Project Deidre (I) -
Modeling Human Facial Expressions
Daniel D. Hung and
Szu-Wen (Steven) T. Huang
CS 718 Topics in Computer Graphics
Introduction
Human faces are among some of the most difficult objects to model in
computer graphics, and have been in the attention of numerous attempts,
some almost as old as computer graphics itself. Facial expressions are
the result of the movement of skin layered atop muscles and bone
structures, and may have thousands of combinations. As part of our
coursework, we propose to focus on this subject. Our work will be divided
into three parts: a survey of various techniques developed throughout the
years, an implementation of one such technique, and the presentation of
our results.
Limitations
Due to time constraints, it is unlikely that a detailed implementation of
any of the models is possible. The goal of the project is thus to produce
a technology demonstration of a technique. The availability of suitable
input devices or a pre-defined face mesh would limit the accuracy and
aesthetic quality of the finished product. We wish to be able to produce,
at the minimum, a wire frame animation of a face mesh. The results from
the survey and the implemented model will be presented in class.
Prior Art
General Structure
Among the various approaches taken over the years, a distinctive
generalization can be made. There generally exists a low-level muscle
motion simulator, called variably as action units, abstract
muscle action procedures, or minimum perceptible actions.
This layer enables the generation of expressions that are not necessarily
humanly possible, such as asymmetric movement of two sides of the face.
On top of the "muscle" layer, we find an abstraction for humanly
significant expressions. This layer might include smile,
frown, horror, surprise, and other expressions.
As the objective of many projects was to emulate the human face during
speech, there usually exists another layer above expressions that include
phonemes (speech primitives). A complete data set of phonemes
allow the synthesized face to look like it is coordinated with speech
which is played back separately.
Keyframing
Keyframing was one of the earliest approaches taken, and
involved linear transformations from one face mesh to another. The amount
of computations were extensive and the data set large. The approach was
rather inflexible because the range of expressions that can be generated
are limited by the keyframes that were previously digitized. It is also
difficult to generalize work on one face mesh to another.
Parametric Deformations
A second approach was to model the human face as a parametric surface and
record the transformations as movements of the control points to minimize
the data storage requirements. These approaches are still difficult to
generalize over different face meshes.
One such attempt utilizes B-spline patches that were defined manually
on an actual digitized face mesh [Nahas88].
The control points for the B-spline patches were moved to effect the
distortion of the face. While this method is powerful, we know of no
automatic way of defining the relevant control points for the B-spline
patches.
Another such attempt used Rational Free-form Deformations to move points
inside a defined volume [Kalra92] based on the
control points placed at the edges of the volume. The volume could be
distorted by either changing the position of a control point or increasing
the weight of the control point. The method has similar shortcomings as
the B-spline approach, though the definition of the volumes are more
intuitive than B-spline control points.
Anatomically-correct muscles
A natural approach would be to simply simulate actual human muscles [Waters87, Platt81], if
not for the amazingly little understanding we have of them. The human
facial muscles are capable of a large amount of expressions, and realistic
simulation of muscles requires the simulation of muscle action wrapping
around the skull structure, jaw rotation, and folding and stretching
properties of skin.
Pseudo-muscles
A hybrid approach would be to simulate muscles that are not necessarily
anatomically-correct [Thalmann88]. This
allows muscle vectors to protrude out of the face if needs be. The
advantage of this approach is that we are more interested in getting a
meaningful expression than actually simulating all the muscles. This
approach is easier and possibly yields a smaller data set than realistic
muscles because some insignificant muscles that contribute to a certain
expression could be generalized by a vector. This is the approach we
selected to follow.
Implementation
The first part of our work involves the tedious task of manually digitizing
a face mesh into a machine-readable form. We decided early on that dealing
with the large data set generated by a scanner would be difficult, and we
were able to generate a face composed of some 400 vertices and some 600
polygons. Some vertices were discarded during polygonalization because
they were too detailed for our purposes.
The face is then divided into regions defined by rectangular bounding
boxes. Each muscle is associated with a bounding box, and could not
affect any point outside that box. A sample bounding box (for the muscle
vector responsible for raising the edge of the left lip during smiles) is
shown below in yellow with translucent skin:
After the data set was complete, we decided that the available functions
to deform the face mesh were overly complicated, and derived a simple
one:
1
P' = P + ( -------------- ) (Pd - P) t
|Po + P| + 1
where t is the time parameter running from 0 to 1, Po
is the point where the force is applied, Pd is the point of
attractive force, and P is every face vertex in the bounding box.
