Extra keypoints when using Rotate with keypoints

[Chxresubles]

Describe the bug

Using A.Rotate in A.Compose with remove_invisible=False returns 9 times more keypoints than the input number of keypoints.

To Reproduce

Steps to reproduce the behavior:

import numpy as np
import albumentations as A

image = np.random.randint(0, 256, (100, 100, 3), dtype=np.uint8)
keypoints = np.random.randint(0, 100, (50, 2), dtype=np.uint8)

# Check shape
print(keypoints.shape)  # (50, 2)

result = A.Compose([A.Rotate(limit=45, p=1.0)], keypoint_params=A.KeypointParams(format='yx', remove_invisible=False))(image=image, keypoints=keypoints)

# Check shape
print(result['keypoints'].shape)  # (450, 2)

Expected behavior

result['keypoints'] should have the same shape as the input keypoints. (50, 2)

Actual behavior

result['keypoints'] has 9 times more keypoints than the input keypoints. (9 * 50, 2)

Additional context

The bug was not present in albumentations==1.4.18 and was introduced since albumentations==1.4.19

[ternaus]

It is not a bug, but a feature.

When you use border_mode=cv2.BORDER_REFLECT_101 which is the default, you get new keypoints in the reflected regions.

If you change border_mode to cv2.BORDER_CONSTANT, no new keypoints will be added.