Orientation Estimation for Airplane Targets in SAR Images Based on Keypoint Detection

Abstract

Object detection in aerial images is a challenging task since objects distribute in arbitrary orientations under the top view of remote sensing images (RSIs). Recent methods exploit oriented bounding boxes (OBB) to represent oriented objects accurately. However, airplane targets in different orientations appear very differently in Synthetic Aperture Radar (SAR) images due to the complicated imaging mechanism and various scattering conditions. Most SAR image datasets for airplanes are annotated with horizontal bounding boxes (HBB) and related research focuses on HBB without orientation information. This paper proposes a method for predicting the orientation of airplane targets in SAR images based on keypoints detection. Specifically, the object detection module is adopted to generate proposal regions of airplane targets in the first stage. Then, the proposal regions are utilized in the orientation prediction module to detect the head and tail of an airplane. The proposed method provides semantic orientation information for airplane targets in SAR images labeled by HBB. The experiments are conducted on Gaofen-3 (GF-3) dataset with labels of aircraft keypoints, and the accuracy of orientation estimation is achieved 85.8%. The Mean Average orientation Error is 18.24°. Furthermore, the orientation estimation model is applied on SAR Aircraft Detection Dataset (SADD) to supplement orientation information for airplane targets.