Graphics Reference
In-Depth Information
93
92.5
92
91.5
91
90.5
90
8
10
15
20
25 30
Number of trees
35
40
45
50
55
60
Figure 5.20
4D expression recognition results using Random Forest classifier when varying the number
of trees
The subsequences have been defined in Sun and Yin (2008) with a length of six frames with
a sliding step of one frame from one subsequence to the following one. For example, with
this approach, a sequence of 100 frames originates a set of 6
570 subsequences, each
subsequence differing from one frame from the previous one. This accounts for the fact that,
in general, the subjects can come into the system not necessarily starting with a neutral expres-
sion, but with a generic expression. Classification of these short sequences is regarded as an
indication of the capability of the expression recognition framework to identify individual
expressions.
According to this, we first computed for each subsequence the Mean Deformation, which
was then presented to multiclass Random Forest, as outlined in Section 5.5.2. The performance
of the Random Forest classifier varied with the number of trees. Thus, we performed the
experiments with different numbers of trees; the results of this experimentation are shown in
Figure 5.20. As illustrated by this figure, the average recognition rate rose with the increasing
number of trees until 40, when the recognition rate reached 93
×
95
=
21%, and then became stable.
Thus, in the following paragraphs, we considered 40 trees and the detailed results (confusion
matrix) with this number of trees are shown in Table 5.10. We recall that the rates were obtained
by averaging the results of the 10-independent and arbitrarily run experiments (10-fold cross
validation).
.
Table 5.10
Confusion matrix for Mean Deformation and Random Forest classifier (for 6-frames
window)
%
An
Di
Fe
Ha
Sa
Su
An
93.11
2.42
1.71
0.46
1.61
0.66
Di
2.3
92.46
2.44
0.92
1.27
0.58
Fe
1.89
1.75
91.24
1.5
1.76
1.83
Ha
0.57
0.84
1.71
95.47
0.77
0.62
Sa
1.7
1.52
2.01
1.09
92.46
1.19
Su
0.71
0.85
1.84
0.72
1.33
94.53
Average recognition rate
=
93.21%
