The fundamental matrix is a two-view tensor playing a central role in Computer Vision geometry. We address its robust estimation given pairs of matched image features, affected by noise and outliers, which searches for a maximal subset of correct matches and the associated fundamental matrix. Overcoming the broadly used parametric RANSAC method, ORSA follows a probabilistic a contrario approach to look for the set of matches being least expected with respect to a uniform random distribution of image points. ORSA lacks performance when this assumption is clearly violated. We will propose an improvement of the ORSA method, based on its same a contrario framework and the use of a non-parametric estimate of the distribution of image features. The role and estimation of the fundamental matrix and the data SIFT matches will be carefully explained with examples. Our proposal performs significantly well for common scenarios of low inlier ratios and local feature concentrations.
Laboratoire MAP5 (Mathématiques appliquées Paris 5), CNRS UMR8145 Université Paris V - Paris Descartes
