Gated Feature Pyramid Network for Object Detection

Feature pyramid is a basic component in recognition systems for detecting objects of different scales. In order to construct the feature pyramid, most existing deep learning methods combine features of different levels based on a pyramidal feature hierarchy (e.g. SSD, Faster-RCNN). However, it lacks attention to those informative features. In this paper, we propose a gated feature pyramid network (GFPN) extracting informative features to enhance the representation ability of feature pyramid. GFPN consists of gated lateral modules and a top-down structure. The former automatically learns to focus on informative features of different scales, and the latter is used to combine the refined features. By using GFPN on SSD, our method achieves 80.1 mAP on VOC 2007 with an inference time of 11.9 ms per image, which improves the accuracy of FPN applied to SSD by 0.5% and adds marginal efficiency cost.