A semi-supervised approach to monocular depth estimation, depthrefinement, and semantic segmentation of driving scenes using a Siamesetriple decoder architecture