Computer Vision
Boxes and Overlap
A detector does not hand you one box for one object. It hands you a heap of overlapping guesses with scores on them, and two rules turn that heap into an answer: a number for how much two boxes are the same box, and a procedure that keeps the best and deletes its neighbours. Step through it one decision at a time.
a real objectkeptdeleted for overlapping a better onestill waiting its turn
Stage 2 of 2: thin the heap · step through the decisions, or click any box to compare it yourself
After the overlap rule—
—
Values
boxes in the heap—
kept—
deleted—
objects found—
false alarms—
how well they sit—
overlap being judged—
What to observe
- This is stage 2 of the pipeline thatFinding an Objectstarts: a heap of scored, overlapping guesses (drawn here on a robot scene of its own). Step through non-maximum suppression: keep the best box left, delete what overlaps it by more than the threshold.
- Raise the overlap that counts as the same box to 0.8 and only near-identical boxes delete each other, so each robot keeps a stack.
- Drop it to 0.2 and the small robot in front of the first disappears: their real boxes already have an IoU of 0.37. No single threshold handles crowds.
- The two false alarms overlap nothing, so no overlap rule removes them. Raisethe score needed to enter to 0.6 and they are gone.
Shortcuts:space run/pause · s step · r reset · f fullscreen