Graphics Reference
In-Depth Information
(a)
(b)
(c)
(d)
(e)
(f )
Figure 1.15
Comparison of four different stereo matching algorithms. (a) Top left non-pattern frame,
captured in ambient lighting conditions. (b) Sequence of top left pattern frames, captured under patterns
projections. (c) Reconstructed face using traditional stereo matching with a [15
×
15] window achieved
on non-pattern left stereo frames. The result is noisy due to the lack of color variation on the face. (d)
Reconstructed face using pattern frames (examples are given in (b)) using stereo matching with a [15
×
15] window. The result is much better because the projected stripes provide texture. However, certain face
details are smoothed out due to the need for a large spatial window. (e) Reconstructed face using local
spacetime stereo matching with a [9
×
5
×
5] window. (f) Reconstructed face using the global spacetime
stereo matching with a [9
×
5
×
5] window. Global spacetime stereo matching removes most of the
striping artifacts while preserving the shape details [from http://grail.cs.washington.edu/projects/stfaces/]
given in Figure 1.15
d
, where
active stereo
is used. The projected colored stripes generate
texture on the face, which helps the spatial matching process. However, certain facial shape
details are smoothed out because of the largeness of the used spatial window (15
×
15).
Frames shown in Figure 1.15
b
) illustrate pattern projections on the face across time.
•
Temporal stereo matching.
In this stereo-matching schema, establishing correspondence for
a pixel (
x
l
,
,
M
is based, this time, on temporal neighborhood
V
t
=
t
0
±
t
,
instead of the spatial window
V
s
. Thus, one can define the matching function
F
(
x
r
)as
follows,
y
t
0
)inframe
t
0
)))
2
F
(
x
r
)
=
(
I
l
(
V
t
(
x
l
,
t
0
))
−
I
r
(
V
t
(
x
r
,
(1.36)
V
t
The previous equation is analogous to Equation 1.35 except that now instead of a spatial
neighborhood, one must consider a temporal neighborhood
V
t
around some central time
t
0
.
Because of the changing of the light patterns over time, this temporal window works. This
