Graphics Reference
In-Depth Information
may seem almost humorously low-tech; however, it illustrates the simplicity of
Nayar's method and how it can be done with little specialized equipment.
Because of the nonzero width of the penumbra, the shadow of a pole is not
a true brightness step: there is a continuous transition from total shadow to full
illumination. This suggests that a smoother illumination pattern might be used
in place of the binary checkerboard pattern. The authors propose the use of a
wave-like pattern computed from trigonometric functions. The brightness of the
illumination pattern varies according to a sine or cosine function. In this method,
three different wave functions are created by shifting the phase, and an image
is captured for each. Separation is performed as in the moving shadow method,
i.e., by setting
L
−
and
L
+
to the minimum and maximum values, respectively, of
corresponding pixels in the three images. The authors found that the separation
can be performed using these three images directly; there is no need for the 25
separately captured images required for each checkerboard pattern.
The authors conclude the paper with one more method: the separation of di-
rect and global components using a
single
image. The method works by illumi-
nating the scene with a high frequency binary pattern in the form of parallel strips
a few pixels wide. Then the image is separated into square tiles, each of which
is large enough to contain several cycles of the illumination pattern
(
Figure 8.47
)
.
These tiles form a kind of lower resolution version of the image; each tile is a
“pixel” of the smaller image. But unlike the original pixels, the tiles contain high
frequency lighting information. The maximum and minimum pixel values are
averaged across the tile, and using these as the
L
+
and
L
−
values allows the sep-
aration to be performed as usual. However, this only works if the detail of the
object changes slowly enough to be considered approximately constant over each
Tile
i
Patch
i
L
min
→
L
−
L
max
→
L
+
Tile
i
Figure 8.47
Separation using only one image can be accomplished by splitting the image into tiles and
treating the tiles as individual pixels in a smaller image. (From [Nayar et al. 06] c
2006
ACM, Inc. Included here by permission.)
