Generalized Intersection Over Union: A Metric and a Loss for Bounding Box Regression
Introduces Generalized IoU (GIoU), a metric and loss for bounding box regression that handles non-overlapping boxes and improves object detectors.
Intersection over Union (IoU) is the standard object-detection metric, yet common distance-based losses for regressing bounding boxes do not directly optimize it. IoU can act as a regression loss for axis-aligned 2D boxes, but its plateau makes optimization infeasible for non-overlapping boxes. The authors introduce Generalized IoU (GIoU) as both a new metric and a new loss that fixes this weakness. Adding GIoU as a loss to state-of-the-art detectors yields consistent improvements on PASCAL VOC and MS COCO under both IoU- and GIoU-based measures.
Based on: Generalized Intersection Over Union: A Metric and a Loss for Bounding Box Regression · Computer Vision and Pattern Recognition
Curated by Aramai Editorial
Read summary →