Decoupling Localization and Semantics for Open-Set Object Detection

Open-set object detection (OSOD) is an important research direction in computer vision, focusing on enhancing a model’s ability to detect unknown categories. Current methods are overly dependent on supervision from known categories, resulting in detection bias that substantially impairs the model’s capacity to recognize unknown classes. In this study, we propose a decoupled localization and semantic OSOD method (DLS-OSOD) that refines supervision granularity to improve unknown category perception. Specifically, to reduce the impact of inaccurate localization on classification, we propose a class-agnostic region proposal network (CA-RPN), which removes the binary classification module in the traditional RPN, allowing the model to focus on region positioning. Furthermore, to mitigate misclassification effects on localization, we design a prototype-based region filtering module (PBF), which constructs a compact prototype space using category semantics during training and pre-filters unknown regions before classification based on region-prototype distance during inference. Additionally, we propose the Unknown Feature Expansion (UFE) and Known Feature Preservation (KFP) modules. UFE enhances supervision for unknown categories by synthesizing unknown category features, improving the model’s ability to detect unknown regions. KFP constrains known-category features through textual anchors, preserving the detection performance of known categories. Experiments on benchmark datasets demonstrate the superior performance of our method.