Cluster-Based Landmark and Event Detection for Tagged Photo Collections

TitleCluster-Based Landmark and Event Detection for Tagged Photo Collections
Publication TypeJournal Article
Year of Publication2011
AuthorsPapadopoulos, Symeon, Christos Zigkolis, Yiannis Kompatsiaris, and Athena Vakali
JournalIEEE MultiMedia
Volume18
Pagination52-63
Abstract

The rising popularity of photosharingapplications on the Webhas led to the generation of hugeamounts of personal image collections.Browsing through image collections ofsuch magnitude is currently supported by theuse of tags. However, tags suffer from severallimitations—such as polysemy, lack of uniformity,and spam—thus not presenting an adequatesolution to the problem of contentorganization. Therefore, automated contentorganizationmethods are of particular importanceto improve the content-consumptionexperience. Because it’s common for users to associatetheir photo-captured experiences withsome landmarks—for example, a tourist site oran event, such as a music concert or a gatheringwith friends—we can view landmarks andevents as natural units of organization forlarge image collections. It’s for this reasonthat automating the process of detecting suchconcepts in large image sets can enhance theexperience of accessing massive amounts ofpictorial content.In this article, we present a novel scheme forautomatically detecting landmarks and eventsin tagged image collections. Our proposal isbased on the simple yet elegant concept ofimage similarity graphs as a means of combiningmultiple notions of similarity betweenimages in a photo collection; in our case, weuse visual and tag similarity. We perform clusteringon such image similarity graphs bymeans of community detection,1 a processthat identifies on the graph groups of nodesthat are more densely connected to eachother than to the rest of the network. In contrastto conventional clustering schemes suchas k-means or hierarchical agglomerative clustering,community detection is computationallymore efficient and doesn’t require thenumber of clusters to be provided as input. Subsequently,we classify the resulting image clustersas landmarks or events by use of featuresrelated to the temporal, social, and tag characteristicsof image clusters. In the case of landmarks,we also conduct a cluster-merging stepon the basis of spatial proximity to enrich ourlandmark model.

PDF: 

auth logo

Location & Contact

Department of Informatics
Aristotle University of Thessaloniki
Thessaloniki GR-54124

t  | (+30) 2310 998415
e | oswinds@csd.auth.gr