To synthesize images with different emotions for certain person by multi-domain image-to-image emotion transfer on FER2013 dataset, and even emotion transfer from real human to virtual character on FER2013 and FERG-DB datasets.

Our contribution can be listed as followings:

• Perceptual loss in D
• Skip (residual) connection in G
• Some GAN tricks (Dropout + LReLu in G)

Implementation details and qualitative results please refer to our report.

1. Download FER2013 and FERG-DB datasets

2. Use python convert_fer2013.py to convert csv file to images with certain folder structure, and FERG-DB also need to create the folder structure as fer2013 one.

3. Train our modified StarGAN

bash run_train.sh
bash run_train_both.sh

3. Test our modified StarGAN

bash run_test.sh
bash run_test_both.sh

4. Facial expression recognition evaluation on generated images from FER2013 test dataset by training ResNet18 classifier on FER2013 training set.

cd eval_code
python run_eval_train.sh
python run_eval_test.sh

[1] PyTorch-StarGAN

[2] Tips and tricks to make GANs work