SIGGRAPH 2023 Outstanding Doctoral Dissertation Award: Zhang
Awardee(s):
Award:
- Outstanding Doctoral Dissertation Award
Description:
Differentiable rendering addresses the problem of differentiating complex light-transport effects (including soft shadows, interreflection, or subsurface scattering) of a computer-generated image with respect to all the parameters describing the scene being rendered – such as the shape of an object, the color of a surface, or the optical density of the surrounding medium. A core issue in differentiable rendering is the handling of discontinuities due to object boundaries and occlusion: the differentiation of these discontinuities leads to Dirac delta distributions inside the rendering equation integral, whose integration via Monte Carlo sampling is notoriously difficult and computationally expensive.
Zhang proposes an ingenious approach to significantly improve the efficiency of discontinuity handling. Starting from the path space formulation of rendering (rather than the one based on integration over solid angle), Zhang’s observation that the typical integrand required to evaluate light transport contains moving discontinuities and that its domain of integration in this formulation is parameter dependent. By importing the Reynolds transport theorem (a staple of continuum mechanics) to rendering, he is able to introduce a change of variables from the parameter-dependent path space to a parameter-independent material space, as well as an integration over discontinuity surfaces in their material path space to account for parameter-dependent discontinuities. This rigorous formulation of differential light transport in path space leads quite elegantly to Monte Carlo methods that estimate the resulting boundary and interior integrals for global illumination.
In addition to his accomplishments in devising the theoretically important differential radiative transfer and path integral formulations, he also makes other significant practical contributions in his dissertation. For instance, Zhang noticed that derivatives in differentiable rendering are often odd functions. He exploits this property to produce negative correlation between samples for variance reduction, which greatly improves the efficiency of Monte-Carlo-based differentiable rendering.
Zhang’s dissertation establishes both sound mathematical foundations and practical algorithms that significantly improve the efficiency of Monte-Carlo-based differentiable rendering. Both for the high quality of his contributions to the field and the exciting developments they are bound to generate in future years, the SIGGRAPH community recognizes Cheng Zhang with the 2023 ACM SIGGRAPH Doctoral Dissertation Award.
Additional Information:
Launched in 2016, the Doctoral Dissertation Award is awarded annually to recognize a recent doctoral candidate who has successfully defended and completed his or her Ph.D. dissertation in computer graphics and interactive techniques. Recognizing young researchers who have already made a notable contribution very early during their doctoral study, the award is presented each year at the SIGGRAPH Conference and is accompanied by a plaque, complimentary full conference registration and travel to the award ceremony. Honorable Mentions may also be awarded.
