Detecting the long-tail of Points of Interest in tagged photo collections

The paper tackles the problem of matching the photos of a tagged photo collection to a list of “long-tail” Points Of Interest (PoIs), that is PoIs that are not very popular and thus not well represented in the photo collection. Despite the significance of improving “long-tail” PoI photo retrieval for travel applications, most landmark detection methods to date have been tested on very popular landmarks. In this paper, we conduct a thorough empirical analysis comparing four baseline matching methods that rely on photo metadata, three variants of an approach that uses cluster analysis in order to discover PoI-related photo clusters, and a real-world retrieval mechanism (Flickr search) on a set of less popular PoIs. A user-based evaluation of the aforementioned methods is conducted on a Flickr photo collection of over 100, 000 photos from 10 well-known touristic destinations in Greece. A set of 104 “long-tail” PoIs is collected for these destinations from Wikipedia, Wikimapia and OpenStreetMap. The results demonstrate that two of the baseline methods outperform Flickr search in terms of precision and F-measure, whereas two of the cluster-based methods outperform it in terms of recall and PoI coverage. We consider the results of this study valuable for enhancing the indexing of pictorial content in social media sites.