Joint confidence for semisupervised object detection in remote sensing imagery

To address the limitations of existing semi-supervised object detection methods, which rely on classification scores for pseudo-label filtering and overlook localization precision, a joint-confidence-based SSOD framework is proposed for remote sensing imagery. The framework integrates a student-teacher architecture with the Double Heads detector, enhanced by an IoU-aware branch to predict localization quality. A joint confidence metric is formulated by combining classification scores and IoU-based localization scores, enabling holistic pseudo-label selection. Weak and strong augmentation strategies are applied to unlabeled data, and the framework is trained using supervised and unsupervised losses. Experiments on NWPU VHR-10 and RSOD datasets demonstrate improvements: at 10% labeled data, the method achieves 41.6% AP, 66.7% AP₅₀, and 51.1% AP₇₅ on NWPU VHR-10. The experiment result reveals that joint confidence metrics mitigate misjudgments in complex scenarios by prioritizing classification accuracy and spatial alignment. This work advances pseudo-label quality in SSOD, particularly for remote sensing applications with sparse labeled data.