An empirical study of multi-scale object detection in high resolution UAV images

Object detection in images collected by Unmanned Aerial Vehicles (UAVs) constitutes a challenging task in computer vision, due to difficulties of learning a well-trained object detection model for handling instances in UAV images with arbitrary orientations, variation in different scales, irregular shapes, etc. In order to facilitate object detection research and extend its applications in natural scenarios by using UAVs, this paper presents a large-scale benchmark dataset, MOHR, aiming at performing multi-scale object detection in UAV images with high resolution. A total of 90,014 object instances with labels and bounding boxes were annotated. In order to build a baseline for object detection on the MOHR dataset, we performed an empirical study by evaluating six state-of-the-art deep learning-based object detection models trained on our proposed dataset. Experimental results show promising detection performance, but also demonstrate that the dataset is quite challenging for adopting natural image-based object detection models for UAV images.