Improvement of the Block Matched Parallax Field by Non-Linear Diffusion

This paper introduces an improved block matching algorithm for the estimation of the parallax field between the left and the right image of a stereo pair. When the block size is small, e.g. for computer vision purposes, traditional block matching alone fails. Most of the small blocks do not expose enough structure to be reliably matched. Therefore, in this paper a parallax field estimated with the traditional block matching algorithm is enhanced by a non-linear diffusion process. This process smoothens the parallax field where the estimated samples are unreliable, and leaves the reliable samples intact. First the principles behind the method are explained. Then the performance of the method is shown on real stereo pairs.