Game Development Reference
In-Depth Information
share the same image plane. This process is known as stereo rectification
(Hartley, 1999) and is intended to eliminate the photometric and orientation
differences between the two source cameras to simplify the correspon-
dence estimation into a 1-D search problem along the scan line and at the
same time to provide parallel processing possibilities for later steps.
2)
X-interpolation
: Given the disparity information, the two parallel views
V
rectiL
and
V
rectiR
are combined by interpolation or extrapolation (Seitz &
Dyer, 1995) to produce another parallel view
V
x
. The corresponding
camera
C
x
is located at
[
]
cD
x
with the same rotation and intrinsic
parameters as
C
rectiL
and
C
rectiR
. The
y
coordinate of each pixel remains the
same, while the
x
coordinate is transformed by
0
0
1 −
x
x
X
p
x
rectiL
p
cD
d
LR
p
=
+
2
and/or (in case of occlusion)
1
+
x
X
p
rectiR
p
cD
RL
p
x
=
x
+
d
,
2
where
x
p
is the
x
coordinate of pixel
p
*
in view
V
*
(* =
X, rectiL, rectiR
).
p
rectiL
and
p
rectiR
are projections of the same 3-D point.
LR
p
RL
p
d
d
and
are
LR
p
rectiR
p
rectiL
p
disparities of
p
rectiL
and
p
rectiR
respectively, where
d
=
x
−
x
and
RL
p
rectiL
p
rectiR
p
d
=
x
−
x
. Note that in the case of occlusion, either
p
rectiL
or
p
rectiR
is not available (Lei & Hendriks, 2002). Through this step, the
difference in
x
position with the final view
V
D
is eliminated.
3)
Y-extrapolation
: The X-interpolated view
V
x
is extrapolated (Scharstein,
1999) by shifting pixels in the
y
direction to produce the view
V
Y
, which
comes from a virtual camera
C
Y
located at
[
]
cD
yx
with the same
rotation and intrinsic parameters as
C
x
. In this process, the
x
coordinate of
each pixel remains the same while the
y
coordinate is transformed by
X
p
0
cD
s
Y
p
X
p
x
y
=
y
−
y
⋅
⋅
d
*
p
y
, where
is the
y
coordinate of pixel
p
*
in view
V
*
cD
s
y
X
p
X
p
LR
p
d
d
=
d
/
2
(
* =
X
,
Y
).
is the disparity of
p
X
, where
or (in case of
X
p
RL
p
occlusion)
. Through this step, the difference in y position with
the final view
V
D
is eliminated.
d
=
−
d
/
2
4)
Z-transfer
: The Y-extrapolated view
V
Y
is transferred along the
z
direction
to generate a closer or more distant look
V
Z
. The corresponding camera
C
Z
is located at
[
]
yx
with the same rotation and intrinsic param-
eters as
C
Y
. Both the
x
and
y
coordinates of each pixel would be transformed
z
cD
cD
cD
Search WWH ::
Custom Search