FoP: Never-Ending Learner for Multimedia Knowledge Extraction

In this paper we present our system Faces of Politics (henceforth FoP), that is able to continuously learn multimedia knowledge of Web multimedia resources about the presence of person(s) in a pictures and to leverage this knowledge to the Linked Open Data cloud (LOD-cloud). FoP promotes both scalability of the data lift process for this domain and a structured knowledge representation for complex queries. The system was boots traped using Freebase data about politicians and their pictures, and we show that the model provides a good generalization with an error rate below 7%. Meantime, FoP not only relates a person to a multimedia resource, but it also detects and publishes metadata on the position of the person in the picture. Moreover, it supports the presence of several persons in the picture. At this step, FoP is also giving data in return to the LoD cloud that fed him in the first place: it leverages Linked Data on people recognized in these pictures, and on which rectangle area. This allows fine-grained queries like creating a curation of documents in which a person is depicted relatively to another for instance. On a technical point-of-view, we also provide a Website for browsing FoP knowledge base as Web users, and we also offer a public SPARQL endpoint for robots or other Web applications.