CR-FPN: channel relation feature pyramid network for object detection

Abstract

Object detection is an important computer vision task, which aims to locate each object and classify them correctly. Since Convolution Neural Network (CNN) high-level feature map contains more semantic information and low-level feature map contains detail information which helps to locate precisely, fusing high-level feature maps with low-level feature maps is proven to be essential for multi-scale object detection. Recent work usually direct combine feature maps from different levels of CNN. However, as each feature map has different extents of discrimination, combining them directly may turn some valuable information into noise and further degrade performance. To address this issue, we focus on the channel relationship from different level of feature maps and propose a novel network, named Channel Relation Feature Pyramid Network (CR-FPN), that captures long-range relation of channel from different level of feature maps by similarity measure function and further to magnify the most relevant channels and suppress the irrelevant channels. Extensive experiments conducted on MS COCO and Pascal VOC datasets demonstrate the effectiveness of our CR-FPN also with competitive performance. Noting that our CR-FPN can apply to several typical state-of-the-art detectors. Using CR-FPN in a basic Cascade R-CNN detector, our method achieves state-of-the-art single model results on the COCO dataset.