the fact that there is not enough 3D data to train these new models on to begin with. Images, text, or videos are abundant, the researchers explain, and they can be used to train "traditional ...