Geo-temporal structuring of a personal image database with two-level variational-bayes mixture estimation

Abstract

This paper addresses unsupervised hierarchical classification of personal documents tagged with time and geolocation stamps. The target application is browsing among these documents. A first partition of the data is built, based on geo-temporal measurement. The events found are then grouped according to geolocation. This is carried out through fitting a two-level hierarchy of mixture models to the data. Both mixtures are estimated in a Bayesian setting, with a variational procedure: the classical VBEM algorithm is applied for the finer level, while a new variational-Bayes-EM algorithm is introduced to search for suitable groups of mixture components from the finer level. Experimental results are reported on artificial and real data. © 2010 Springer-Verlag Berlin Heidelberg.