We are delighted to announce that 1 paper from VI Lab (Prof. Jongwon Choi) has been accepted for publication in Pattern Recognition Letters.

Title: 

Annotation-free Confidence Calibration of Object Detectors via Object Tracking

Authors: 

Dongyoung Lee, Juyong Lee, Jaeyoon Lee, Jongwon Choi

Abstract: 

Post-hoc calibration of object detectors typically requires ground-truth boxes, which may be unavailable when models are deployed in a new domain. This study replaces manual localization annotations with temporally consistent bounding boxes obtained from video object tracking. Detector class predictions and tracking geometry provide supervision for a confidence calibrator, while an intersection-over-union-aware weighted loss reduces the influence of tracking errors. Experiments show improved localization-aware calibration in unlabeled target domains and competitive results against supervised calibration baselines in the original domain.