This is keras implemtation for JAA-NET:Deep Adaptive Attention for Joint Facial Action Unit Detection and Face Alignment
Use the command python train_model.py
*AU_detect model did not contains attention block
*AU_detect2 model contains attention block,and the init weight for attention is propotional to distance of landmark points
*AU_detect2 model contains attention block,and the init weight for attention is implementation of paper
landmarks and Multilabel AUs for person,i.e if the AUs base dataset is BP4D,
the corresponding AU number is label_disfa = [1, 2, 4, 5, 6, 9, 12, 15, 17, 20, 25, 26].
For example , the AUs result is [0,1,1,0,0,0,0,0,0,0,0,1],the people have AUs[2,4,26]
run python test_JAA.py
##the labels introductions You can also make datasets yourself ,the labels contains image_id and AUS and 68 landmarks
For CK+ dataset
label.txt is the CK+ dataset with only 936 images selected from label_org.txt
label_org is the CK+ dataset with only 2250 images
label_line.txt is the CK+ dataset with labels only choosing the first data of one line
label_v.txt is the unique prefix of CK+ dataset and to be used for labels_per_video.txt
labels_per_video.txt is the sequence of CK+ dataset ,in which the frames is 24 for each video