HandOccNet: Occlusion-Robust 3D Hand Mesh Estimation Network

Introduction

This repository is the offical Pytorch implementation of HandOccNet: Occlusion-Robust 3D Hand Mesh Estimation Network (CVPR 2022). Below is the overall pipeline of HandOccNet.

Quick demo (update soon)

Directory

Root

The ${ROOT} is described as below.

${ROOT}  
|-- data  
|-- demo
|-- common  
|-- main  
|-- output

data contains data loading codes and soft links to images and annotations directories.
demo contains demo codes.
common contains kernel codes for HandOccNet.
main contains high-level codes for training or testing the network.
output contains log, trained models, visualized outputs, and test result.

Data

You need to follow directory structure of the data as below.

${ROOT}  
|-- data  
|   |-- HO3D
|   |   |-- data
|   |   |   |-- train
|   |   |   |   |-- ABF10
|   |   |   |   |-- ......
|   |   |   |-- evaluation
|   |   |   |-- annotations
|   |   |   |   |-- HO3D_train_data.json
|   |   |   |   |-- HO3D_evaluation_data.json
|   |-- DEX_YCB
|   |   |-- data
|   |   |   |-- 20200709-subject-01
|   |   |   |-- ......
|   |   |   |-- annotations
|   |   |   |   |--DEX_YCB_s0_train_data.json
|   |   |   |   |--DEX_YCB_s0_test_data.json

Download HO3D(version 2) data and annotation files [data][annotation files]
Download DexYCB data and annotation files [data][annotation files]

Pytorch MANO layer

For the MANO layer, I used manopth. The repo is already included in common/utils/manopth.
Download MANO_RIGHT.pkl from here and place at common/utils/manopth/mano/models.

Output

You need to follow the directory structure of the output folder as below.

${ROOT}  
|-- output  
|   |-- log  
|   |-- model_dump  
|   |-- result  
|   |-- vis

Creating output folder as soft link form is recommended instead of folder form because it would take large storage capacity.
log folder contains training log file.
model_dump folder contains saved checkpoints for each epoch.
result folder contains final estimation files generated in the testing stage.
vis folder contains visualized results.

Running HandOccNet

Start

Install PyTorch and Python >= 3.7.4 and run sh requirements.sh.
In the main/config.py, you can change settings of the model including dataset to use and input size and so on.

Train

In the main folder, set trainset in config.py (as 'HO3D' or 'DEX_YCB') and run

python train.py --gpu 0-3

to train HandOccNet on the GPU 0,1,2,3. --gpu 0,1,2,3 can be used instead of --gpu 0-3.

Test

Place trained model at the output/model_dump/.

In the main folder, set testset in config.py (as 'HO3D' or 'DEX_YCB') and run

python test.py --gpu 0-3 --test_epoch {test epoch}

to test HandOccNet on the GPU 0,1,2,3 with {test epoch}th epoch trained model. --gpu 0,1,2,3 can be used instead of --gpu 0-3.

For the HO3D dataset, pred{test epoch}.zip will be generated in output/result folder. You can upload it to the codalab challenge and see the results.
Our trained model can be downloaded from here

Results

Here I report the performance of the HandOccNet.

Reference (update soon)

@InProceedings{Park_2022_CVPR_HandOccNet,  
author = {Park, JoonKyu and Oh, Yeonguk and Moon, Gyeongsik and Choi, Hongsuk and Lee, Kyoung Mu},  
title = {HandOccNet: Occlusion-Robust 3D Hand Mesh Estimation Network},  
booktitle = {Conference on Computer Vision and Pattern Recognition (CVPR)},  
year = {2022}  
}

Acknowledgements

For this project, we relied on research codes from:

zhang0557kui / HandOccNet