Generative Models in Computer Vision

Andreas Geiger

University of Tübingen, DEU

Abstract

Today, generative models form the foundation for many downstream tasks. In this lecture, I will first introduce and motivate the utility of generative models, and then cover the fundamentals of three generative models that had a profound impact on computer vision: variational auto-encoders, generative adversarial networks and diffusion models. I will introduce the models formally and highlight connections between. For each model, I will demonstrate some notable applications in 2D and 3D computer vision. In particular, I will discuss (text-conditioned) image and 3D shape generation models as well as 3D-aware models for novel-view synthesis. I will also cover models that are able to synthesize multi-object 3D scenes, 3D human bodies and 3D traffic scenes. Finally, I will briefly introduce scholar-inbox.com, a new paper recommender tool developed by my group.