Analysis and Synthesis of Facial Expressions - 3D Modeling and Animation: Synthesis and Analysis Techniques for the Human Body - page 239

et al., 1989; Li, 1993; Choi et al., 1994; and Hoch et al., 1994). According to that
scheme, any facial expression results from the combined action of the 268
muscles in the face. Ekman and Friesen discovered that the human face
performs only 46 possible basic actions. Each of these basic actions is affected
by a set of muscles that cannot be controlled independently. To obtain the
deformation of the facial skin that is caused by a change of an action unit, the
motion of the muscles and their influence on the facial tissue can be simulated
using soft tissue models (Terzopoulos et al., 1993; Lee et al., 1995). Due to the
high computational complexity of muscle-based tissue simulation, many applica-
tions model the surface deformation directly (Aizawa et al., 1989; Choi et al.,
1994) using heuristic transforms between action units and surface motion.
Very similar to the FACS is the parameterization in the synthetic and natural
hybrid coding (SNHC) part of the MPEG-4 video coding standard (MPEG,
1999). Rather than specifying groups of muscles that can be controlled indepen-
dently and that sometimes lead to deformations in larger areas of the face, the
single parameters in this system directly correspond to locally limited deforma-
tions of the facial surface. There are 66 different facial animation parameters
(FAPs) that control both global and local motion.
Instead of using facial expression descriptions that are designed with a relation
to particular muscles or facial areas, data-driven approaches are also used for
the modeling. By linearly interpolating 3-D models in a database of people
showing different facial expressions, new expressions can be created (Vetter et
al., 1998; Blanz et al., 1999). Ortho-normalizing this face-space using a KLT
leads to a compact description that allows the representation of facial expres-
sions with a small set of parameters (Hölzer, 1999; Kalberer et al., 2001).
Facial Expression Analysis
Synthesizing realistic head-and-shoulder sequences is only possible if the facial
animation parameters are appropriately controlled. An accurate estimation of
these parameters is, therefore, essential. In the following sections, different
methods are reviewed for the estimation of 3-D motion and deformation from
monoscopic image sequences. Two different groups of algorithms are distin-
guished: feature-based approaches, which track distinct features in the images
and optical flow based methods that exploit the entire image for estimation.
Feature-based estimation
One common way for determining the motion and deformation in the face
between two frames of a video sequence is the use of feature points (Kaneko