Dynamic generalized IoU threshold: optimizing AP and mAP for object detection using a data-driven approach

Object detection, a key technology in computer vision applications such as facial recognition, self-driving vehicles, and industrial surveillance, is widely utilized in research and industrial settings. The evaluation of object detection algorithms relies heavily on performance metrics, but challenges arise with models like the Single Shot Detector (SSD), which features variable bounding box areas and diverse aspect ratios. These properties complicate the application of traditional metrics such as Receiver Operating Characteristic (ROC), Area Under the Curve (AUC), and Youden's J Statistic, which are better suited for classification, as well as the Jaccard Index (IoU), a localization measure. This paper introduces a data-driven approach to address these challenges by generating dynamic IoU threshold constants tailored to different object categories, guided by three constraints derived from Machine Learning and Statistics. Experiments using the 2017 COCO validation dataset demonstrate that this method enhances Average Precision (AP) and mean Average Precision (mAP) compared to fixed threshold approaches.