Source-Grouped Recall-Matched Calibration for Reducing Detector Review Queues

Abstract

A detector threshold chosen on the evaluation set gives an optimistic account of deployment burden, whereas a universal threshold ignores shifts in score scale. We study a source-grouped, recall-matched policy that uses 30% of acquisition groups for calibration, freezes the highest observed threshold that reaches fixed-threshold baseline recall, and evaluates the remaining 70%. Across five paired YOLOv10m seeds, the policy reduced false positives per image by 19.1% on public HeinSight4.0 data and by 51.0% on a difficult acquisition holdout. Candidate output fell by 6.9% and 21.6%, respectively. Group-aware confidence intervals resolved both false-positive reductions and both candidate reductions; the difficult-set recall difference remained unresolved. The procedure changes only the operating threshold, requires no retraining, and makes the calibration-evaluation boundary explicit. It therefore provides a practical route from detector scores to a smaller review queue without selecting the threshold on reported outcomes.