Improving Human-Object Interaction Detection via Streamlining Matching Procedure and Enhancing Interaction Query

The domain of Human-Object Interaction (HOI) detection has captured considerable attention within the computer vision research community, given its focus on grasping the action connections between humans and objects. While various methodologies, including sequential, parallel, and end-to-end techniques, have been introduced to address this challenge, recent attention has shifted toward the exploration of end-to-end systems. This paper unveils and addresses the limitations inherent in conventional multi-branch end-to-end methods. Specifically, the output matching process between decoders is identified as a potential source of additional errors in the final predictions. Furthermore, we introduce a novel cross-attention mechanism for human and object representations to mitigate duplication issues in interaction predictions. Our proposed modules are integrated into the HOTR baseline, and experiments show substantial result improvements compared to the original HOTR model. This positions our model competitively among existing methods, as evidenced by evaluations on a widely-used dataset: V-COCO.