[Colloquium] Sminchisescu talk April 19th at TTI

[Meridel Trimble]

Toyota Technological Institute Talk

Speaker:  Cristian Sminchisescu, University of Toronto 
Speaker's Homepage:  http://www.cs.toronto.edu/~crismin

Date:  Monday, April 19, 2004
Time:  2:45 PM
Place:  TTI-C, 1427 E. 60th St., 2nd Floor (University Press Building)
Refreshments Provided

Title: "Density Propagation and Manifold Learning Algorithms for Visual Inference" 

Abstract: 
Many visual perception problems, including object tracking or scene
reconstruction can be formulated as inference using non-linear generative
models, defined over high-dimensional state spaces. Suboptimal modeling,
model-image matching ambiguities or occlusion lead, however, to representations
that are weakly constrained by the image evidence. Therefore, robust solutions
typically involve estimating and propagating highly multimodal posterior state
distributions. Trapping in local optima represents a significant difficulty, and
it is important to locate and, for dynamic scenes, track a set of representative
configurations, over time. Learning models well-adapted to the task, but with
good generalization power is also a necessary step, to reduce the search
complexity, and obtain reliable estimates. 

In the talk, I will present inference and learning algorithms, and demonstrate
them on the problem of monocular tracking and 3D motion reconstruction of a
35-dimensional articulated human model. Besides high-dimensionality and matching
difficulties, other challenges are the presence of nonlinear physical
constraints and ill-conditioning, because the depth related state variables are
highly uncertain in any single monocular image. Many of our algorithms apply
quite generally, to any differentiable multi-modal probability density
(Covariance Scaled Sampling / Hyperdynamic Sampling, Eigenvector Tracking /
Hypersurface Sweeping, Generalized Darting, Variational Mixture Smoothing),
whereas `Kinematic Jump Sampling' further exploits the implicit symmetries
encountered in the monocular 3D human articulated problem. I will also describe
an `Embedded Visual Density Propagation' algorithm, that restricts inference to
automatically learned, non-linear manifolds. I will show that under the proposed
construction, one can learn an embedded, globally differentiable generative
model, where all the above methods apply, with the the advantage of reduced
uncertainty, and low-dimensional computational costs. 

If you have questions, or would like to meet the speaker, please contact Carole
at 773.702.5033 or cfkipp at tti-c.org.

For information on future TTI-C talks or events, please go to the TTI-C Events
page at http://www.tti-c.org/events.shtml
-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://mailman.cs.uchicago.edu/pipermail/colloquium/attachments/20040413/a4ab22d0/unnamed.htm