Analysis and Synthesis of Facial Expressions - 3D Modeling and Animation: Synthesis and Analysis Techniques for the Human Body - page 240

et al., 1991; Terzopoulos et al., 1993; Gee et al., 1994; Huang et al., 1994; Lopez
et al., 1995; and Pei, 1998). Highly discriminant areas with large spatial
variations, such as areas containing the eyes, nostrils, or mouth corners, are
identified and tracked from frame to frame. If corresponding features are found
in two frames, the change in position determines the displacement.
How the features are searched depends on properties such as color, size, and
shape. For facial features, extensive research has been performed, especially in
the area of face recognition (Chellappa et al., 1995). Templates (Brunelli et al.,
1993), often used for finding facial features, are small reference images of
typical features. They are compared at all positions in the frame to find a good
match between the template and the current image content (Thomas et al.,
1987). The best match is said to be the corresponding feature in the second
frame. Problems with templates arise from the wide variability of captured
images due to illumination changes or different viewing positions. To compensate
for these effects, eigen-features (Moghaddam et al., 1997; Donato et al., 1999),
which span a space of possible feature variations or deformable templates
(Yuille, 1991) and reduce the features to parameterized contours, can be utilized.
Instead of estimating single feature points, the whole contour of features can also
be tracked (Huang et al., 1991; Pearson, 1995) using snakes . Snakes (Kas et al.,
1987) are parameterized active contour models that are composed of internal and
external energy terms. Internal energy terms account for the shape of the
feature and smoothness of the contour, while the external energy attracts the
snake towards feature contours in the image.
All feature-based algorithms have in common that single features, like the eyes,
can be found quite robustly. Dependent on the image content, however, only a
small number of feature correspondences can typically be determined. As a
result, the estimation of 3-D motion and deformation parameters from the
displacements lacks the desired accuracy if a feature is erroneously associated
with a different feature in the second frame.
Optical flow based estimation
Approaches based on optical flow information utilize the entire image informa-
tion for the parameter estimation, leading to a large number of point correspon-
dences. The individual correspondences are not as reliable as the ones obtained
with feature-based methods, but due to the large number of equations, some
mismatches are not critical. In addition, possible outliers (Black et al., 1996) can
generously be removed without obtaining an underdetermined system of equa-
tions for the determination of 3-D motion.