RFP-Net: Receptive field-based proposal generation network for object detection

Abstract Recently, object detection has achieved great improvements due to deep CNNs. In this paper, we propose a novel proposal generation network named RFP-Net by mimicking human visual system for high-quality proposals generation. Specifically, RFP-Net takes receptive fields (RFs) as reference boxes to remove many hyper-parameters of anchor boxes that have large sensibility to object detection results. During network training, we select positive samples using an effective RF (eRF) rule instead of the Intersection-over-Union (IoU) rule, which only requires the centroid of a ground truth box to be within the eRF region. This renders RFP-Net learn the representation of region proposals not limited to be of a fixed range of scales and accurately localize the bounding boxes of region proposals around the eRF. RFP-Net also solves the imbalance problem between negative and positive samples with less computational cost. The proposed RFP-Net significantly improves multiply state-of-the-art two-stage and multi-stage detectors. For example, it achieves 43.1% AP by combined it with Cascade RCNN on MS COCO dataset, outperforming previous approaches.