Database Reference
In-Depth Information
where
ʵ
1
, and
ʵ
2
are the positive constants. Note that, all other features of the query
h
j
,
j
∈{
1
,...,
M
}
remain unchanged if these features are not members of the feature
subsets
Y
and
Y
.
The query modification obtained by Eq. (
5.28
) is analogous to the pseudo-
relevance feedback method discussed in Chap.
3
. The relevant features are added to
the original query, while the features with a lower degree of relevancy are subtracted
from the query. In this way, the structure of the modified query will be more similar
to the relevant samples, so that the cosine similarity function will be high and the
distance in Eq. (
5.18
) will be low. This results in a reduction of the loss,
L
< L
(5.29)
N
x
=
1
R
(
x
)
D
d
(
I
x
,
I
q
)
L
=
(5.30)
x
=
1
R
(
x
)
D
d
I
x
,
I
q
N
L
=
(5.31)
where
I
q
denotes the query image which has the BoW features modified according
to Eq. (
5.28
).
Fig. 5.3
Flowchart for feature selection and query modification
Search WWH ::
Custom Search