Source: ACM SIGGRAPH Citation
Tzu-Mao Li’s dissertation explores the connections between visual computing, programming systems, and statistical learning. He connects classical computer graphics and image processing algorithms with modern data-driven methods to facilitate physical understanding. He uses mathematical tools from statistics and machine learning to develop new algorithms that solve graphics and vision problems. He also develops programming systems that simplify the efficient implementation and mathematical derivations of learnable visual computing algorithms. The dissertation’s contributions provide the foundation of the newly emerging research field of differentiable computer graphics. In particular, Tzu-Mao is a pioneer of the new field of physically-based differentiable rendering.
The main theme of Tzu-Mao’s dissertation is about addressing the challenges of computing and applying the derivatives of complex graphics pipelines, in order to use them for fitting and sampling parameters or solving inverse problems. The contributions go beyond traditional automatic differentiation by addressing discontinuities in graphics algorithms and massive parallelism for modern hardware.
A major contribution of the dissertation is the derivation and implementation of the first comprehensive differentiable rendering solution, that computes correct derivatives of scalar functions over a rendered image with respect to arbitrary scene parameters such as camera pose, scene geometry, materials, and lighting. This enables a wide variety of graphics and vision algorithms for analyzing 3D properties of images using the derivatives, including 3D reconstruction and adversarial examples generation.
Another major contribution is the introduction of a domain-specific automatic differentiation compiler that builds on Halide for differentiating image processing algorithms. The compiler enables automatic generation of the gradients of complex image processing programs, at high performance, with little programmer effort. This contribution opens up new ways to develop efficient and accurate data-driven image processing algorithms using flexible building blocks, as opposed to the coarse-grained operators often used in deep learning.
Finally, the dissertation introduces the first MCMC rendering algorithm that uses the second derivatives of the light path contribution to speed up rendering of challenging effects such as moving caustics or multi-bounce illumination with glossy materials. In particular, the Hessian of the light transport contribution is used to capture the strong anisotropy of the integrand.
The research code used for the dissertation is open-source and has numerous users in academia and industry, including graduate students at MIT, UC Berkeley, Cornell, and Brown, and industrial research labs such as Adobe, Google, Technicolor, and InterDigital. In particular, the differentiable renderer, “redner”, written by Tzu-Mao for the dissertation is downloaded more than 100,000 times so far.
Tzu-Mao is currently a postdoctoral researcher at MIT CSAIL working with Jonathan Ragan-Kelley. He did a six-month postdoc with Jonathan Ragan-Kelley at UC Berkeley. Before the postdoc, he did his Ph.D. in the computer graphics group at MIT CSAIL, advised by Frédo Durand. He received his B.S. and M.S. degrees in computer science and information engineering from National Taiwan University in 2011 and 2013, respectively. During his time at National Taiwan University, he was a member of the graphics group at Communication and Multimedia Lab, where he worked with Yung-Yu Chuang.