Small Object Detection in Unmanned Aerial Vehicle Images Leveraging Density-Aware Scale Adaptation and Knowledge Distillation

Small object detection remains a challenging task despite the impressive progress in general object detection over the past decade. The difficulties are attributed to the small size, severe occlusion, and variant scales. To address these challenges, we propose a novel method for small object detection, employing a coarse-to-fine approach that integrates density-aware scale adaptation and knowledge distillation. Firstly, we comprise a coarse network for global sketchy predictions in large scenes, generating adaptively scaled block regions containing potential objects. Subsequently, a fine network is utilized for precise detection in densely packed areas, where objects exhibit approximately unified scales. To enhance scale transformation, a density map, providing object distribution information, is introduced as an auxiliary task for scene classification. Moreover, a grid masked knowledge distillation strategy is proposed to facilitate the student network's feature recovery, thus promoting detection accuracy without directly mimicking the teacher network. Extensive experiments on renowned benchmark datasets, including VisDrone and UAVDT, substantiate the superiority of the proposed method for small object detection in unmanned aerial vehicle images.