Graphics Reference
In-Depth Information
Curless and Levoy's space-time analysis to improve stripe localization by smooth-
ing the color stripe pattern and shifting it across the object's surface (this is only
applicable to static scenes).
13
Several researchers have extended the basic idea proposed earlier. Je et al. [
217
]
analyzed the problem of color selection for the stripes and suggested that patterns
that used only red, green, and blue were most reliable. Pagès et al. [
359
] proposed
a hybrid color stripe pattern that contains both color edges and a square-wave pat-
tern of intensity; employing multiple intensity levels allows fewer colors to be used.
Schmalz and Angelopoulou [
429
] described a fast method for decoding the stripe
index based on aggregating similarly colored regions of the camera image into super-
pixels and applying a graph-based algorithm to the entire image (instead of decoding
each scanline independently).
In general, correctly identifying stripe colors and color transitions is difficult when
the scene to be scanned is also colorful. For example, a red stripe falling on a green
object surface and a green stripe falling on a red object surface may both present the
same dark color in the camera image. Thus, as with the other methods discussed so
far, a neutral-colored, matte surface is the best-case scenario.
It's important to note that an algorithm that computes 3D point locations based
on a one-shot approach is not necessarily real time. Indeed, Zhang et al.'s original
one-shot algorithmhad a reported running time of aminute per frame (later reduced
to five seconds per frame using a modern processor [
429
]). Rusinkiewicz et al. [
410
]
proposed a real-time 3D data acquisition method based on the binary patterns in
Figure
8.20
that operated at sixty frames per second with a latency of four frames.
In order to allow object motion during scanning, they tracked the stripe bound-
aries across the frames to maintain the correct correspondence of points on the
object surface. Koninckx and Van Gool [
250
] achieved real-time scanning at around
twenty frames per second using a one-shot technique. They used a pattern com-
posed of periodic black and white stripes, cut diagonally by a series of colored lines.
The intersections of the diagonal lines with the bars provide uniquely identifiable
points for triangulation; the stripe labels are refined with a graph-cut algorithm.
An interesting feature of this approach is that the vertical stripe frequency and the
color and slope of the diagonal lines are adapted to the current scene over time,
instead of always using the same pattern. Algorithms from the last class of structured-
light approaches we discuss, fringe-based methods, are also well suited to real-time
acquisition.
8.2.4
Fringe Projection Methods
The stripe-basedmethods we've discussed so far implicitly assume that the projected
pattern looks like a step function in color space. The final class of methods we discuss
is based on the projection of a continuous sinusoidal pattern of intensity, also known
as a
fringe pattern
. This pattern is swept across the scene in discrete steps of phase,
so such methods are also known as
phase-shifting methods
.
13
We discuss similar phase shifting techniques in the next section.
