Pursuing face identity from view-specific representation to view-invariant representation.