Temporal Enhance and Spatial Gated Network for Group Activity Recognition

Tiansheng Sun,Jianning Chi,Chengdong Wu

DOI: https://doi.org/10.1109/ccdc55256.2022.10034360

2022-01-01

Abstract:The main goal of Group Activity Recognition(GAR) is to analyze the group behavior in a multi-actors scene. It is a challenging task that integrates group activity features from individual actors who frequently interact with each other. However, existing state-of-the-art methods do not consider temporal information and take full advantage of spatial relations. Moreover, existing methods utilize max-pooling operation to generate group activity features, which introduces noise into group activity features and leads to bad performance. To tackle these problems, a temporal enhance and spatial gated network is proposed. To fully utilize the rich temporal features and spatial features, two branches are set up to extract temporal features and actor interactive relations, respectively. Then, a feature aggregation module is used to generate group activity features, which can decrease the noise introduced by the max-pooling operation. We conduct a series of experiments on two stand datasets. Experimental results on these benchmarks demonstrate that the proposed approach achieves state-of-the-art performance. Compared with previous work, the recognition accuracy increases by 0.5%~1.0%.