Speaker: Daniel Cornel
Stochastic sampling is an indispensable tool in computer graphics which allows to approximate complex functions and integrals in finite time. Applications which rely on stochastic sampling include ray tracing, remeshing, stippling and texture synthesis. In order to cover the sample domain evenly and without regular patterns, the sample distribution has to guarantee spatial uniformity without regularity and is said to have blue noise properties. Additionally, the samples need to be distributed according to an importance function such that the sample distribution satisfies a given sampling probability density function globally while being well-distributed locally. The generation of optimal blue noise sample distributions is expensive, which is why a lot of effort has been devoted to finding fast approximate blue noise sampling algorithms. Most of these algorithms, however, are either not applicable in real-time or have weak blue noise properties.
Forced Random Sampling is a novel algorithm for real-time importance sampling. Samples are generated by thresholding a precomputed dither matrix with the importance function. By the design of the matrix, the sample points show desirable local distribution properties and are adapted to the given importance. In this thesis, an efficient and parallelizable implementation of this algorithm is proposed and analyzed regarding its sample distribution quality and runtime performance. The results are compared to both the qualitative optimum of blue noise sampling as well as the state of the art in real-time importance sampling, which is Hierarchical Sample Warping. With this comparison it is investigated whether Forced Random Sampling is competitive with current sampling algorithms. The analysis of sample distributions includes several discrepancy measures and the sample density to evaluate their spatial properties as well as Fourier and differential domain analyses to evaluate their spectral properties.
With these established methods it is shown that Forced Random Sampling generates samples with approximate blue noise properties in real-time. Compared to the state of the art, the proposed algorithm is able to generate samples of higher quality with less computational effort and is therefore a valid alternative to current importance sampling algorithms.