Reversible Resize

[Dipet]

I tried to implement the interface for Test Time Augmentations (TTA) using ReplayCompose API.

If some transform could support reverse operation we must implement these functions:

• get_reverse_args - to save all needed parameters to do reverse transformation
• reverse_image - to reverse images and masks
• reverse_bbox - to reverse bboxes
• reverse_keypoint - to reverse keypoints.

Example:

import numpy as np
import albumentations as A

t = A.ReversibleCompose(
    [A.Resize(512, 512, p=1)], bbox_params=A.BboxParams("pascal_voc"), keypoint_params=A.KeypointParams("xy")
)
img = np.empty([100, 100, 3], dtype=np.uint8)
bboxes = [(0, 0, 10, 10, 1)]
keypoints = [(10, 20)]

res = t(image=img, bboxes=bboxes, keypoints=keypoints, test="test")
res = t.reverse(**res)

assert np.allclose(np.array(bboxes), np.array(res["bboxes"]))
assert np.allclose(np.array(keypoints), np.array(res["keypoints"]))

[glenn-jocher]

@Dipet thanks for the PR! This would be super useful.

For the TTA use case we do not need to reverse the images, only the boxes/keypoints. What's the use case for reversing image transforms? Aren't many image transforms non-reversible in any case?

[Dipet]

At the moment, this is just a blank.

For the TTA use case we do not need to reverse the images, only the boxes/keypoints. What's the use case for reversing image transforms? Aren't many image transforms non-reversible in any case?

We usually use the same function for segmentation masks, this function is only needed for them.

[glenn-jocher]

@Dipet oh got it, ok! The main thing i was worried about was the extra cost in reversing a full image transform, but if only the keypoints are reversed and not the full images that's perfect.

[Dipet]

@Dipet oh got it, ok! The main thing i was worried about was the extra cost in reversing a full image transform, but if only the keypoints are reversed and not the full images that's perfect.

Yes, only provided data will be reversed. Perhaps I will remove the reverse_image and leave only the reverse_mask.

[agporto]

Is this still planned? It would super useful to have this feature.

[hsuominen]

+1, this would be great to have