In this paper we consider large image collections and their organization into meaningful data structures upon which applications can be build (e.g. navigation or reconstruction). In contrast to structures that only reflect local relationships between pairs of images we propose to account for the information an image brings to a collection with respect to all other images. Our approach builds on abstracting from image domains and focusing on image regions, hereby reducing the influence of outliers and background clutter. We introduce a graph structure based on these regions which encodes the overlap between them. The contribution of an image to a collection is then related to the amount of overlap of its regions with the other images in the collection. We demonstrate our graph based structure with several applications: image set reduction, canonical view selection and image-based navigation. The data sets used in our experiments range from small examples to large image collections with thousands of images.     «

In this paper we consider large image collections and their organization into meaningful data structures upon which applications can be build (e.g. navigation or reconstruction). In contrast to structures that only reflect local relationships between pairs of images we propose to account for the information an image brings to a collection with respect to all other images. Our approach builds on abstracting from image domains and focusing on image regions, hereby reducing the influence of outlier...     »