SC-RPN: A Strong Correlation Learning Framework for Region Proposal

Current state-of-the-art two-stage detectors heavily rely on region proposals to guide the accurate detection for objects. In previous proposal approaches, interaction between different functional modules is correlated weakly, which limits or decreases performance of approaches. this paper, we propose a novel strong correlation learning framework, abbreviated as SC-RPN, aims set up stronger relationship among in task. Firstly, Light-weight IoU-Mask branch predict intersection-over-union (IoU) mask and refine classification scores well, it used prevent high-quality from being filtered. Furthermore, sampling strategy named Size-Aware Dynamic Sampling (SADS) proposed ensure consistency stages. addition, point-based representation exploited generate with fitting ability. Without bells whistles, SC-RPN achieves AR 1000 14.5% higher than that Region Proposal Network (RPN), surpassing all existing We also integrate into Fast R-CNN Faster test its effectiveness object task, experimental results achieve gain 3.2% 3.8% terms mAP compared original ones.