Visual Understanding via Multi-Feature Shared Learning With Global Consistency