An efficient refinement for relative pose estimation with unknown focal length from two views

This paper describes the relative pose estimation problems in a semi-calibrated case. It is well known that the fundamental matrix completely encapsulates the epipolar geometry between two perspective views. Through the parameterization of the fundamental matrix with the minimal parameters of pose and focal length, we formulate the problem as the minimum cost estimation problem with the aid of the principle of maximum likelihood. The corresponding analytical differentiation and optimization algorithm are proposed as means for efficient computation. Experiments on simulated and real data show that the accuracy of this approach is more or less comparable to bundle adjustment and its implementation performs much better in terms of computational cost.