Stable Structure from Motion for Unordered Image Collections

Abstract

Abstract. We present a non-incremental approach to structure from motion. Our solution is based on robustly computing global rotations from relative geometries and feeding these into the known-rotation framework to create an initial solution for bundle adjustment. To increase robustness we present a new method for constructing reliable point tracks from pairwise matches. We show that our method can be seen as maximizing the reliability of a point track if the quality of the weakest link in the track is used to evaluate reliability. To estimate the final geometry we alternate between bundle adjustment and a robust version of the known-rotation formulation. The ability to compute both structure and camera translations independent of initialization makes our algorithm insensitive to degenerate epipolar geometries. We demonstrate the performance of our system on a number of image collections. 1

Point Triangulation through Polyhedron Collapse Using the l∞ Norm

“…1 We also present results for the datasets of Alcatraz,Église du Dôme, Orebro castle, Stockholm town hall and the Vercingetorix statue from [4] and [15]: the results for the datasets as shown in Figures 10 through 15 illustrate that our proposed method executes faster than the existing methods (summarised in Table 1). All of the techniques were evaluated in MATLAB using SeDuMi as the solver for the LP and SOCP problems [19], and SBA for bundle adjustment [13].…”

Radial Distortion from Epipolar Constraint for Rectilinear Cameras

“…This is, however, a known problem for which solutions exist and beyond the scope of this paper. Considering the future work, one promising approach to decrease computation time and to further increase the method's robustness with respect to F computation stability is to use graph-based point tracking [31] that minimizes the amount of initial data by reducing false matches from it.…”

“…However, if false matches are so abundant that the fundamental matrix calculation results in a wrong F with significantly displaced epipoles, this (obviously) prevents the algorithm convergence by making the majority voting for the center of distortion location fail. Beyond the measures taken here, this risk may further be reduced by preparing the initial data with, e.g., graph-based point tracking [31].…”

Hierarchical structure-and-motion recovery from uncalibrated images

“…Existing pipelines either assume known internal parameters [26,27,19,28,29,15,24], or constant internal parameters [30], or rely on EXIF data plus external information (camera CCD dimensions) [31,32]. Methods working in large scale environments usually rely on a lot of additional information, such as camera calibration and GPS/INS navigation systems [2,33] or geotags [17].…”