Organizing a personal image collection with statistical model-based ICL clustering on spatio-temporal camera phone meta-data

This paper addresses the issue of automated organization of a personal image collection, in particular to respond to the emerging needs from a mobile camera phones, and proposes a collection organization technique based on picture geolocation and timestamps.

Abstract

This paper addresses the issue of automated organization of a personal image collection, in particular to respond to the emerging needs from a mobile camera phones. The issues related to browsing through large image collections acquired from such devices are first discussed. In contrast with retrieval in meta-data-less collections, which necessarily relies on image content, we propose a collection organization technique based on picture geolocation and timestamps. These descriptors are indeed available and generally reliable in the proposed context. Collection organization is formulated as an unsupervised classification problem, in both space and time. The statistical integrated completed likelihood criterion is chosen, providing effective solutions both to model complexity determination and the cluster separability objective, in a setting which limits arbitrary algorithm parametrization. Reliability of space and time partitions obtained are then assessed to select a segmentation, which may then provide a calendar-type structured view for navigating in the picture collection.