Action Recognition From Variable Viewpoints