ACM SIGGRAPH has honoured MIT CSAIL postdoctoral researcher Li Tzu-Mao with its 2020 Doctoral Dissertation Award for his PhD thesis Differentiable Visual Computing. Launched in 2016, The Doctoral Dissertation Award is presented annually and recognizes young researchers who make notable contributions during their doctoral studies. ACM SIGGRAPH says Li’s dissertation provides a foundation for the emerging differentiable computer graphics research field and hails him as a “pioneer of the new field of physically-based differentiable rendering.”
In his 148-page dissertation, Li addresses challenges involved in obtaining and applying derivatives for complex graphics algorithms and investigates the use of derivatives in the context of visual computing. He introduces three tools related to computing and the application of derivatives for computer graphics, image processing and deep learning applications: differentiable image processing, differentiable Monte Carlo ray tracing, and Hessian-Hamiltonian Monte Carlo rendering.
Efficient Automatic Differentiation for Image Processing and Deep Learning
Previous research efforts had to compose programs with limited coarse-grained operators or hand-deriving derivatives. Li’s dissertation extends the image processing language Halide with reverse-mode automatic differentiation, which can automatically optimize the gradient computations and promises automatic generation of the gradients of arbitrary Halide programs with high performance and little effort from programmers.
Differentiable Monte Carlo Ray Tracing through Edge Sampling
Performing 3D rendering requires gradients related to variables such as camera parameters, light sources, geometry, and appearance. The gradient calculation is however challenging because the rendering integral includes visibility terms that are not differentiable. Li’s dissertation proposes the first general-purpose differentiable ray tracer, which solves the full rendering equation while correctly taking geometric discontinuities into account.
Hessian-Hamiltonian Monte Carlo Rendering
This dissertation also demonstrates that the derivatives of light path throughput — especially those that are second-order — can be useful for guiding sampling in forward rendering. It extends the Metropolis Light Transport algorithm by adapting to the local shape of the integrand using second-order Taylor expansion, thereby increasing sampling efficiency.
Li is a postdoctoral researcher at the MIT Computer Science & Artificial Intelligence Laboratory (CSAIL) who did his PhD in the computer graphics group at MIT CSAIL and a six-month postdoc at UC Berkeley. He received BS and MS degrees in computer science and information engineering from National Taiwan University in 2011 and 2013.
SIGGRAPH started in 1969 and has grown into a leading global community of researchers, artists, developers, filmmakers, scientists and businesspersons interested in computer graphics and interactive techniques. It is part of the Association for Computing Machinery (ACM), the world’s first and largest computing society.
The paper Differentiable Visual Computing is on the MIT CSAIL website.
Author: Yuqing Li | Editor: Michael Sarazen & Fangyu Cai