Every box, score and overlap here is a measurement, read from the detector's cached output. Gold, dashed, is the ground truth: two boxes drawn by hand for this book, on a photograph that holds more than two things. The nine numbered boxes are everything torchvision's Faster R-CNN R50-FPN v2 emits above its default score floor of 0.05 — blue where the class agrees and the overlap clears 0.5, grey where it does not. Point at a row and its box lights up; point at a box and its row does. That is the only way to separate boxes 1 and 8, which sit within a couple of pixels of each other on the ball: the orange scores a better overlap than the correct sports-ball detection and is still worth nothing, because a match requires the class to agree. Four of the nine are stacked on one rock.