CascadeFormer 🌊

A Family of Two-stage Cascading Transformers for Skeleton-based Human Action Recognition

News/Updates

• [March 31, 2026] paper is accepted at ICPR 2026 🥳🎉!
• [August 31, 2025] paper available on arXiv!
• [July 19, 2025] model checkpoints are publicly available on HuggingFace for further analysis/application!

Evaluation Results

Overall accuracy evaluation results of CascadeFormer variants on three datasets. CascadeFormer 1.0 consistently achieves the highest accuracy on Penn Action and both NTU60 splits, while 1.1 excels on N-UCLA. All checkpoints are open- sourced for reproducibility.

Citation

Please cite our work if you find it useful/helpful:

@misc{peng2025cascadeformerfamilytwostagecascading,
      title={CascadeFormer: A Family of Two-stage Cascading Transformers for Skeleton-based Human Action Recognition}, 
      author={Yusen Peng and Alper Yilmaz},
      year={2025},
      eprint={2509.00692},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2509.00692}, 
}

Contacts

If you have any questions or suggestions, feel free to contact:

• Yusen Peng (peng.1007@osu.edu)
• Alper Yilmaz (yilmaz.15@osu.edu)

Or describe it in Issues.