Annotation-free confidence calibration of object detectors via object tracking

Citations

WEB OF SCIENCE

0
Citations

SCOPUS

0

초록

Reliable deployment of object detectors requires confidence scores that reflect actual detection quality, but existing calibration methods have limitations in optimizing without ground truth. We propose an annotation-free calibration framework for object detection that exploits temporal consistency in video to train a calibrator. Because detection boxes can be unreliable in unlabeled domains, the proposed framework uses tracking boxes temporally consistent localization references for calibration and combines them with predicted class labels the detector to form calibration targets. We also introduce an IoU-aware weighted loss to mitigate noisy supervision from tracking errors by assigning larger importance to training samples with higher IoU between the detection and tracking boxes. Experiments show that the proposed method enables effective calibration in unlabeled target-domain, yielding large improvements on localization-aware calibration metrics. In the in-domain setting, the proposed method remains competitive with supervised calibration baselines across CNN-based and Transformer-based detectors.

키워드

Post-hoc calibration; Object detection
제목
Annotation-free confidence calibration of object detectors via object tracking
저자
Lee, Dongyoung; Lee, Juyong; Lee, Jaeyoon; Choi, Jongwon
DOI
10.1016/j.patrec.2026.09.016
발행일
2026-11
유형
Article
저널명
Pattern Recognition Letters
권
209
페이지
194 ~ 201