Non-Incremental Derivation of Scale and Pose from a Network of Relative Orientations

Abstract. The majority of approaches to Structure from Motion apply an incremental triangulate-and-resect strategy in order to reconstruct camera motion and scene structure in a common reference frame. The sequential addition of images may cause a drifting behaviour during the reconstruction, in some cases causing the process to fail. Over the last decade, more attention has come to non-incremental approaches, which exploit the network characteristics arising from the 2- or 3-view relations, given for a set of images through relative orientations. Most approaches employ rotation registration, followed by translation registration. The latter being carried out with or without simultaneous scene reconstruction. We suggest an approach which starts by estimation of relative scales, followed by simultaneous registration of rotation and translation. The latter is achieved by employing a path-finding algorithm based on Ant- Colony-Optimization. For scale estimation we propose a window-search adaption of Levenberg-Marquardt, which avoids unnecessary matrix inversions. We also suggest a simple method for detection of outlier orientations.