Refinement of depth maps by fusion of multiple estimates

摘要

Computing depth maps from a stereo pair is a well-studied problem of computer vision, and a large number of methods and cost functions have been proposed. The methods have different strengths and weaknesses and different error characteristics. That is, a pixel could be assigned an erroneous depth value by some methods, but other methods could assign the correct depth value to the same pixel. We describe a method that can make use of multiple depth estimates of low quality and fuse them, by trying to retain the correct depth values and rejecting the incorrect depth values, in order to obtain a more accurate result. We observe that depth values of pixels located in smooth areas of a depth estimate and depth values which survive left-right cross validation tend to be more accurate. Our method makes use of a reliability criterion-based upon the smoothness and cross-validation of the depth estimates that allows us to patch the estimates together and obtain a higher quality result.