IoU-Balanced loss functions for single-stage object detection

Single-stage object detectors have been widely applied in computer vision applications due to their high efficiency. However, the loss functions adopted by single-stage hurt localization accuracy seriously. Firstly, cross-entropy for classification is independent of task and drives all positive examples learn as scores possible regardless accuracy. Thus, there exist many detections with but low IoU or IoU. Secondly, smooth L1 loss, gradient dominated outliers poor In this work, IoU-balanced consisting are proposed solve these problems. pays more attention enhances correlation between tasks. decreases increases IoU, which improves models. Extensive experiments on MS COCO, PASCAL VOC, Cityscapes WIDERFace demonstrate that losses can substantially improve popular detectors, especially On COCO test-dev, methods AP 1.0%∼1.7% AP75 1.0%∼2.4%. Cityscape WIDERFace, it also 1.0%∼1.5% AP80, AP90 ∼3.9%. The source code will be made publicly available.