An attention-based feature pyramid network for single-stage small object detection

Recently, single-stage detection methods have made great progress, achieving comparable accuracy to two-stage detection methods. However, they have poor performance over small object detection. In this work, we improve the performance of the single-stage detector for detecting objects of small sizes. The proposed method makes two major novel contributions. The first is to devise an attention-based feature pyramid network (aFPN) by introducing a learnable fusion factor for controlling feature information that deep layers deliver to shallow layers. The design of a learnable fusion factor could adapt a feature pyramid network to small object detection. The second contribution is to propose a soft-weighted loss function, which reduces the false attention during network training. To be specify, we reweight the contribution of training samples to the network loss according to their distances with the boundaries of the ground-truth box, leading to fewer false-positive detections. To verify the performance of the proposed method, we conduct extensive experiments on different datasets by comparing including RetinaNet, ATSS, FCOS, FreeAnchor, et al. Experimental results show that our method can achieve 44.2% AP on MS COCO dataset, 23.0% AP on VisDrone dataset, which significantly gains improvements with nearly no computation overhead.