8 Discussion

Stereo matching is greatly simplified if the epipolar lines are parallel and horizontal in each image, i.e. if the images are rectified. We have developed a rectification algorithm which imposes explicitly all the constraints sufficient to obtain two unique projection matrices as the solution of four constrained, homogeneous linear systems; the resulting algorithm is quite simple. The correct behaviour of the algorithm has been demonstrated with both synthetic and real images. Interestingly enough, reconstruction can be performed directly from the disparities of the rectified images, using the rectified projection matrices. Our tests show that this process does not introduce appreciable errors compared with reconstructing from the original images. We believe that a general rectification algorithm, together with the material we have made available from the URL given in Section 6, can prove a useful resource for the research and application communities alike.