Database Reference
In-Depth Information
Table 10.4 Average precision (%), obtained by retrieving 40
queries, measured from the top 16 retrievals, using KFVI
and TFM methods with user-controlled relevance feedback,
automatic relevance feedback, and semi-automatic retrieval
methods
Methods
Average precision (%)
KFVI
SKF
39.22
MKF
62.34
TFM
Initial result without RF
73.59
User controlled-RF
90.47
Automatic-RF
80.12
Semi-automatic retrieval
92.03
The system was tested using 40 sample video clips chosen randomly from
fourteen movies. Table 10.4 shows the retrieval results averaged over the 40 queries.
From the table, it is observed that the iARM system had a very high precision of
73.6 % at the initial stage (i.e., more than 11 relevant video clips were retrieved out
of the top 16 best matches). It also saw a significant improvement of 90.51 % in
precision after a single feedback cycle. This implies that the TFM method is highly
effective in capturing spatio-temporal information from video. This also indicates
that TFM is efficient and highly adaptable, as only single user feedback was required
for significant improvement.
Table 10.4 also shows the comparison between TFM and other video indexing
methods that use video clustering strategies for the video content characterization.
The compared methods are denoted as KFVI (key-frame based video indexing).
Here, the KFVI employed video clustering approaches discussed in [ 320 - 322 ]for
selection of the representative frames from video clips. In this way, for each video
clip, frames are clustered based on frame descriptors, and frames that are close to the
cluster centriods are selected as key frames. The k -mean algorithm and clustering
validity method demonstrated in [ 321 ] were employed for the selection of key
frames.
KFVI process began by extracting 48-bin color histogram vectors from each
video frame in a given video clip. Then, it applied the clustering algorithm to
the resulting histogram vectors to obtain k -means with different values of k ,for
k
10. The k -means was run multiple times for each k , and the best of
these was selected based on the sum of squared errors. Finally, the Davies-Bonldin
index [ 319 ] was calculated for each k , k
=
1
,
2
,...,
, and the k that gave the
smallest Davies-Bonldin index was chosen. In doing so, the optimum number of
clusters will vary according to the cluster validity analysis of the resulting clusters.
The closest frames to the clusters (one frame from each cluster) were selected as
∈{
1
,
2
,...,
10
}
 
Search WWH ::




Custom Search