DA-FPN: Deformable Convolution and Feature Alignment for Object Detection

This study sought to address the problem of insufficient extraction shallow object information and boundary when using traditional FPN structures in current detection algorithms, which degrades accuracy. In this paper, a new structure model, DA-FPN, is proposed. DA-FPN replaces 1 × convolution used conventional for lateral connection with 3 deformable adds feature alignment module after 2x downsampling operation connection. design allows framework extract more accurate about object, particularly small objects. A bottom-up was also added incorporate accurately into high-level map, module, thereby improving The experimental results show that can improve accuracy single-stage algorithms FoveaBox GFL by 1.7% 2.4%, respectively, on MS-COCO dataset. model found two-stage algorithm SABL 2.4% offer higher better robustness.