Whichever has the most information wins. When the information has structure you can heavily exploit it for generating synthetic data. For this I point you to Ap

Whichever has the most information wins. When the information has structure you can heavily exploit it for generating synthetic data. For this I point you to Apple Sim. It’s a repository of 3D models for interiors. You can generate many layers of information by controlling the renderer and then use it on real photos. That’s done all over images so vectorial spaces are pretty natural for embeddings. You don’t need to add much structure algebraically speaking. If your domain is heavily algebraic, you might even be able to generate correct examples arbitrarily, which is a situation I recommend anyone to be in.