SynFace: Face Recognition with Synthetic Data

This is the Pytorch implementation of our ICCV 2021 paper

SynFace: Face Recognition with Synthetic Data.
Haibo Qiu, Baosheng Yu, Dihong Gong, Zhifeng Li, Wei Liu and Dacheng Tao

Requirements

Main packages:

• python=3.6.7
• pytorch=1.8.1
• torchvision=0.9.1
• cudatoolkit=10.2.89

Or directly create a conda env with

conda env create -f environment.yml

Data preparation

1. Clone this repo:

   git clone https://github.com/haibo-qiu/SynFace.git
   

2. Clone the DiscoFaceGAN and insert our files as the mixup face generator (To run DiscoFaceGAN, you also need to satisfy its requirements.):

   git clone https://github.com/microsoft/DiscoFaceGAN.git data/DiscoFaceGAN
   cp data/syn_factors.py data/DiscoFaceGAN/
   cp data/syn_images.py data/DiscoFaceGAN/
   

3. Generate the face images with identity mixup, following with face alignment and crop:

   bash data/syn.sh
   

4. (Optional) Check our generated synthetic dataset via this google drive link.

5. Download the real face data CASIA and LFW from this link

6. Put all these data into data/datasets/

Training

Simply run the following script:

bash run.sh

Testing

To reproduce the results in our paper, please download the pretrained models and put them in pretrained/, then run:

bash eval.sh

Acknowledgement

The code of face alignment and crop data/imgs_crop/ is borrowed from face.evoLVe and re-written with multi-processing for acceleration.

Citation

If you use our code or models in your research, please cite with:

@inproceedings{qiu2021synface,
  title={SynFace: Face Recognition with Synthetic Data},
  author={Qiu, Haibo and Yu, Baosheng and Gong, Dihong and Li, Zhifeng and Liu, Wei and Tao, Dacheng},
  booktitle={Proceedings of the IEEE/CVF International Conference on Computer Vision},
  pages={10880--10890},
  year={2021}
}