This paper presents a novel method for facial expression recognition using only sequence level labeling. With facial image sequences containing multiple peaks of expression, our method aims to label these sequences and identifies expressional peaks automatically. We formulate this weakly labeled expression recognition as a multi-instance learning (MIL) problem. First, image sequences are clustered into multiple segments. After segmentation, image sequences are regarded as bags in MIL and the segments in the bags are viewed as instances. Second, bags data are used to train a discriminative classifier which combined multi-instance learning and discriminative Hidden Markov Model (HMM) learning. In our method, HMM is used to model temporal variation within segments. We conducted experiments on CK+ database and UNBC-McMaster Shoulder Pain Database. Experimental results on both databases show that our method can not only label the sequences effectively, but also locate apex frames of multi-peak sequences. Besides, the experiments demonstrate that our method outperforms state of the art.
展开▼
