Matching quality-guided model-free satellite pose estimation

With the development of learning-based techniques and large-scale public datasets, satellite pose estimation has seen significant progress for the past several years. However, most of current methods still rely on a known three-dimensional (3D) model of the object for the pose estimation, limiting its generalization and wide application. To this end, we propose a model-free pose estimation method, which takes as input only a set of images. The proposed method consists of two stages, i.e. the reconstruction stage and pose estimation stage, of which the former reconstructs a 3D model from the input images and the pose is estimated via two-dimensional (2D)-3D feature matching by the latter. More importantly, a matching quality guidance strategy is introduced to further improve the robustness to in-plane rotation during feature matching. Additionally, since no known 3D model is assumed, our method generalizes well to novel objects without retraining. We provide evaluation results on datasets of several satellites with different structures, which demonstrate impressive performances without known models against current methods.