Abstract:
Obtaining accurate depth maps from multi-view configurations is an essential component for dense scene reconstruction from images and videos. After a brief introduction about research activities of the NetCentric Warfare-oriented Sensor Data Processing working group of the Fraunhofer IOSB institute, a fast plane sweep algorithm based on image warping will be presented. As input, images with precomputed orientations as well as several scene points are assumed to be given. The algorithm comprises two modules: Multi-view cost aggregation and non-local optimization. During multi-view cost aggregation, care must be taken while handling occlusions and slanted, non-fronto-parallel surfaces. This is achieved by taking into account redundant views and, respectively, triangles spanned from already available 3D points. The non-local optimization step presupposes minimization of a cost function penalizeing radiometric inconsistences and depth jumps between neighboring pixels. Six optimization algorithms will be discussed.
In the outlook, solution techniques for the generalization of the optimization problem from second-order to higher-order potentials will be presented.
Speaker Bio:
Dimitri Bulatov graduated 2004 in mathematics at the University of Würzburg, Germany. Since 2005, he worked as a scientist in the department of Scene Analysis of the Research Institute of Optronics and Pattern Recognition in Ettlingen, near Karlsruhe, Germany, which in 2009 has become a part of the Fraunhofer Company as Institute of Optronics, System Technologies and Image Exploitation. His main area of research includes structure-from-motion, dense reconstruction and semantic urban terrain modeling algorithms. In 2011, Dimitri Bulatov defended his PhD about textured 3D reconstruction from UAV-borne video sequences.