ProCal Improves Open-Vocabulary Object Detection via Proposal Calibration
July 3, 2026
ProCal enhances open-vocabulary object detection by calibrating classification scores during inference. It uses a localization-aware foreground score and a background-aware suppression score to align VLM classification with object position and scale.
HOW THIS AFFECTS YOU
●
builderThis provides a method to improve detection accuracy in open-vocabulary systems using simple inference-time adjustments.
●
researcherYou can use this method to improve detection quality without retraining frozen VLM backbones.