Spectral Latent Variable Models for Perceptual Inference
Speaker: Cristian Sminchisescu , Professor, University of Toronto and TTI-C
Date: November 2 2007
Time: 4:00PM to 5:00PM
Location: Star Seminar Room (32-D463)
Host: C. Mario Christoudias, Gerald Dalley, MIT CSAIL
Contact: C. Mario Christoudias, Gerald Dalley, 3-4278, 3-6095, cmch@csail.mit.edu, dalleyg@mit.edu
Relevant URL: I will discuss a recently introduced class of non-linear generative models
referred to as Sparse Spectral Latent Variable Models (SLVM), that
combine the advantages of spectral embeddings with the ones of parametric
latent variable models, as follows: (1) provide latent spaces that
preserve geometric properties -- either global or local -- of the data
distribution; (2) offer low-dimensional spaces with probabilistic,
bi-directional mappings to and from the data space, (3) are
probabilistically consistent (reflect the data distribution, both jointly
and marginally) and can be learned efficiently. We show that SLVMs provide
competitive priors versus methods based on PCA, GPLVM (Gaussian
Process Latent Variable Model) or GTM (Generative Topographic Mapping),
for the visual tracking of human motions like walking, running, pantomime
or dancing in a benchmark dataset. Empirically, we show that SLVMs are
effective at the task of 3d human motion reconstruction in movies like
Run Lola Run. Time allowing, I will discuss discriminative continuous
time series estimators based on conditional Bayesian Mixture of Experts
Markov Models (the BM^3E model) and illustrate their potential for
entirely automating the tasks of 3d reconstruction and visual tracking.
Acknowledgements: Allan Jepson, Atul Kanaujia, Dimitris Metaxas.
See other events that are part of MIT Machine Vision Colloquium 2007/2008
See other events happening in November 2007
