VarifocalNet: An IoU-aware Dense Object Detector

Accurately ranking a huge number of candidate detections is a key to the high-performance dense object detector. While prior work uses the classification score or the combination of it and the IoU-based localization score as the ranking basis, neither of them can reliably represent the rank, and this harms the detection performance. In this paper, we propose to learn IoU-aware classification scores (IACS) that simultaneously represent the object presence confidence and localization accuracy, to produce a more accurate rank of detections in dense object detectors. In particular, we design a new loss function, named Varifocal Loss, for training a dense object detector to predict the IACS, and a new efficient star-shaped bounding box feature representation for estimating the IACS and refining coarse bounding boxes. Combining these two new components and a bounding box refinement branch, we build a new dense object detector on the FCOS architecture, what we call VarifocalNet or VFNet for short. Extensive experiments on MS COCO benchmark show that our VFNet consistently surpasses the strong baseline by $\sim$2.0 AP with different backbones and our best model with Res2Net-101-DCN reaches a single-model single-scale AP of 51.3 on COCO test-dev, achieving the state-of-the-art among various object detectors. Code is available at https://github.com/hyz-xmaster/VarifocalNet .