Graphics Reference
In-Depth Information
facial expression recognition. In Gong et al. (2009), the shape of an expressional 3D face
is approximated as the sum of a basic facial shape component, representing the basic face
structure and neutral-style shape, and an expressional shape component that contains shape
changes caused by facial expressions. The two components are separated by first learning a
reference face for each input non-neutral 3D face then, on the basis of the reference face and
the original expressional face, a facial expression descriptor is constructed, which accounts for
the depth changes of rectangular regions around eyes and mouth. Average recognition rates
of 71
22% have been reported on the BU-3DFE database, respectively, not using
and using a reference neutral scan for each subject.
Approaches in the second category, extract features from 3D scans and use these features to
classify the facial scans into different expressions. In Wang et al. (2006), a feature-based facial
expression descriptor is proposed and the BU-3DFE database is used for the first time. The face
is subdivided into seven regions using manually annotated landmarks and classified primitive
surface features are classified into basic categories, such as ridge , ravine , peak , and saddle ,
using surface curvatures and their principal directions. They reported the highest average
recognition rate of 83.6% using the primitive facial surface features and an LDA classifier.
The facial expressions of happiness and surprise were reported to be the best well-identified
with accuracies of 95% and 90.8%, respectively. Comparison with the results obtained using
the Gabor-wavelet and the Topographic Context 2D appearance feature-based methods on the
same database showed that the 3D solution outperforms the 2D methods. Soyel and Demirel
(2007) also performed 3D facial expression recognition on the BU-3DFE database. Among
the 83 facial landmarks labeling the 3D faces of the BU-3DFE database, only six distance
measures maximizing the differences of facial expressions were selected. These six distance
values were used to form a distance vector for the representation of facial expressions as
defined by the MPEG-4 Facial Definition Parameter Set (Pandzic and Forchheimer, 2005).
The results obtained from a neural-network classifier using the 3D distance vectors reached up
to 98.3% in the recognition of the surprise facial expression, whereas the average recognition
performance is 91.3%. Tang and Huang (2008) first extracted a set of candidate features
composed of normalized Euclidean distances between the 83 facial landmarks of the BU-
3DFE database. Then they used a feature-selection method on the basis of the maximizing
the average relative entropy of marginalized class-conditional feature distributions to retain
only the most informative distances. Using a regularized multiclass AdaBoost classification
algorithm, they obtained a 95.1% average recognition rate for the six basic facial expressions
on a subset of the BU-3DFE database. The neutral facial expression was not classified rather,
as a preprocessing step, its features served as fiducial measures that were subtracted from the
features of the six basic facial expressions of the corresponding subject. The approach proposed
in Venkatesh et al. (2009) on the other hand, used a modified PCA to classify facial expressions
using only the shape information at a finite set of fiducial points that were extracted from the
3D neutral and expressive faces of the BU-3DFE database. The approach used 2D texture
images of the face to mark interest regions around the eyebrows, eyes, nose, and mouth, and
extracted facial contours in those regions with the help of an active contour algorithm. Then,
these contours were uniformly sampled, and the sampled points were mapped onto the 3D data
set to generate a shape and color descriptor of the interest-regions. An average recognition
rate of 81.67% was reported. Maalej et al. (2010) proposed an approach based on the shape
analysis of local facial patches. The patches were extracted around the 83 manually annotated
facial landmarks of the BU-3DFE database, and the shape of each patch described by a set
.
63% and 76
.
Search WWH ::




Custom Search