Non-Maximum Suppression: How Object Detectors Pick One Box

A convolutional object detector rarely produces a single tidy box per object — it produces a cloud of overlapping candidates at slightly different positions, sizes and confidence scores, all pointing at the same thing. This simulator renders that cloud as a 3D "skyline" over a scene: each candidate detection is a tower whose height and colour encode its confidence, floating above the ground-truth object it's trying to describe. Tune the confidence threshold to prune weak proposals outright, tune the IoU (intersection-over-union) threshold to control how aggressively overlapping boxes get merged, then run the greedy non-maximum-suppression algorithm and watch it eliminate duplicates one step at a time until only the strongest box per object survives — the exact mechanism running inside every real-time object detector, from YOLO to Faster R-CNN.