Speaker: Simon Lucey

Presenter: Simon Lucey (Carnegie Mellon University) 

Enforcing Convex Local Discriminative Responses for Generic Non-Rigid Object Alignment

A common method in computer vision for non-rigidly aligning an object template to a 
source image is through the application of active appearance model (AAM) fitting. This 
approach employs a generative model of the object, which is decoupled into appearance 
and non-rigid shape variation. It then employs an iterative Gauss-Newton minimization to 
fit the non-rigid object template to a given source image. Unfortunately, this approach 
encounters a similar problem to many generative model style approaches in vision, in that 
the generative appearance model generalizes poorly to unseen appearance variation.   We 
refer to this poor generalization effect as the œGeneric AAM problem.  
In this talk, we will demonstrate an approach to non-rigid alignment which employs a 
hybrid generative and discriminative object model. Our approach, which we refer to as a 
constrained local model (CLM), employs a discriminative appearance model and a 
generative shape model. We will demonstrate that since our object model is no longer 
completely generative, traditional convex optimization methods (like Gauss-Newton) can 
no longer be easily applied. To circumvent this problem, we present an approach we refer 
to as robust convex quadratic fitting (RCQF) which forces the local discriminative 
appearance responses (for each point in the non-rigid object) to be convex allowing us to 
fit the whole object through efficient convex optimization methods.  We demonstrate the 
utility of our approach on faces, making our technique an ideal front-end to any 
automatic face analysis system that is trying to recognize identity, expression, speech, 
Also, if there is time we will quickly showcase some extensions to the approach of 
œcongealing we have recently developed that allows for the efficient unsupervised 
automatic alignment of an ensemble of images stemming from the same object class.  
Speaker Bio: Simon Lucey has been a Systems Scientist in the Robotics Institute at 
Carnegie Mellon University since October 2005. Before that he was a Post-Doc in the 
Electrical and Computer Engineering (ECE) department at Carnegie Mellon University 
since June 2002. Dr. Lucey™s research interests are in computer vision, pattern 
recognition and machine learning with specific interests in their application to face 
analysis and multi-modal biometrics. He received his Ph.D. in 2002 on the topic of audio- 
visual speaker and speech recognition from the Queensland University of Technology 
(QUT), Australia. To his credit he has over 30 publications in international conferences, 
journals  and book chapters. He has been a reviewer for a number of international 
journals and conferences in vision, learning, pattern recognition and multimedia.

TASC-1 9204
Wednesday, June 11, 2008 - 13:30