Reconstructing Humans with a Biomechanically Accurate Skeleton

<a href="https://isshikihugh.github.io/HSMR/"><img src="https://img.shields.io/website?url=https%3A%2F%2Fisshikihugh.github.io%2FHSMR%2F&label=Project%20Page&up_message=Online&up_color=CAB7A5&down_message=Offline&down_color=%23FF3F4D&logo=googlechrome&logoColor=white"></a> <a href="https://arxiv.org/abs/2503.21751"><img src="https://img.shields.io/badge/arXiv-2503.21751-%23B31C1C?logo=arxiv&logoSize=auto"></a> <a href="https://www.cs.utexas.edu/~pavlakos/hsmr/resources/hsmr_suppmat.pdf"><img src="https://img.shields.io/badge/SupMat-PDF-%2347A141?logo=overleaf&logoColor=white"></a> <a href="https://colab.research.google.com/drive/1RDA9iKckCDKh_bbaKjO8bQ0-Lv5fw1CB?usp=sharing"><img src="https://img.shields.io/badge/Demo-Open%20In%20Colab-blue?logo=googlecolab"></a> <a href="https://huggingface.co/spaces/IsshikiHugh/HSMR"><img src="https://img.shields.io/badge/%F0%9F%A4%97%20Demo-Open%20In%20HF-blue"></a> <br> <a href="#"><img src="https://img.shields.io/badge/HSMR%20Demo-Released-green?logo=github"></a> <a href="#"><img src="https://img.shields.io/badge/HSMR%20Evaluation-Released-green?logo=github"></a> <a href="#"><img src="https://img.shields.io/badge/HSMR%20Training-Released-green?logo=github"></a> <a href="#"><img src="https://img.shields.io/badge/SKELify-Released-green?logo=github"></a>

<!--  --> <!-- Teaser Parts -->

Reconstructing Humans with a Biomechanically Accurate Skeleton <br> Yan Xia, Xiaowei Zhou, Etienne Vouga, Qixing Huang, Georgios Pavlakos <br> CVPR 2025 (Oral)

📢 News

• [2025.04.04] HSMR has been accepted as Oral paper at CVPR 2025! 🎉
• [2025.03.28] Release HSMR demo code, evaluation code and training code.

⚒️ Setup

1. 🌏 Environment Setup
2. 📦 Data Preparation

🚀 Demo & Quick Start

<!-- **[<img src="https://i.imgur.com/QCojoJk.png" width="30"> Google Colab demo](#) | [<img src="https://s2.loli.net/2024/09/15/aw3rElfQAsOkNCn.png" width="20"> HuggingFace demo](#)** -->

python exp/run_demo.py --help

Quick start with images:

# Folders wil be identified as image folders by default if `--input_type` is not specified.
python exp/run_demo.py --input_path "data_inputs/demo/example_imgs"

Quick start with videos:

# Single file wil be identified as a video by default if `--input_type` is not specified.
python exp/run_demo.py --input_path "data_inputs/demo/example_videos/gymnasts.mp4"

Tips: Rendering skeleton meshes is pretty slow. For videos, adding --ignore_skel or decrease --max_instances could boost the speed. Check lib/kits/hsmr_demo.py:parse_args() for more details.

🧱 Reproducibility

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

We also provide the SKELify optimization pipeline, which optimizes SKEL parameters according to trustworthy 2D keypoints. Please refer to docs/OPTIM.md for more details.

🗓️ TODOs

• [x] Release Colab demo.
• [x] Release Huggingface demo.
• [x] Release SKELify pipeline.
• [x] Release training code & data.

📝 Citation

@inproceedings{xia2025hsmr,
  title={Reconstructing Humans with a Biomechanically Accurate Skeleton},
  author={Xia, Yan and Zhou, Xiaowei and Vouga, Etienne and Huang, Qixing and Pavlakos, Georgios},
  booktitle={CVPR},
  year={2025},
}

📜 Acknowledgement

Parts of the code are adapted from the following repos: SKEL, 4D-Humans/HMR2.0, SMPLify-X, SPIN, ProHMR, ViTPose, Detectron2, and GVHMR.