Globally Optimal Rotation Estimation for a Generic Spherical Camera Model

Robust pose estimation in visual localization remains a critical challenge, especially in dynamic occlusion, lowtexture, and long-term environmental changes. Current methods are limited by the camera's field of view (FOV), which prevents capturing broader scenes to reduce dynamic occlusion effects and extract richer scene features. Outlier rejection algorithms, such as Random Sample Consensus (RANSAC) and global optimization techniques, also suffer from robustness and real-time performance limitations. To address these issues, we propose a novel global optimal rotation estimation method based on a generalized spherical camera model. Specifically, we first unify a wide FOV spherical projection model that is applicable to both fisheye and panoramic cameras. Building on this, we transform the maximum consensus set problem into a geometric intersection problem of spherical regions, which is further reduced to a one-dimensional angular overlap problem. We then design a votingbased algorithm to determine the optimal rotation by selecting the maximum angular interval. Simulation results show that our method remains robust even under extreme outlier conditions (95 % outliers), and real-world fisheye image experiments validate the robustness and efficiency of our proposed algorithm.