Computer Vision Seminar Series: Dr. Qixing Huang, "Enforcing 3D Inductive Bias via Network Design and Regularization Losses."

Abstract: Generative models map a latent parameter space to instances in an ambient space and have found broad applications in 3D vision and related domains. A standard probabilistic framework seeks to align the ambient distribution induced by a generative model, given a prior distribution in the latent space, with the empirical distribution of training instances. While this paradigm has been highly successful in image generation, its application to 3D shape generation faces fundamental challenges due to limited training data and difficulties in generalization. A key distinction between image and shape generation is that 3D shapes exhibit rich geometric, topological, and physical priors that should be preserved during generation. Existing probabilistic approaches to 3D generation often fail to explicitly encode these priors, resulting in synthesized shapes with geometric distortions, topological inconsistencies, or physically implausible properties.

In this talk, I will discuss recent work toward a geometric framework for learning 3D shape generators. The central idea is to incorporate geometric, physical, and topological priors directly into generative models through carefully designed network architectures and regularization losses, enabled by computational tools from differential geometry and computational topology. I will discuss applications to deformable shape generation, latent-space design, joint shape matching, 3D man-made shape generation, and novel-view synthesis.

Speaker Bio.: Qixing Huang (https://www.cs.utexas.edu/~huangqx/) is an Associate Professor in the Department of Computer Science at The University of Texas at Austin, where he directs the Visual Computing Center. His research lies at the intersection of computer graphics, geometry, optimization, computer vision, and machine learning. He has published more than 150 papers in leading venues across these areas. His research has received multiple honors, including several best paper awards, the Best Dataset Award at the 2018 Symposium on Geometry Processing, the IJCAI 2019 Early Career Spotlight, and the 2021 NSF CAREER Award. His recent research has been supported by awards from Adobe, Google, Amazon, and the National Science Foundation. Dr. Huang has served as an area chair or senior area chair for ICLR, ICML, NeurIPS, CVPR, ECCV, and ICCV, and has served on the program and technical paper committees of SIGGRAPH and SIGGRAPH Asia. He also co-chaired the 2020 Symposium on Geometry Processing.

Food and beverages will be served at 11:30 AM before the talk.

There will also be a student roundtable session with Qixing from 3-4 PM at Siebel 3403. Feel free to join!