Location recognition on lifelog images via a discriminative combination of generative models

Abstract

This paper presents a generative framework aimed at the analysis of a "visual lifelog" captured by wearing a camera for long periods of time. Here, we focused on location recognition and we propose the use of an ensemble of heterogeneous generative models able to capture the different aspects that characterize each location. We defined the likelihood of the ensemble as the likelihood of a mixture model whose components are the individual models themselves. Our results set the new state of the art on all the tasks associated with the SenseCam-32 dataset and outperform Bayesian model averaging and several other discriminative combination techniques. From a theoretical perspective, this paper proposes a principled (discriminative) combination of heterogeneous generative models able to cope with extremely challenging classification tasks and it demonstrates that combining such diverse heterogeneous models is indeed advantageous.