Abstract

In this talk I give an overview of my recent research activities in computer vision during my post-doctoral stays at UNC -Chapel Hill and ETH Zurich. The primary goal of generating faithful virtual representations of real environments from image data led to research in two methodologically very different fields: first, I investigated in the structure-and-motion problem for large image collections. A lot of my research on this topic addresses the question of how difficult and ambiguous visual structures—potentially leading to non-recoverable incorrect 3D models—can be detected. The proposed probabilistic models go beyond the standard reprojection error to assess the quality of the structure-and-motion solution. The demand for 3D mesh models from images with optimality guarantees initiated a second line of research rooted in continuous optimization. The proposed, relatively simple energy formulation for surface reconstruction yields high-quality results even for inputs strongly contaminated with noise and outliers. The utilized methodology also extends to different application areas such as medical imaging and fast inference in graphical models. Efficiency of these methods is assured by utilizing graphics processing units.