Audio-Visual Fusion for Film Database Retrieval and Classification - Multimedia Database Retrieval: Technology and Applications - page 288

Table 10.5 Average
precision (%), obtained by
retrieving 40 queries, using
four sets of templates for
video indexing

Average precision (%)

Number of templates Initial result After 1 RF

T c = 2 , 000 73.59 90.47

T c = 1 , 500 72.81 89.69

T c = 1 , 000 72.50 88.13

T c = 500 68.75 81.25

Fig. 10.4 iARM's user interface for retrieval of movie clips using the TFM method

files were indexed, and the resulting feature databases were used by the retrieval
system. Table 10.5 shows the retrieval results obtained by retrieving 40 queries.
It is observed that the iARM system was fairly robust compared to the deviation
in the number of templates T c . The results in the second column show the small
variation in retrieval performance (i.e., 1.1 % reductions of precision) as the number
of templates was reduced from 2,000 to 1,000. In this case, video matching used
the non-adaptive matrix. However, the performance gap is increased to 2.3 % after
relevance feedback as we had anticipated, since human interpretation has been
addressed to some extent by the interactive algorithm. The results show that a set of
templates with T c ≥ 000 is sufficient to model the video content in this database.

This value of T c is approximately 0.1 % of the entire video data in this collection.

1 ,