Mutual Learning Networks Of Actor Relation Graph For Group Activity Recognition

Abstract

The Actor Relation Graph (ARG) is an effective method for detecting group behaviour but still needs improvement in some areas. In this paper, we propose using the sum of absolute differences (SAD) to compute the similarity of characters' appearance, introduce deep mutual learning to support the network's training, and add a visualization model. By training with the extended dataset, the results show that our improved network can achieve the expected better prediction accuracy of group activities.