Revisiting representation models for visual recognition: Social and Cultural Perception

Funder: MINISTERIO DE CIENCIA, INNOVACIÓN Y UNIVERSIDADES

Call: Subprograma de Proyectos de Investigación Fundamental no orientada (IFNO 2012) (MINISTERIO DE CIENCIA, INNOVACIÓN Y UNIVERSIDADES)

Of National scope.

Despite recent advances, visual recognition of objects, scenes and events still remains one of the most challenging tasks of computer vision. From an historical perspective, computer vision research in this topic has been developing along two complementary directions: representation models and machine learning methods. Representation models of visual data were the focus of research during the early days of computer vision but faded away twenty years ago because of the practical success of machine learning methods applied to simple image representations. This learning-based approach has been hegemonic during the last years in most of the conference and journal publications but it is getting into a maturity state where a disruptive advance seems very unlikely. The aim of this three-year project is to shift the focus, again, on the representation problem by offering a new perspective to an old problem. Existing representation schemes are mainly based on local image appearance. This represents a serious drawback, since the interpretation we can perform on these data is very limited. This time however, we plan to add a new dimension to these models which will include a mid-level layer to cope with the large variability in appearance and structure which is common to a lot of visual objects and events. Through this approach, we expect to witness an improvement in task performance and goal achievement. Our solution builds on previous knowledge and experience on local appearance/shape modeling, machine learning and their application to large scale visual databases.