Graphics Reference
In-Depth Information
(a)
(b)
Figure 8.19. (a) Two images in which the projector is fully on and totally off allow the determi-
nation of a per-pixel threshold. In this case the (on,off) intensities at the (darker) red pixel are
(137, 3) and at the (brighter) green pixel are (226, 18), leading to per-pixel thresholds of 70 and
122, respectively. (b) Alternately, each binary pattern and its inverse can be projected, and we
interpreta1ifthepattern at a pixel is brighter than its inverse.
1
2
3
4
Figure 8.20. Hall-Holt and Rusinkiewicz's proposed patterns for stripe boundary coding.
that we can use to determine the on/off state for each pixel in the subsequent binary
patterns.
Alternately, Scharstein and Szeliski [ 426 ] suggested projecting each binary pattern
followed by its inverse, as illustrated in Figure 8.19 b. The codeword bit is assigned
as 1 if the pixel's intensity is brighter in the original pattern compared to the inverse
pattern, and 0 in the opposite case. They claimed this approach was more reliable
than using all-on and all-off images; on the other hand, it requires projecting twice as
many patterns. Regardless of the approach, scanning objects that are shiny or contain
surfaces with very different reflectances can be difficult. As with LiDAR scanning, the
best-case scenario is a matte object with uniform reflectance.
Hall-Holt and Rusinkiewicz [ 185 ] advocated the use of stripe boundary codes .
That is, instead of trying to detect the center of each stripe to use in triangulation,
they proposed to detect the boundary between stripes, which can bemore accurately
located. Thus the changing pattern of on/off illumination of the stripes on each side
of the boundary generates the codeword. In particular, they proposed the set of four
patterns illustrated in Figure 8.20 ; each pattern contains 111 stripes, and each of the
110 quadruples of on-off patterns across the stripe boundaries occurs only once. 11
Clearly, we can reduce the number of projected patterns required to define a
codeword if we allow the patterns to have more than two states. One possibility is
to allow grayscale values in the projections, and another is to allow colored stripes.
In either case, if each stripe in a pattern can be in one of N states, then N M unique
stripes can be coded with M patterns. Horn and Kiryati [ 204 ] explored the grayscale
approach, usingHilbert space-filling curves to producewell-separated codewords for
a user-specified number of patterns or stripes. However, the more gray-level states
11 Note that we can't directly resolve a boundary when the illumination is constant across it (e.g.,
white-white). However, the sequence is designed so that the patterns both before and after such
an occurrence have a visible illumination change, which localizes the boundary.
 
Search WWH ::




Custom Search