ul 2004
are the
' gray-
Te we
| cost,
t. The
be the
ve use
Other
n one
nages.
value
check
(3)
npling
of the
Fa
3). In
ieness
(4)
S bi-
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B3. Istanbul 2004
3.3.2 Smoothness Constraint: The matching pixels must
have similar intensity values(i.e. differ lower than a specified
threshold) or the matching windows must be highly correlated.
Thus, the next probability is proposed by using this
characteristic. We calculate a variance in a disparity map that
satisfies uniqueness constraint, note (Equation 4). Each value is
extracted from weight-function:
0.857
w.(x,y)-
575. (5)
Consequently, we define local variance in the extracted
disparity map:
w Ww
2 1 2 2 - . . . 2
Gy(x.y)-7 N. S NS hw. Gr x, jt y) -x,j- yr
V I. Ww s Ww
S if Ss
l x d - . . *
> S ow Grex jy) dex jy) >
(vy ime w ha
Au. (6)
Ww Ww
2
N, = > Sw +x. +).
wow
pd
where 21-7
Using these conventions, we define the smoothness reliability:
fonte, (7)
where 4 is an damping parameter to control the decreasing
. "pn en ~ + ^r . d 2 .
velocity of fp> . The function value of /p2 is large if op is small,
whereas fz, will take small value in opposite case.
3.3.3 Total Reliability
The final reliability function is defined as a linear combination
of f», and fs» :
doe! = À Fr + As fis , (8)
where À, and A, are weight coefficients satisfying the relation
A + À, =],
Here, we consider local intensity and smoothness constraint.
The number of upper disparity vectors is 9 and that of current
State vectors is 4. Each disparity map is obtained by calculating
the mean absolute difference (MAD) and smoothness of upper
disparity vectors between a pair of stereo. In the modified
hierarchical BMA, we define cost functions S; and D; as:
|
1 ;
S'(x, y,d) = S a^ (x, y )- D' :
pee x =e)
9)
771
wo wo
D'(x,y.d)=are mi ] : A aC ?
(x,y,d)=arg min | — > > Felx +x, p+")
de(-») mn 0
Mam T mm 5
-Fuxtdex,y-y)
(10)
and we can compose our cost function as
ai i i
E=snD+nsS, (11)
where Di denotes the cost function at level of the ith pyramid,
the other parameters are the same as Equation(2), 7, and 7 are
weight coefficients satisfying the relation y, + y,=1.
4. EXPERIMENTS
The simulation result indicates that proposed algorithm shows
outstanding ability in low texture and projective distorted
region such as the nose and cheek in Figure 9 and Figure 13.
Image
Items
“Man” image
“Claude” image
Image size
Actual disparity
Searching range
Block size
384x384
About -100—100
-120~120
7x5
720x288
About -30-30
-45~45
7x5
T 6 S
Weights (4, / 4) 0.5/0.5 0.5/0.5
Weights (y;/ 72) 0.4/0.6 0.4/0.6
Table 1. Parameters for correspondence estimation
o 6
Figure 8. The original stereo images of ‘man’, (a) Left image,
(b) Right image.
(a)
Figure 9. “claude ” image pair. (a) Left image, (b) Right image.
