Enhanced Sparse Detection for End-to-End Object Detection

Abstract

In this paper, we propose an enhanced end-to-end object detector based on Sparse R-CNN (EnSparse R-CNN), which aims at backbone, neck and head of object detector. For the backbone network, we constructure an aggregate residual network (ARNet) with the aggregate connection, which can achieve global features to improve the ability of feature extraction by extracting the salient feature from the feature maps. For the neck network, we build an attention context feature pyramid network (ACFPN) to fuse the global features with the context features, which can integrate multi-layer features and attention features to make full use of global context features. For the detection head, we design a gradually refined detection head (ReHead) to detect objects, which adopts the sparse detection with the refined bounding box regression to improve the confidence of the regression. The experimental results on COCO dataset show that EnSparse R-CNN outperforms Sparse R-CNN by 2.5 % and the anchor-free one-stage detector AutoAssign by 1.1%.

Keywords

• EnSparse R-CNN
• ARNet
• ACFPN
• ReHead
• object detection