RSSD: Object Detection via Attention Regions in SSD Detector

This paper designs a module of attention regions in SSD detector for accurate and efficient object detection (RSSD). Different from previous one-stage detection method like SSD which just simply applied the multi-scale head-features and directly extracted from backbone network, for classification and regression, our method aims to strengthen the characterization of head-features further. The parallel encode-to-decode structure is constructed and a computation method of regional distribution on features (R-Softmax) is proposed. What’s more, in order to reduce time-costs, the down-sampling layers are shared with the multi-scale layers from backbone network. Our detector performs better on PASCAL VOC datasets (e.g., 78.4% mAP V.S. SSD 76.4% on VOC 07test) and costs 0.001s per image more than SSD.