Is Dense Optic Flow Useful to Compute the Fundamental Matrix?

Estimating the fundamental matrix from a pair of stereo images is one of the central problems in stereo vision. Typically, this estimation is based on a sparse set of point correspondences that has been obtained by a matching of characteristic image features. In this paper, however, we propose a completely different strategy: Motivated by the high precision of recent variational methods for computing the optic flow, we investigate the usefulness of their dense flow fields for recovering the fundamental matrix. To this end, we consider the state-of-the-art optic flow method of Brox et al. (ECCV 2004). Using non-robust and robust estimation techniques for determining the fundamental matrix, we compare the results computed from its dense flow fields to the ones estimated from a RANSAC method that is based on a sparse set of SIFT-matches. Scenarios for both converging and ortho-parallel camera settings are considered. In all cases, the computed results are significantly better than the ones obtained by the RANSAC method --- even without the explicit removal of outliers.