Abstract
Abstract. The generalized camera model allows handling a broad range of imaging systems for which the common central perspective model no longer applies. While offering greater modeling flexibility, designated orientation procedures need to support its application. In that respect, models that facilitate the computation of the bundle-adjustment solution are imperative as a means to solve the motion and structure of a large set of images. Being non-linear, the bundle adjustment solution requires good initial estimates and the filtering of outlying matches. To facilitate such a solution, we propose in this paper a method for estimating the global rotations of a set of views independent of their translation and the scene structure. This method is then used to improve the robustness, efficiency, and convergence of the solution when used as a refinement to the global rotations. The proposed pipeline is evaluated by experiments on synthetic and real-world data of axial flat-refractive cameras. It shows that the proposed method produces more accurate and optimal initial estimates of the global rotations than the state-of-the-art rotation averaging method.