Learning to spell out the unknown in semantic segmentation


Hanno Gottschalk, Bergische Universität Wuppertal, Fakultät für Mathematik und Naturwissenschaften


Learning to spell out the unknown in semantic segmentation



In modern computer vision, neural segmentation networks are capable to identify and localize objects from categories they have previously been trained on. As a consequence, the semantic world of such networks is closed. In this talk I present approaches, how to teach segmentation networks to identify objects from outside of the semantic space of the training data and to extent their semantic space dynamically.


Short Biography

Prof. Dr. Hanno Gottschalk grew up in Leverkusen and studied mathematics and physics in Freiburg and Bochum. After completing his doctorate in mathematics at the Ruhr University in 1999, he researched as a DAAD scholarship holder at the University ‘La Spienza’ in Rome. This was followed by three years as a PostDoc at the University of Bonn. There he habilitated in mathematics in 2003 and became a university lecturer in 2006. In 2007, he temporarily left the academic career and worked as Core Competency Owner for Probabilistic Design in Gas Turbine Engineering at Siemens Energy. After a call to the Bergische Universität Wuppertal, Hanno Gottschalk has been teaching and researching as a professor of stochastics in Wuppertal since 2011. In addition to modeling the reliability of mechanical components, Prof. Gottschalk researches the uncertainty and errors of artificial intelligence. In June 2018, he became the founding director of the Interdisciplinary Center for Machine Learning & Data Analytics with Anton Kummert.