For each image, a variable-length list of (box, class, confidence). Three design questions follow from that one fact:

• · How many boxes to output?
• · Which prediction matches which ground truth?
• · How to merge nearby duplicate boxes?

Different families of detectors answer these three questions in different ways.