Generative Geometrical Learning: Injecting structure in 3D and 4D Generation

Abstract: Generative AI has increasingly extended to 3D data, offering unprecedented opportunities for the synthesis and manipulation of shapes. Such advancements, driven by large-scale datasets and substantial computational power, appear to reinforce the “bitter lesson” that scale is the key driver of progress. However, how good are these models at inferring and preserving structures in the data? Several studies indicate that even foundational vision models trained on billions of images lack a basic understanding of geometry. This is further exasperated when the aim is to synthesize 4D assets, where shapes are supposed to evolve over time, respecting physical laws. By incorporating in the learning geometric inductive biases and structure insights into the learning, it not only improves performance but also opens up new avenues for applications and research.

Bio: Riccardo is a researcher and interim professor at the Computer Vision Group of the Technical University of Munich (TUM), part of the Munich Center for Machine Learning (MCML), and a member of the European Laboratory for Learning and Intelligent Systems (ELLIS). Previously, he was a Marie-Curie postdoc at the University of Tubingen and a postdoc at Sapienza University of Rome. Riccardo obtained his PhD from the University of Verona, collecting the Best PhD Thesis Award in Computer Graphics from the Italian Chapter of EuroGraphics. His research focuses on 3D Geometry Processing, Spectral Shape Analysis, and, in particular, on Shape Matching and Virtual Humans applications.