[Colloquium] TTIC Talks: Dahua Lin, MIT

[Liv Leader]

REMINDER:

When:     Wednesday, March 14 @ 11 a.m.

Where:    TTIC, 6045 S. Kenwood Avenue, Room 526

Who:       Dahua Lin, MIT

Title:       Visual Modeling with Bayesian Nonparametrics

Modeling and understanding complex visual scenes is one of the
fundamental goals of computer vision. My thesis research aims to
develop a generative model
of dynamic scenes that integrates different aspects, notably
appearance, motion, and
layering. This model not only provides coherent interpretation of the
scenes by bringing
different views together, but also leads to effective algorithms for
solving practical tasks. In this
talk, I will describe the overall framework of the model and then
focus on a specific component,
namely the model of image appearance. In particular, this appearance
model incorporates
a patch manifold with a conditional Markov random field in a probabilistically
consistent way, achieving great capability of expressing rich details
while maintaining global coherence.
Furthermore, the appearance model, which considers each video frame as
a composite of local patches, provides a useful representation for semantic
analysis. A mixture of visual topics is then built thereon to capture
the semantic aspect. The need
to evolve the topic mixture model over time motivates an important
work in my thesis
research, namely the construction of dependent Dirichlet processes
(DDPs). Dirichlet
processes are widely used in nonparametric Bayesian models to provide
priors over mixture
components, and how to incorporate temporal dependency between DPs has
been an open problem.
To address this, I develop a new construction of DDPs that exploit the
connections
between Poisson and Dirichlet processes, enabling the use of various
forms of dependency
between nonparametric mixture models.
Both the generative model of visual scenes and the new DDP
construction have been
tested on a variety of applications. I will present the results on
image/video denoising and
dynamic topic analysis (on both videos and documents). The models and
techniques to be presented constitute an initial but significant step
towards the goal
of modeling complex phenomena. I will also discuss several directions
that I wish to
follow to advance this work.

Host: Greg Shakhnarovich, greg at ttic.edu

-- 
Liv Leader
Human Resources Coordinator

Toyota Technological Institute Chicago
6045 S Kenwood Ave
Chicago, IL 60637
Phone- (773) 702-5033
Fax-     (773) 834-9881
Email-  lleader at ttic.edu
Web-   www.ttic.edu