SV-FPN: Small Object Feature Enhancement and Variance-Guided RoI Fusion for Feature Pyramid Networks

Small object detection is one of the research difficulties in detection, and Feature Pyramid Networks (FPN) a common feature extractor deep learning; thus, improving results small based on FPN great significance this field. In paper, SV-FPN proposed for task, which consists Object Enhancement (SOFE) Variance-guided Region Interest Fusion (VRoIF). When using as extractor, an SOFE module designed to enhance finer-resolution level maps from features are extracted. VRoIF takes variance RoI data driver learn completeness several different layers, avoids wasting information introducing noise. Ablation experiments three public datasets (KITTI, PASCAL VOC 07+12 MS COCO 2017) demonstrate effectiveness SV-FPN, mean Average Precision (mAP) achieves 41.5%, 53.9% 38.3%, respectively.