Unsupervised domain adaptation for semantic segmentation of urban street scenes reflected by convex mirrors