Domain randomization (https://arxiv.org/abs/1703.06907) is used to create a synthetic dataset with enough variance that it will encompass unseen real data, as just one variation.

I am trying to understand how this is different from applying data augmentation techniques to a synthetically generated dataset.

  • $\begingroup$ "domain randomization" strikes me as a special kind of data augmentation. A kind of difference is that data augmentation seeks to expand the training data to include invariances you expect/know are there. Where the domain randomization approach kind-of "overwhelms" the network with all kinds of "irrelevant" variability that it's "forced" (hopefully) to learn the relevant things. In that paper, how to detect objects. $\endgroup$
    – bogovicj
    Jun 16 '20 at 17:11

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.