Graphics Reference
In-Depth Information
(a)
(b)
(c)
(d)
Figure 2.6
An average range face (a) and the first three eigenfaces (eigenvectors of the covariance
matrix), (b), (c), and (d), computed from 200 facial range images of different people
E
f
p
. The PCA coefficients of the range image
c
p
can be matched against
another PCA coefficients of another unseen range image
c
g
(found in a similar way) on the
basis of any metric distance (dissimilarity measure). Typical dissimilarity measures used along
with PCA are the Euclidean
projection
c
p
=
c
p
c
g
c
p
−
c
g
and the cosine
distances.
c
p
c
g
Comments on the use of PCA in 3D face recognition:
The use of PCA in recognition
requires the pose-correction of the facial surfaces (to a predefined pose, typically the frontal
one), the cropping of the facial surface. The range images should also be of the same resolution.
Variations that could be introduced by these factors affect the computation of the PCA sub-
space and/or the PCA features which ultimately may undermine the recognition performance.
Another aspect of using PCA features for 3D face recognition is that they are sensitive to facial
expressions in that the values of the features vary according to facial expressions. However, as
