Abstract

Abstract: Humans excel at the task of image understanding. When we see a face we immediately infer if the person is male or female, attractive or not, young or old, hostile or friendly or just a boring speaker. For computers this is much harder, but some progress has been made. In this talk I’ll argue that for image understanding strong prior knowledge is needed. Humans have acquired this knowledge over the course of their phylogenesis and ontogenesis, while we are still very much trying to force-feed computers by hand. I’ll introduce you to generative face models to be used as prior knowledge, and I’ll demonstrate a range of different uses of these models which I’ve encountered during my thesis. I hope to give you an overview of the state of the art such that you know the possibilities and limitations of generative face models for image understanding.

Biography: In 2005 Brian Amberg achieved a masters degree in Computer Science (Dipl. Inf.) at the University of Freiburg (Germany) with a minor in Cognitive Science. Since then he works as a PhD student at the University of Basel (Switzerland), creating a system for expression manipulation in video streams and dabbling in many related topics. His research focuses on computer vision, and so far mainly on face understanding with the help of generative models.