Aligning where to see and what to tell: image caption with region-basedattention and scene factorization