In case you haven’t heard, the UNM Advanced Graphics Lab has a new
paper that has been accepted to the ACM Transactions on Graphics
journal and will be presented as a technical paper at the ACM SIGGRAPH
conference next year in Los Angeles.
The paper, titled “On Filtering the Noise from the Random Parameters
in Monte Carlo Rendering” examines an old problem in computer graphics:
the noise that appears in renderings computed with the Monte Carlo
techniques. We have posted a pre-print of the paper here:
You see, rendering systems sometimes use Monte Carlo techniques to
approximate the multidimensional integrals that model light transport
in complex environments. Because of their statistical nature, the
estimates from Monte Carlo have some variance which appears as noise in
the final image. Typically, a significant amount of time (or computing
power) is needed for the Monte Carlo algorithm to converge to the point
where it is suitable for high-quality rendering.
Researchers in graphics have been studying this problem since the mid
1980’s when Monte Carlo techniques were first introduced as a solution
to the global illumination problem. Although many algorithms and
solutions have been proposed, none of them are able to deal with the
noise in an efficient manner. The fundamental problem is that it is
hard for an algorithm to examine the final image and determine what
is noise to be removed and what is scene detail (such as noisy detail
on the floor).
In this paper, my student (Soheil Darabi) and I propose a new way to
distinguish noise from scene detail in Monte Carlo rendering. To do this,
we treat the rendering system as a black box and use concepts from
information theory to see if the output sample values are dependent on
the random parameters used in the Monte Carlo simulation.
Our basic observation is that the noise we want to remove will be
dependent on these random parameters (after all, the random parameters
act as a noise source in our system), while scene detail does not. This
allows us to implement a simple filter that detects this dependency and
removes only the unwanted noise while preserving the desired detail. This
simple algorithm produces results that would have normally taken 100x
longer to produce.
We are currently in the process of working with companies like Sony
Pictures Imageworks that use Monte Carlo-based rendering systems to
see if we can integrate our algorithm into their rendering pipeline.
We hope that this would allow them to produce results in a fraction
of the time, saving them considerable computing resources and money.
Please check out the video to learn more about the paper!
Thanks to Pradeep Sen of the Dept. of Electrical & Computer Engineering
University of New Mexico for the information.