Subject adaptation in spontaneous facial behavior recognition using style transfer functions

The appearance of facial Action Units (AUs) and painful expression may significantly vary for different people. Thus the probability distribution of both test and training data is not the same for person-independent facial behavior recognition. Some researchers have proposed methods to bring the performance of a person-independent system closer to a person-dependent one. Subject style is the cause of inter-personal variations. With this in mind, we propose methods to increase the generalization ability of facial AUs and pain detection through style transfer functions. We conducted extensive experiments on spontaneous UNBC-McMaster database to compare supervised methods. The results show that our approach can effectively perform the task of pain and AUs detection. So that the best average recognition rate of action units was 96.84 (with AUC criterion) and the same method in terms of low adaptation data and appropriate adaptation data had pain recognition rates of 87.30 and 93.26 (with AUC criterion), respectively.