The formula moves every vertex in the bounding box towards the attraction
point Pd by a magnitude inversely proportional to its distance
from the point of force application Po. An example of this
muscle model is shown below:
This formula had an unacceptable side effect of seemingly to "pinch" the
points close to Po towards Pd. We abandoned it because
we felt we needed a "smoother attraction", which led us to:
|Po - P|
1 + cos( ---------- pi )
P' = P + ( R ) (Pd - Po) t
------------------------
2
where we introduced a new parameter R. R defines the
radius of influence, and while closely related to the size of the bounding
box, is a separate parameter. In general, however, R should
cover roughly the same area as the bounding box. Note that singularities
appear if R is much smaller than the bounding box. Note that the
Pd - Po term in this formula moves all the vertices in
the bounding box and the radius of influence parallel to the vector
between the point of force application towards the point of attraction.
In effect, the point of attraction actually defines a plane of
attraction for the bounding volume. This is the formula we settled with,
and the cosine term does provide us with a smooth deformation. A sample
of this muscle model is shown below:
Not all muscles on the face apply force linearly. Sphincter muscles around
the lips, for example, do not work well with this formula unmodified. We
thus defined a special case of the formula to deal with these muscles:
|Po - P|
1 + cos( ---------- pi )
P' = P + K ( R ) (P - Pd) t
------------------------
2
There are two important differences. The K term was introduced
to allow the muscle to contract in one axis and expand in another, instead
of a uniform contraction or expansion. The direction of motion is now
defined by P - Pd, which is in the direction of the
point of attraction (similar to formula 1). The reason we inverted the
sign of this formula was because we felt that a positive K should
denote contraction and a negative K should denote expansion, in
order to be consistent with formula 2. The example on the left below
shows a contracting sphincter muscle and the one on the right shows an
expanding sphincter muscle:
Putting it all together
Armed with the face mesh and the muscles, what remains is the definition
of parameters, which is largely a tedious trial-and-error task. We
present a few expressions generated by IBM DataExplorer
below (for mpeg demos click on the links below the images):
MPEG demos of our work
Click Here to see the smile animation.
Click Here to see the kiss animation.
Data Explorer files of our work
- Click here for the connections
list of the mesh.
- Click here for the DX macro which models a
linear contract/relax muscle on the face.
- Click here for the DX macro which models the
facial mesh. Note that this requires a vertex list, connections, and lip
connections file (similar to the ones above) in order to work.
Ahead
The effects we have achieved are far from realistic. The complexities
of facial animation have not yet been fully explored, but we believe we
have come up with a reasonable model for manipulating data (less than
500 vertices) versus the much larger data set acquired by machines with
good quality. Other areas may be explored, and those we feel are
interesting are listed below:
- Texture-mapped skin [Nahas90]
- Coordination of lip muscles with speech [Nahas88]
- Motion of other facial elements such as eyeball or jaw (click
here for the work we did in this area)
- Modelling of the interior of the mouth such as teeth and gums
(click here for the work we did
in this area)
- Modelling of hair features such as hair, moustache, beard, or
eyebrows
(click here for the work we did
in this area)
Project Deidre (II). Check it out!
~
References
Arad N., Dyn N., Reisfeld D., and Yeshurun Y., Image Warping by
Radial Basis Functions - Application to Facial Expressions,
CVGIP-Graphical Models and Image Processing, 1994
Badler N. and Platt S., Simulation of Facial Muscle Actions Based on
Rational Free-form Deformations, Computer Graphics, September
1992
Kang C., Chen Y., and Hsu W., Automatic Approach to Mapping a Lifelike
2.5D Human Face, Image and Vision Computing, 1994
Magnenat-Thalmann N., Minh H., de Angelis, M., and Thalmann D., Design, Transformation, and Animation of Human
Faces, Visual Computer
Magnenat-Thalmann N., Primeau E., and Thalmann D., Abstract Muscle Action Procedures for Human
Face Animation, Visual Computer, 1988
Nahas M., Huitric H., Rioux M., and Domey J., Facial Image Synthesis
Using Skin Texture Recording, Visual Computer, December 1990
Nahas M., Huitric H., and Saintourens M., Digital Actors
for Interactive Television, Proceedings of the IEEE, 1995
Tost D. and Pueyo X., Human Body Animation: A Survey, Visual
Computer, March 1988
Waters K., A Muscle Model for Animating
Three-dimensional Facial Expressions, Computer Graphics, July
1987
Maintained by: Steven Huang
and Dan Hung
Last Modified: December 11, 1995