[4eyes] Talk on Feb. 16 - Model-Based Perceptual Grouping and Shape Abstraction

[Matthew Turk]

Although Monday, Feb. 16 is a holiday, I have a visitor from the University of Toronto who will give a talk that day, and all are invited:

Time and location: 2:00pm, CS Conference Room (1132 HFH)
Title: Model-Based Perceptual Grouping and Shape Abstraction

Sven Dickinson
Department of Computer Science
University of Toronto
http://www.cs.toronto.edu/~sven/

Abstract:

For many object classes, shape is the most generic feature for object categorization. However, when a strong shape prior, i.e., a target 
object, is not available, domain independent, mid-level shape priors must play a critical role in not only grouping causally related 
features, but regularizing or abstracting them to yield higher-order shape features that support object categorization. In this talk, I will 
present a framework in which mid-level shape priors take the form of a vocabulary of simple, user-defined 2-D part models.  From the 
vocabulary, we learn to not only group oversegmented regions into parts, but to abstract the shapes of the region groups, yielding a set of 
abstract part hypotheses.  However, the process of shape abstraction can be thought of as a form of "controlled hallucination", which comes at 
the cost of many competing 2-D part hypotheses.  To improve part hypothesis precision, we present two approaches that exploit the context 
of the hypotheses.  In the first approach, we exploit spatiotemporal coherence (temporal context) of part hypotheses in a dynamic 
environment, and formulate hypothesis selection in a graph-theoretic, probabilistic framework. In the second approach, we assume that the 2-D parts represent the component faces of aspects that model a vocabulary of 3-D part models. We then exploit the relational structure (spatial 
context) of the faces encoded in the aspects, and again formulate hypothesis selection in a graph-theoretic, probabilistic framework.  
Finally, we introduce a technique that is able to recover the pose and shape of a volumetric part from a recovered aspect, yielding a framework 
that revisits the classical problem of recovering a set of qualitative 3-D volumetric parts from a single 2-D image.