Submodular Attribute Selection for Action Recognition in Video