On Dynamic Scene Geometry for View-invariant Action Matching