Self-adaptation in Image and Video Retrieval - Multimedia Database Retrieval: Technology and Applications - page 95

Fig. 3.15 Some representative key-frames extracted from a subset of 844 video shots in the test
video database obtained from CNN broadcast news

Figure 3.17 shows precision results as a function of top matches, averaged over
all 25 queries. It can be observed that TFM performed substantially better than KFVI
for every setting of the number of top matches (the average precision was higher
by more than 18 %). It is also observed that TFM is very effective in capturing
spatio-temporal information from video, as seen in Fig. 3.18 which depicts retrieval
results from the top sixteen best matches. It was observed that TFM allows similarity
matching based on video contents, whereas the KFVI emphasis is on the content of
the key-frame. There was a dominant brown color on the key-frame, degrading the
performance of KFVI on this query.
Next the adaptive cosine network is applied to improve retrieval accuracy.
The structure of the information in the video database can be represented by a
network with 5,844 nodes and 14,800 connections. The results of three tests are
shown: letting the activation spread for one, three, and twenty iterations. The
parameters were set at

05. Figure 3.19 shows
the improvement of the average precision in retrieving 25 queries.
The following observations were made from the results. Firstly, the adaptive
cosine network was very effective in improving retrieval performance—the average

ʾ = 0 . 1 , and ʱ = 0 . 95 and ʲ = − 0 .