Julie Digne
LIRIS - CNRS - Université Claude Bernard Lyon 1

Wednesday, July 1, 09:30

Geometric not-so-deep learning (Video)

Over the past decade, deep learning for geometric data processing has advanced significantly, with numerous methods proposed to handle irregular, non-Euclidean data. However, 3D objects databases are scarce and only partially cover the variety of shapes practitioners want to analyze. As a consequence many shapes fall out-of-distribution. However, many tasks, such as compression, denoising or resampling, can already benefit from leveraging statistical geometric features, without requiring shape space priors. In this talk, I will focus on lightweight methods that are computationally efficient, run on standard hardware by operating directly on individual shapes. Through a series of projects, I will show how modern optimization techniques, with or without neural networks, can address geometric challenges effectively, without relying on large datasets or heavy computational resources.

Julie Digne is a Senior Researcher at CNRS, affiliated with the LIRIS laboratory in Lyon (Origami team). She earned her PhD in Applied Mathematics from École normale supérieure de Cachan in 2010. Following a postdoctoral position at INRIA, she joined LIRIS in 2012, where her research focuses on geometry processing and machine learning for geometric data.

Justin Solomon
MIT CSAIL

Thursday, July 2, 09:30

Computational Optimal Transport: From Low to High Dimension and Back (Video)

The optimal transport problem asks a simple geometric question: What is the most efficient way to transform one probability distribution into another along a piece of geometry? Beyond its mathematical interest, optimal transport underlies a variety of applications, from supply chains to mesh processing, statistics, and even generative AI. Over the past two decades, research in geometry processing has played a central role in shaping algorithms for optimal transport, with foundational advances emerging from the SGP community. At the same time, popular problems in applied optimal transport have shifted from low-dimensional settings in graphics and imaging to high-dimensional settings in machine learning.

In this retrospective keynote, I will trace how my team’s work in computational optimal transport was shaped by studying its applications to geometry processing—even as the landscape of research in this area shifted in dimensionality and application. Ultimately, this journey illustrates the broader value of “Geometric Data Processing” as a discipline: identifying shared geometric and variational principles across domains that differ dramatically in dimension, scale, and data fidelity.

Justin Solomon is an associate professor of Electrical Engineering and Computer Science in MIT’s Computer Science and Artificial Intelligence Laboratory (CSAIL). He leads the CSAIL Geometric Data Processing group, which studies problems at the intersection of geometry, large-scale optimization, and applications.

Bernhard Kerbl
University of Copenhagen / Vienna University of Technology (TU Wien)

Friday, July 3, 09:30

Resource-Efficient Visual Computing – Frontiers and Applications of Real-Time Visual AI

Modern visual computing is transforming the means and ways by which we map, understand and interact with the physical world. 3D and 4D reconstructions of real artefacts are now viable from just a handful of casual camera observations; object recognition and classification can be done with unprecedented accuracy and reliability. However, a key requirement for the overall usefulness of these methods is their efficiency: Efficiency dictates whether a solution can run in real-time; it governs the hardware requirements for execution, and wether it can be used without requiring massive, industry-grade infrastructure. Real-time performance enables crucial emerging trends, such as robots interacting with the real world, or visual AI providing on-line assistance in medical treatments on patients. Resource-efficiency, on the other hand, ensures that these breakthrough technologies can be employed by almost anyone. In this talk, Dr. Kerbl will discuss key challenges and opportunities of real-time, resource-efficient visual computing and AI, focusing on open tasks in fundamental research and applied fields, including biomedicine and robotics.

Dr. Kerbl is a Tenure-track Assistant Professor at the University of Copenhagen and a Principal Project Investigator associated with the Vienna University of Technology (TU Wien). Before that, he was a Visiting Researcher at the Robotics Institute, Carnegie Mellon University, in the Human Sensing Lab under Fernando de La Torre. He obtained his PhD at Graz University of Technology in 2018. In 2019 he pursued a postdoc at TU Wien in 2019, followed by another in 2022 with INRIA in George Drettakis’ GraphDeco group. His research focuses on real-time graphics, parallel processing, point-based rendering, image-based rendering, radiance fields and novel-view synthesis. Through collaboration with mentors, colleagues and students, Dr. Kerbl’s research has been honored with Best Paper awards at major graphics conferences, including SIGGRAPH, High-Performance Graphics, Pacific Graphics, EGPGV, and GRAPP. He has lectured on the topics of GPU programming, real-time rendering, physically-based rendering, game physics and scientific working at TU Wien, Graz University of Technology and FH Salzburg.