Hello community. I have been googling for long time but I didn’t find any clear answer.
I have big images in 1200x1200 and I need to resize them to 288x288. What do you recommend me in order to (1) best quality and (2) best quality-time balance?
As far as I Know, in this cases people usually uses Image.Antialias in torchvision.transform.Resize. I saw that Image.Antialias was changed by Lancoz (supported in torchvision.transforms.InterpolationMode)
So I think this topic is between lanczos and bicubic?
Thanks all !
EDIT1: In my application, I’m using a convolution model that usually need to be sensible to medium and small details.
EDIT2: Currently I’m using the default InterpolationMode.BILINEAR
T.Resize(288),
T.RandomHorizontalFlip(0.5),
T.CenterCrop(256),
T.ToTensor(),
T.Normalize([.485, .456, .406], [.229, .224, .225])