1921e127cdc2836606e0ed136a41c8e1abf982c511
vith.pt	8f1e65520974b4a5c613309f3d3d95d2651c826cae4ed73a4711c237a21882e0

Each file is a PyTorch dict with state_dict, epoch and val_metrics (no optimizer state). Load with the model classes in the GitHub repository (strict=True):

git clone https://github.com/Bottle101/Vid-HACO && cd Vid-HACO/inference
python checkpoints/fetch_checkpoints.py
./run.py --video my_clip.mp4 --out-dir out/my_clip --model vitl

Training data

HOT3D, ARCTIC (egocentric and allocentric), HOI4D, DECAF, Hi4D, InterHand2.6M, H2O, EPIC-Contact, MOW, PROX and HIC — 1,553,353 training / 172,877 validation samples. See the repository for per-dataset counts and how the contact labels were produced.

License and intended use

CC BY-NC 4.0, non-commercial research use only. The weights derive from HACO (CC BY-NC 4.0) and were trained on datasets with their own, mostly research-only, terms; use must comply with those terms as well.

Citation

If you use these checkpoints, please cite CARD (project page):

@article{dai2026card,
    title = {CARD: Contact-Anchored Retargeting for Dexterous Manipulation},
    author = {Dai, Cunxi and He, Botao and Hogan, Francois R. and Pan, Wei and Fu, Rao and Hsieh, Wen-Han and Liu, Ruoshi and Malik, Jitendra and Shi, Guanya and Sharma, Akash},
    journal = {arXiv preprint arXiv:2610.36539},
    year = {2026}
}

Acknowledgement

Vid-HACO builds on HACO by Daniel Sungho Jung and Kyoung Mu Lee. We are grateful to the authors for releasing their code and data.