Database Reference
In-Depth Information
Fig. 10.5
Architecture of SVM-based late fusion method. In the fusion module, the vector
x
is formed by the fusion of similarity matching scores in audio and visual domains, which are
respectively denoted by
d
a
and
d
v
. In the SVM, the input
x
and the support vector
x
i
are nonlinearly
mapped (by
) into a feature space
H
, where dot products are computed. By the use of the kernel
k
,
there two layers are in practice computed in one single step. The results are linearly combined by
weights
v
i
, found by solving a quadratic program. The linear combination is fed into the function
˃
(
ˆ
x
)=
sgn
(
x
+
b
)
,where
b
is a bias
database, a subset of samples of
x
i
,
N
T
is used for training SVM to
learn a specific semantic concept defined by the positive class. This is described in
the following sections.
i
=
1
,
2
,...,
10.4.2
Support Vector Machine Learning
A mapping function
f
:
ˇ
ₒ{±
1
}
is estimated by the input-output training data
[
306
,
307
],
(
x
1
,
y
1
)
,...,
(
x
i
,
y
i
)
∈
ˇ
×{±
1
}
(10.20)
is some nonempty set that the patterns
x
i
are taken from, and
the
y
i
are the corresponding labels. It is assumed that the data were generated
independently from some unknown probability distribution
P
The domain
ˇ
(
,
)
x
y
. The goal here
(
,
)
(
)=
is to learn a function that will correctly classify a new example
x
y
, i.e.,
f
x
y
for examples
(
x
,
y
)
that were also generated from
P
(
x
,
y
)
. In other words, we choose
(
,
)
y
such that
is in some sense similar to the training examples. To this end, we
need a similarity measure in
x
y
ˇ
, i.e.,
Search WWH ::
Custom Search