OGMN: Occlusion-guided multi-task network for object detection in UAV images

Occlusion between objects is one of the overlooked challenges for object detection in UAV images. Due to variable altitude and angle UAVs, occlusion images happens more frequently than that natural scenes. Compared scene images, with feature confusion problem local aggregation characteristic. And we found extracting or localizing beneficial detector address this challenge. According finding, localization task introduced, which together constitutes our occlusion-guided multi-task network (OGMN). The OGMN contains two interactions. In detail, an estimation module (OEM) proposed precisely localize occlusion. Then utilizes results implement One interaction guide decoders problem, decoupling head (ODH) replace general head. Another designed process according characteristic, a two-phase progressive refinement (TPP) optimize process. Extensive experiments demonstrate effectiveness on Visdrone UAVDT datasets. particular, achieves 35.0% mAP dataset outperforms baseline by 5.3%. provides new insight accurate competitive performance.