Nicolas Galoppo von Borries

1
Facial Animation Synthesis

• Physically based modeling
• Class Presentation
• April 5, 2004

2
Why is it an interesting topic?

• The human face anatomy is very complex
• We are very sensitive to facial appearance, so it
  is very hard to make it look real!
• The facial modeling and animation process is very
  cumbersome for the animator
• Nowadays, the faces of characters in movies such
  as Toy Story, Shrek, Final Fantasy typically have
  several thousand control points

3
Goal in face synthesis research

• Create realistic animation
• Operate in real time
• Very much automated, make the animators job
  easier
• Adapt easily to newly introduced faces
• Animate/deform according to intrinsic
  parameters/expressions, rather than to
  representation parameters

4
The mathematical background

• We use statistics and take a geometrical approach
• Do statistical analysis on a large number of
  example faces
• Animate between faces based on the distribution
  of the example faces
• Extract pose, expression, emotion, illumination,
  from newly introduced face.

5
Linear Statistical Models

• Aim - to model object x as a function of
  parameters b xM(b)

Example face model
Frown Smile Wink

• Input - database of objects x1,x2,...,xm

6
Linear Statistical Models
We build linear models
using statistical data X x1 x2 ... xm
as follows

• Simple linear model

or

• PCA-based model Orthogonal decomposition

7
Linear Statistical Models

• PCA-based model
• where
• F f1 f2 ... fk, k ? m,

f1,..., fk eigenvectors corresponding to the k
largest eigenvalues of the covariance matrix
8
Linear Statistical Models

• PCA-based model bounding the parameter space
  This is the main advantage of PCA over simple
  linear model!
• To account for 98 of variation
• or
• Estimating the probability of lying in the
  object space

9
Linear Statistical Models
PCA-based model. Bounding the parameter space
illustration.
Principal axis Of variation
We take the k first principal components to
account for the main modes of variation in the
representation space