Intersection Over Union

Intersection over Union (IoU), a metric measuring the overlap between predicted and ground-truth bounding boxes or segmentation masks, is central to evaluating object detection and segmentation models. Current research focuses on improving IoU's use in model training, including developing differentiable surrogates for direct optimization and exploring variations like mask-aware IoU for instance segmentation and rotation-decoupled IoU for 3D object detection. These advancements aim to enhance model accuracy and efficiency, particularly in real-time applications. However, recent work highlights the limitations of solely relying on IoU, emphasizing the need for more holistic evaluation methods that incorporate human perception of localization quality.