Weiqing Min , Institute of Automation, CAS, Beijing
Bing-Kun Bao , Institute of Automation, CAS, Beijing
Changsheng Xu , Institute of Automation, CAS China-Singapore Institute of Digital Media, Beijing Singapore
DOI Bookmark: http://doi.ieeecomputersociety.org/10.1109/MMUL.2014.1
In this paper, we propose a topic model, namely multi-modal Spatio-Temporal Theme Modeling (mmSTTM), which enables the differentiation among three kinds of landmark themes, i.e., temporal themes happened at a specific moment, local themes characterizing local characteristics of landmarks and general themes shared by most of all the landmarks. Furthermore, mmSTTM can incorporate both textual and visual information to explore multimedia data for representing themes. Based on the model, we formulate a framework, which consists of three modules: data preparation, theme modeling and theme analysis. The landmark photos are firstly downloaded from Flickr, and pre-processed in the module of data preparation. Secondly, the proposed mmSTTM is used to discover general, local and temporal themes in theme modeling. Lastly, we analyze the discovered themes including landmark-specific time distribution of themes and time-specific location distribution of themes.
Weiqing Min, Bing-Kun Bao, Changsheng Xu, "Multi-modal Spatio-Temporal Theme Modeling for Landmark Analysis", IEEE MultiMedia, , no. 1, pp. 1, PrePrints PrePrints, doi:10.1109/MMUL.2014.1