Discovering Latent Clusters from Geotagged Beach Images Yang Wang 1 and Liangliang Cao 2 1 Department of Computer Science, University of Manitoba, Canada ywang@cs.umanitoba.ca 2 IBM T.J.Watson Research Center, USA liangliang.cao@us.ibm.com Abstract. This paper studies the problem of estimating geographical locations of images. To build reliable geographical estimators, an impor- tant question is to ﬁnd distinguishable geographical clusters in the world. Those clusters cover general geographical regions and are not limited to landmarks. The geographical clusters provide more training samples and hence lead to better recognition accuracy. Previous approaches build geographical clusters using heuristics or arbitrary map grids, and can- not guarantee the eﬀectiveness of the geographical clusters. This paper develops a new framework for geographical cluster estimation, and em- ploys latent variables to estimate the geographical clusters. To solve this problem, this paper employs the recent progress in object detection, and builds an eﬃcient solver to ﬁnd the latent clusters. The results on beach datasets validate the success of our method. 1 Introduction Geotagged images are receiving more and more research attentions in recent years. A geotagged image is associated with a two dimensional vector, latitude and longitude, representing a unique location on the Earth. The goal of this paper is to use the visual information to estimate the geographical locations even when they are not provided. As evidenced by the success of Google Earth, there is great need for such geographic information among the mass. Many web users have high interests on not only the places they live but also other interesting places around the world. Geographic annotation is also desirable when reviewing the travel and vacation images. For example, when a user becomes interested in a nice photo, he or she may want to know where exactly it is. Moreover, if a user plans to visit a place, he or she may want to ﬁnd out the points of interest nearby. Recent studies suggest that geo-tags expand the context that can be employed for image content analysis by adding extra information about the subject or environment of the image. Estimating the geolocation of images is not an easy task. As the earlier work shown in [10] [6], only a quarter of the test images can be located subject to a rough region (approximately 750 km) near their true location. At the metropoli- tan scale, visual feature based annotations perform no better than chance. S. Li et al. (Eds.): MMM 2013, Part II, LNCS 7733, pp. 133–142, 2013. c  Springer-Verlag Berlin Heidelberg 2013