We describe a solution to the problem of stereo matching for non-copla
nar points and wide angular separation of uncalibrated cameras. The im
age projections of four corresponding coplanar points, with no three c
ollinear, are assumed to be known. The 'virtual image' that results by
2D warping of the first image towards the second one according to the
homography of a known plane in 3D is used. Then, instead of trying to
match points between the two images, we exploit the fact that all cor
responding points between the virtual and the second image are in alig
nment. It is shown how the matching problem can be cast into an optimi
sation one, and a cost function for evaluating each possible solution
is derived. For increased robustness to outliers a Hough-Transform-lik
e algorithm is suggested for obtaining the optimal solution, and a ran
domised scheme is shown to be efficient for practical applications. Ex
perimental results on both synthetic and real data are also included.
(C) 1998 Elsevier Science B.V.