Toward Efficient Object Detection in Aerial Images Using Extreme Scale Metric Learning

In aerial image object detection, how to efficiently detect different size objects in input images of scales and obtain a unified multi-scale representation the is an important issue. Existing methods rarely consider connection between training inference, do not well optimize constraint samples process, which limits performance representation. this study, efficient detection algorithm for proposed alleviate problem. Firstly, we propose use metric learning scale boundary each class, reduce support indistinguishable at extreme enhance effect Secondly, small are merged into regions, these regions trained recommend detector on following high-resolution scale. Thus, reasonable association inference established, efficiency considerably improved. The has been tested three popular datasets, including VisDrone, DOTA UAVDT. Experimental results show that it can improve accuracy number processing pixels.