Generating representative images for landmark by discovering high frequency shooting locations from community-contributed photos

Representative images generation offers a comprehensive knowledge for landmark and is a hot research area recent years. This paper presents a representative images generation system by discovering high frequency shooting locations from geo-tagged community-contributed photos. We discover that the views (e.g. far and near, front, back and side) of the photos taken in the same location are usually similar and but different in different shooting locations. Our system is realized by three steps: 1) Landmark dataset is filtered from social media by the combination of tags and geo-tags. 2) High frequency shooting locations are mined by geo-tag cluster. 3) Visual feature is then used for removing irrelevant images and ranking by intra and inter SIFT matching. This work is the first attempt to generate representative images by high frequency shooting locations mining. Evaluating on ten landmarks shows its effectiveness .