Vehicle Detection in High-Resolution Aerial Images with Parallel RPN and Density-Assigner

Vehicle detection in aerial images plays a significant role many remote sensing applications such as city planning, road construction, and traffic control. However, detecting vehicles remains challenging due to the existence of tiny objects, scale variance within same type vehicle dense arrangement some scenarios, parking lots. At present, state-of-the-art object detectors cannot generate satisfactory results on images. The receptive field current detector is not fine enough handle slight variance. Moreover, densely arranged will introduce ambiguous positive samples label assignment false predictions that be deleted by NMS. To this end, we propose two-stage framework for better leverages prior attribution knowledge First all, design Parallel RPN exploits convolutional layers different fields alleviate variation problem. tackle vehicles, density-based sample assigner vehicle-intensive areas reduce low-quality occluded training process. In addition, scale-based NMS proposed filter out redundant proposals hierarchically from levels feature pyramid. construct two datasets based AI-TOD xView which contain objects. Extensive experiments these demonstrate effectiveness our method.