Graphics Reference
In-Depth Information
(a)
(b)
(c)
Figure 7.8
Distribution of samples (point light source) resulting from (b) the
k
-means algorithm, and
(c) structured importance sampling. The corresponding environment map is shown in (a).
The
k
-means algorithm does not place enough samples across the very bright window at
the far right of the map, which is the primary source of light. (From [Agarwal et al. 03]
c
2003 ACM, Inc. Included here by permission.)
samples constructed by the weighted
k
-means algorithm (via LightGen); those in
the right column were rendered using the structured importance sampling algo-
rithm. Both used the same number of samples (300) to represent the environment
map. The banding effect visible in the
k
-means images is a result of the sudden
appearance of samples representing the bright lights, which occurs because not
enough samples were used for those parts of the environment map. Figure 7.8(b)
and (c) illustrate the distribution of sample points generated by the weighted
k
-
means algorithm and the stratified importance sampling algorithm, respectively,
compared to the environment map itself shown in Figure 7.8(a). The
k
-means
algorithm fails to concentrate samples in the bright windows, which results in the
banding artifacts. The authors point out that in addition to producing better results
with the same number of samples, their algorithm requires much less preprocess-
ing time than is needed for the
k
-means algorithm.
7.2.4 Environment Maps and Prefiltering
Another problem that arises with point sampling HDR environment maps comes
in capturing “high frequency” detail. If the environment map has a lot of detail,
a small variation in a sampling direction can produce a large difference in the
