Precision, recall, mAP and choosing the threshold

Evaluate on the test split: precision (how many found boxes are correct), recall (how many real objects were found), mAP@0.5 (overall detection quality with loose boxes) and mAP@0.5:0.95 (with tight boxes), per class. Look at the confusion matrix and PR curve. Choose the confidence threshold on the validation set so 'head' recall ≥ 0.90, then report test results at that threshold.