Skip to Main Content
The goal of this work is to automatically detect frequently occurring groups of media in a user's collection that have a unifying theme. These groups provide a narrative structure that ties in images that are temporally far apart and cannot be browsed easily. The media in the collection is analyzed by a variety of algorithms to generate metadata of different types. The media and associated metadata are represented as a transactional database, and frequent item set mining is employed to detect frequently occurring groups of images that share several metadata in common. It is expected that a user's primary picture-taking interests (e.g., baby, garden, school sports, etc.), will appear as groups based on some combination of underlying metadata. A confidence and interest measure relevant to the consumer domain is used to determine the quality of the frequent item sets and create a list of the top "themes" within the collection. We also detect annually recurring groups in multi-year collections, as these capture common themes such as birthdays and holidays. Because the detected recurring groups are strictly data-driven (with no a priori assumptions about a user's collection), they are customized to the type of content in specific user's collections. Experiments with large user collections show the usefulness of our approach.