Graphics Reference
In-Depth Information
frame in second video
Figure 5.20.
Fitting a quadratic function to frame-to-frame matching costs for video matching.
The white dot is the current estimate of the matching frame's position in the second video. We
evaluate the frame-to-frame matching cost at this estimate, and 1 and 5 frames on either side
(gray dots). By fitting a quadratic function to the costs, we obtain a newestimate for thematching
frame at the minimizer of the quadratic (striped dot).
5.7
MORPHING
One of the most compelling visual effects created using dense correspondence
between an image pair is
morphing
, also known as
image metamorphosis
. Morph-
ing uses an estimated dense correspondence field to create a smooth transformation
from one image into another, and was used to great effect in films like
Terminator
2
,
Indiana Jones and the Last Crusade
, and the video for Michael Jackson's
Black or
White
.
Unlikeour assumptions for optical flowand stereo, the two images in themorphing
problem typically contain different objects (for example, two different faces). Since
these images significantly violate the brightness constancy assumption, the dense
correspondence is typically estimated from a hand-selected set of feature matches,
using methods from Section
5.2
. Correspondence fields for morphing applications
generally don't take into account occlusions or discontinuities, instead resembling a
deformed “rubber sheet.” That is, we require each point in the first image to have a
correspondence in the second image and vice versa.
We beginwith two images,
I
1
and
I
2
, and two dense correspondence fields
fwd
(
u
,
v
)
from
I
1
to
I
2
and
(
u
,
v
)
I
2
,
and the intermediate images create a natural transformation from one image to the
other.
A naïve solution to morphing is to simply
cross-dissolve
between the two images,
that is, letting
{
M
t
,
t
∈[
0,
t
,2
t
,
...
,1
]
so that
M
0
=
I
1
,
M
1
=
M
t
(
x
,
y
)
=
(
1
−
t
)
I
1
(
x
,
y
)
+
tI
2
(
x
,
y
)
(5.60)
However, as we can see from Figure
5.21
, this approach generates poor results,
since corresponding structures in the two images are not aligned. In particular,
21
A consistent
fwd
if we only computed the flow in
one direction. However, it is sometimes useful to allow the two fields to be inconsistent (as is the
case for the field morphing algorithm discussed shortly).
bwd
can be constructed from a given
(
u
,
v
)
(
u
,
v
)
