Anchor boxes are a technique used in some computer vision object detection algorithms to help identify objects of different shapes.

Anchor boxes are hand-picked boxes of different height/width ratios (for 2-dimensional boxes) designed to match the relative ratios of the object classes being detected. For example, an object detector that detects cars and people may have a wide anchor box to detect cars and a tall, narrow box to detect people.

The Fast R-CNN paper introduced the idea of using the \(k\)-means-clustering to automatically determine the appropriate anchor box dimensions for a given \(k\) number of anchor boxes.