TextonBoost: Joint Appearance, Shape and Context Modeling for Multi-class Object Recognition and Segmentation