Robust Two-View Reconstruction Procedure for Geometrical Model Retrieving

Abstract

In this paper we present the enhancement of the standard two-view reconstruction procedure. An ordinary approach assumes determination of points’ correspondences followed by projection matrix estimation, to finally refine results with bundle adjustment taking as a cost function reprojection error. Our contribution is realized in two manners: introducing an additional step of outliers rejection, changing cost function of bundle adjustment process to Relative Reprojection Error (ER) and applying central difference as a method for Jacobian matrix approximation. Tests revealed gain in average ER with lower variance, for confidence level of 0.95. Besides accuracy improvement, the suggested modifications supply the final result in the time virtually independent on initial object’s complexity and, in most cases, shorter than the standard approach.