TITLE: Analysis,
Synthesis and Retargeting of Facial Expressions
SPEAKER: Erika Chuang [Stanford University]
DATE: 2:00 - 3:00 P.M., ThursdaySeptember 4, 2003
LOCATION: Sigma, 1L (PA)
HOST: Vinay Deolalikar
ABSTRACT:
Computer animated characters have recently gain popularity in many
applications, including web pages, computer games, movies, and
various human computer interface designs. In order to make these
animated characters lively and convincing, they require sophisticated
facial expressions and motions. Traditionally, these animations are
produced entirely by skilled artists. Although the quality of the
animation produced this way remains the best, the process is slow and
costly. Motion capture performance of actors/actresses is one
technique that attempts to speed up this process. One problem for
this technique is that the captured motion data is not easily
editable. In recent years, statistical techniques have been used to
address this problem by learning the mapping between audio speech and
facial motion. New facial motion can be synthesized for novel audio
data by reusing the motion capture data. However, since facial
expressions are not modeled in these approaches, the resulting facial
animation is realistic, yet expressionless.
This work takes an expressionless talking face and creates an expressive
facial animation. This process consists of three parts: expression
synthesis, blendshape retargeting, and head motion synthesis. Expression
synthesis uses a factorization model to describe the interaction between
facial expression and speech content underlying each particular facial
appearance. A new facial expression can be applied to novel input video,
while retaining the same speech content. Blendshape retargeting maps
facial expressions onto a 3D face model using the framework of blendshape
interpolation. Three methods of sampling the key shapes, or prototype
shapes, from data are evaluated. In addition, the generality of blendshape
retargeting is demonstrated in three different domains. Head motion
synthesis uses audio pitch contours to derive new head motion. The global
and local structure of the pitch statistics and the coherency of head
motion are utilized to determine the optimal motion trajectory.
Finally, expression synthesis, blendshape retargeting, and head
motion synthesis are combined into a prototype system and
demonstrated through an example.
|
|
|
printable version
