Automatic region of interest detection in tagged images

On the Web, tagging is the preferred approach to describing multimedia items in order to make them searchable. The information value of tags can be significantly enhanced if they are linked to specific image regions. In this paper, we describe an approach to automatically detect regions of interest (ROIs) that are visually related to a given tag. Our technique is domain independent and works unsupervised, just by leveraging the knowledge from large-scale collections of tagged images. The ROIs are obtained by local feature matching between similarly tagged images. We demonstrate the performance and high accuracy of our approach in experiments on a set of 41 different topics and more than 9000 images.