A Transformer Based Network in Monocular Satellite Pose Estimation

Deep learning methods for optical camera-based pose position estimation are becoming more and more widely used with the improvement of computer hardware, such as target classification, target detection and other fields. One of the roles of deep neural network in pose estimation is to estimate the coordinates of keypoints. However, in some previous methods and applications, there are some drawbacks. For example, it can only be applied to scenes with small changes in distance and rotation angle or road 3D target detection that contains only one axis of rotation, and the accuracy of target estimation for multi-axis rotation is not high. In this paper, we present a monocular pose estimation network that integrates YOLO-v5 with a Transformer module. This approach transforms the pose estimation task into a keypoint Gaussian kernel regression problem, enhancing accuracy and robustness to scale variations.