Graphics Reference
In-Depth Information
(a)
(b)
(c)
(d)
(e)
(f)
Figure 2.5
The top row shows two neutral facial scans (of the same subject) and their registration using
ICP (c). The bottom row shows the registration of a neutral facial scan to a non-neutral scan of the same
subject, which appears to be of a higher registration error than the former case
n
i
=
1
f
i
. The covariance matrix
The average range face is then computed,
f
=
1
/
is next
computed as in Equation 2.50.
n
f
)(
f
i
−
f
)
.
=
(
f
i
−
(2.50)
i
=
1
The lower dimensional subspace is given by the
k
eigenvectors of
with the highest eigen-
values
E
e
k
], where
k
is much lower than the number of range pixels (fixed in each
image). This is because the extent of data variation along an eigenvector is indicated by the
corresponding eigenvalue,
=
[
e
1
...
e
i
. See Figure 2.6.
The compact representation of an unseen range image (after vectorization and subtracting
the average face)
f
p
is called the PCA coefficients
c
p
(the feature vector) and is given by the
α
i
e
i
=
