Global Context Interactive Learning for Group Activity Recognition

Group activity recognition(GAR) aims to understant the activity in multi-person scenes. Previous methods have focused on individual feature inference but have not fully exploited individual-specific contextual information, which is important and useful in computer vision tasks. In this paper, we propose a global context interaction network(GCIN) with embedded Attention Pooling(AP) module and Context Encoding(CE) modules to bridge the gap between group activities and the visual environment. The Attention Pooling module extracts more representative individual features from extracted features, while the CE module encodes global contextual information into a individual feature and refines aggregated information to enhance individual representations. A graph network is then constructed through a multi-graph spatiotemporal graph convolutional network to generate semantic representations. Additionally, we further design a dual-branch model that integrates the designed modules into a pipeline. The effectiveness of each module on each branch is demonstrated through extensive experiments. Furthermore, our method achieves state-of-the-art results on two widely-used benchmark datasets.