IoU-uniform R-CNN: Breaking through the limitations of RPN

• We reveals the importance of solving limitations RPN and our proposed IoU-uniform R-CNN can alleviate IoU distribution imbalance inadequate training samples by generating with uniform distribution. improve performance prediction branch eliminating feature offsets RoIs at inference. Our method consistently obtains significant improvements over multiple state-of-the-art detectors. Specially, it achieves 2.4 AP improvement than Faster (with ResNet 101-FPN backbone) on MS COCO dataset. Region Proposal Network (RPN) is cornerstone two-stage object It generates a sparse set proposals alleviates extrem foreground-background class problem during training. However, we find that potential detector has not been fully exploited due to quantity generated RPN. With increasing intersection union (IoU), exponentially smaller numbers positive would lead skewed towards lower IoUs, which hinders optimization high levels. In this paper, break through RPN, propose IoU-Uniform R-CNN, simple but effective directly for regression as well branch. Besides, inference, helps NMS procedure preserving accurately localized bounding box. Extensive experiments PASCAL VOC dataset show effectiveness method, its compatibility adaptivity many detection architectures. The code made publicly available https://github.com/zl1994/IoU-Uniform-R-CNN .