Non-Maximum Suppression

The problem, from first principles

A detector fires many overlapping boxes around the same object, each with a confidence score. Reporting all of them would count one cat five times. Non-maximum suppression (NMS) keeps the best and discards its near-duplicates: sort the boxes by score; take the highest-scoring box and keep it; remove every remaining box whose IoU with it exceeds a threshold; repeat with what is left. A box far from the kept one survives and may be kept later, so two different cats are both detected. The threshold is a trade-off: too low suppresses neighbouring objects, too high leaves duplicates.

From theory to code

Implement nms.

Constraints

• boxes has shape (N, 4) as (x1, y1, x2, y2), scores has shape (N,) and iou_thresh is a float.
• Process boxes in order of descending score (ties by lower index first). Keep a box unless a previously kept box has IoU strictly greater than iou_thresh with it.
• Return the indices of the kept boxes (into the original arrays) as a list in the order they were kept.
• Include your own IoU computation. Do not modify the inputs.

Hints