Graphics Reference
In-Depth Information
(a)
(b)
(c)
(d)
(e)
(f)
(g)
(h)
Figure 4.8
Correspondence estimation. By individually fitting components (b,c) to scan data (a),
artifacts may occur at the borders of
S
fine
(d). The maximum displacement error (
r
in mm) for a vertex is
an indication of the local mesh distortion (e), toward purple/black means a higher distortion. Using this
error, surrounding vertices are selected that all contribute to the voting for a new position of the vertex.
After the voting, each vertex is repositioned on the basis of the weighted votes for close by components
(f). The final model fit
S
final
(g) presents smooth transitions between the different components. Five
iterations of Laplacian smoothing (h) was not enough to repair the artifact close to the eyebrow whereas
the face already lost most of its detail. Copyright
C
2009, IEEE
the mean face
S
of the morphable model and include it in the example set of 3D faces from
which the morphable face model was built. Then, we can recompute the PCA model using
100
k
principal eigenvectors and eigenvalues of the face
(see Section 4.4). This way the face properties of a new example face can be added to the
statistical face model. With this enhanced model, we should be able to produce accurate model
fits with only a single component to face scans similar to those
k
example scans. In the end,
each face scan can be described using
m
+
k
example faces, and keep the
m
+
+
k
model coefficients, and face identification can
be performed on the basis of such a
m
k
feature vector.
The addition of
k
extra example faces to the current face model, causes an increase of
computation costs for both the model-fitting and face identification method. So, it is important
not to add example faces that are already covered in the current morphable face model.
+
