An extension of PatchMatch Stereo for 3D reconstruction from multi-view images

Abstract

PatchMatch Stereo is a method generating a depth map from stereo images by repeatedly applying spatial propagation and view propagation to the depth map. The extension of PatchMatch Stereo for multi-view 3D reconstruction has been recently proposed. This extension is very ad hoc and does not fully utilize the potential of multi-view images, since the method generates a 3D point cloud by combining a set of depth maps obtained from each binocular stereo image pair. This paper proposes a multi-view 3D reconstruction method using PatchMatch Stereo. To fully utilize the impact of multi-view images, the proposed method have two key ideas: (i) integrate matching scores from multiple stereo image pairs and (ii) perform view propagation among multi-view images. The use of multi-view images makes it possible to generate a reliable depth map by reducing occlusions. Through a set of experiments, we demonstrate that the proposed method generates more reliable depth map from multi-view images than the conventional method.