Information Technology Reference
In-Depth Information
3.1
Colour Segmentation
Considering a pair of colour images that can be used to extract a 3-D structure, we
mainly focus on the region edges where most likely embed depth discontinuities. To
extract homogeneous regions mean shift based colour segmentation [21] is applied
to search for a maxima in a density function. This process is demonstrated in Fig. 7,
where (a) and (b) are original colour images, and (c) is the colour segmentation by
mean shift.
(a)
(b)
(c)
Fig. 7
Colour images and the segmentation by mean shift: (a) Left image, (b) right colour
and (c) segmentation result
A surface comprises a number of patches that can be represented by a dispar-
ity plane:
d
=
c
1
x
+
c
2
y
+
c
3
,where(
x
,
y
) refers to image pixel coordinates, and
(
c
1
,
c
2
,
c
3
) are used to determine a disparity
d
. Without further process, the available
disparity planes will be redundant and sometimes appear to be “noisy”. A number of
approaches can be used to reduce the noise. Klaus
et al.
[8] utilised a self-adapting
dissimilarity measure that integrates the sum of absolute intensity differences (SAD)
and a gradient based measure which is defined as
∑
F
SAD
(
x
,
y
,
d
)=
I
1
(
i
,
j
)
−
I
2
(
i
+
d
,
j
)
(1)
(
i
,
j
)
∈
N
(
x
,
y
)
and
∑
F
GRAD
(
x
,
y
,
d
)=
(
i
,
j
)
∈
N
x
(
x
,
y
)
|
x
I
1
(
i
,
j
)
−
x
I
2
(
i
+
d
,
j
)
|
∑
+
N
y
(
x
,
y
)
|
y
I
1
(
i
,
j
)
−
y
I
2
(
i
+
d
,
j
)
|
,
(2)
(
i
,
j
)
∈
where
N
(
x
,
y
) is a 3
3 window surrounding position (
x
,
y
).
N
x
(
x
,
y
) is a window
without the rightmost column,
N
y
(
x
,
y
) is a window without the lowest row,
×
x
is
the forward gradient to the right and
x
is the forward gradient to the left.
between
F
SAD
and
F
GRAD
can be used to maximise the
number of reliable correspondences that are handled by a cross-checking scheme in
An optimal weight
ω