Why do we need expand2sqaure in scorers?

[Hoi2022]

Hello, thanks for your great works.

However, I find that the preprocess inside this looks weird.

The input image is first expanded to a square one, with extra mean-color margins by applying the expand2sqaure function.

I visualized a preprocessed image, which looks like this:

while the original image is:

My teammate also reports that larger aspect ratio results in lower aesthetic score. So we are wondering if this preprocess is correct.

[CarlCloudWang]

In my survey, I found that no expandsqure can promote the capicity of whole model...
u can only use preprocessing from CLIP_ImageProcessor, and this one can cause real improvement.

[elvindp]

In my survey, I found that no expandsqure can promote the capicity of whole model... u can only use preprocessing from CLIP_ImageProcessor, and this one can cause real improvement.

hi, without expandsqure, do you directly resize to 224? I wonder whether directly resize would defect the original image.

[Hoi2022]

In my survey, I found that no expandsqure can promote the capicity of whole model... u can only use preprocessing from CLIP_ImageProcessor, and this one can cause real improvement.

hi, without expandsqure, do you directly resize to 224? I wonder whether directly resize would defect the original image.

CLIPImageProcessor includes a CenterCrop and a resize, rather than directly resizing the image to a square shape.

[FengqiZhang0329]

In my sight, I concern the padding part may infect the image quality in some way.