That would not seem computationally efficient if for every class (i.e. object) you were to create a 3D modeling and paste it into a training set