The Six Point Algorithm Revisited

摘要

This paper presents an algorithm for estimating camera focal length from tentative matches in a pair of images, which works robustly in practical situations such as automatic computation of structure and camera motion from unknown photographs, e.g. from the web or from various instruments mounted on a vehicle. We extend the standard 6-pt algorithm based on the observations: (i) the quality of the estimation of this algorithm is strongly correlated with the ratio of the singular values of the essential matrix computed from inliers, which is calibrated by using the estimated focal length, returned by RANSAC and (ii) the reprojection error of the affine camera model, fit to the inliers, predicts the uncertainty in the estimated focal length. Furthermore, for scenes with dominant plane we propose a novel algorithm calculating relative orientation and unknown focal length given a plane homography and a single off the plane point correspondence. The performance of the proposed algorithm is demonstrated on a set of real images having different focal lengths.