A branch and contract algorithm for globally optimal fundamental matrix estimation

We propose a unified branch and contract method to estimate the fundamental matrix with guaranteed global optimality, by minimizing either the Sampson error or the point to epipolar line distance, and explicitly handling the rank-2 constraint and scale ambiguity. Based on a novel denominator linearization strategy, the fundamental matrix estimation problem can be transformed into an equivalent problem that involves 9 squared univariate, 12 bilinear and 6 trilin-ear terms. We build tight convex and concave relaxations for these nonconvex terms and solve the problem deterministically under the branch and bound framework. For acceleration, a bound contraction mechanism is introduced to reduce the size of the branching region at the root node. Given high-quality correspondences and proper data normalization, our experiments show that the state-of-the-art locally optimal methods generally converge to the globally optimal solution. However, they indeed have the risk of being trapped into local minimum in case of noise. As another important experimental result, we also demonstrate, from the viewpoint of global optimization, that the point to epipolar line distance is slightly inferior to the Sampson error in case of drastically varying object scales across two views.