In this paper, we describe our approach and its results as part of the MediaEval 2013 Retrieving Diverse Social Images Task evaluation. We illustrate a content - based technique relying on a single type of visual descriptors that makes it possible to identify groups of similar instances of a given landmark and select the most representative images from each such group out of a set of relatively noisy or redundant images. This method builds for each landmark a matching graph through quantized interest points similarities and identifies groups of similar instances of the landmark as connected components. In this approach we do not make use of the textual metadata of the images or any other external source of information.