Sub-pixel disparity search for binocular stereo vision

The idea of cost or similarity volume may date back to Marr and Poggio´s cooperative algorithm. One of the challenging difficulties in applying their algorithm to recovering sub-pixel disparity is 1) how to take advantages of the edge location in the images for computing reliable cost between the corresponding sub-pixel coordinates. In addition, it should be noted that 2) accidental correspondences will produce many confusing sub-pixel disparity candidates with small cost. This study proposes a novel method where a reliable subpixel cost volume is computed though the combination of an interpolation method and an edge-preserving filtering method. In order to avoid confusing disparity candidates, the search range of the sub-pixel disparity will be limited according to the initial disparity which can be computed robustly but not so precisely in advance. All the procedures can be implemented in a highly-parallel style, therefore the proposed method seems promising for real-time applications.