Information Technology Reference
In-Depth Information
wise layer similarities. The normalized edge weight between a layer
L
t
−
1
i
of
F
t
−
1
,
and a backward motion-compensated layer
L
t
,
w
j
of
F
t
is defined as
L
t
−
1
i
L
t
,
w
j
|
∩
|
,
L
t
,
j
)=
E
(
L
t
−
1
i
)
,
(1)
L
t
,
w
j
L
t
−
1
i
min
(
|
|
,
|
|
L
t
,
w
j
where
L
t
−
1
i
denotes the
area of a region. The initial graph constructed with the above similarity weights has
an edge between every layer in
U
and every layer in
V
. This redundancy is elimi-
nated by deleting the links having weights below a predefined threshold. However,
the links, whose source or target vertices has only one link left, are retained to ensure
that every vertex is connected to the graph.
∩
denotes the overlapping region of the layers and
|
.
|
3.2
Motion Model Parameter Interpolation
Model parameter interpolation refers to estimating forward and backward motion
models for a set of layers corresponding to the interpolation frames by using the
backward models from
F
t
to
F
t
−
1
. Suppose that only a single frame
F
t
−
Δ
t
,cor-
responding to time instant
t
t
<
1), is to be interpolated between the
original frames. Given a set of backward layer motion models
−
Δ
t
(0
<
Δ
P
1
,
P
2
,...,
P
n
}
,rep-
resenting the motion from
F
t
to
F
t
−
1
, model parameter interpolation problem can
be defined as estimating the parameters of forward models
{
P
t
−
Δ
t
1
,
f
,
P
t
−
Δ
t
2
,
f
,...,
P
t
−
Δ
t
n
,
f
{
}
P
t
−
Δ
t
1
,
b
,
P
t
−
Δ
t
2
,
b
, ...,
P
t
−
Δ
t
n
,
b
from
F
t
−
Δ
t
from
F
t
−
Δ
t
to
F
t
−
1
. The problem is depicted in Fig. 2 for a single layer. In order to compute
a meaningful parameter set to define the flow between a layer in the frame to be
to
F
t
and backward models
{
}
Fig. 2
Motion model parameter interpolation