Statistical foreground modelling for object localisation

Abstract

A Bayesian approach to object localisation is feasible given suitable likelihood models for image observations. Such a likelihood involves statistical modelling—and learning—both of the object foreground and of the scene background. Statistical background models are already quite well understood. Here we propose a “conditioned likelihood” model for the foreground, conditioned on variations both in object appearance and illumination. Its effectiveness in localising a variety of objects is demonstrated.