Computing the gradients of computer graphics models has become increasingly crucial for computer vision and machine learning in solving inverse problems, inference, or synthesizing images. When differentiating a light simulator/renderer, several challenges arise. Firstly, object boundaries and occlusion introduce discontinuities. Naive automatic differentiation would fail to account for the resulting Dirac delta signals from the differentiation of discontinuities, leading to incorrect results. Secondly, light transport simulation and its differentiation require solving a high-dimensional integral using stochastic estimators. Naive estimators of derivatives may exhibit high variance and cause divergence in gradient-based optimization. In this talk, I will discuss our recent work on new numerical methods and programming languages for addressing these challenges.
Bio: Tzu-Mao Li
Tzu-Mao Li is an assistant professor at the CSE department of UCSD, working at the Center for Visual Computing. He explores the connections between visual computing algorithms and modern data-driven methods and develops programming languages and systems for facilitating the exploration. He did a 2-year postdoc at both MIT CSAIL and UC Berkeley. He did his Ph.D. in the computer graphics group at MIT CSAIL. He received his B.S. and M.S. degrees in computer science and information engineering from National Taiwan University, respectively, where he worked at the Chuang at the Communication and Multimedia Lab. He received the ACM SIGGRAPH 2020 Outstanding Doctoral Dissertation Award and the NSF CAREER Award.
See upcoming and previous presentations at CS Seminar Series