bounding box
/ BOWND-ing boks /
A bounding box is the simplest way to say where an object is in an image: an upright rectangle drawn just large enough to contain it. It is described by four numbers — typically the position of one corner plus a width and height, or the coordinates of two opposite corners. When an object detector announces "there's a car here," the "here" is a bounding box. It is the basic unit of localization in computer vision, deliberately crude so it is cheap to predict, store, and reason about.
Because boxes are so simple, they have a clean way to be scored. To check whether a predicted box is good, we compare it to the human-drawn correct box using Intersection over Union (IoU): the area where the two boxes overlap, divided by the total area they jointly cover. IoU is 1 for a perfect match and 0 for no overlap; a detection usually counts as correct above some threshold like 0.5. This single number is how detection accuracy is measured across the whole field.
The thing to internalize is what a box throws away. An upright rectangle cannot hug a tilted, round, or sprawling shape, so it always swallows some background — a box around a bicycle is mostly air between the frame and wheels. It cannot represent a precise outline, separate two objects that overlap inside the same rectangle, or capture orientation. When those things matter, you reach for richer outputs: rotated boxes, keypoints, or pixel masks from segmentation. The bounding box endures not because it is accurate but because it is a fast, good-enough handle on "where."
A detector predicts a box for a dog at corner (100, 60) with width 80 and height 50. The human-labeled correct box is at (105, 65), width 75, height 48. They overlap heavily, giving an IoU of about 0.82 — well above the usual 0.5 threshold, so this counts as a correct detection.
Four numbers locate the object; IoU against the true box turns "close enough" into a precise score.
A bounding box is always an axis-aligned rectangle, so it inevitably includes background around non-rectangular objects and cannot separate overlapping things or represent orientation. It is a deliberately coarse handle on location — for exact shape you need segmentation masks, not boxes.