Non-Parametric Latent Variable Models for Shape and Motion Analysis

Speaker: Dr. Raquel Urtasun, MIT
Host: Prof. Fei Sha, Prof. Ram NevatiaAbstract:
Dimensionality reduction is a popular approach to dealing with high dimensional data sets. It is often the case that linear dimensionality reduction, such as principal component analysis (PCA), does not adequately capture the structure of the data. In this talk I will discuss Probabilistic Non-linear Latent Variable models in the context of 3D human body tracking, 3D shape recovery from single images, character animation and classification.First, I will describe how to use Gaussian Process Latent Variable models (GPLVMs) for learning human pose and motion priors for 3D human body tracking from monocular images. I will then show how to combine multiple local models to model the space of possible deformations of objects of arbitrary shapes, but made of the same material. This will allow us to perform monocular 3D shape recovery in the presence of complex deformations of poorly textured objects.In dimensionality reduction approaches, the data is typically embedded in a Euclidean latent space. However for some data sets, such as human motion, this is inappropriate. We present a range of approaches for embedding data into non-Euclidean latent spaces that incorporate prior knowledge. This allows us to learn models suitable for motion generation with good generalization properties. Finally, I.ll show how to extend these models to be discriminative, resulting in accurate classification even when dealing with high dimensional input spaces and very small training sets.Biography:
Raquel Urtasun is currently a Research Scientist at the International Computer Science Institute at Berkeley, working with Prof. Trevor Darrell and will be a Visiting Scholar at UC Berkeley EECS. Raquel’s main research areas are computer vision, machine learning and computer graphics. During 2006-2008, she was a postdoctoral associate at MIT-CSAIL. She earned her PhD at EPFL (Switzerland) in 2006 under the supervision of Prof. Pascal Fua on Motion Models for Robust 3D Human Body Tracking.

Location: Charles Lee Powell Hall (PHE) - 223

Audiences: Everyone Is Invited

Contact: CS Colloquia