Focal Loss for Region Proposal Network

摘要

Currently, most state-of-the-art object detection models are based on a two-stage scheme pioneered by R-CNN and integrated with region proposal network (RPN), which is served as proposal generation. During the training of RPN, only a fixed number of samples with a fixed objectot-object ratio are sampled to avoid class imbalance problem. In contrast to the sampling strategies, focal loss is utilized to solve the class imbalance problem by down-weighting the losses of vast number of easy samples, which is encountered in one-stage detection methods. Inspired by this, we investigate the adaptation of focal loss to RPN in this paper, which allow us to train RPN free of the sampling process. Based on Faster R-CNN, we adapt focal loss to RPN and the experimental results on PASCAL VOC 2007 and COCO datasets outperform the baseline, which shows the efficiency of the proposed method and implies that focal loss can be applied to RPN directly.