Parallel Residual Bi-Fusion Feature Pyramid Network for Accurate\n Single-Shot Object Detection

Abstract

This paper proposes the Parallel Residual Bi-Fusion Feature Pyramid Network\n(PRB-FPN) for fast and accurate single-shot object detection. Feature Pyramid\n(FP) is widely used in recent visual detection, however the top-down pathway of\nFP cannot preserve accurate localization due to pooling shifting. The advantage\nof FP is weakened as deeper backbones with more layers are used. In addition,\nit cannot keep up accurate detection of both small and large objects at the\nsame time. To address these issues, we propose a new parallel FP structure with\nbi-directional (top-down and bottom-up) fusion and associated improvements to\nretain high-quality features for accurate localization. We provide the\nfollowing design improvements: (1) A parallel bifusion FP structure with a\nbottom-up fusion module (BFM) to detect both small and large objects at once\nwith high accuracy. (2) A concatenation and re-organization (CORE) module\nprovides a bottom-up pathway for feature fusion, which leads to the\nbi-directional fusion FP that can recover lost information from lower-layer\nfeature maps. (3) The CORE feature is further purified to retain richer\ncontextual information. Such CORE purification in both top-down and bottom-up\npathways can be finished in only a few iterations. (4) The adding of a residual\ndesign to CORE leads to a new Re-CORE module that enables easy training and\nintegration with a wide range of deeper or lighter backbones. The proposed\nnetwork achieves state-of-the-art performance on the UAVDT17 and MS COCO\ndatasets. Code is available at https://github.com/pingyang1117/PRBNet_PyTorch.\n