Instance_augmentation on different image domains.

[pILLOW-1]

Hi, great work!
The core idea in the paper is to use large-scale pre-trained generative models to regenerate the object in an image. However, in other image domains such as x-ray security images, there is no large-scale data to train a large generative model. So, how can i borrow the idea presented in your paper to implement instance augmentation on the data-insufficient image domains like x-ray?
Looking forward to your answers!

[sidgairo18]

@pILLOW-1 not sure if this would work for very specific cases like x-ray images (where the diversity in images is very specific and limited). Imho if this does have to work, potentially one would first have to train a Text2Image generative model to be able to have such augmentations.

Additionally, with medical image datasets one has to be very careful, since different artifacts introduced during training might actually corrupt the data. Thus, I would be wary before using this in sensitive settings such as medical data.

@KupynOrest could have more insights here :)

Best,
Sid