Database Reference
In-Depth Information
Fig. 3.12 Representation of
encoding x m with the set of
labels
l x m
1
l x m
2
l x m
5
{
,
,...,
}
M
m = 1
i = 1 I l x m
ʷ
l r
=
=
F jr
(3.33)
i
The function I is equal to 1 if the argument is true, and 0 otherwise. In addition,
the maximum in Eq. ( 3.32 ) is computed over all templates mentioned in the content
of the video
V
j ; N denotes the total number of videos in the system; and n r denotes
the number of videos in which the index template g r appears.
The weight w jr balances two effects for clustering purposes: intra-clustering
characterization and inter-clustering characterization. First, the intra-clustering
similarity provides one measure of how well that template describes the video
contents in the desired class, and it is quantified by measuring the raw frequency of a
template g r inside a video
V
j . Second, the inter-clustering dissimilarity is quantified
by measuring the inverse of the frequency of a template g r among the videos in the
collection, thereby specifying that the templates which appear in many videos are
not very useful for the discriminant analysis.
3.5.2
Adaptive Cosine Network
3.5.2.1
Network Architecture
The TFM video indexing characterizes the j -th video by using numerical weight
parameters, w jr ,
R , each of which characterizes a degree of importance
of the templates presented in the video. In this section, an adaptive cosine network
demonstrated in [ 100 , 101 ] is adopted to re-organize the weight parameters on a
per query basis. Using these weight parameters, video clusters, which maximize the
similarity within a cluster while also maximizing the separation from other clusters,
can be formed based on content identifiers, to initialize the ranking for answering
a query. This ranking is now adopted to re-organize the degree of importance of
the templates through the following process. First, the process identifies effective
r
=
1
,...,
Search WWH ::




Custom Search