Open Access

Jan Stodt, Christoph Reich, Nathan Clarke

• Data scientists, researchers and engineers want to understand, whether machine learning models for object detection work accurate and precise. Networks like Yolo use bounding boxes as a result to localize the object in the image. The principal aim of this paper is to address the problem of a lack of an effective metric for evaluating the results of bounding box regression in object detection networks when boxes do not overlap or lie completely within each other. The standard known metrics, like IoU, lack of differentiating results, which do not overlap but differ in the distance between predicted bounding box and label. To solve this challenge, we propose a new metric called UIoU (Unified Intersection over Union) that combines the best properties of existing metrics (IoU, GIoU and DIoU) and extends them with a similarity factor. By assigning weight to each component of the metric, it allows for a clear differentiation between the three possible cases of box positions (not overlapping, overlapping, boxes inside each other). The result of this paper is a new metric that outperforms the existing metrics such as IoU, GIoU and DIoU by providing a more understandable measure of the performance of object detection models. This provides researchers and users in the field of explainable AI with a metric that allows the evaluation and comparison of prediction and label bounding boxes in an understandable way.