Efficient Spatialtemporal Context Modeling for Action Recognition