Enhanced Interactive Understanding for Human-Object Interaction Detection

Identifying and classifying the interaction between people and objects in images plays an important auxiliary role in the task of visual understanding. However, it is still a challenging task due to the diversity of human movements and object types. In this paper, a single-stage three-branch HOI detection model is proposed, which divides and fuses the difference in the relationship between the subtasks of the model, uses the three-branch to deal with the difference of subtasks, and enhances the interaction understanding by constructing the association module between the subtasks. In addition, by introducing a powerful visual language model, we designed a visual language knowledge transfer strategy, and used its visual language prior knowledge to assist HOI detection, which enhanced the understanding ability of the model. Finally, the proposed model was evaluated on two published standard datasets, HICO-Det and V-COCO. The experimental results show that the proposed network has better performance than other state-of-the-art methods.