The CMU/Pitt Automated Facial Image Analysis System
T. Kanade1 and J.F. Cohn2
1The Robotics Institute, Carnegie Mellon University,
Pittsburgh, PA, USA
2Department of Psychology, University of Pittsburgh, Pittsburgh,
PA, USA
Both the configuration and the timing of facial actions are important
in emotion expression and recognition. To investigate the timing and configuration
of facial actions, our interdisciplinary group of behavioral and computer
scientists developed and applied a computer-vision based approach, the
CMU/Pitt Automated Facial Image Analysis (AFA) System. AFA is capable
of automatically recognizing facial action units and analyzing their timing
in facial behavior.
The latest version of the system is based on Active Appearance Models
(AAMs). AAMs are generative, parametric models and consist of a shape
component and an appearance component. The shape component is a triangulated
mesh that deforms in response to changes in the parameters corresponding
to a face undergoing both rigid motion (head pose variation) and non-rigid
motion (expression). The appearance component of the AAM is an image of
the face, which itself can vary under the control of the parameters. As
the parameters are varied, the appearance varies so as to model effects
such as the emergence of furrows and wrinkles and the visibility of the
teeth as the mouth opens.
Traditional AAMs are actually 2D in the sense that rigid
head motion and non-rigid facial motion are confounded in the 2D-mesh
shape model. To address these problems, we use an extension to AAMs that
augments the usual 2D mesh model with an actual 3D shape model, thus separately
and explicitly modeling the 3D rigid motion of the head and 3D non-rigid
facial expression into two disjoint sets of parameters. This advancement
allows us to extract and separate the pose, 3D shape deformation, and
appearance change of the face, which are then input to the facial action
recognizer.
In initial testing, this version of the system has demonstrated concurrent
validity with human-observer based facial expression recognition and both
human-observer and EMG based analysis of timing.
Paper presented
at Measuring Behavior 2005
, 5th International Conference on Methods and Techniques
in Behavioral Research, 30 August - 2 September 2005, Wageningen, The
Netherlands.
© 2005 Noldus
Information Technology bv
|