关键词: 人-物交互, 上下文, 注意力机制, 图注意力网络

Abstract:

A graph-based Human-Object Interactions（HOIs） recognition method is proposed. In order to effectively represent the rich interactions between humans and objects in static images, the graph structure with powerful relational modeling capabilities is used to generate corresponding human-object interaction graph for the image. Considering the good performance of context information on various image recognition tasks, for utilizing the context information in the image, a Feature Processing Network（FPNet） that introduces an attention mechanism is proposed. The detection and recognition of real HOIs is done through a Graph Attention Network（GAT）. The method is validated on the V-COCO dataset and HICO-DET dataset, and compared with other methods. The results show that the proposed method has good results.

Key words: human-object interactions, context, attention mechanism, Graph Attention Network（GAT）