MotionColBert checkpoints

Checkpoints for Fine-grained Motion Retrieval via Joint-Angle Motion Images and Token-Patch Late Interaction (ACCV 2026).

Code, data preparation and evaluation instructions: https://github.com/yaozhang182/MotionColBert

File	Model	Dataset	Backbones
motioncolbert_humanml3d.pt	MotionColBert	HumanML3D	ViT-B/16 + DistilBERT
motioncolbert_kit.pt	MotionColBert	KIT-ML	ViT-B/16 + DistilBERT
motioncolbert_l_humanml3d.pt	MotionColBert-L	HumanML3D	ViT-L/16 + RoBERTa-Large
motioncolbert_l_kit.pt	MotionColBert-L	KIT-ML	ViT-L/16 + RoBERTa-Large

The MotionColBert checkpoints reproduce Table 2 of the paper exactly. The MotionColBert-L checkpoints were retrained with the released code (see the GitHub README for their results).

huggingface-cli download zyyy12138/MotionColBert --local-dir checkpoints
python scripts/test.py --config-name=motioncolbert_humanml3d eval.checkpoint=checkpoints/motioncolbert_humanml3d.pt

Citation

@inproceedings{zhang2026motioncolbert,
  title     = {Fine-grained Motion Retrieval via Joint-Angle Motion Images and Token-Patch Late Interaction},
  author    = {Zhang, Yao and Liu, Zhuchenyang and He, Yanlan and Ploetz, Thomas and Xiao, Yu},
  booktitle = {Proceedings of the Asian Conference on Computer Vision (ACCV)},
  year      = {2026}
}