FER-global-local-network

Mingjing Yu, Huicheng Zheng, Zhifeng Peng, Jiayu Dong, and Heran Du. 2020. "Facial expression recognition based on a multi-task global-local network". in Pattern Recognition Letters, Volume 131, Pages 166-171. Paper

Download data

Download CK+ and Oulu-CASIA databases from network

CK+ : http://www.pitt.edu/~emotion/ck-spread.htm

Oulu : https://www.oulu.fi/cmvs/node/41316

Preprocess datasets

The main code is in 'Data_preprocess' directory.

1. We use Dlib tools to detect landmarks and extract the eyes, nose, mouth regions. Download the dat file from http://dlib.net/files/shape_predictor_68_face_landmarks.dat.bz2 and unzip the shape_predictor_68_face_landmarks.dat to 'Data_preprocess' directory.

2. We give the grouping lists in 'Data_preprocess/data_ck+/CK+_Ten_group.txt' and 'Data_preprocess/data_oulu/Oulu_CASIA_Ten_group.txt'

3. Run 'productPKLforCKplus.py' to preprocess CK+ dataset and save in a '.pkl' file.

4. Run 'productPKLforOuluCasIA.py' to preprocess Oulu-CASIA dataset and save in a '.pkl' file.

We develop part of the preprocessing code from Tang Yan's work

Network Training

The main code is in 'Train_FER' directory.

1. We use deformable cnn in our network. We thank OuYang Wei for the pytorch implementation.

2. We use Mixup for data augmentation. We thank Zhang hongyi for the pytorch implementation.

3. Run 'train_ck.py' for training and testing CK+ datasets.

4. Run 'train_Oulu.py' for training and testing Oulu-CASIA datasets.