Research News: Early-career research (ECR) funding - Control variates for efficient rendering
Yifan Zhou, PDRA, in the signal processing group, has recently won funding from the university’s Early Career Research fund which will enable him to continue his research on using control variates for efficient computer graphical rendering.
Background
While being a PDRA funded by the £2.5m EPSRC project, “Big Hypotheses”, Yifan has had a paper published in IEEE Signal Processing Letters that described a way of using Control Variates to solve problems which have stochastic variables in constrained domains. This method has been tested on several statistical models and reported improvements in accuracy of orders of magnitude with negligible additional computational cost. The objective of the ECR funded research is to investigate the use of this method to improve the efficiency of Monte-carlo Light Transport (for rendering).
Importance of the research
There is a significant market for advanced rendering in the context of both high-end special effects (for blockbuster films) and computer games. While there have been recent advances in the application of neural networks (eg Neural Radiance Fields, NeRFs) in these contexts, such approaches fail to adequately capture the physical understanding of how light interacts with the environment to produce realistic images involving complex optical effects (e.g. caustics). Monte-carlo light transport is a state-of-the-art approach for physics-based rendering. Since it relies on Monte-carlo simulation, it is computationally costly and real-time high-fidelity physical simulation is beyond what is believed to be possible. Were it possible to dramatically alter the trade-off between accuracy and computational cost, there would be significant potential for high profile publications, knowledge exchange and impact.
What comes next?
The success of winning the Early-career research funding will enable the researcher to continue the research and visit the collaborators. It is anticipated that a testing environment for rendering will be implemented. The research will include investigate which control variates can perform well in the context of rendering regrading both accuracy and reducing computational cost.
Related paper can be found here: https://ieeexplore.ieee.org/stamp/stamp.jsp?arnumber=9944852