HandAugment

The winner method of HANDS19 Challenge: Task 1 - Depth-Based 3D Hand Pose Estimation

The code for paper: "HandAugment: A Simple Data Augmentation Method for Depth-Based 3D Hand Pose Estimation"

https://arxiv.org/abs/2001.00702

Updates !!!!

(2020-June-30) upload test script and pretrained model.

Required libraries

Python 3.6
Numpy 1.17.2
PyTorch 1.0.1
OpenCV 4.1

Usage

Clone this repo

git clone https://github.com/wozhangzhaohui/HandAugment.git
cd HandAugment

Download Hands19 dataset from HANDS19 website. Replace spaces in file path with underscores "_" and link HANDS19 folder by mkdir dataset; ln -s your-hands19-folder-path dataset/HANDS19_Challenge
Run the test script by command: bash run_test.sh, the result is saved in output folder "output/stage1/result.txt".
The result file in "output/stage1/result.zip" can be submitted directly to Hands19Task1

Pre-trained model

We provide two stages pre-trained model for Hands19Task1 dataset.

Intermediate score at stage0 can reach 14.06

Final score at stage1 can reach 12.99

HandAugment Architecture

Quantitative Comparison

HANDS19 Task1 Dataset

###NYU Dataset

Citation

@article{zhang2020handaugment,
  title={HandAugment: A Simple Data Augmentation Method for Depth-Based 3D Hand Pose Estimation},
  author={Zhang, Zhaohui and Xie, Shipeng and Chen, Mingxiu and Zhu, Haichao},
  journal={arXiv},
  pages={arXiv--2001},
  year={2020}
}

WangHui19/HandAugment