Geo-clustering of Images with Missing GeoTags

Images with GPS coordinates are a rich source of information about a geographic location. Innovative user services and applications are being built using geotagged images taken from community contributed repositories like Flickr. Only a small subset of the images in these repositories is geotagged, limiting their exploration and effective utilization. We propose to use optional meta-data along with image content to geo-cluster all the images in a partly geotagged dataset. We formulate the problem as a graph clustering problem where edge weights are vectors of incomparable components. We develop probabilistic approaches to fuse the components into a single measure and then, discover clusters using an existing random walk method. Our empirical results strongly show that meta-data can be successfully exploited and merged together to achieve geo clustering of images missing geotags.