An User-Driven Tool for Interactive Retrieval of Non Annotated Videos

A prototype to retrieve videos from non-annotated video databases is proposed. We focus on the problem of retrieving relevant videos from the audiovisual signal when the query is unknown for the system, since it is assumed that most of the available annotations are useless, as it is the case for most of the videos from common users in Internet. The approach presented is defined inside of the on-line learning paradigm where user and system collaborate to improve alternative rankings of the items dataset. The user guides the system in the semantic level and the system tries to adapt the low-level similarity distance between items according to the user preferences. The user interacts with the system until a prefixed number of relevant items is retrieved. The video database is represented as a dense graph where a semi-supervised algorithm is used to propagate the user feeedback.