Dynamic Generalized IoU Threshold: Optimizing AP and mAP for Object Detection Using a Data-Driven Approach

Object detection as a pivot technology plays a vital role in computer vision applications like self-driving vehicles, virtual reality, and various surveillance applications. For implementation, object detection needs to recognize multiple objects and their localization in an image, which complicates the recognition in the models. However, multiple objects and their localization in an image are determined by the Intersection of Union (IoU), which is empirically set to a constant value of 0.5 universally to all object classes. In this paper, aiming at optimizing performance metrics (including Average Precision (AP) and mean Average Precision (mAP)) for object detection, three constraints from the perspective of both Machine Learning and statistics are considered to get a dynamic IoU threshold for class-specific object detection. With the datasets of COCO 2017 validation, the experiments have verified that the optimized AP and mAP can be achieved by dynamic generation of the IoU thresholds for class-specific object detection.