Progressive Linear Search for Stereo Matching and Its Application to Interframe Interpolation

摘要

In this paper we present a stereo matching strategy that represents disparity as a linear piecewise function. The function is obtained by recursively subdividing intervals in corresponding scanline pairs. Each subdivision step delineates new intervals by explicitly searching for breaks of disparity. In contrast to most approaches, we do not assume a constant disparity within a region, but we define disparity values by a linear model. A disparity model provides strong constraints in the estimation problem giving spatial coherence. Parametric models are estimated by minimizing the similarity error via the Hough transform. A regularization cost is included during the subdivision process by considering disparity values between consecutive intervals. Experiments on synthetic and real images show that our adaptive matching strategy is capable of obtaining good detail with a small number of spurious points even if scanlines are processed independently and without using any postprocessing smoothing. We have successfully applied our matching results to create realistic image sequences using pixel-based interpolation. Occluded regions are identified by overlapping intervals and they are displayed by using a back-to-front strategy.