A Meta-Technique for Increasing Density of Local Stereo Methods through Iterative Interpolation and Warping

Abstract

Despite much progress in global methods for computing depth from pairs of stereo images, local block matching methods are still immensely popular largely due to low computational cost and ease of implementation. However, such methods usually fail to produce valid depths in several image regions due to various reasons such as violations of a fronto-parallel assumption and lack of texture. In this paper, we present a simple and fast meta-technique for increasing the percentage of valid depths (depth map density) for local methods while keeping the percentage of pixels with erroneous depths, low. In the method, the original disparity map computed by a local stereo method is iteratively improved through a process of depth interpolation and image warping based on the interpolated depth. Image warping gives a mechanism for testing the validity of the interpolated depths allowing for incorrect depths to be discarded. Our results on the KITTI stereo data set demonstrate that, on average, we can increase density by 7-13% after a single iteration, for a 15-29% increase in computation and only a slight change in the outlier percentage, depending on the cost function used for matching.