Graphics Reference
In-Depth Information
Figure 7.4.
Comparison of the proposed approach with geometric-only method in person-
dependent test.
In the second experiment, we compare the classification performances of
ratio-image based appearance feature and non-ratio-image based appearance
feature. The goal of this test is to see whether the proposed appearance feature
is less person-dependent. The non-ratio-image based feature does not consider
the neutral face texture, and is computed as instead of using
equation (7.1). To show the advantage of ratio-image based feature in the ability
to generalize to new people, the test is done in a person-independent way. That
is, all data of one person is used as test data and the rest as training data. This
test is repeated 47 times, each time leaving a different person out (leave one out
cross validation). The person-independent test is more challenging because the
variations between subjects are much larger than those within the same subject.
To factor out the influence of geometric feature, only the appearance feature
is used for recognition in this experiment. The average recognition rates are
shown in Table 7.4 and Figure 7.5. We can see that ratio-image based feature
outperforms non-ratio-image based feature significantly. For individual subject,
we found that the results of the two features are close when the texture does not
have much details. Otherwise, ratio-image based feature is much better.
The third experiment again uses person-independent setting and leave one
out cross validation. For each test, we use 50% of the data of the test person as
adaptation data and the rest as test data. Without applying adaptation algorithm,
Search WWH ::
Custom Search