This paper presents a novel approach in a rarely studied area of computervision: Human interaction recognition in still images. We explore whether thefacial regions and their spatial configurations contribute to the recognitionof interactions. In this respect, our method involves extraction of severalvisual features from the facial regions, as well as incorporation of scenecharacteristics and deep features to the recognition. Extracted multiplefeatures are utilized within a discriminative learning framework forrecognizing interactions between people. Our designed facial descriptors arebased on the observation that relative positions, size and locations of thefaces are likely to be important for characterizing human interactions. Sincethere is no available dataset in this relatively new domain, a comprehensivenew dataset which includes several images of human interactions is collected.Our experimental results show that faces and scene characteristics containimportant information to recognize interactions between people.
展开▼
