Prof. Bernhard Egger | "Vision as Inverse Rendering?"

Inverse-rendering–based computer vision systems are becoming increasingly influential. Most recently, Sam3D was released and has overshadowed much of the previous state of the art. In this talk, I will discuss inverse rendering and, in particular, analysis-by-synthesis methods from a computer vision perspective. Using Mooney faces as an example, I will illustrate why I believe that inverse rendering - or more generally analysis-by-synthesis - is the right way to think about human vision.

Assuming we accept inverse rendering as a key principle for 3D perception, the next question is how such a system can be learned with minimal inductive bias. I will present some recent experiments in which we pretrained feature detectors and proto–optic-flow mechanisms on retinal-wave data. The results suggest that this biologically inspired pretraining strategy is a promising avenue for initializing an efficient vision system.