Sifting through visual arts collections

Abstract

We introduce a visualization system for large image sets which combines a distance function, a clustering and a projection method. The distance function, the clustering and the projection methods run so fast that they can calculate new results during the interaction with the user and can therefore be adapted dynamically to the context of the investigation and the requests made by the user at any given moment. The system aims to facilitate investigations which take similarity between images in terms of human perception into account. Similarity in terms of human perception is highly context and task dependent and cannot be described by a metric in the mathematical sense. Functions reflecting similarity in terms of human perception have to be adapted dynamically to the context of the investigation as well as to the tasks assigned at any given time. Our system thus shows that these requirements can be met in principle, and we propose it as a basis for developing specific applications and suitable surfaces in collaboration with experts for whom such tools are useful, as for instance experts of art theory.