We computed MEF's and MDF's, respectively, using 50 sequences (10 for each signs). These signs are obtained from different subjects and the viewing positions are slightly different. Fig. 5 (a) shows the samples in the subspace spanned by the first two MEFs and Fig. 5 (b) shows them in the subspace spanned by the first two MDFs.
Figure 5:
The difference between MEF and MDF in representing samples.
(a) Samples represented in the subspace spanned by the first two
MEFs. (b) Samples represented in the subspace spanned by the first two
MDFs. The numbers in the plot are the class labels of the samples.
As clearly shown, in the MEF subspace, samples from a single class spread out widely and samples of different classes are not far apart. In fact, some samples from different classes mingle together. However, in the MDF subspace, samples of each class are clustered more tightly and samples from different classes are farther apart. This shows that the MDFs are better in terms of classification of signs.