Analysis and Synthesis of Facial Expressions - 3D Modeling and Animation: Synthesis and Analysis Techniques for the Human Body - page 255

Figure 7: Reconstruction quality in PSNR over bit-rate needed for encoding
the animation parameters.
to exclude effects from the background, which is not explicitly modeled. The
trade-off between bit-rate, which can be controlled by changing the quantifying
values for the FAPs, and reconstruction quality is shown in Figure 7.
Facial Animation
The use of different head models for analysis and synthesis of head-and-shoulder
sequences is also interesting in the field of character animation in film produc-
tions or web applications. The facial play of an actor sitting in front of a camera
is analyzed and the resulting FAPs are used to control arbitrary 3-D models. This
way, different people, animals, or fictitious creatures can be animated realisti-
cally. The exchange of the head model to animate other people is shown in Figure
8. The upper row depicts some frames of the original sequence used for facial
expression analysis. Instead of rendering the sequence with the same 3-D head
model used for the FAP estimation, and thus reconstructing the original se-
quence, the head model is exchanged for image synthesis leading to new
sequences with different people that move according to the original sequence.
Examples of this character animation are shown in the lower two rows of Figure
8. In these experiments, the 3-D head models for Akiyo and Bush are derived
from a single image. A generic head model whose shape is controlled by a set
of parameters is roughly adjusted to the outline of the face and the position of
eyes and mouth. Then, the image is projected onto the 3-D model and used as