Image segmentation 'cropping after reszing' vs 'resizing after cropping'

vision
1

Hi, although it’s not a topic about pytorch, I wrote to get your help.
I am studying image segmentation and I have to do cropping images and do resizing.

I have two datasets which have different image size. The reason I concern the order of these functions is because of ‘scale problem’

If cropping is done prior to resizing, I think the scale problem will be caused.

For example, I have two datasets that have the sizes ‘300x300’ and ‘600x600’ then, I crop these images with size of ‘256x256’. In the former case, the ratio is 300/256 and latter case, the ratio is 600/256. I think that it will make degrading performance.

If you have some tips about this case. please give me some advice thanks!

2

I would also share your concern.
However, given that torchvision.transforms.RandomResizedCrop uses a scale of [0.08, 1] and seems to work fine, I would recommend to just try out both approaches.
If I’m not mistaken, these settings were used in the original Inception model.