Information Technology Reference
In-Depth Information
Fig. 6
Results of filtering of flat blocks and outliers. Sequence “Claire” c
LABRI.
Fig. 7
Object extraction at the top of the pyramid by merging frame segmentation and motion
masks. Sequence “Lance trousse” c
LABRI.
MRF based adjustment of object borders with LL and HF subbands
When projected between two successive levels of the pyramid, one wavelet coef-
ficient of a higher
W
k
(
i
,
j
) level in each subband is transformed into four coef-
ficients
W
k
−
1
(2
i
,
2
j
),
W
k
−
1
(2
i
,
2
j
+ 1),
W
k
−
1
(2
i
+ 1
,
2
j
),
W
k
−
1
(2
i
+ 1
,
2
j
+ 1) of
lower level. To avoid this aliasing in resulting object border a morphological erosion
of the projected mask is performed thus defining an uncertainty area. Then pixel-
coefficients in it are assigned according to the minimal energy adapting the classical
MRF formulation with Maximal A posteriori Probability (MAP) criterion. Gaussian
multivariate distribution of LL coefficients in each segmented region is supposed.
For a coefficient in LL subband at position
p
the label
l
is assigned according to the
minimum of the energy:
[1
,
L
]
U
1
(
l
,
p
)+
U
2
(
l
,
p
)
l
(
p
)=arg min
l
(11)
∈
The first potential
U
1
expresses the Gaussian assumption of distribution of the coef-
ficients in LL subband: