Box-aware feedback with structured attention for HOI detection

Human-Object Interaction (HOI) detection aims to simultaneously achieve precise localization of humans and objects in images and recognition of interaction semantics. Although Transformer-based methods have made progress, they still face challenges such as insufficient task collaboration, difficulty in extracting fine-grained features, and low computational efficiency. This paper proposes a model integrating Structured Attention with Box-Aware Feedback (BAF-Satt). Its core is a box-aware feedback modulation mechanism, which coordinates localization and classification tasks by dynamically injecting geometric information of detection boxes. To capture key interaction details that are easily overlooked, we design a Cross-Region Knowledge Learning Module (CRLM), which focuses on subtle interaction cues by enhancing salient region features and integrating visual priors. Furthermore, to address the high computational cost of Transformers, we introduce an Axial-Diagonal Attention Module, which reduces computational complexity significantly through axial-diagonal decomposition of attention and embedding of positional encodings. Experiments on the V-COCO and HICO-DET datasets demonstrate that the proposed method outperforms existing state-of-the-art models, validating its effectiveness.