Graphics Reference
In-Depth Information
of curves representing the surface points at the same Euclidean distance from the landmark.
A Riemannian framework was then applied to compare the shape of curves undergoing to
different facial expressions. The length of the geodesic path that separates corresponding
curves provided quantitative information about their shape similarity. The best expression
recognition results on the BU-3DFE database have been obtained using these measures as
entries of a MultiBoost classifier. An improved version of this approach is reported in Maalej
et al. (2011) with new experiments and results given. The work presented in Berretti et al.
(2010e) exploits the local characteristics of the face around a set of facial landmarks to classify
facial expressions. In particular, the facial landmarks are a subset of the 83 facial landmarks
of the BU-3DFE plus a set of facial keypoints automatically identified starting from the given
landmarks. SIFT descriptors are computed around the facial keypoints, combined together and
used as feature vector to represent the face. Before to perform classification of the extracted
descriptors, a feature selection approach is used to identify a subset of features with
minimal-
redundancy
and
maximal-relevance
among the large set of features extracted with SIFT. The
set of selected features is finally used to feed a set of classifiers on the basis of
support vector
machines
(SVM).
From the previous discussion, it emerges that the large part of existing works on 3D
facial expression recognition relies on the presence of landmarks accurately identified on the
face. Methods based on the
generic facial model
use landmarks to establish correspondences
between faces in the construction of a deformable template face. Usually, these approaches
are also computationally demanding because of the deformation process. Solutions based
on
feature classification
in many cases compute distances between landmarks and evaluate
how these distances change between expressional and neutral scans. That several landmarks
are not automatically detectable and the precision required for their positioning demand for
manual annotation in both training and testing stages. Furthermore, several solutions require
a neutral scan for each subject in order to evaluate the differences generated in the 3D scans
by facial expressions with respect to neutral reference scans. In practice, these factors limit
the applicability of many approaches.
In the following sections, we provide more details on two facial expression recognition
methods that are semi-automatic and fully automatic.
5.4.2 Semi-automatic 3D Facial Expression Recognition
In the following paragraphs, we discuss an approach that uses local descriptors called
local
patches
of the face represented by a set of local curves to perform person independent 3D facial
expression recognition. This approach was originally proposed in Maalej et al. (2010, 2011).
In this work, sets of level curves
c
l
{
λ
}
1
≤
λ
≤
λ
0
are associated to
N
reference points (landmarks)
{
r
l
}
1
≤
l
≤
N
(Figure 5.11(a)) (Figure 5.11(b)).
These curves are extracted over the patches centered at these points. Here
stands for the
value of the distance function between the reference point
r
l
and the point belonging to the
curve
c
l
λ
λ
. Accompanying each facial model
are 83 manually picked landmarks; these landmarks are practically similar to the MPEG-4
feature points and are selected on the basis of the facial anatomy structure. Given these points,
the feature region on the face can be easily determined and extracted. We were interested in
a subset of 68 landmarks laying within the face area, discarding those marked on the face
, and
λ
0
stands for the maximum value taken by
λ
