Group Activity Recognition by Exploiting Position Distribution and Appearance Relation

Group activity recognition in multi-person scene videos is a challenging task. Most previous approaches fail to provide a practical solution to describe the person relations and distribution within the scene, which is important for understanding group activities. To this end, we propose a two-stream relation network to simultaneously deal with both position distribution information and appearance relation information. For the former, we build Position Distribution Network (PDN) to obtain the spatial position distribution. For the latter, we propose Appearance Relation Network (ARN) to explore the appearance relation of the individuals in scene. We fuse the two clues, i.e. position distribution and appearance relation, to form the global representation for group activity recognition. Extensive experiments on two widely-used group activity datasets demonstrate the effectiveness and superiority of the proposed framework.