Localizing activity groups in videos

摘要

Beyond recognizing actions of individuals, activity group localization in videos aims to localize groups of persons in spatiotemporal spaces and recognize what activity the group performs. In this paper, we propose a latent graph model to simultaneously address the problem of multi-target tracking, group discovery and activity recognition. Our key insight is to exploit the contextual relations among people. We present them as a latent relational graph, which hierarchically encodes the association potentials between tracklets, intra-group interactions, correlations, and inter-group compatibilities. Our model is capable of propagating multiple evidences among different layers of the latent graph. Particularly, associated tracklets assist accurate group discovery, activity recognition can benefit from knowing the whole structured groups, and the group and activity information in turn provides strong cues for establishing coherent associations between tracklets. Experiments on five datasets demonstrate that our model achieves both significant improvements in activity group localization and competitive performance on activity recognition.