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 297

the generated 3D model could not be immediately animated, since the underlying
animation structure is missing.
An alternative approach is the use of a generic 3D face model with a built-in
animation structure. Action Units from the Facial Action Coding System
(Ekman & Friesen, 1977), MPEG-4 facial animation parameters (FAP)
(Sarris, Grammalidis & Strintzis, 2002) or muscle contraction parameter s
(Fischl, Miller & Robinson, 1993) from a model of facial muscles can be used as
an animation structure for facial expression. A limited number of characteristic
feature points on a generic face model are defined, e.g., the tip of the chin or the
left corner of the mouth. At the first step of 3D modeling using a generic 3D face
model, those defined feature points are detected in facial images. Then, the
characteristic feature points of the generic 3D face model are adapted using the
detected feature points. This process is also called face model adaptation .
According to available input resources, 3D face model adaptation approaches
can be categorized as follows: (a) range image : An approach using range image
to adapt a generic face model with a physics-based muscular model for animation
in 3D is proposed in Lee, Terzopoulos & Waters (1995). From the generic 3D
face model, a planar generic mesh is created using a cylindrical projection. Based
on the range image, the planar generic face mesh adaptation is iteratively
performed to locate feature points in the range image by feature-based matching
techniques; (b) stereoscopic images/videos : An approach to using stereoscopic
images/videos for face model adaptation is proposed in Fua, Plaenkers &
Thalman (1999). Information about the surface of the human face is recovered
by using stereo matching to compute a disparity map and then by turning each
valid disparity value into a 3D point. Finally, the generic face model is deformed
so that it conforms to the cloud of those 3D points based on least-squares
adjustment; (c) orthogonal facial images : Orthogonal facial images are used
to adapt a generic face model in Lee & Magnenat-Thalmann (2000) and Sarris,
Grammalidis & Strintzis (2001). They all require two or three cameras which
must be carefully set up so that their directions are orthogonal; (d) monocular
images/videos : For face model adaptation using monocular images/videos,
facial features in the facial images are determined and the face model is adapted
(Kampmann, 2002). Since no depth information is available from monocular
images/videos, depth information for feature points is provided only in advance
by a face model and is adapted in relation to the determined 2D feature points.
In the following, we concentrate on animated faces for applications in the field
of visual communication where only monocular images are available. For visual
communication applications, like video conferencing or video telephony, a virtual
face clone represents the human participant in the video conference or in the
videophone call. Movements and facial expressions of the human participants
have to be extracted and transmitted. At the receiver side, the virtual face model
is animated using the extracted information about motion and facial expressions.