Structure-aware distillation via graph consistency for source-free object detection

Object detection in IoT and smart sensing systems often encounters significant domain shifts caused by varying environments, heterogeneous sensors, and dynamic operating conditions. In many practical deployments, source data cannot be accessed due to privacy, storage, or bandwidth constraints, making source-free domain adaptation essential for maintaining reliable performance. To address these challenges, this paper presents a structure-aware graph distillation framework designed to enhance domain-robust detection in distributed sensing environments. The proposed method integrates three mutually reinforcing components. A Category-Adaptive Pseudo-label Weighting (CAPW) module learns class-specific confidence thresholds and soft weights to suppress label noise and rebalance rare categories. A Foreground-Guided Occlusion Augmentation (FOA) strategy applies targeted occlusions on teacher-detected foreground regions and enforces occlusion-aware consistency to improve robustness to partial visibility. Furthermore, a Distribution-Spatial Graph Distillation (DSGD) module aligns semantic and spatial dependencies between teacher and student models through graph-level consistency, supported by an auxiliary structural inference network. Comprehensive experiments on various benchmark datasets under multiple domain shifts—such as weather changes, sensor discrepancies, and synthetic-to-real scenarios—show that the proposed framework delivers stable and significant gains compared with cutting-edge source-free and unsupervised domain adaptation methods, confirming its utility in IoT and intelligent sensing contexts.