Every blue box here is the detector calling the same animal a dog. These are
its raw outputs with suppression switched off — no filtering beyond the score
floor, which you can raise. A confident object does not produce one box, it
produces a cluster, because a candidate region that nearly matches the right
answer covers nearly the same pixels and scores nearly as well. Gold, dashed,
is where the dog actually is. Raise the floor and the crowd thins, but it does
not resolve into a single answer: even at a high floor several boxes remain,
and every one of them beyond the first counts as a false positive. Something
has to choose. Every box, score and overlap is a measurement, read from the
detector's cached output.