Unveiling the Unknown: A SAM Guided Open World Object Detection Method for Remote Sensing

Despite the remarkable success of remote sensing object detection, these methods primarily operate under a closed-world paradigm, as they often misclassify or ignore novel objects in real world scenarios. To address this limitation, Open World Object Detection (OWOD) has emerged, enabling the discovery and incremental learning of new categories. However, existing OWOD approaches typically distinguish between known and unknown categories based on feature distance, overlooking the inherent challenge in remote sensing: large intra-class variation versus small inter-class variation. To bridge the gap, we propose a novel framework for SAM guided OWOD tailored for remote sensing imagery. Our approach is designed to leverage SAM’s capabilities for discovering new categories while systematically handling the noisy labels produced by SAM. We introduce four key components: (I) a Multi-scale Feature Fusion Perception (MFFP) module to enhance the detection of unknown objects across various scales in remote sensing; (II) a Cross-layer Cascaded Decoupling Decoder (CCDD) to alleviate the optimization conflicts between objectness and classification tasks for similar known and unknown classes in remote sensing images; (III) a Label Mapping Alignment (LMA) mechanism to adaptively filter background noisy proposals from SAM. And (IV) an Active Learning (AL) strategy is proposed to intelligently select exemplars for robust incremental learning. Extensive experiments on benchmark remote sensing datasets, including DIOR, DOTA, and NWPU demonstrate that our method significantly improves the recall of unknown objects while maintaining robust detection performance for known classes, demonstrating the effectiveness and potential of the SAM guided OWOD paradigm.