ECCV 2022 - Wednesday

Gerard, can you tell us about your work? We work on roughly three areas. One is on modeling the geometry and appearance of people. More recently, we’ve been working on capturing behavior and how we interact with the 3D world around us. There are two phases to this. The capture part is where you have observations and want to understand what’s going on, like in images, point clouds, or whatever data you have. The other side is wanting a compact model you can control to generate data. These two problems are interrelated because you can build better synthesis models with a good capture model. You can compare your beliefs against the observations and iterate this loop with a good synthesis model. We also work on general 3D representation learning, 3D scenes, and 3D rendering. Everything to do with 3D excites me! What are the applications of all this work in the real world? There are many. Imagine you deploy a robot, and the robot needs to understand what humans are doing in the scene. At the very coarsest level, it needs to know how to avoid them. All humans have these predictive models of the world, and we need to build these into intelligent systems. That’s the most basic need, but there are more concrete ones. For example, any 3D content generation application, like 3D entertainment or video games, needs a way to synthesize humans and 10 Gerard Pons-Moll Gerard Pons-Moll is a Professor at the University of Tübingen and a Senior Researcher at the Max Planck Institute for Informatics.