A Distance-Based Attention Mechanism for Object Detection
Xiao Hui Bai, Chengzhi Liang, Jianqun Zhou · 2022
Attention mechanisms have been widely studied and applied in various computer vision tasks because of their ability to establish inter-dependencies between channels and spatial positions. Recent works show it necessary to provide light-weight and effective implementations for higher computation efficiency. To this end, this paper proposes to restrict the attention to a local region for object detection and presents a distance-based attention mechanism to incorporate contextual information from regional areas, instead of the full image, in an efficient and effective way. Furthermore, we implement the proposed approach in an object detection model that combines tricks from several state-of-the-art models. Specifically, the popular one-stage RetinaNet is used as the backbone and the two-step classification and regression, feature enhancement module, max-out operation and multi-scale testing are adopted to obtain a high detection accuracy. Experimental results on WIDER FACE show that the proposed detection model outperforms many state-of-the-art models and the proposed attention mechanism can greatly increase the detection accuracy without introducing extra computational overheads.