SKEL-CF: Coarse-to-Fine Biomechanical Skeleton and Surface Mesh Recovery

TL;DR: SKEL-CF is a coarse-to-fine transformer framework for estimating anatomically accurate SKEL parameters from 3D human data. By converting 4DHuman to SKEL-aligned 4DHuman-SKEL and incorporating camera modeling, it addresses data scarcity and depth/scale ambiguities. SKEL-CF outperforms prior SKEL-based methods on MOYO (85.0 MPJPE / 51.4 PA-MPJPE), offering a scalable, biomechanically faithful solution for human motion analysis.

---

Da Li^1,2,*,   Jiping Jin^1,3,*,   Xuanlong Yu¹,   Wei Liu⁵,   Xiaodong Cun⁴,  
Kai Chen⁵,   Rui Fan³,   Jiangang Kong⁵,   Xi Shen^1,†

*Equal contribution    †Corresponding author

¹Intellindust AI Lab   ²ShenZhen University    ³ShanghaiTech University   
⁴Great Bay University    ⁵Didi Chuxing Co.Ltd

If you like our work, please give us a ⭐!

🗓️ Updates

• [2025.11.26] Release SKEL-CF.
• [2025.11.27] Release checkpoints and labels on Hugging Face.

🧭 Table of Content

• 1. ⚒️ Setup
• 2. 🚀 Demo & Quick Start
• 3. 🧱 Reproducibility
• 4. 👀 Visual Results
• 5. 📝 Citation
• 6. 📜 Acknowledgement
• 7. 🌟 Star History

⚒️ Setup

1. 🌏 Environment Setup
2. 📦 Data Preparation

🚀 Demo & Quick Start

Quick start with images:

bash vis/run_demo.sh

Quick start with videos:

bash vis/run_video.sh

🧱 Reproducibility

For reproducing the results in the paper, please refer to docs/EVAL.md and docs/TRAIN.md.

👀 Visual Results

Per-Layer Refinement

Sports Video

💡 Tip: Click the buttons above to watch videos, or visit our project page for more visual results.

📝 Citation

If you use SKEL-CF or its methods in your work, please cite the following BibTeX entries:

@article{li2025skelcf,
  title={SKEL-CF: Coarse-to-Fine Biomechanical Skeleton and Surface Mesh Recovery},
  author={Li, Da and Jin, Jiping and Yu, Xuanlong and Cun, Xiaodong and Chen, Kai and Fan, Rui and Kong, Jiangang and Shen, Xi},
  journal={arXiv},
  year={2025}
}

📜 Acknowledgement

Parts of the code are adapted from the following repos: SKEL, CameraHMR, HSMR, ViTPose, Detectron2

🌟 Star History