Bounding Boxes: Saying Where an Object Is | Visual AI Learning Guide

Each rectangle in the diagram is defined by four numbers: where one corner sits, and how wide and tall the box is. Notice that the two dogs get two separate boxes even though they share a class label — the box is what keeps them apart. Now look at the cluster of overlapping boxes around the car. A detector proposes many boxes during its search, and most of them are redundant. When two boxes overlap heavily and agree on the class, the weaker one is dropped. That step is what turns a messy pile of candidate boxes into one clean box per object.

0:00 / 0:00

Auto-play

A bounding box is the smallest upright rectangle that contains an object, described by four numbers: the horizontal and vertical position of one corner, and the width and height of the box. A detector predicts one box for each object it finds, together with a class label and a confidence score. Because each object gets its own box, two objects of the same class — two dogs, for example — are reported as two separate entries rather than merged into one. To decide which candidate boxes to keep, detectors compare overlapping boxes and suppress duplicates: when two boxes overlap heavily and predict the same class, the one with lower confidence is discarded. This is how a detector avoids reporting the same dog five times. The box is a coarse description of location; it does not trace the object's outline, but it is enough to say where the object is and to separate it from its neighbors.