Multi-dimensional convolution transformer for group activity recognition

Group activity recognition, which aims to understand the activity performed by a group of people, has attracted growing attention in the realm of computer vision over the past decade. In this paper, we propose a novel multi-dimensional convolution Transformer network for group activity recognition, which not only models spatial-temporal feature representations, but also combines channel information to analyze the spatial-temporal dependencies of individual actors. Specifically, we first construct a multi-scale feature extraction module in the feature extraction stage, which can exploit discriminative high-level and low-level feature representations. The multi-branching strategy combined with the dilated convolution can further capture multi-scale feature information in complex group scenarios. Then, to construct the inter-dependence among involved actors from different dimensions, we design a multi-dimensional convolution Transformer in the relational reasoning stage, which consists of the following three parts: a channel attention module, a spatial-temporal convolutional Transformer, and a spatial-temporal attention module. Finally, the final activity recognition result is obtained by using a softmax classifier. Extensive experiments on two public GAR datasets demonstrate that the recognition accuracy on the Volleyball Dataset and Collective Activity Dataset can reach 92.8% and 96.1%, respectively, which is a significant improvement compared with the mainstream methods in recent years.