EHGFormer: An efficient hypergraph-injected transformer for 3D human pose estimation

This is the official implementation of the approach described in the paper:

EHGFormer: An efficient hypergraph-injected transformer for 3D human pose estimation,
Siyuan Zheng, Weiqun Cao
Image and Vision Computing, 2025

Network Architecture

Environment Prepare

Train and Evaluation

• one GPU RTX 3090(24GB)
• Python 3.8.0
• cuda 11.1

Detail libraries will be installed by running following command:

pip install -r requirements.txt

Inference and Application

See details in target document.

3D Human Pose Estimation

See details in reference document.

Skeleton-based Action Recognition

See details in reference document.

Mesh Recovery (Image-based Refine)

See details in reference document.

Wild Video Inference

See details in reference document.

Application

Simple application: Motion capture for skeletal animation

See details in reference document.

Complex application: Animated Drawings

Detailed instructions see in Simplified Animated Drawings.

Citation

If you find our work useful in your research, please consider citing:

@article{ZHENG2025105425,
title = {EHGFormer: An efficient hypergraph-injected transformer for 3D human pose estimation},
journal = {Image and Vision Computing},
pages = {105425},
year = {2025},
issn = {0262-8856},
doi = {https://doi.org/10.1016/j.imavis.2025.105425},
url = {https://www.sciencedirect.com/science/article/pii/S0262885625000137},
author = {Siyuan Zheng and Weiqun Cao},
keywords = {Estimation of human pose in 3D, Transformers, Hypergraph, Efficient inference}
}

Acknowledgement

Our code is extended from the following repositories. We thank the authors for releasing the codes.

• MixSTE

• MotionBERT

• HyperFormer

• VITKD

• easy_ViTPose

• live_mocap

• ROMP

• HybrIK

• nosmpl

• pytorch3d

• bvhio

• Simplified_Animated_Drawings