Audio-Visual Fusion for Film Database Retrieval and Classification - Multimedia Database Retrieval: Technology and Applications - page 297

at different settings of training data. The ground truth for the Love Scene concept
(explained in Table 10.6 ) was utilized to test the SVM classifier in different
training conditions. First, the number of positive samples in the training set was
fixed to 22 samples (these were selected from the total of 66 samples). Then the
negative samples were added to the training set one at a time and used for training.
Figure 10.7 shows the recognition accuracy of the system. As we may anticipate,
without negative samples included in the training set, the classifier has the highest
false positive error rate. Adding more negative samples to the training set also
increased the accuracy of the classifier, but with the cost of higher false negative
errors.
Figure 10.7 also shows that the training size was increased to more than 2 %
of the total number of samples stored in the database. When the system was
allowed to learn more negative samples it produced a high false negative error rate.
In order to achieve a good compromise between performance and error rates, the
system required approximately 0.1 % of the total negative samples for training. In
comparison, we observed from a new experiment that the system required a large
number of positive samples, i.e., more than 22 % of the total positive samples, for
training in order to obtain a good tradeoff between accuracy and error rate.

100

90 False Negative Rate

False Positive Rate
Accuracy

80

70

60

50

40

30

20

10

0

20 40 60 80 100 120 140

Number of Training Samples

Fig. 10.7 Classification results obtained by the SVM-based decision fusion model for the “Love
Scene” concept, at a different setting from the training set