Generalized Intersection Over Union: A Metric and a Loss for Bounding Box Regression

The paper addresses a mismatch in object detection: Intersection over Union (IoU) is the most popular evaluation metric, but the distance-based losses commonly used to regress bounding box parameters do not directly optimize it. The authors note that the optimal objective for a metric is the metric itself, and show that for axis-aligned 2D bounding boxes IoU can be used directly as a regression loss. However, IoU has a plateau that makes it infeasible to optimize when bounding boxes do not overlap, which motivates a generalized formulation.

To address this weakness, the authors introduce Generalized IoU (GIoU) as both a new loss and a new metric for bounding box regression. By incorporating GIoU as a loss into state-of-the-art object detection frameworks, they demonstrate consistent performance improvements on popular benchmarks such as PASCAL VOC and MS COCO, measured with both the standard IoU-based and the new GIoU-based performance measures. This showed that GIoU serves as an effective drop-in loss for training object detectors.