random crop

[masaisai111]

Why crop the image with noise, but do not crop the reference image when extracting features, which will not result in feature mismatch

random crop

    delta_h = image_tensor.shape[1] - self.size
    delta_w = image_tensor.shape[2] - self.size
    assert not all([delta_h, delta_w])

    if self.center_crop:
        top = delta_h // 2
        left = delta_w // 2
    else:
        top = np.random.randint(0, delta_h + 1)
        left = np.random.randint(0, delta_w + 1)
    image = transforms.functional.crop(
        image_tensor, top=top, left=left, height=self.size, width=self.size
    )

[xiaohu2015]

in fact, we only use center crop during training.

[masaisai111]

There are some differences between the photos obtained after clipping and the feature pictures extracted by clip, for example, the edge area is clipped off. Then why can the photos obtained after clipping be used for noise pictures? Won't there be some information conflict

[xiaohu2015]

what do you mean "clipping"

[masaisai111]

this，If the size of the training data I set is not square, the phenomenon of cropping will occur when training XL, and the edge information of the picture will be cropped，