Multimodal Semantic-Probabilistic Objectness for Open World Object Detection

MSPO adds a lightweight semantic calibration module to the PROB open-world object detector. For each currently known category, it creates an extended language description covering attributes, appearance, typical scenes, and functional use, then encodes it with a frozen CLIP text encoder. Decoder-query features are projected into the same semantic space, and this known-category evidence is fused with PROB’s class-agnostic probabilistic objectness. The method does not use future-category names and does not convert OWOD into open-vocabulary classification. According to the abstract, experiments on M-OWODB and S-OWODB improve aggregate metrics, early unknown-confusion measures, and final PASCAL VOC mAP by up to 2.7 points.