Removal-based multi-view stereo using a window-based matching method

Abstract In this paper, we present a very simple and effective removal-based multi-view stereo (MVS) algorithm for application to the images of large outdoor objects. Owing to the complexity of outdoor illumination and occlusion, it is challenging to guarantee the accuracy of the estimated depth of every pixel in a stereo image. The key idea of the proposed algorithm for addressing this issue is to calculate the 3D coordinates of every pixel by attempting to reconstruct highly accurate point-cloud data (PCD) using only the appropriate corresponding points from a stereo pair of two views. All erroneous corresponding points in the pair are removed. The proposed algorithm was validated by conducting a set of experiments using several datasets for different environments. The algorithm for removing incorrectly matched points, which is referred to as R-NCC (removal normalized cross-correlation), was found to have a high reconstruction accuracy comparable to those of current state-of-the-art sparse algorithms.