Orientation-guided geodesic weighting for PatchMatch-based stereo matching

We propose an orientation-guided geodesic weighting (OGGW) strategy for local stereo matching.We propose a method of cost volume filtering combining a multipoint LPA method with our OGGW strategy.We propose a PatchMatch filter with curved surface fitting (PMF-CS) to obtain a disparity map with sub-pixel accuracy. Recently, PatchMatch-based methods for local stereo matching are experiencing great progress with the use of compact and over-segmented regions that have similar intensities or colors. Using patches as support regions, this paper proposes an orientation-guided geodesic weighting (OGGW) strategy to search for an approximate shortest path from a support pixel in the patch to a pixel of interest along a guided orientation. The OGGW is computed by accumulating intensity differences or color dissimilarities between connected pixels along the path. After obtaining matching cost updates by model fitting, the OGGW is used for weighted averaging on the updated costs to obtain a filtered cost volume. In addition, a new filtering method that combines the PatchMatch filter with curved surface fitting (PMF-CS) is presented in this paper. Curved surface fitting along with outliers removal is carried out to seek for a reliable regression model for estimating the disparities on a patch and to achieve a disparity map with sub-pixel accuracy. We conduct a number of experiments to evaluate the performances of OGGW and PMF-CS on cost volume filtering and disparity estimation. Experimental results show that our algorithm produces accurate stereo matching results and outperforms the current state-of-the-art PatchMatch-based methods.