Shape-IoU: More Accurate Metric Considering Bounding Box Shape and Scale
This paper proposes the Shape-IoU method, which analyzes the impact of bounding box shape and scale on regression results, designs a new loss function, and achieves state-of-the-art performance on multiple detection tasks, surpassing existing methods.
Existing bounding box regression losses typically only consider the geometric relationship between predicted and ground-truth boxes, ignoring the influence of the bounding box's own shape and scale. This paper analyzes regression characteristics and proposes Shape-IoU, incorporating shape and scale factors into the loss calculation for more accurate regression. Experiments show that this method effectively improves detection performance and achieves optimal results across different tasks.
The Shape-IoU method first analyzes bounding box regression characteristics and finds that shape and scale factors significantly affect regression results. Based on this, a new loss function is designed that introduces shape and scale weights when calculating IoU, making the model focus more on the inherent properties of bounding boxes. Experiments on multiple datasets (e.g., COCO, PASCAL VOC) show that this method outperforms existing methods such as GIoU, DIoU, and CIoU in tasks like object detection and instance segmentation, with notable improvements on small objects and irregularly shaped objects. The code is open-sourced.
This work has direct value for improving the accuracy of object detection systems in industry, especially in scenarios with high precision requirements such as autonomous driving, security surveillance, and industrial quality inspection. It can replace existing loss functions to enhance model performance.
It is recommended to replace existing IoU losses with Shape-IoU in object detection model training. It can be integrated into existing detection frameworks (e.g., MMDetection, Detectron2) to improve detection accuracy, especially for small objects and irregularly shaped objects.
Future work can verify the performance of Shape-IoU in more complex scenarios (e.g., occlusion, dense objects) and its compatibility with other detection frameworks (e.g., YOLOv8, DETR).