Weakly-Supervised Cross-Domain Dictionary Learning for Visual Recognition