Subspace Representations And Learning For Visual Recognition