Kwai Project for Hand Pose Estimation

This is a project for hand pose estimation on top of Simple Baselines for Human Pose Estimation and Tracking, you can find more details in the original paper.

For more detailed information, please view this sheet.

Requirements

Python == 3.5
PyTorch >= 0.4.0
torchvision >= 0.2.0
OpenCV > 2.4.9

Datasets

Images

The current in-use and potentially to-be-used dataset is located at liujintao@gpu110:/data/Kwai.

Currently, the project is using:

CMU_Panoptic: HandDB powered by CMU
RHD: Rendered Handpose Dataset (in the wild) powered by Freiburg Univ.
union/real/kwai_*.jpg: data provided by Kwai Inc., please keep it confidential

Labels

The label files can be find at liujintao@gpu110:/home/liujintao/app/Hand-Keypoint-Estimation/data/labels.

Usage

Set up

Go to the root directory of the project, and follow the next steps:

Compose the python environment using the environment.yml:
```
conda create env -f environment.yml
```

Set up environment variables for CUDA configuration:

export LD_LIBRARY_PATH=/usr/local/cuda/lib64${LD_LIBRARY_PATH:+:${LD_LIBRARY}}
export PATH=/usr/local/cuda/bin${PATH:+:${PATH}}
export CUDA_HOME=/usr/local/cuda${CUDA_HOME:+:${CUDA_HOME}}

Note that this is not required but recommended.

Create a directory (mkdir data) in the root directory of the project, and arrange the layout like this:

├── cache
├── images -> /data/Kwai
└── labels
    ├── real
    │   ├── all_test_labels.txt
    │   ├── all_train_labels.txt
    │   ├── kwai_label_file.txt
    │   └── union_train_labels.txt
    └── synth
        ├── all_test_labels.txt
        ├── all_train_labels.txt
        ├── RHD_label.txt
        ├── union_train_labels.txt
        ├── vis_test_labels.txt
        └── vis_train_labels.txt

Two more things to note:

The program will automatically generate a directory named cache and save pickle files under it for the first time you run (it will take a couple of minutes).
images is a symbolic link to where you store the images.

Train

python main.py --config config/train.yml --gpu 0

Test

Change EVALUATE to True and modify MODEL.RESUME in config/train.yml and run:

python main.py --config config/train.yml --gpu 0

ParanoidW/Hand-Keypoint-Estimation