Generalized Intersection over Union (GIoU): A better metric and loss for object detection

Hamid Rezatofighi, Nathan Tsoi, JunYoung Gwak, Amir Sadeghian, Ian Reid, Silvio Savarese

2019

Comparison between IoU and Generalized IoU for bounding box regression

Publication summary

This paper introduces Generalized Intersection over Union (GIoU), an extension of the widely used Intersection over Union (IoU) metric for object detection and bounding box regression.

Traditional IoU performs poorly when predicted and ground-truth boxes do not overlap because the metric becomes zero and provides no gradient for optimization. GIoU addresses this limitation by incorporating the smallest enclosing region around both boxes, enabling meaningful optimization even in non-overlapping cases.

The authors demonstrate that replacing standard regression losses with GIoU loss consistently improves performance across major object detection architectures including YOLOv3, Faster R-CNN, and Mask R-CNN on PASCAL VOC and MS COCO benchmarks.

“GIoU transformed bounding box regression by directly optimizing a localization metric that remains informative even when predictions and targets do not overlap.”

Editorial research summary

OrthoAI research content adaptation

Key findings

+6.69%

YOLOv3 achieved up to 6.69% AP improvement on MS COCO validation when trained with GIoU loss instead of MSE loss.

+5.95%

Faster R-CNN achieved notable AP improvements on PASCAL VOC using GIoU-based bounding box regression.

-1 to 1

Unlike IoU, GIoU provides meaningful values for non-overlapping boxes with a symmetric score range.

Method

Generalized Intersection over Union

GIoU extends IoU by measuring not only overlap but also the normalized empty space between two bounding boxes using the smallest enclosing box.

This modification provides informative gradients even when bounding boxes do not intersect, solving one of the primary optimization limitations of IoU.

Generalized IoU formulation and bounding box comparison

Direct optimization for object detection

The authors derive an analytical formulation allowing GIoU to be directly used as a regression loss for axis-aligned bounding boxes.

Experiments replace traditional L1, smooth-L1, and MSE regression losses with GIoU loss in YOLOv3, Faster R-CNN, and Mask R-CNN training pipelines.

Object detection training using GIoU loss

Key contributions

GIoU became one of the most influential localization losses in computer vision and serves as the foundation for later losses such as DIoU, CIoU, and EIoU.

Introduces Generalized Intersection over Union as an extension of IoU.
Provides a metric that remains informative for non-overlapping bounding boxes.
Derives a differentiable loss suitable for deep learning optimization.
Improves localization performance across YOLOv3, Faster R-CNN, and Mask R-CNN.
Maintains scale invariance while better capturing spatial relationships between boxes.
Established a new standard for bounding box regression in modern object detection systems.

Similar publications

CVPR 2021

DoDNet: Learning to Segment Multi-Organ and Tumors from Multiple Partially Labeled Datasets

Jianpeng Zhang, Yutong Xie, Yong Xia, Chunhua Shen

view

CVPR 2024 · IEEE/CVF Conference on Computer Vision and Pattern Recognition

PairAug: What Can Augmented Image-Text Pairs Do for Radiology?

Yutong Xie, Qi Chen, Sinuo Wang, Minh-Son To, Iris Lee, Ee Win Khoo, Kerolos Hendy, Daniel Koh, Yong Xia, Qi Wu

view

view all