SAFPN: spatial adaptive feature pyramid network for small object detection

In recent years, Feature Pyramid Network (FPN) and its variants have significantly improved the performance of object detection by fusing multi-scale features. However, current FPN-based methods still face challenges such as information loss and feature confusion, which result in insufficient discriminative information for small objects. In this paper, we propose a novel spatial adaptive feature pyramid network (SAFPN), whose core innovation lies in the utilization of spatial information to drive the adaptive enhancement and fusion of multi-scale features. Specifically, SAFPN contains three functionally complementary modules: Lateral Adaptive Representation Module (LARM), Spatial Feature Correction Module (SFCM), and Dynamic Feature Fusion Module (DFFM). LARM is designed to expand the receptive field and enhance channel representation. SFCM is established to compensate for the loss of spatial information caused by upsampling and to correct the spatial misalignment of adjacent feature maps simultaneously. DFFM aims to alleviate feature confusion by generating adaptive fusion factors during the feature fusion process. Experimental results demonstrate that SAFPN significantly improves the performance on small objects on MS COCO, PASCAL VOC, and TT-100K datasets and is also highly competitive compared to other advanced FPN methods. In addition, our proposed SAFPN can be integrated into various FPNs and object detectors.