How do I fill the black or "blank" corners after I use rotation data augmentation?

1 
transform = transforms.Compose([
        transforms.Resize(64),
        transforms.RandomRotation(360, resample=Image.BICUBIC),
        transforms.ToTensor(),
        transforms.Normalize(mean=(0.5, 0.5, 0.5), std=(0.5, 0.5, 0.5))
])

After I call RandomRotation I get this image

Screenshot_43

Its screwing up the GAN I am trying to train.

Screenshot_44

Whats the easiest way to fill the corners white?

2 Likes
2

some fill with nearest pixels, but that may be hard to implement. alternatively, you can increase the size (e.g., by using nearest pixels to fill in blank areas. this is easy because it is shifting in axes-aligned directions), and then rotate.

3

So let me get this straight I increase the white border size but not the content so when I rotate/crop I would still get to keep the corners? If so how do I increase the size is there a direct function you can call in pytorch?

4

Ideally you would want to do that in dataset.__getitem__, where you can operate on np arrays and/or PIL images. There are plenty of useful functions there.

1 Like
5

Your images seem to have a white background. In this case you can set the fill parameter of torchvision.transforms.RandomRotation with transforms.RandomRotation(360, fill=255) in order to fill the black corners. For more information please have a look at the official documentation: RandomRotation — Torchvision main documentation

I hope this helps.
All the best,
Stefan

1 Like