Regularizing phase-based stereo

Wavelet-based techniques proved to be a promising approach for estimating the disparity between two stereo images. The complex-valued Gabor filter responses reduce the ambiguity of raw image intensities, and their phase differences between left and right image provide a direct measure for the disparities. Experience shows that such phase based measurements are reliable near edges but yield poor results between them. To improve these phase-based disparity estimations, a regularization scheme is proposed which directly compares possible matching pairs. Unreliable regions are filled by using a simple smoothness constraint. In the spirit of Markov random fields, we propose a probabilistic lattice model which describes the complete disparity distribution instead of representing only a single configuration. Experimental results are presented for an artificial image pair generated by computer graphics.