Automatic 3D Face Model Adaptation with Two Complexity Modes for Visual Communication - 3D Modeling and Animation: Synthesis and Analysis Techniques for the Human Body - page 298

Since such information can be coded with a limited amount of bits, a video
conference or a videophone system with very low bit rates is possible (Musmann,
1995). To implement such a coding system, a generic 3D face model has to be
adapted to the particular face of the participant involved in the monocular video
telephony or video conferencing call. This adaptation must be carried out at the
beginning of the video sequence. Instead of achieving the highest 3D modeling
accuracy, it is the quality of the animated facial expressions in 2D images that
is more important for visual communication at the receiver side.
Face model adaptation for visual communication differs from other applications
in that it has to be done without human interaction and without a priori
information about the participant's face and its facial features. It is unrealistic
to assume that a participant always has a particular facial expression, such as a
neutral expression with a closed mouth or a particular pose position, in a 3D
world. An algorithm for 3D face model adaptation should not only adapt the face
model to the shape of the real person's face. An adaptation to the initial facial
expression at the beginning of the video sequence is also necessary.
Furthermore, an algorithm for 3D face model adaptation should be scalable, since
a number of different devices will likely be used for visual communication in the
future. On the one hand, there will be small mobile devices like a mobile phone
with limited computational power for image analysis and animation. The display
size is restricted, which results in less need for high quality animation. On the
other hand, there will be devices without these limitations like stationary PCs. In
the case of a device with limitations regarding power and display, the face model
adaptation algorithm would need to switch to a mode with reduced computational
complexity and less modeling accuracy. For more powerful devices, the algo-
rithm should switch to a mode with higher computational complexity and greater
modeling accuracy.
Some algorithms in the literature deal with automatic face model adaptation in
visual communication. In Kampmann & Ostermann (1997), a face model is
adapted only by means of eye and mouth center points. In addition, nose position,
and eye and mouth corner points are also used in Essa & Pentland (1997). A 3D
generic face model onto which a facial texture has previously been mapped by
hand is adapted to a person's face in the scene by a steepest-gradient search
method (Strub & Robinson, 1995). No rotation of the face model is allowed. In
Kuo, Huang & Lin (2002), a method is proposed using anthropometric informa-
tion to adapt the 3D facial model. In Reinders et al. (1995), special facial features
like a closed mouth are at first estimated and the face model is then adapted to
these estimated facial features. Rotation of the face model is restricted.
Furthermore, no initial values for the facial animation parameters like Action
Units (Reinders et al., 1995) or muscle contraction parameters (Essa &
Pentland, 1997) have been determined by the adaptation algorithms. An ap-