How INRIA/IRISA identifies Geographic Location of Videos

In this paper, we describe our approaches and their results as part of the MediaEval 2012 Placing Task evaluation. We present two different tag-based techniques. Both first preselect one or several geographic area of interest and then perform a deeper analysis inside the selected area(s) to return the coordinates more likely to be related with the input tags. In addition, we also implement a content-based method that uses aggregated local images descriptors (VLAD) to find the video’s visual nearest neighbors and infer its coordinates. In this work we do not use gazetteers or any other external information.