From this paper in section 4.1:
To reduce the risk of overfitting, we performed data augmentation by randomly translating and rotating character images. We also created new classes through 90◦, 180◦ and 270◦ rotations of existing data.
I looked through 2 papers after this one came out and they also say that they rotate images in this same manner (although not explicitly stated in those papers, if I understand correctly, they didn't bother translating images because they use convolutional networks that are translation-invariant)
How does rotating images prevent overfitting in a few-shot classifier, and why choose (90, 180, 270) degree rotations? Why would those specifically count as a "new class?"