An effective multiview stereo method for uncalibrated images

Abstract

For most dense multi-view stereo methods, the process of finding correspondences is the basis and is independent of acquiring 3D information, and this often brings about erroneous correspondences followed by erroneous 3D information. To tackle this problem, by expanding matched points and by expanding 3D patches, this paper proposes an effective approach to acquire dense and accurate point clouds from multi-view uncalibrated images. In the approach, two novel algorithms are newly designed and are placed before and after the Bundler: 1) the match expansion algorithm, which generates evenly distributed correspondences with geometric consistency; after using Bundler to produce geometry estimation and quasi-dense point clouds which are not dense and accurate, 2) the point-cloud expansion algorithm, which is proposed to improve the density and accuracy of point clouds by optimizing the geometry of each 3D patch and expanding each good patch to its neighborhood. A large number of experimental results demonstrate the proposed approach get more accurate and denser point clouds than the state-of-the-art methods. A quantitative evaluation shows the accuracy of the proposed method favorable to PMVS.