Database Reference
In-Depth Information
Table 3.8
Summary of competitive learning algorithm for template generation for video database
indexing
Input:
Query vector
=
v q = {
w qr |
r
[
1
,...,
R
] }
Maximum number of iterations, I max
Output:
The final retrieval set, containing k -relevant samples
=
S k (
v q )
Initialization:
Threshold value
ʾ
Learning parameters,
ʲ
Activation level of the query node,
ʱ
and
1
,
if w qr
=
0
a r =
0
,
Otherwise
Calculate a ( t )
a ( q )
Computation:
mesg r ( q ) r ( t ) =
×
w qr ,for r
=
1
,
2
,...,
R
,
r
r
Calculate mesg r ( t ) j ( v ) = a ( t )
× w jr ,for j = 1 , 2 ,..., N ,
r
Repeat
1. Calculate activation level of the video nodes, for j = 1 , 2 ,..., N ,
R
r = 1 mesg r ( t ) j ( v ) =
R
r = 1 a ( t )
a ( v )
j
=
× w jr
r
2. Obtain a positive set, Pos ,andanegativeset, Neg ,
ja ( v )
j
ja ( v )
j
Pos =
ʾ
,
Neg =
< − ʾ
3. Calculate the activation level of the video template node,
l r
a ( t )
ₐ−
r
1 l r
2
r
=
Pos a ( v )
j
Neg a ( v )
j
where l r = w qr + ʱ ∑
w jr + ʲ ∑
w jr
j
j
Until:
Iteration
=
I max
a ( v k
where S k ( v q ) is the set of top- k vectors most similar to the query vector and a ( v )
k
v
a ( v )
Return:
Top- k retrieval set, S k
(
v q
)=
|
is the activation level of the k -th nearest neighbor of v q .
was described by its associated weight vector. This was generated by the template
models, using neighborhood
5 [cf. Eq. ( 3.30 )].
A total of 25 queries were made and the judgments on the relevance of each video
to each query shot were evaluated. In general, the relevance judgment of videos is
difficult because two video clips may be related in terms of the story context, and
not just visual similarity. This fact was taken into account in this experiment, so a
criterion employed here is a very subjective judgment of relevance: only retrieved
video shots from the same stories were judged to be relevant. For example, four
video shots shown in Fig. 3.16 a were judged to be relevant because they were parts
of the same stories. Similarly, the four video shots shown in Fig. 3.16 b are relevant
to each other.
ʷ =
 
Search WWH ::




Custom Search