SRGT: Spatial-Resolution-Guided Teacher for Robust Semi-Supervised Aerial Object Detection

Semi-supervised object detection (SSOD) in aerial imagery is significantly hampered by extreme object scale variations, primarily due to diverse spatial resolutions inherent in aerial data, measurable through Ground Sample Distance (GSD). Existing SSOD methods often overlook this crucial GSD information, leading to unreliable pseudo-labels when GSD distributions mismatch between labeled and unlabeled data. To address this issue, we propose the Spatial-Resolution-Guided Teacher (SRGT), a novel framework that explicitly leverages GSD metadata. SRGT incorporates a temporal incremental weighted updating mechanism to efficiently track category-wise GSD distributions and integrates two GSD-based weighting strategies into the loss computation. These mechanisms dynamically adjust pseudo-label importance and unsupervised loss contribution, fostering a virtuous cycle of improved pseudo-label quality and enhanced teacher model's robustness with minimal overhead. Extensive experiments on multiple challenging benchmarks show significant performance gains of SRGT when built upon three baseline SSOD methods, validating its efficacy in mitigating GSD-induced noise and enhancing semi-supervised training for robust aerial object detection.