Transductive Zero-Shot Learning with Visual Structure Constraint