Research on Human-Object Interaction Detection Algorithm Based on Relational Reasoning

In recent years, Human-Object Interaction Detection (HOI) has gradually become a research hotspot with the progress of visual relationship detection techniques in the fields of target detection, action recognition, and scene segmentation. The HOI task aims at recognizing the interaction between people and objects in a scene, and advancing the understanding of complex scenes. Traditional HOI detection methods are usually categorized into one-phase and two-phase methods, which extract features from global features or local person-object pairs for interaction inference, respectively. However, existing methods are difficult to fully utilize the interaction information between people-objects, resulting in unsatisfactory suppression of non-interacting pairs. To this end, this paper proposes an interaction-aware filtering module (IAF) based on relational reasoning, which filters out instances with interaction relationships through graph neural networks, thus reducing the interference of non-interaction pairs. In addition, this paper designs a novel single-stage HOI detection framework (lAF-Net) based on relational reasoning, which consists of instance branching and interaction branching, and aggregates the branching information through the instance-aware attention module (AIN) to improve the detection performance of the model. Experimental results show that the method in this paper achieves excellent performance on both HICO-DET and V-COCO datasets.