Solving Perspective-n-Point Problem with Spherical Regression

This paper proposes a spherical regression relaxation solution for perspective-n-point problem. The problem is formulated as a minimization problem using angle error based spherical cost. With the relaxation of scale constraint, translation parameters can be calculated in closed form. Then, rotation parameters are immediately solved in closed form using spherical regression. The scale, translation and rotation parameters are alternatively estimated while keeping the others fixed until convergence. The proposed method is simple and able to cope with arbitrary central camera models. Experiment results show that the proposed method achieves accuracy almost the same as the maximum-likelihood estimator, and its computational efficiency is even comparable with some state-of-the-art non-iterative methods.