View-dependent features for recognition