Modeling dynamical and multi-modal computer vision data via non-linear probabilistic dimensionality reduction

When?

Thursday 14 June 2012, 11:00 to 12:00

Where?

CVSSP Seminar Room (40bAB05)

Open to:

Public, Staff, Students

Speaker:

Mr Andreas Damianou, University of Sheffield

Abstract: Real world data encountered in computer vision applications are usually difficult to deal with due to their high-dimensional and complex nature. One line of work aims at modelling such data via probabilistic dimensionality reduction, representing the original observations in a low dimensional latent space. This talk briefly reviews the principles behind such approaches and presents a new Bayesian method of this type which is based on Gaussian processes. The basics of Gaussian process modelling will also be discussed. The Bayesian framework allows us to incorporate in the low-dimensional data representation some structure or prior assumptions that we consider appropriate for the problems at hand. For example, by including a temporal prior distribution for the latent space we constrain it in a way that better suits dynamical data. This is demonstrated in experiments with high resolution video sequences. Another extension is concerned with automatically segmenting the latent space, therefore leading to a model suitable for multi-modal datasets, such as silhouette and pose pairs of motion capture data or RGB and depth images recorded with Kinect.

Biography: I am a PhD candidate at The University of Sheffield,  under the supervision of Prof. Neil Lawrence. I am split between the dept. of Neuroscience (Sheffield Institute for Translational Neuroscience (SITraN)) and dept. of Computer Science. Prior to that, I received my MSc (Artificial Intelligence and Machine Learning) from the Department of Informatics at the University of Edinburgh and my BSc from the Department of Informatics & Telecommunications at the University of Athens. My area of interest is Machine Learning (with an emphasis on Gaussian processes and Bayesian nonparametics) with applications (mainly) in computer vision and computational biology.

Webpage: http://staffwww.dcs.shef.ac.uk/people/A.Damianou/index.html