Graphics Reference
In-Depth Information
advantage of using 2D pseudorandom arrays is that the correspondence problem is
relatively unambiguous. A related idea proposed by Salvi et al. [
417
] is to use a grid
generated by a (red, green, blue) de Bruijn sequence for the horizontal lines and a
(cyan, magenta, yellow) de Bruijn sequence for the vertical lines. The intersections of
the grid thus uniquely encode a location in the projected image. However, these 2D
approaches generally have lower spatial resolution in the image plane than the other
methods discussed in Section
8.2
. Lanman et al. [
262
] described a clever approach for
360
◦
structured light scanning of small objects using an array of mirrors to surround
the object with light and enable viewing it from all sides with a single camera.
Koppal et al. [
251
] noted that a high-speed DLP projector could be reverse engi-
neered to determine the dithering pattern it uses to display a grayscale intensity —
that is, the on/off flipping each micromirror uses to modulate the light from the
always-on projector bulb. They exploited this dithering pattern as ameans of creating
coded structured light for a very high-speed camera. Other extensions to structured
light include modifications for underwater scanning, in which the water significantly
scatters the light beam [
345
] or for scanning difficult refractive objects (e.g., a crystal
goblet) by immersing them in a fluorescent liquid [
209
].
While most structured light methods operate in the visible spectrum, Früh and
Zakhor [
157
] suggest using infrared light instead, which is likely to be much less
perceptually distracting for a human being scanned. Since there is no natural sense
of “color” in the infrared spectrum, they used a pattern of vertical stripes along with a
periodically sweeping horizontal line that cuts across the stripes to disambiguate the
stripe index. The
Microsoft Kinect
also uses a proprietary structured light approach
in the infrared spectrum to generate real-time depthmaps. A pseudorandompattern
of infrared dots is projected into the scene, which plays the same role as a coded light
stripe pattern since the local dot neighborhoods are unique.
The proprietary
MOVA Contour
system used in the visual effects community for
markerless facial motion capture also works in the nonvisual spectrum. Phospho-
rescent makeup is applied to a performer that is invisible under white light but
phosphoresces as a blotchy, mottled pattern under fluorescent light. The two types of
light are toggled rapidly and the visible and fluorescent images are separated by mul-
tiple high-speed video cameras surrounding the performer. A proprietary multi-view
stereo technique fuses the depth maps obtained from the fluorescent images.
In 1996, Debevec et al. presented the Façade system for 3D modeling and ren-
dering of architectural scenes [
117
], a great example of research from the academic
community that had a major impact on visual effects production in Hollywood. This
image-basedmodeling approach has been used inmanymovies including the
Matrix
and
Transformers
series. A user sketches edges on a small set of wide-baseline pho-
tographs of a piece of architecture, and interactively specifies how the edges match
to block primitives of a 3D model. The user can specify how blocks are related (for
example, that a roof must sit squarely on its base) as well as symmetries in the model
(for example, that a clock tower has four identical sides). The 3D model then highly
constrains a structure from motion problem in which the parameters of the archi-
tectural model and the camera locations are estimated. To create a new view of the
piece of architecture, the 3D model is texture-mapped with an appropriate set of
input images, which can produce a visually compelling result even when the under-
lying model is fairly simple. If the model needs to be refined, a higher-quality depth
