Graphics Reference
In-Depth Information
have sharp changes, the geometry smoothing term
u
ij
+
v
ij
is given a position-
dependent weighting function that de-emphasizes the value at points near an ap-
parent discontinuity. This ensures that the reconstructed height field preserves the
important discontinuities, i.e., is not too smooth.
The reconstructed height field corresponds to the small surface sample of the
BTF data. This is not large enough to use by itself, so it must be expanded. As
noted above, tiling the surface with the height field of the small sample will not
be visibly plausible because of the obvious repetition of the sample. Instead, a
basic 2D texture-synthesis algorithm applied to the height field after it has been
converted to a grayscale image; the gray value is a function of the height. This re-
sults in a general synthesized height field for the entire surface; however, it is not
suitable for direct use in BTF synthesis. Imprecision in the geometry reconstruc-
tion is propagated by the texture-synthesis algorithm. Height fields reconstructed
from different viewpoints cannot be expected to have the same geometry, so the
synthesized height fields will vary significantly. The BTF synthesis works by
matching features of a synthesized texture with a captured image having a similar
lighting/viewing directions.
Conceptually, the synthesis algorithm works by combining the important fea-
tures of the reference image with the geometry of the template image. To perform
the BTF synthesis at a desired viewpoint and lighting direction, the authors create
a “template” image by rendering the synthesized height field with shadows from
the given lighting/viewing direction. Next, they find a “reference” image in the
captured image data set that has a similar orientation. The first step is to warp the
reference image to the given viewpoint. If the reference image is sufficiently close
to the desired lighting/viewing direction, this can be done using a simple projec-
tive transformation. Otherwise it becomes a more difficult structural morphing
problem, because the apparent light direction has to be appropriately adjusted.
The authors describe an algorithm for this in the paper, which is valuable in itself
from a practical standpoint, as it provides a kind of direct BTF sample interpola-
tion.
The process of combining the template image and the reference image amounts
to replicating detail from the reference image into the template image. The method
works by selecting blocks of pixels from the reference image and distributing
them over the template image. The blocks are chosen according to a measure
of importance: blocks that have the most significant detail are copied first, then
blocks of less visual importance are copied. The template image is of course much
larger than the reference image, so each block needs to be copied multiple times.
When there are no spaces in the synthesized image large enough to accommo-
date a full block, smaller blocks are copied that can overlap with existing blocks.
Finally, a 2D texture-synthesis algorithm is applied to fill in the remaining holes.
