Automatic computation of fundamental matrix based on voting

Abstract

To reconstruct point geometry from multiple images, a new method to compute the fundamental matrix is proposed in this paper. This method uses a new selection method for fundamental matrix under the RANSAC (Random Sample And Consensus) framework. It makes good use of some low quality fundamental matrices to fuse a better quality fundamental matrix. At first, some fundamental matrices are computed as candidates in a few iterations. Then some of the best candidates are chosen based on voting the epipoles of their fundamental matrices to fuse a better fundamental matrix. The fusion can be simple mean or weighted summation of fundamental matrices from the first step. This selection method leads to better result such as more inliers or less projective errors. Our experiments prove and validate this new method of composed fundamental matrix computation.