Browsing visual collections using graphs

Abstract

Category search can be supported by methods that allow intelligent selection of potentially relevant images. This paper explores the use of a nearest neighbor network in the selection process. We created a prototype that visualizes the network of images. As in the nearest neighbor network the images are connected to similar images we assume that if an image is selected or deselected, the same action can be performed on its neighbors. This results in five possible actions: selecting an image, selecting an image with its nearest neighbors, deselecting an image, deselecting an image with its nearest neighbors and growing the selection with all the nearest neighbors. Using these actions four different interaction scenarios are defined which are evaluated using experiments. Our experiments show that the nearest neighbor network can have a positive effect on the interaction effort needed to select images, compared to the baseline of sequentially selecting images. The experiments further show that the different interaction scenarios can cater for different values of allowed interaction effort with respect to the requirements on precision and recall.