Talk by Guandao Yang
Geometry Processing with Discretization-Free Representation and Prior
Geometry processing algorithms provide a pipeline to acquire, manipulate, and analyze geometry data. These algorithms have been an indispensable part of our life, powering many applications including creating animation and industrial designs. Traditionally, geometry processing is mostly done with explicit representations such as polygon meshes and point clouds, which require some predetermined discretization of a continuous surface. In this talk, I will show that one can avoid such discretization by using deep neural networks to represent geometry data, encode geometry prior, and perform tasks such as shape deformation and smoothing. The talk will first cover some background on neural fields, deep generative models, and geometry processing. After that, I will introduce my works on using deep generative models to create a continuous representation and a learning-based prior for point clouds (Yang et. al., 2019, Cai et. al. 2020). Such representation and prior allow us to acquire and complete geometry data from observation. Then I will show how to manipulate such geometry representation can without discretizing the surface (Yang et. al. 2021). Finally, I will describe the potential pathway to do geometry analysis with such discretization-free representation.
Guandao Yang is a Ph.D. student at Cornell University, advised by Prof. Serge Belongie and Prof. Bharath Hariharan. His research focuses on shape analysis, which lies in the intersection of computer vision, graphics, and machine learning.