Quick Reconstruction of Cameras and Quasi-dense Geometry from Unordered Photos

This paper presents a novel approach to quick reconstruction of cameras and quasi-dense geometry from unordered photos of a scene. With three reconstruction goals: reliable camera postures, high-density scene geometry, and a low cost of time, our approach consists of two stages: pairwise relating and two-tier reconstruction. At the first stage, to obtain better putative matches, we impose bidirectional feature matching and disparity constraint; to achieve better normalization, we perform approximate self-calibration using EXIF tag. The second stage includes a sparse and a quasi-dense reconstruction with different pivot respectively, adding selection strategies to avoid degenerate cases and improve time efficiency.