An Efficient and Accurate Method for Detecting Objects in Aerial Images

Aerial image object detection in aerial images is a hot and challenging task in computer vision, due to the bird-view perspective, complex backgrounds, variant scales and appearance of objects and extremely dense objects distribution. It has previously been observed that existing methods cannot meet the application requirements of accuracy and speed at the same time. In this paper, we propose an efficient and accurate method for aerial image object detection. The pipeline of our method has only two stages. The oriented bounding boxes of objects are predicted by utilizing a simple network in the first stage, and the rotated bounding box predictions are then sent to non-maximum suppression (NMS) to produce final detection results. Besides, atrous spatial pyramid pooling (ASPP) network is added to the pipeline to extract multi-scale features, and Bi-directional long short term memory network (BiLSTM) is adopted to improve detection performance of long and slender instances. Experiments on the challenging DOTA dataset have shown the propose method outperforms existing methods in terms of detection rate and speed.