This is a brilliant idea, IMO, and a real game changer (in principle, not necessarily tied to the implementation). Without this, CV problems are largely only li

This is a brilliant idea, IMO, and a real game changer (in principle, not necessarily tied to the implementation). Without this, CV problems are largely only limited by how much data you can get. If you can simulate arbitrary scene configurations it's now possible to solve almost any non-human CV problem, even ones with sparse data. For example, in security applications you have basically zero data on the sorts of anomalous events you want to trigger on. If you can simulate example cases of those things that means you can actually start applying supervised machine learning.