Zoom: https://illinois.zoom.us/j/82486445903?pwd=aHBGSWFwMHE4R2J2SXlpUlhMcUlsUT09
Abstract:
Physics-based forward rendering, a central topic in computer graphics, focuses on simulating the flow of light in fully described virtual scenes. On the other hand, inverse rendering is concerned with inferring scene parameters based on measured images of the scene and has numerous applications in science and engineering.
In practice, a widely adopted framework for solving inverse rendering problems is to formulate them as numerical optimizations that seek scene parameters minimizing the difference between simulations and measurements. To solve these optimizations, a key ingredient is to compute gradients of forward rendering results with respect to the scene parameters---a process known as differentiable rendering.
Unfortunately, despite long-lasting research efforts, grand challenges remain:
• Generality: Most existing differentiable rendering techniques consider only simple light transport models and offer limited support for differentiating with respect to object shapes.
• Robustness: Differentiable rendering methods capable of handling complex light transport phenomena (e.g., caustics) have been lacking.
• Ambiguities: Inverse rendering problems are typically under-constrained, which can cause the optimization results to generalize poorly to novel conditions.
My research aims to enable differentiable and inverse rendering that is efficient, physically accurate, and enjoy the generality to handle arbitrary scene parameters under a wide variety of light transport phenomena.
In this talk, I present some of our recent works that have greatly advanced physics-based differentiable and inverse rendering:
• We devised new mathematical tools to describe how infinitesimal changes of a virtual scene affect the distribution of light. These formulations offer the generality to handle a variety of light transport models (e.g., steady-state, and transient).
• Based on the new formulations, we introduced new Monte Carlo differentiable rendering algorithms that enjoy the generality of our formulations while providing low-variance derivative estimates.
• Utilizing our formulations and algorithms, we developed new physics-based inverse rendering pipelines that offer a new level of generality and accuracy.
In addition, I discuss some remaining challenges as well as important future research topics.
Bio:
Shuang Zhao is an Associate Professor of Computer Science at the University of California, Irvine (UCI) and co-directs UCI's Interactive Graphics & Visualization Lab (iGravi). Before joining UCI, Shuang was a postdoctoral associate at MIT. Shuang obtained his Ph.D. in computer science from Cornell University in 2014. Shuang received the NSF CAREER award in 2023.
Faculty Host: David Forsyth
Meeting ID: 824 8644 5903; Password: csillinois
