A STEREO MATCHING DATA COST ROBUST TO BLURRING

Most modern stereo matching algorithms involve solving an optimization problem where the objective function includes a data cost term and a smoothness term. The data cost term measures how well corresponding pixels match between the left and right images. In this paper a new stereo matching data cost is proposed which is robust to variations in blurring between the images caused by camera focus. In our method, each image is blurred once with a large filter. By comparing the original and blurred versions of each image we obtain a range of possible values each pixel could take on for different levels of blurring. Based on this range we construct a blur robust data cost for comparing pixels between two images. Experimental results show our proposed method greatly improves stereo matching accuracy when the left and right images in a stereo pair are focused differently.