IEEE Workshop on Models versus Exemplars in Computer Vision 
to be held in conjunction with CVPR 2001 in Kauai, Hawaii 
(http://www.cs.berkeley.edu/~leungt/cvpr01_workshop.html) 

CALL FOR PAPERS 

The variety of models of shape and intensity/color distributions used in 
computer vision could be regarded as lying along a spectrum. At one 
extreme, the model space is continuous, consisting of objects with 
explicit, continuous parameterizations. Examples include: spline-based 
models of shape and eigenspace models of shading. At the other extreme 
are discrete sets of "exemplars." Examples include: modeling texture as 
random samples from a collection of exemplar patches and modelling a 
walking person as a set of exemplar outlines. As an illustration of the 
two paradigms, consider how one might represent a dancer: a model-based 
approach might construct an articulated 3D model with texture-mapped 
limbs and hand-specified degrees of freedom; an exemplar-based 
representation, on the other hand, might consist of images of the dancer 
in M different poses from N vantage points. 

Each kind of representation has its advantages and disadvantages. 
Parameterized models have the advantage of sparse representation and 
clear correspondence with semantics and underlying physical phenomena; 
on the other hand, they need to be laboriously constructed by hand or 
built from pre-annotated training data, and search over parameters 
suffers combinatorial growing pains with increasing model 
dimensionality. Exemplar-based representations, in contrast, are easy 
to create, and they are capable of representing nonlinear manifolds 
efficiently. Disadvantages include a lack of continuous structure to 
represent physical variables and inefficiency in space-complexity. 

The goal of this workshop is to examine the inherent strengths and 
weaknesses of these two kinds of representation. We invite papers that 
address issues related to the following (with preference for domains 
other than object recognition, where the "object-based versus 
view-based" discussion has covered similar ground): 
* Automated construction of models from unannotated or partially 
annotated training data. 
* Algorithms for embedding exemplars in structured spaces. 
* Algorithms which alleviate the problems of one or the other 
approach. 
* Comparisons of model-based and exemplar-based techniques applied to 
the same task. 
* Analysis of algorithm complexity, for models or exemplars, with 
increased dimensionality. 
* Hybrid techniques for combining model- and exemplar-based methods. 

In order to incite debate and discussion, we will allocate time for 
multiple panel discussions during the workshop, with paper presenters 
and invited speakers as panelists. 

IMPORTANT DATES (subject to minor changes): 
August 17, 2001 Electronic paper submission [PDF] 
(one week after CVPR 2001 notifications) 
Send submissions to mailto:exemplar@microsoft.com 
8 pages maximum. 
September 14, 2001 Notification of acceptance/rejection 
October 12, 2001 Final electronic submission due [PDF] 
December 14, 2001 Workshop (following CVPR 2001, Kauai, Hawaii) 

ORGANIZING COMMITTEE: 
Andrew Blake Microsoft Research Cambridge ablake@microsoft.com 
Thomas Leung Compaq Cambridge Research Lab tleung@crl.dec.com 
James Rehg Compaq Cambridge Research Lab rehg@crl.dec.com 
Kentaro Toyama Microsoft Research Redmond kentoy@microsoft.com 

PROGRAM COMMITTEE (for affiliations, please see web page): 
George Bebis Larry Davis Jitendra Malik 
Matt Brand Brendan Frey Jianbo Shi 
Tim Cootes Dariu Gavrila Thomas Vetter 
Trevor Darrell David Kriegman Song-Chun Zhu