“Can Machines Learn to Generate 3D Shapes?”

by Professor Richard Zhang

Simon Fraser University

Director of graphics (GrUVi) lab

Day 2, June 12 — 9:15~10:00

Abstract:

At heart, computer graphics is about synthesis and creation by computing machinery. Early success has been obtained on training deep neural networks for speech and image syntheses, while similar attempts on learning generative models for 3D shapes are met with difficult challenges. In this talk, I will first go over how the sub-field of 3D shape modeling and synthesis in computer graphics has evolved, from early model-driven approaches to recent data-driven paradigms, and highlight the challenges we must tackle. I would argue that the ultimate goal of 3D shape generation is not for the shapes to look right; they need to serve their intended (e.g., functional) purpose with the right part connection, arrangements, and geometry. Hence, I advocate the use of structural representations of 3D shapes and show our latest work on training machines to learn one such representation and an ensuing generative model. Finally, I would like to venture into creative modeling, perhaps a new territory in machine intelligence: can machines learn to generate 3D shapes creatively?

Biography:

Hao (Richard) Zhang is a professor at Simon Fraser University (SFU), Canada, where he directs the graphics lab. He has also been a visiting professor at Stanford, Shandong University, and Shenzhen University. Richard obtained his Ph.D. from the University of Toronto, and MMath and BMath degrees from Waterloo. He works in computer graphics with special interests in geometry modeling, shape analysis, machine learning, and computational design and fabrication, and has published more than 100 papers on these topics. Richard is an editor-in-chief of Computer Graphics Forum and has been a program chair for SGP’13, SIGGRAPH Asia Course’14, GI’15, and will be conference chair for Geometry Summit’19. He is an IEEE Senior Member and his awards include an NSERC DAS Award (2014), best paper awards from SGP 2008 and CAD/Graphics 2017, an SFU Research Excellence Award (2014), and a National Science Foundation of China Overseas Outstanding Young Researcher Award (2015).

“Simulating fluids with the “path bundle” method.”

by Dr. Bruno Levy

Inria Research Director

Day 3, June 13 — 9:15~10:00

Abstract:

In this presentation, I’ll introduce the “path bundle” method, a method for simulating fluid dynamics. Recent advances in computational mathematics (Gallouet-Merigot scheme, Brenier projection) had spectacular applications in computer graphics (DeGoes et.al, power particles). The “path bundle” method elaborates on these approaches, and aims at simulating fluid behaviors such as turbulence in incompressible fluids based on very simple considerations (Newton laws for a single Lagrangian particle). I consider a continuum of particles with their trajectories parameterized by time (characteristics). The “path bundle” method considers groups of characteristics, described by a small number of parameters (e.g. center of mass, momentum, angular momentum). The PDE that governs the time evolution of these parameters is deduced from a principle of least action under the constraint that each characteristic remains in the same group. The numerical problem is solved by specific algorithms that take into account the geometric nature of the equation. The benefits of the method is that interface tracking (including free surface) becomes very easy. More importantly, since the method is purely Lagrangian, another benefit is the absence of field interpolation that introduce dispersion/damping when fields are transfered between Lagrangian particles and the Eulerian grid (here there is no Euler grid), thus fine scale motion of the fluid is accurately represented. Finally, the formulation makes it easier to enforce conservation laws (mass, linear and angular momentum, energy).

Biography:

Bruno Levy is a senior researcher with Inria, and the head of the ALICE group (geometry processing and computer fabrication) that he created in 2004 (now eight faculties). He received the Inria/French Academy of Sciences young researcher award in 2011. He is associate editor for The Visual Computer, ACM TOG and Graphical Models, and he is a member of the steering committee of SMA/SPM. He was paper co-chair of Eurographics 2014, Pacific Graphic 2013, SGP 2010, SPM 2008 and 2007. From 1998 to 2006, he focused on mesh parameterization, texture mapping and conversion between representations (e.g. mesh to Splines). From 2007 to 2014, in the frame of his ERC projects GOODSHAPE and VORPALINE, he worked on sampling (vector quantization) and meshing (isotropic, anisotropic, hex-dominant mesh generation). His latest research concerns computational physics, more specifically numerical algorithms for solving some partial differential equations and practical geometric algorithms for generating structured and unstructured meshes. All of his research results are available as open-source software (Graphite/Geogram) or proprietary software (Vorpaline).