Incremental Learning of Gaussian Mixture Models

Yulia Hicks

(Cardiff University)

Dr Yulia Hicks will give a quick overview of her research over the recent years, which involved using and developing statistical models for audio, image and video processing. The applications include tracking articulated human motion in video, modelling and generating human interactions, animated speech driven models of human faces, blind source separation and biological models. The main part of the talk will be devoted to her most recent research in the area of incremental learning of Gaussian Mixture Models (GMMs) and Hidden Markov Models (HMMs) and application of these methods to image segmentation and learning dynamical models of faces. GMMs and HMMs became very popular for modelling data sets in the area of Computer Vision in recent years. However, learning GMMs and HMMs on large data sets can be both time and memory intensive. In addition, not all data may be available at once. Thus the possibility of splitting the data sets into smaller ones, learning the models separately on them, possibly at different times and sites, and then merging the models would be useful. The proposed methods allow for this. Applications and experiments demonstrating the effectiveness of the methods will be
presented.
Tuesday 1st December 2009, 14:00
Robert Recorde Room
Department of Computer Science