Side-path FPN-based multi-scale object detection

Multi-scale object detection faces the problem of how to obtain distinguishable features. Feature pyramid network (FPN) is the most typical work to construct a feature pyramid to obtain multi-scale features, and is beneficial for multi-scale object detection tasks to improve the mean average precision (mAP) of the detectors. However, due to the lack of feature selection to eliminate redundant information, FPN cannot make full use of multi-scale features. In this paper, side-path FPN is proposed to address this problem. Side-path FPN contains two components: feature alignment and feature fusion. The feature alignment component uses the best operator to extract features. The feature fusion component can enhance features that are helpful for detection and reduce redundant information. With ResNet-50 as the backbone, compared to the original FPN, side-path FPN improves mAP by 1.8 points on the VOC2007 test dataset and 1.0 point on the COCO 2017 test dataset with MS COCO metrics.