Build from first principles

Build an Object Detector's Back Half

Write everything that happens after a detector's network, from box formats, IoU and decoding to anchor matching, NMS, and average precision. Then run it on a real trained detector and find the bug that silently sinks mAP.

What happens after the network?Anchors, offsets, scores, and why most detection bugs live here

1Compute an IoU Matrix in Any Box FormatBoxes and overlap

2Decode Predicted Offsets into BoxesAnchors and decoding

3Label Anchors as Positive, Background, or IgnoredTraining targets

4Remove Duplicate Detections with Class-Aware NMSRemoving duplicates

5Compute Average Precision for One ClassEvaluating a detector

From Pieces to a Working DetectorSee how your functions fit around a real network before you run it

Run, Score, and Debug a DetectorEvaluate a trained detector, find a decoding bug from the numbers, and tune NMS for crowds