Reconfigurable models for scene recognition