EFR-FCOS: enhancing feature reuse for anchor-free object detector.

Abstract

In this paper, we propose enhancing feature reuse for fully convolutional one-stage object detection (EFR-FCOS) to aim at backbone, neck and head, which are three main components of object detection. For the backbone, we build a global attention network (GANet) using the block with global attention connections to extract prominent features and acquire global information from feature maps. For the neck, we design an aggregate feature fusion pyramid network (AFF-FPN) to fuse the information of feature maps with different receptive fields, which uses the attention module to extract aggregated features and reduce the decay of information in process of the feature fusion. For the head, we construct a feature reuse head (EnHead) to detect objects, which adopts the cascade detection by the refined bounding box regression to improve the confidence of the classification and regression. The experiments conducted on the COCO dataset show that the proposed approaches are extensive usability and achieve significant performance for object detection.