Unsupervised Face Normalization with Extreme Pose and Expression in the Wild

This repository contains training code for the paper

Unsupervised Face Normalization with Extreme Pose and Expression in the Wild (The paper will be released soon.)
IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019

The model learns to generate a frontal, neutral expression, photorealistic face image from one face image in the wild.

Contact: Yichen Qian (mx54039q@bupt.edu.cn), Weihong Deng (whdeng@bupt.edu.cn), Jiani Hu (jnhu@bupt.edu.cn)

Usage

Prerequisites

Python 2.7
Tensowflow 1.8.0

Prepare Datasets

We used CASIA-Webface for non-normal face set and Multi-PIE for normal face set in the unconstrained experiment. We can not redistribute the original dataset due to copyright. For normal face set, you can use other public-available datasets that has highly-normalled faces.

All face images are detected by MTCNN, and then cropped.

Pretrained models

Face expert network (VGG-Face2)
FNM

Train and test

Train

After unnormal face images and normal face images are ready, prepare image list respectively. Modify image path and list path and other settings in config.py.

python main.py

Test

python test.py --test_path your_test_path --test_list your_test_list --checkpoint_ft pretrained_model_path

Adjust face attention region

If you use other dataset for normal face set, please modify the attention discriminators (WGAN.py) according to the face region.

Notes

The identity perception loss (L_ip) is slightly modified. Cosine distance is used other than mean square error.

Citation

If you find FNM useful in your research, please consider citing:

@inproceedings{FNM,
  author = {Qian, Yichen and Deng, Weihong and Hu, Jiani},
  booktitle = {IEEE Conference on Computer Vision and Pattern Recognition},
  title = {Unsupervised Face Normalization with Extreme Pose and Expression in the Wild},
  year = {2019}
}

CharlesNord / fnm