Modeling and Synthesis of Realistic Visual Speech in 3D - 3D Modeling and Animation: Synthesis and Analysis Techniques for the Human Body - page 277

Figure 7. The selection of RBF type is adapted to the local geometry. The
figure shows the improvement that results from switching from exponential
to hermite for the central point on the forehead.
Similarly, there are places where an exponential is much more effective than a
hermite RBF. If the generic head, which is of a rather Caucasian type, has to be
mapped onto the head of an Asian person, hermite functions will tend to copy the
shape of the mesh around the eyes, whereas one wants local control in order to
narrow the eyes and keep the corners sharp. The size of the region of influence
is also determined by the scale r . Three such scales were used (for both RBF
types). These scales and their spatial distribution over the face are shown in
Figure 8(1). As can be seen, they vary with the scale of the local facial
structures.
Figure 8. (1) The RBF sizes are also adapted to local geometry. There are
three sizes, where the largest is applied to those parts that are the least
curved. (2,3) For a small subset of points lying in a cavity the cylindrical
mapping is not carried out, to preserve geometrical detail at places where
captured data quality deteriorates.