Unsupervised Domain Adaptation for Aerial Object Detection with Self-Distillation

The domain shift problem remains a significant obstacle in aerial object detection. Unmanned aerial vehicles (UAVs) have become a vital component of modern visual sensing systems, yet their performance in object detection remains highly sensitive to environmental changes. A model trained in one domain often struggles when deployed in a different visual context. This work addresses that challenge through a self-distillation-based unsupervised domain adaptation (UDA) framework for UAV object detection. Instead of relying on adversarial training, our approach adopts a teacher-student paradigm, where the teacher network generates soft pseudo-labels on the unlabeled target data and the student iteratively refines its predictions through self-distillation. The framework integrates feature alignment and an exponential moving average (EMA) teacher update to maintain stability and reduce label noise throughout training. Experiments across three UAV benchmarks, VisDrone, UAVDT, and DOTA-Drone, demonstrate the robustness of the proposed method. It consistently improves detection accuracy by 30-35% mAP over non-adapted baselines, achieving 83.9% mAP0.5 on VisDrone and 78.5% on UAVDT without using a single labeled target image.