Data-driven visual similarity for cross-domain image matching