img has been cut off abnormally

[TITC]

assume We have an image, with a size of (352,31), and the min_dimensions is (32,32). Then it's qualified any judgment. Image (352,31) is being pasted to a new blank Image whose size is (32,32).

        if any([s < min_dimensions[i] for i, s in enumerate(img.size)]):
            padded_im = Image.new('L', min_dimensions, 255)
            padded_im.paste(img, img.getbbox())
            img = padded_im

[TITC]

Here is the picture we have.

and output as beow

[TITC]

And I wonder why we need the minmax_size function? Does this is used to prevent downsample process dimension not meeting the requirements?

[lukas-blecher]

Thank you very much!
The idea behind the minmax_size function is that the ViT can accept only a very discrete amount of image sizes. The formula you showed above is for CNNs right?
That's what pad is for but it doesn't cover the edge case of a too large image.
Come to think about it the function would be fine without any of the min_dimensions stuff.

[TITC]

Thanks for the explanation, seems I need to reread ViT's paper carefully and its GitHub repository.