For a project I want to evaluate a custom computer-vision system I made for which I can't really use libraries and therefore want to implement the mAP.50 method by myself. As I have been doing some digging regarding the mAP.50 metric for object detection, I found a lot of inconsistent information as to what constitutes a false negative. Some sources say that ground truth (GT) bounding boxes are considered only a false negative if there is no prediction bounding box sufficiently close (IoU < IoU threshold). The sources I found agreed unanimously on that, but what if there is an predicting bounding box with sufficient IoU and confidence, but with the wrong class predicted? The prediction will get treated as a false negative, but what about the GT bounding box? Do I label that one as a false negative? I am curious as to what you guys think about this. Continue reading...
