Database Reference
In-Depth Information
key frames, and this method is denoted as a MKF (multiple-key frame) method.
For comparison, a single frame which is the closest frame to a cluster centriod
was selected as a key frame, and this method is denoted as a SKF (single key
frame) method. Video content similarity matching used by the SKF was obtained
by comparing the descriptor vectors of the selected key frames of the query and the
target videos. However, for the MKF method, similarity measures are obtained by
matching multiple key frames of the query against multiple key frames in the target
video clips. To be precise, let
S
be a similarity score. The similarity was obtained by:
N
i
=
1
s
i
=
S
(10.15)
s
i
=
min
M
{
d
[
i
,
j
]
}
(10.16)
j
=
1
,...,
where
d
is the distance between the
i
-th key-frame of the query and the
j
-th
key-frame of the target video;
N
and
M
are the total number of key-frames of the
query and target videos, respectively.
From the results, it is observed that although the SKF method can be used for
retrieval of video shots, SKF is less effective in characterizing video content of
video clips. The SKF result achieved 39.22 % precision. By considering multiple
key frames as in the MKF method, the performance of the key-frame based video
indexing method can be improved to 62.34 %. However, this result is approximately
10 % less precise than that of TFM.
In order to achieve high retrieval performance, the iARM system was imple-
mented using the automatic and semi-automatic retrieval algorithms. The pseudo-
relevance feedback using the adaptive cosine network architecture (discussed in
Chap.
3
) was employed. In this case, depending on the internet traffic conditions,
users can submit automatic and semi-automatic queries, and the automatic query
can avoid the transmission of training sample video files over the internet. Using
the same set of queries as in the previous results, this system first performed an
automatic retrieval for each query to adaptively improve its performance. After three
iterations of signal propagation in the adaptive cosine network, the system was then
assisted by users. Table
10.4
provides the summary of the retrieval results, obtained
by automatic and semiautomatic methods. It is observed that the semiautomatic
method was superior to the automatic method and the user interaction method.
The best performance was achieved at 92.03 % precision. In addition, the moderate
performance of the automatic method can be beneficial to the user when internet
resources are limited.
The strength of the iARM system was evaluated against a variety of templates
used by TFM for indexing video clips. Specifically, three sets of templates at
T
c
=
[
i
,
j
]
500, were generated using a competitive learning algorithm,
where
T
c
denotes the number of templates. These are approximately 3 %, 6 %,
and 9 % of the training sample set, respectively. For each set of templates, video
500, 1
,
000 and 1
,
Search WWH ::
Custom Search