Compass Clustering: A New Clustering Method for Detection of Points of Interest Using Personal Collections of Georeferenced and Oriented Photographs

摘要

Knowledge discovery in large online photographic repositories has been an active area of research in recent years. This is due to the great popularization of devices equipped with image capture, such as digital cameras, smartphones and tablets. Moreover, the image files generated by those devices are easily spread out on the Web through social networking sites. Typically, the photos stored in these repositories bear valuable metadata, such as, geographic coordinates, timestamp, and camera orientation. This information can be used for many interesting data mining tasks, such as detection of points-of-interest (POls) and trip planning. This paper introduces Compass Clustering, a new clustering algorithm for detecting POls in georeferenced and oriented photo repositories. Most of the state-of-the-art approaches for POI detection cluster photos based solely on their geographic proximity. However, in many cases, the POIs are within a certain distance from the point where the photo was taken, that is, not in the exact camera location but in the direction it is pointing to, and thus many photos would be erroneously classified by existing methods. Therefore, we propose to exploit the camera orientation in order to identify more reliable POIs that reflect the real intention of people when taking photos. We evaluated our approach on a collection of more than 8,000 georeferenced and oriented photos collected from Flickr.