Tag recommendation for georeferenced photos

Abstract

This paper presents methods for annotating georeferenced photos with descriptive tags, exploring the annotations for other georeferenced photos which are available at online repositories like Flickr. Specifically, by using the geospatial coordinates associated to the photo which we want to annotate, we start by collecting the photos from an online repository which were taken from nearby locations. Next, and for each tag associated to the collected photos, we compute a set of relevance estimators with basis on factors such as the tag frequency, the geospatial proximity of the photo, the image content similarity, and the number of different users employing the tag. The multiple estimators can then be combined through supervised learning to rank methods such as Rank-Boost or AdaRank, or through unsupervised rank aggregation methods well-known in the information retrieval literature, namely the CombSUM or the CombMNZ approaches. The most relevant tags are finally suggested. Experimental results with a collection of photos collected from Flickr attest for the adequacy of the proposed approaches.