Using Multiple Segmentations to Discover Objects and their Extent in Image Collections