Drop-relationship learning for semi-supervised facial action unit recognition

•We propose a novel Drop-relationship learning module, which uses prior knowledge to construct AU relationship units. We randomly drop AU relationship units during training to suppress co-adaptation. To the best of our knowledge, we are the first to consider both AU prior knowledge and complex co-adaptation relationship between AUs;•We propose a new end-to-end AU semi-supervised framework, which uses consistency regularization with pseudo-multi-labeling method to extract learning signals from unlabeled facial images;•The proposed method outperforms the state-of-the-art semi-supervised and supervised methods on two widely used AU datasets (BP4D and DISFA).