Distance-IoU Loss: Faster and Better Learning for Bounding Box Regression

The paper improves bounding box regression, a crucial step in object detection. The authors note that the commonly used ln-norm loss is misaligned with the Intersection over Union (IoU) evaluation metric, and that recent IoU and generalized IoU (GIoU) losses still suffer from slow convergence and inaccurate regression. They propose a Distance-IoU (DIoU) loss that incorporates the normalized distance between the predicted box and the target box, which converges much faster in training. They further identify three geometric factors, overlap area, central point distance, and aspect ratio, and combine them into a Complete-IoU (CIoU) loss.

Incorporating DIoU and CIoU losses into state-of-the-art detectors such as YOLO v3, SSD, and Faster R-CNN yields notable performance gains measured by both IoU and GIoU metrics. Moreover, DIoU can be adopted as the criterion within non-maximum suppression (NMS), further boosting detection performance, and the authors released source code and trained models.