Vortrag von Professorin Dr. Simone Schaub-Meyer, TU Darmstadt, Fachbereich Informatik, Fachgebiet Image and Video Analysis

Abstract

Recent developments in deep learning have led to significant advances in many areas of computer vision. Yet, improvements in benchmark accuracy often provide only a partial picture of a model’s capabilities.

In this talk, I will discuss how carefully analyzing and understanding model performance and behavior can yield both methodological and conceptual advances in visual perception. In the first part, I will demonstrate how detailed performance analyses can expose structural inefficiencies, enabling the design of more computationally efficient and conceptually simpler deep vision models.

In the second part, I will discuss the role of visual explanations in understanding and improving classification, and present a method to obtain such explanations in a way that is both effective and computationally feasible. Together, these perspectives highlight how deeper insight into model behavior can drive the next generation of advances in visual perception.

Short bio

Simone Schaub-Meyer is an assistant professor at the Technical University of Darmstadt, as well as a member of the Hessian Center for Artificial Intelligence (hessian.AI). The focus of her research is on developing efficient, robust, and understandable methods and algorithms for image and video analysis. She recently got the renowned Emmy Noether Programme (ENP) fund of the German Research Foundation (DFG) supporting her research on Interpretable Neural Networks for Dense Image and Video Analysis. Before starting her own group, she was a postdoctoral researcher in the Visual Inference Lab of Prof. Stefan Roth.

Prior to joining TU Darmstadt, she was a postdoctoral researcher at the Media Technology Center at ETH Zurich working on augmented reality. She obtained her doctoral degree from ETH Zurich, advised by Prof. Dr. Markus Gross and in collaboration with Disney Research Zurich. In her doctoral thesis, awarded with the ETH Medal, she developed novel methods for motion representation and video frame interpolation.