Graphics Reference
In-Depth Information
the corner of the mouth superiorly and posteriorly to produce a
smile, can effectively discriminate valance (pleasure) and intensity of
emotional states.
In general, a slow low- or band-pass fi lter is applied together with
a linear piecewise detrending of the time series at a 10-s basis. From
the subject's respiration, the following features (Boiten et al., 1994) are
computed (low-pass fi ltered at 0.15 Hz): mean and standard deviation
of the fi rst derivatives (10-s time window), breathing volume, mean and
standard deviation of breathe intervals and Poincaré plots (30-s time
window each). The EMG signals were used to compute the following
features (band-pass fi ltered at 20-120 Hz, piecewise linear detrend):
mean of fi rst and second derivatives (5-s time window) (Picard, 2003)
and power spectrum density estimation (15-s time window) (Welch,
1967). The following features are extracted from the skin conductance
(SCL) (low-pass fi ltered at 0.2 Hz): mean and standard deviation of
fi rst and second derivative (5-s time window).
3.3 Statistical evaluation
In this section, the statistical evaluation of the architecture is described
for the mentioned data. All reported results originate from leave one
subject out experiments.
3.3.1 EmoRec II
Classification of Spoken Utterances
The MFCC features have been calculated using 40-ms windows and
were averaged to form 200-ms blocks such that all features have
a uniform alignment. The three available features were separately
classified using a SVM with an RBF kernel function and a probabilistic
output function. For the individual features, accuracies from 52.5% to
57.9% were accomplished. Furthermore, these results were combined
using the average of the confidence values and a temporal fusion of
10 s was conducted. This resulted in an accuracy of 55.4% (compare
Table 1).
Classification of Facial Expressions
We classified the facial expressions using a multivariate Gaussian.
In order to render a stable classification, a bagging procedure was
conducted and a reject option was implemented: hereby 99% of the test
frames were rejected with respect to the confidence of the classifier.
An accuracy of 54.5% was achieved and only 52.3% without reject
option (see Table 1).
Search WWH ::
Custom Search