High Fidelity 3D Hand Shape Reconstruction via Scalable Graph Frequency Decomposition

Introduction

This repo is official implementation of High Fidelity 3D Hand Shape Reconstruction via Scalable Graph Frequency Decomposition(CVPR 2023).

Install environment

conda env create -f environment.yml

Dataset

We use subdivided MANO to register 3D hand Mesh from "DeepHandMesh". Download joint annotation and our processed dataset from here. After downloading, put the contents in $root/data directory.

Download images for DeephandMesh and following the unzip instructions. Then, put the images folder under $root/data/DeepHandMesh directory

Assets

Download subdivided MANO from here and put the contents under $root/assets/. Download MANO from here. Put models folder under $root/assets/mano.

Pretrained Model

Download pretrained model from here and put the contents under $root/model.

Testing

python main.py --test --nmp --nrd

Training

python main.py --nmp --nrd

Reference

@InProceedings{Luan_2023_CVPR,
    author    = {Luan, Tianyu and Zhai, Yuanhao and Meng, Jingjing and Li, Zhong and Chen, Zhang and Xu, Yi and Yuan, Junsong},
    title     = {High Fidelity 3D Hand Shape Reconstruction via Scalable Graph Frequency Decomposition},
    booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
    month     = {June},
    year      = {2023},
    pages     = {16795-16804}
}

Acknowledgements

This repo inherited code from DeepHandMesh and S2HAND.

tyluann / FreqHand