What the detector hands to non-maximum suppression for the two classes that are really in this photograph: 59 candidate boxes from one run of Faster R-CNN R50-FPN v2. That same run proposes 90 boxes over 7 classes; the other 31 are for classes the scene does not contain, and are left out of this demonstration. Press step. The algorithm takes the highest-scoring candidate still in the queue (blue), deletes every remaining box that overlaps it by more than the IoU threshold (gold), and repeats until the queue is empty. The ground truth is one dog and one ball, drawn by hand for this book, so any survivor past two is a duplicate. Predict before you drag: which end of the threshold hurts? From 0.73 up, duplicates survive. The other end costs nothing among these candidates — the largest overlap between any dog candidate and any sports ball candidate is only 0.016, so the two objects can suppress each other only with the class filter off and the threshold below that: 0.00 or 0.01 on this slider.