Graphics Reference
In-Depth Information
exploiting reject options in order to circumvent false classifications,
can be mitigated. Another aspect is the usage of partially supervised
learning, either to support annotation or to improve classifiers. Parts
of these aspects are then exemplified using two recent examples of
emotion recognition, showing a successful realization of MCSs in
emotion recognition.
Research in affective computing has made many achievements
in the last years. Emotions begin to play an increasingly important
role in the field of human-computer interaction, allowing the user to
interact with the system more efficiently (Picard, 2003) and in a more
natural way (Sebe et al., 2007). Such a system must be able to recognize
the users' emotional state, which can be done by analyzing the facial
expression (Ekman and Friesen, 1978), taking the body posture and
the gestures into account (Scherer et al., 2012) and by investigating the
paralinguistic information hidden in the speech (Schuller et al., 2003;
Oudeyer, 2003). Furthermore, biophysiological channels can provide
valuable information to conclude to the affective state (Cannon, 1927;
Schachter, 1964).
However, the emotions investigated so far were in general acted
and the larger part of research was focused on a single modality,
albeit the problem of emotion recognition is inherently multi-modal.
Obviously, the entire emotional state of an individual is expressed and
can be observed in different modalities, e.g. through facial expressions,
speech, prosody, body movement, hand gestures as well as more
internal signals such as heart rate, skin conductance, respiration,
electroencephalography (EEG) or electromyogram (EMG). Recent
developments aim at transferring the insights obtained from single
modalities and acted emotions to more natural settings using multiple
modalities (Caridakis et al., 2007; Scherer et al., 2012; Zeng et al., 2009;
Chen and Huang, 2000). The uncontrolled recording of non-acted data
and the manifold of modalities make emotion recognition a challenging
task: subjects are less restricted in their behavior, emotions occur more
rarely and the emotional ground truth is difficult to determine, because
human observers also tend to disagree about emotions.
In this chapter, MCSs for the classification of multi-modal features
are investigated, the numerical evaluation of the proposed emotion
recognition systems is carried out on the data sets of the 1st AVEC
challenge (Schuller et al., 2011) and a data set recorded in a Wizard-
of-Oz scenario (Walter et al., 2011). Combining multi-modal classifiers
is a promising approach to improve the overall classifier performance
(Schels and Schwenker, 2010). Such a team of classifiers should be
accurate and diverse (Kuncheva, 2004). While the requirement to the
Search WWH ::




Custom Search