RGB-D joint modelling with scene geometric information for indoor semantic segmentation