Graphics Reference
In-Depth Information
Figure 2.7
Sample point clustering and ways to avoid it. Random samples sometimes end up too
close to each other (top left). This can be mitigated by first partitioning (stratifying) the
region and choosing a fixed number of samples in each subregion (top right and middle).
In contrast, quasi-Monte Carlo sampling uses a fixed sequence of sample points designed
to adequately cover the sampling region (bottom). (Courtesy of Alexander Keller.)
them. One approach is to start with a coarse sampling on the hemisphere and then
use that to guide the importance sampling.
A nonuniform sample distribution is governed by an
importance function
.The
BRDF is largest where the reflection is most significant, so it is a natural choice for
an importance function in GI integration. However, the BRDF says nothing about
the directions of the strongest incoming light. Photon simulation can be used to
precompute the dominant light directions. When this information is used in con-
junction with the BRDF, better sampling can be achieved. Figure 2.8 illustrates
such an approach. In the figure, the hemisphere is divided into smaller patches
