observation is consistent with [Cohen et al., 2003]. Therefore, in this section
we choose to omit the results for “happiness” and “surprise” for conciseness.
The confusion matrix for the geometric-feature-only method is presented in
Table 7.1. In Table 7.2, we show the confusion matrix for the method using
both appearance and geometric features. Note that we omit the rows of “hap-
piness” and “surprise” for conciseness. The analysis of the confusion between
different expressions shows that the proposed appearance features help to sig-
nificantly reduce the confusion among expressions, especially the four more
easily confused expressions: anger, disgust, fear and sadness. For example, the
confusion rate between “anger” and “disgust” is reduced more than 10 percent.
This improvement demonstrate the effectiveness of the proposed appearance
features in handling motion details.
In Table 7.3, we show the comparison of the recognition rates for the four
easily confused expressions. A more intuitive comparison is illustrated in Fig-