Automatic CenterCrop to the image smallest size

[nicolas-dufour]

🚀 The feature

The idea here is that centercrop allows for automatic size inference and center_crop to the smallest of HxW.

Motivation, pitch

That allows to keep the most information before resizing the image. We maximize the amount of pixels kept in the center_crop for a given aspect ratio.

Alternatives

No response

Additional context

No response

cc @vfdev-5

[pmeier]

Thanks @nicolas-dufour for the request. Could you elaborate what you want? It is not clear to me from the description you gave above. What would the function signature look like? Could you give a few examples of input images and parameters?

[NicolasHug]

Hi @nicolas-dufour , nice to see you around!

To clarify the feature request, you're looking for something like CenterCrop(size="shortest_edge") such that when passing an image whose size is (100, 50) or (50, 100), we end up with (50, 50) in both cases? Is that correct?

[nicolas-dufour]

Hi @NicolasHug and @pmeier.
Yes the shortest_edge solution is basically what i had in mind for square center crops. For rectangular ones, i guess it would depend on the aspect ratio.

[nicolas-dufour]

Hey @NicolasHug ,
Here is the code i had in mind for this:

import torch.nn.functional as F
import torchvision
import torch

class CenterCrop(torch.nn.Module):
    def __init__(self, size=None, ratio="1:1"):
        super().__init__()
        self.size = size
        self.ratio = ratio
    def forward(self, img):
        """
        Args:
            img (PIL Image or Tensor): Image to be cropped.

        Returns:
            PIL Image or Tensor: Cropped image.
        """
        if self.size is None:
            if isinstance(img, torch.Tensor):
                h, w = img.shape[-2:]
            else:
                w, h = img.size
            ratio = self.ratio.split(":")
            ratio = float(ratio[0]) / float(ratio[1])
            # Size must match the ratio while cropping to the edge of the image
            print(ratio, w, h)
            ratioed_w = int(h * ratio)
            ratioed_h = int(w / ratio)
            if w>=h:
                if ratioed_h <= h:
                    size = (ratioed_h, w)
                else:
                    size = (h, ratioed_w)
            else:
                if ratioed_w <= w:
                    size = (h, ratioed_w)
                else:
                    size = (ratioed_h, w)
        else:
            size = self.size
        return torchvision.transforms.functional.center_crop(img, size)

    def __repr__(self) -> str:
        return f"{self.__class__.__name__}(size={self.size})"