Singular Vector Methods for Fundamental Matrix Computation

The normalized eight-point algorithm is broadly used for the computation of the fundamental matrix between two images given a set of correspondences. However, it performs poorly for low-size datasets due to the way in which the rank-two constraint is imposed on the fundamental matrix. We propose two new algorithms to enforce the rank-two constraint on the fundamental matrix in closed form. The first one restricts the projection on the manifold of fundamental matrices along the most favorable direction with respect to algebraic error. Its complexity is akin to the classical seven point algorithm. The second algorithm relaxes the search to the best plane with respect to the algebraic error. The minimization of this error amounts to finding the intersection of two bivariate cubic polynomial curves. These methods are based on the minimization of the algebraic error and perform equally well for large datasets. However, we show through synthetic and real experiments that the proposed algorithms compare favorably with the normalized eight-point algorithm for low-size datasets.