Transform domain similarity measures in stereo matching

Abstract

A Fourier-domain representation of an image exhibits a property where a translation of the image is included in the phase term. This property extends to the Discrete Cosine Transform, where the translation is encoded in the signs of the coefficients. An interpretation of this property for use as a similarity measure in local stereo matching is explored with an efficient way of comparing the transformed blocks to generate a dense disparity map. The method is empirically demonstrated to work also with the Haar wavelet transform, which offers faster computation and improved quality. Results show that presented transform based similarity measures provide better disparity estimates than box aggregated Sum of Absolute Differences and the Census transform when using the percentage of bad pixels as the quality measure. In terms of mean squared error, they achieve better results than SAD but are marginally below Census. The simplicity of making the comparisons results in good scaling with the number of disparity estimates, making the suggested method perform better than SAD also computationally.