This document summarizes a team's participation in the MediaEval 2015 Workshop on Retrieving Diverse Social Images. Their approach used a supervised maximal marginal relevance method (sMMR) to jointly optimize relevance and diversity when retrieving images. They tested sMMR using different combinations of visual and textual features. sMMR builds a refined result set incrementally by selecting images that score highest based on relevance to the query and diversity from images already selected. The team trained relevance classifiers using ground truth image labels for queries. Their work was supported by the USEMP project and they provided more details in a poster session.