Manipulating 3D Objects with Machine Learning
Noam Aigerman - University of Montreal
Oct. 11, 2024, 2:30 p.m. - Oct. 11, 2024, 3:30 p.m.
ENGMD 280
Hosted by: Paul Kry
The field of neural geometry processing - applying deep learning to 3D shapes - poses unique challenges: in general, when applied to 3D shapes, standard “deep” architectures are not as successful as they are on, e.g., images or text, due to the unique nature of 3D geometric problems. This in turn requires custom-made solutions, relying on concepts from, e.g., differential and computational geometry.
In this talk I will discuss my ongoing work on applying machine learning to a specific fundamental 3D task, of manipulating 3D objects by “deforming” them, i.e., modifying the object as though it were made out of clay – this task has broad applications, from design of mechanical parts and up to computer animation. As such, it stands to gain immensely from incorporation of deep learning.
I will show how we devised deep learning architectures that achieve much higher levels of accuracy, or provide guarantees on crucial properties, such as ensuring the deformation does not make the object self-intersect (injectivity), as well as how we applied these methods to generative tasks such as text-driven modeling, and generation of repeating design patterns.
Bio: Noam Aigerman is an Assistant Professor at the University of Montreal, and an affiliate member of Mila. Before that, he worked as a research scientist at Adobe Research. His work focuses on applying machine learning to 3D problems, and its use in geometric modelling, vision, and computer graphics.
