Information Technology Reference
In-Depth Information
Under these assumptions, the backward projective homography matrix
P
t
−
Δ
t
b
of
the interpolation frame is reconstructed as
t
b
n
T
)(
K
t
)
−
1
=(1
P
t
−
Δ
t
b
=
K
t
(
I
+
tP
b
,
θ
b
[a]
x
−
−
Δ
t
)
I
+
Δ
(9)
which reveals that the homography decomposition is not required under the men-
tioned assumptions. Finally, the forward homography matrix
P
t
−
Δ
f
can be computed
from the available models by the following simple linear transformation:
P
t
−
Δ
t
f
=(
P
t
)
−
1
P
t
−
Δ
t
b
.
(10)
3.3
Layer Map Interpolation
The interpolation of the motion models is followed by estimation of the correspond-
ing layer support maps at time instants
t
t
. This is achieved essentially
by backward warping the extracted layers (both support maps and intensities) of
F
t
to the two previous time instants, and updating the overlapping and uncovered
regions so as to ensure a single layer assignment for each pixel of
F
t
−
Δ
t
−
1and
t
−
Δ
and
F
t
.
Fig. 4 provides an overview of the layer map interpolation process.
Depth ordering relations of the overlapping layers is extracted by computing a
visual similarity measure between each warped layer and the original frame
F
t
−
1
over the region of overlap. Each pixel of an overlapping region votes for the layer
that gives the minimum sum of absolute intensity differences. Visual similarity of
Fig. 4
Overview of the layer map interpolation process. Top row: warping layer maps at time
instant
t
(right) to the two previous time instances
t
−
1(left)and
t
−
Δ
t
(middle). Bottom
row: updating the layer assignments on overlapping and uncovered regions of the warped
layers at
t
−
1(left)and
t
−
Δ
t
(right). Bottom right: frame
t
= 337 of the original sequence
Flower Garden
.