You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
# 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.
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
Release Colab demo.
Release Huggingface demo.
Release SKELify pipeline.
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},
}