2019/02/25 by Hamid Rezatofighi, Nathan Tsoi, Rezatofighi, Hamid +9 · 107 citations
Computer Science · #Advanced Image and Video Retrieval Techniques #Advanced Neural Network Applications #Artificial Intelligence (cs.AI) #Computer Vision and Pattern Recognition (cs.CV) #Domain Adaptation and Few-Shot Learning #FOS: Computer and information sciences #Machine Learning (cs.LG)
paper · pdf · doi:10.48550/arxiv.1902.09630
openalex publication_date 2019/02/25 · openalex created_date 2025/10/10 · openalex updated_date 2026/07/28
Intersection over Union (IoU) is the most popular evaluation metric used in\nthe object detection benchmarks. However, there is a gap between optimizing the\ncommonly used distance losses for regressing the parameters of a bounding box\nand maximizing this metric value. The optimal objective for a metric is the\nmetric itself. In the case of axis-aligned 2D bounding boxes, it can be shown\nthat IoU can be directly used as a regression loss. However, IoU has a\nplateau making it infeasible to optimize in the case of non-overlapping\nbounding boxes. In this paper, we address the weaknesses of IoU by\nintroducing a generalized version as both a new loss and a new metric. By\nincorporating this generalized IoU (GIoU) as a loss into the state-of-the\nart object detection frameworks, we show a consistent improvement on their\nperformance using both the standard, IoU based, and new, GIoU based,\nperformance measures on popular object detection benchmarks such as PASCAL VOC\nand MS COCO.\n