Graphics Reference
In-Depth Information
Figure 5.14
Selected patches at the early few iterations of MultiBoost classifier for the six facial
expressions (anger, disgust, fear, happiness, sadness, and surprise) with their associated weights
To better understand and explain the results mentioned earlier, the Multiboost algorithm is
applied on feature vectors built from distances between patches for each class of expression.
In this case, we consider these features as weak classifiers. Then, we look at the early iterations
of the MultiBoost algorithm and the selected patches in each iteration.
Figure 5.14 illustrates for each class of expression the most relevant patches. Notice that,
for example, for the
happy
expression the selected patches are localized in the lower part of
the face, around the mouth and the chin. As for the
surprise
expression, we can see that most
relevant patches are localized around the eyebrows and the mouth region. It can be seen that
patches selected for each expression lie on facial muscles that contribute to this expression.
Comparison with Related Work
Table 5.9 shows a comparison of semi-automatic approaches proposed in the state-of-the-art
methods using the same experimental setting (54-versus-6-subject partitions) of the BU-
3DFE database. In general, results are obtained by averaging many trials where the subjects
partitioning between train and test are randomly changed. Although not all the methods use
the same number of trials, and just the work in Maalej et al. (2011) evidences the differences
between different trials, we can assume that the reported results can be compared giving an
idea of the effectiveness of the different solutions.
5.4.3 Fully Automatic 3D Facial Expression Recognition
Some recent works have shown that local descriptors computed at salient keypoints of 3D
objects can be usefully applied to capture relevant 3D shape features. For example, in (Mian
et al., 2008) a 3D keypoint detector and descriptor has been defined for the purpose of 3D
Table 5.9
Comparison of state of the art methods on the BU-3DFE data set
Maalej et al. (2011)
Gong et al. (2009)
Berretti et al. (2011a)
Zhao et al. (2011)
RR
98.0
±
1.6%
76.2%
77.5%
82.0%
