FDA and random image augmentation methods

[Feobi1999]

Hello, thank you for your outstanding work. I would like to know if you have compared the fda method and the random image augmentation method, because from the visual results, there are some similarities and fda may be a more theoretical image enhancement.

[YanchaoYang]

I have not, but your intuition is correct. As FDA does not require any training, i.e., what FDA needs are two different images, so we can always grab a random image, and use it to augment the other without changing the semantic meaning of the other images. I have seen works in cvpr 2021 extending FDA on image classification, and achieve better performance than sota UDA methods that use strong augmentation.