Graphics Reference
In-Depth Information
can reduce the testing error. On the other hand, the variances of the
classification are relatively high.
Rejecting samples when classifying such kind of data turns out to
be a sound approach leading to more robust results, especially when
the distribution of the classes in the data is heavily overlapping. In
future work, it could be promising to implement an iterative classifier
training procedure, where the training data can be rejected.
The results presented in Tables 1 to 3 are preliminary and must
be further evaluated in several directions:
1. Feature extraction techniques as described in the previous
sections have been successfully applied to the recognition of
Ekman's six basic emotions for benchmark data sets consisting
of acted emotional data. In these data sets, emotions shown
by the actors are usually over-expressed and different from
the emotional states that can be observed in the AVEC data set.
2. The classifier architecture is based on the so-called late fusion
paradigm. This is a widely used fusion scheme that can be
implemented easily just by integrating results of the pre-trained
classifier ensemble by fixed or trainable fusion mappings, but
more complex spatio-temporal patterns on an intermediate
feature level cannot be modeled by such decision level fusion
scheme.
3.
The emotional states of the AVEC2011 data set are encoded by crisp
binary labels, but human annotators have usually problems to
assign a confi dent crisp label to an emotional scene (e.g. single
spoken word or a few video frames) or disagree, and thus dealing
with fuzzy labels or labels together with a confidence value
during annotation and classifi er training phase could improve
the overall recognition performance.
Acknowledgments
This chapter is based on work done within the Transregional
Collaborative Research Centre SFB/TRR 62 “Companion-Technology
for Cognitive Technical Systems” funded by the German Research
Foundation (DFG). Tobias Brosch and Miriam Schmidt are supported
by scholarships of the graduate school Mathematical Analysis of
Evolution, Information and Complexity of the University of Ulm.
Search WWH ::
Custom Search