An efficient algorithm for learning-based visual localization.

This paper investigates the problem of visual localization in Global Positioning System (GPS)-denied environments. To achieve high-precision and robust localization performance, we propose a novel optimization algorithm based on the optimal control principle (OCP) that incorporates diagonal Hessian information estimation. By exploiting this curvature information, the proposed algorithm improves the training efficiency of deep neural networks and accelerates optimization convergence. Theoretically, the algorithm achieves a convergence rate of O(1T). Experimental results on public datasets demonstrate the effectiveness of our method. For the task of visual localization, our algorithm achieves a 33.71% improvement in position accuracy and a 15.66% improvement in rotation accuracy compared with Adam on the Great Court scene. Similarly, consistent performance gains are observed across other tasks, demonstrating strong generalization capability.