Sampling of Products of Functions
Advisor: Tomas Akenine-Möller
Completed: March 2005
This thesis presents a novel method for importance sampling
the product of two-dimensional functions. The functions are represented
as wavelets, which enables rapid computation of the product as well as
efficient compression. The sampling is done by computing wavelet coefficients
for the product on-the-fly, and hierarchically sampling the wavelet tree.
The wavelet multiplication is guided by the sampling distribution. Hence,
the proposed method is very efficient as it avoids a full computation
of the product. The generated sampling distribution is superior to previous
methods, as it exactly matches the probability density of the product
of the two functions. As an application, the method is applied to the
problem of rendering a virtual scene with realistic measured materials
under complex direct illumination provided by a high-dynamic range environment
My original idea for this thesis was to implement a PRT (precomputed radiance transfer) system using wavelets to represent the transport vectors, and then try to extend previous work in order to find a way to handle dynamic scenes. The work soon evolved into research of methods for importance sampling according to the wavelet representations. I then realized that it is relatively straightforward to combine wavelet sampling with wavelet products, introduced by Ng et al. in 2004. Thus, efficient sampling of products of functions was made possible. By using this novel sampling method, samples can be distributed according to the product of a BRDF and an environment map. The samples are then used for sampling the visibility through standard ray tracing, hence eliminating the need for precomputed visibility.
Most of my thesis research was conducted at the University of California
at San Diego, where I was visiting during August 2004 to the end of February
2005. Together with Henrik Wann Jensen, Tomas Akenine-Möller, and
Wojciech Jarosz, we further developed the ideas presented in this thesis.
The main extensions were a generalization of the wavelet product to higher
dimensions, and a novel sampling scheme that uses multi-dimensional low-discrepancy
points for reducing the variance. Our efforts resulted in a paper that was
published at SIGGRAPH 2005.
[ thesis.pdf ] (10MB Adobe PDF)
|This page is copyright © 2005 by Petrik Clarberg. All rights reserved.|