Multi-view Stereo Reconstruction for Internet Photos

This paper develops a multi-view stereo approach to reconstruct the shape of a 3D object from a set of Internet photos. The stereo matching technique adopts region growing approach, starting from a set of sparse 3D points reconstructed from structure-from-motion (Sfm), then propagates to the neighbouring areas by the best-first strategy, and produces dense 3D points. View selection and filter algorithms are proposed considering the characteristics of Internet images. Specifically, for each 3D point, we choose a reference image at first which determines the right subset images from unordered sets for optimization. Two filter algorithms, namely Sfm points filter and quality filter which is based on the assumption that depth changes smoothly, are designed to eliminate low-quality reconstructions. We demonstrate our algorithms with several datasets which show that they perform robustly.