The fast development ofWeb2.0 and the widespread use of GPS-equipped mobile smart devices empower people to the use location data from different social networks in various ways, which fosters the emergence of geo-data. For example, people can upload landmark images to Flickr and share their present venue information in Foursquare. Besides the location information, these geo-data are associated with other context information, e.g., the time-stamps and textual meta-data. The large-scale geo-data presents rich content with different modalities and thus serves as a handy resource for various location based applications(e.g. social media organization and recommendation,travel recommendation and media visualization). Therefore, how to effectively process this kind of the geo-data becomes the key problem of location based application. Compared with the traditional multimedia, the geo-data has distinctive characteristics: they are generally associated with the geo-location and presents content with the heterogeneous metadata. Although researchers have done a lot of work in recent years, there are still several key technical issues, such as the fusion of the heterogeneous metadata and the information correlation across different platforms. To cope with the above mentioned issues, we conduct the research on the geo-data processing and applications for social media. The main contributions of this dissertation can be summarized as follows: (1) Scene and viewpoint based summarization for landmarks. Considering the diversity in both scenes and viewpoints, in order to better visually summarize landmarks, we propose a scene-viewpoint based theme model for modeling both scenes and viewpoints. This model is capable of learning the subspace of both the shared scene themes and viewpoint-specific scene-viewpoint themes. We obtain representative images with different scenes and viewpoint via the two kinds of learned subspace. (2) Spatio-temporal theme based landmark analysis. The landmark images from so cial networks are generally associated with other information, such as time and text information. We propose a probabilistic topic model to utilize different multimodal information to learn three kinds of theme subspace, i.e., global themes shared by many landmarks,local themes characterizing local characteristics of one landmark and temporal themes happened at a specific moment for one landmark. In addition, we consider the correlation between the local theme and landmarks, ...
修改评论