Database Reference
In-Depth Information
Table 3.7
Summary of competitive learning algorithm for template generation for video
database indexing
Input:
Set of feature vectors extracted from video frames in a given database:
H
,where
h
j
=
h
j
1
,
h
jP
t
P
is the color
histogram vector of the
j
-th video frame, and
H
is the number of training
samples.
t
=[
h
1
,
h
2
,...,
h
H
]
=[
h
ji
]
h
j
2
,...,
∈
R
P
Output:
The set of weight vectors
=
{
g
r
|
r
=
1
,...,
R
},
g
r
∈
R
,
R
H
Initialization:
Maximum number of iterations
=
t
f
Learning parameter
=
ʷ
0
Weight vectors,
g
1
,
g
2
,...,
g
R
Computation:
h
ji
ₐ−
(
h
ji
−
μ
i
)
˃
i
(normalized all patterns)
where μ
i
and ˃
i
are the mean value and the standard deviation of the i-th
column vector in
H
, respectively.
Repeat
Randomly select a pattern
h
, at iteration
t
r
ₐ−
arg min
r
h
−
g
r
(classify
h
)
g
r
ₐ−
g
r
+
ʷ
(
h
−
g
r
)
(weight update)
ʷ
ₐ−
ʷ
0
1
−
t
f
(parameter update)
t
Until:
t
=
t
f
g
1
,
g
2
,...,
g
R
Return:
templates
that are the common templates among videos in a retrieved set. Then,
those templates considered to be the most significant for reweighting the existing
templates of the initially submitted query are weighted, to improve the ranking
performance. In other words, we allow the templates that are not presented by
the initially submitted query (i.e.,
w
qr
=
), but are common among
the top-ranked videos (i.e., the potentially relevant videos), to
expand
. This results
in reorganization of the degree of importance of the query's templates for better
measurement of video similarity.
The re-ranking process is performed by an adaptive cosine network [
100
,
101
],
with the network architecture presented in Fig.
3.13
. The network is composed of
three layers: one for the query templates, one for the video templates, and the third
for the videos themselves. Each node has a connection weight communicated to its
neighbors via the connection links. The query template nodes initiate the inference
process by sending signals to the video template nodes. The video template nodes
then themselves generate signals to the video nodes. Upon receiving this stimulus,
the video nodes, in turn, generate new signals directed back to the video template
nodes. This process might repeat itself several times, through the second and the
0
,
r
∈
[
1
,...,
R
]
Search WWH ::
Custom Search