Graphics Reference
In-Depth Information
Figure 5.13
Examples of interclass (different expressions) geodesic paths between source and target
patches
using a Riemannian framework for shape analysis of 3D curves. The shortest path between two
patches at landmark
i
, one in a candidate scan and the other in the reference scan, is defined
as the sum of the distances between all pairs of corresponding curves in the two patches as
indicated in Equation 5.9. The feature vector is then formed by the distances computed on
all the patches and its dimension is equal to the number of used landmarks
N
70 (i.e.,
68 landmarks are used out of the 83 provided by BU-3DFED and the two additional cheek
points). The
i
th element of this vector represents the length of the geodesic path that separates
the relative patch to the corresponding one on the reference face scan. All feature vectors
computed on the overall data set will be labeled and used as input data to machine learning
algorithms such as MultiBoosting and SVM, where MultiBoosting is an extension of the
successful Adadoost technique for forming decision committees.
=
Recognition Experiments
To investigate facial expression aforementioned, the above approach is applied to a data set
that is appropriate for this task. In this section, we describe the experiments, obtained results,
and comparisons with related work.
For the goal of performing identity-independent facial expression recognition, the experi-
ments were conducted on the BU-3DFE static database. A data set captured from 60 subjects
were used, half (30) of them were female and the other half (30) male, corresponding to the
high and highest intensity levels 3D expressive models (03-04). These data are assumed to be
