Joint Voxel and Coordinate Regression (JVCR) for 3D Facial Landmark Localization

This repository includes the PyTorch code of the JVCR method described in Adversarial Learning Semantic Volume for 2D/3D Face Shape Regression in the Wild (IEEE Transactions on Image Processing, 2019).

Requirements

• python 2.7

packages

• PyTorch
• NumPy
• Matplotlib
• progress

Usage

Clone the repository and install the dependencies mentioned above

git clone https://github.com/HongwenZhang/JVCR-3Dlandmark.git
cd JVCR-3Dlandmark

Then, you can run the demo code or train a model from stratch.

Demo

1. Download the pre-trained model (trained on 300W-LP) and put it into the checkpoint directory

2. Run the demo code

python run_demo.py --verbose

Training

1. Prepare the training and evaluation datasets

• Download 300W-LP and AFLW3000-3D
• Create soft links to the dataset directories

ln -s /path/to/your/300W_LP data/300wLP/images
ln -s /path/to/your/aflw2000 data/aflw2000/images

• Download .json annotation files from here and put them into data/300wLP and data/aflw2000 respectively

2. Run the training code

python train.py --gpus 0 -j 4

Acknowledgment

The code is developed upon PyTorch-Pose. Thanks to the original author.

Citation

If the code is helpful in your research, please cite the following paper.

@article{zhang2019adversarial,
  title={Adversarial Learning Semantic Volume for 2D/3D Face Shape Regression in the Wild},
  author={Zhang, Hongwen and Li, Qi and Sun, Zhenan},
  journal={IEEE Transactions on Image Processing},
  volume={28},
  number={9},
  pages={4526--4540},
  year={2019},
  publisher={IEEE}
}