Graphics Reference
In-Depth Information
segmentation are supported; Green screening according to Eq.
6.1
, where
R
I
i
,
G
I
i
,
and
B
I
i
are the red, green, and blue components of
I
i
. For clarity the pixel location
(
x
,
y
)
has been omitted.
1
:
G
I
i
>˄
g
·
(
R
I
i
+
G
I
i
+
B
I
i
)
S
i
=
(6.1)
0
:
G
I
i
≤
˄
g
·
(
R
I
i
+
G
I
i
+
B
I
i
)
The second method is able to subtract a real-life background [
18
] according to
Eq.
6.2
, where
I
B
i
is the static background picture and
˄
a
are experimentally
determined thresholds which are subjected to parameter fine tuning. For shadow
removal, the cosine of the angle
I
i
I
B
i
between the color component vectors of the
image pixel
I
i
(
˄
g
,
˄
f
,
˄
b
,
is determined. As a
final step, the silhouette is further enhanced by a single erosion and dilation [
30
].
x
,
y
)
and the static background pixel
I
B
i
(
x
,
y
)
I
i
−
I
B
i
>˄
f
⊧
⊨
:
1
or
I
i
−
I
B
i
≥
˄
b
(
I
i
I
B
i
)
≤
˄
a
and cos
I
i
−
I
B
i
<˄
b
S
i
=
(6.2)
0
:
or
⊩
I
i
−
I
B
i
≤
˄
f
(
I
i
I
B
i
)>˄
a
and cos
Both methods are evaluated on a pixel basis and require very little processing
power, while still being robust against moderate illumination changes.
6.3.2 View Interpolation
To interpolate the desired viewpoint, we adopt and slightly modify a plane sweeping
approach based on the method of [
32
]. As depicted in Fig.
6.7
a, the 3D space is
discretized into
M
planes
parallel to the image plane of the virtual
camera
C
v
. For each plane
D
j
, every pixel
f
v
of the virtual camera image
I
v
is
backprojected on the plane
D
j
by Eq.
6.3
, and reprojected to the input images
I
i
according to Eq.
6.4
. Here,
T
j
is a translation and scaling matrix that defines the
depth and extent of the plane
D
j
in world space. The relationship between these
coordinate spaces is represented in Fig.
6.7
b.
{
D
1
,...,
D
M
}
V
−
1
v
P
−
1
v
f
=
×
×
T
j
×
f
v
(6.3)
f
i
=
P
i
×
V
i
×
f
(6.4)
Points on the plane
D
j
that project outside a foreground silhouette in at least one
of the input images, are immediately rejected—e.g., point
g
in Fig.
6.7
a—and all
further operations are automatically discarded by the GPU hardware. This provides
a means to lever both speed and quality because noise in the segmentation masks
will, with a high probability, not be available in all
N
cameras. Otherwise, the mean
