In the MDFs, factors that are not related to classification are discarded or weighted down, which is accomplished by minimizing the within-class scatter; factors that are crucial to classification are emphasized, which is achieved by maximizing the between-class scatter. In this experiment, we show an example that the MDFs can capture the important geometric features.
Figure 6:
Two sample sequences of sign ``of course'' (left) and
``wrong'' (right).
In our gesture vocabulary, the image sequences of two signs: ``of course'' and ``wrong'' are visually very similar. Fig. 6 illustrates two sample sequences of the above signs. The nearest neighbor approximator generally has difficulty to distinguish them, but not the recursive partition tree approximator in the MDF space. Fig. 7 shows the difference between the MEF and the MDF. The left sequence in Fig. 7 is a reconstruction of the sequence ``of course'' based on the first MDF and the right sequence is a reconstruction of the same sequence using 95% of MEFs. We can see that the MEFs are good in terms of preserve the information but not much help for classification. On the other hand, the first MDF captures the feature locations (edges) because it accounts for the major between-sign variation.
Figure 7:
The difference between the MEF and MDF. (a) Reconstruction based
on the first MDF. (b) Reconstruction based
on 95% MEFs.