Questions about the warping part of the code

[JerryPW]

Hi ! Recently, I've been using these codes to complement my warping constrains. However, I've found something strange.

batch_patch_gt_color = (
    torch.nn.functional.grid_sample(
        batch_patch_gt_color.permute(0, 3, 1, 2),
        batch_patch_uv.view(batch_patch_uv.shape[0], batch_patch_gt_color.shape[0], -1, 2).permute(1, 0, 2, 3),
        mode="bilinear",
        padding_mode='border'
    )
    .permute(2, 0, 3, 1)
    .contiguous()
    .view(-1, 3)
)

Specifically, as I focus on 1 patch of image and tuning the patch_size into the whole picture, the output, i.e. batch_patch_gt_color, suppose to be exactly the same with input image. Whereas, the output is the input with rotate 90 degree and flipped. Here's the input and the output image:

Do you know why is that happened？ Hoping for your reply soon!

[poptree]

Hi,

It seems like the batch_patch_uv is wrong. batch_patch_uv[...,0] should be the x-axis of the image, which is usually along the width.

Please refer to line 113 of utils.py , where the first entry is width.

BTW, LINE 120 has a similar coordinate error on width and height, so I rewrite a new function get_samples_by_indices_batch instead of using get_samples_batch.

[poptree]

You could find more detail on pytorch grid sample.

Hope it will be helpful.

[poptree]

Besides, you know, some individuals may feel uncomfortable when viewing medical images. While I believe the pictures you posted are synthetic, this won't alleviate the discomfort.

🙏Maybe you could replace it with other images next time. I will close this issue since I cannot stay on it anymore.

[JerryPW]

Hi,

It seems like the batch_patch_uv is wrong. batch_patch_uv[...,0] should be the x-axis of the image, which is usually along the width.

Please refer to line 113 of utils.py , where the first entry is width.

BTW, LINE 120 has a similar coordinate error on width and height, so I rewrite a new function get_samples_by_indices_batch instead of using get_samples_batch.

Thanks a lot !!!!!! I've been working on this problem for all day!!! And I will notice the 'medical image problem' you just mentioned next time! Thanks for remind me that!