For machine vision problems, synthetic data with domain randomisation definitely works, provided, that is, you are able to generate a large enough number of different modes of variation.

In fact, it works so well it feels like cheating.