Exploiting photo location and direction for clustering-based points-of-interest discovery

Several works have exploited the geographic information of photos through spatial clustering algorithms aiming at the automatic discovery of points of interest (POIs). The assumption is that dense regions in terms of geographically nearby photos are good POI surrogates. However, this approach fails when: (i) nearby photos point to different POIs, and (ii) POIs lay within a large distance from the camera. In (i) current approaches would erroneously associate nearby photos to the same POI, whereas in (ii) the photos would not be associated to the POI they really point at. In this paper, we propose to address these problems by devising two novel clustering-based strategies that exploit location along-side compass metadata for POI discovery. We use a large collection of geotagged and oriented photos collected from Flickr related to three different cities and show that our approaches can be more accurate than baselines solely based on location metadata.