Segmentation masks shift after A.Resize

[kakukakujirori]

🐛 Bug

I noticed A.Resize shifts the alignment of an RGB image and its corresponding segmentation masks, and especially masks are shifted in the right direction.

The same phenomenon also happens if we use A.RandomResizedCrop.

To Reproduce

import requests
from io import BytesIO
from PIL import Image
import numpy as np
import cv2
import matplotlib.pyplot as plt
import albumentations as A

fig, ax = plt.subplots(1, 2, figsize=(10, 5))

# original image
response = requests.get("https://miro.medium.com/max/1200/1*lHMua9ruClZwHR8O0fyO7A.jpeg")
img = np.array(Image.open(BytesIO(response.content)))
img, mask = np.split(img, 2, axis=1)
ret = cv2.addWeighted(img, 0.5, mask, 0.5, 0)
ax[0].imshow(ret, aspect='auto')

# resized image
aug = A.Resize(256, 256, always_apply=True)
transformed = aug(image=img, mask=mask)
img_aug, mask_aug = transformed["image"], transformed["mask"]
ret_aug = cv2.addWeighted(img_aug, 0.5, mask_aug, 0.5, 0)
ax[1].imshow(ret_aug, aspect='auto')

plt.show()

Result: you can see the man's right shoulder is not aligned with the red mask after resize. (Before Resize <===> After Resize)

Expected behavior

This phenomenon doesn't occur if you use cv2.resize or torchvision.transforms.Resize.

import requests
from io import BytesIO
from PIL import Image
import numpy as np
import cv2
import matplotlib.pyplot as plt
import albumentations as A

fig, ax = plt.subplots(1, 2, figsize=(10, 5))

# original image
response = requests.get("https://miro.medium.com/max/1200/1*lHMua9ruClZwHR8O0fyO7A.jpeg")
img = np.array(Image.open(BytesIO(response.content)))
img, mask = np.split(img, 2, axis=1)
ret = cv2.addWeighted(img, 0.5, mask, 0.5, 0)
ax[0].imshow(ret, aspect='auto')

# resized image
img_aug = cv2.resize(img, (256, 256))
mask_aug = cv2.resize(mask, (256, 256))
ret_aug = cv2.addWeighted(img_aug, 0.5, mask_aug, 0.5, 0)
ax[1].imshow(ret_aug, aspect='auto')

plt.show()

Result: (Before Resize <===> After Resize)

Environment

• Albumentations version: 1.3.0
• Python version: 3.10.5
• OS: Linux or Mac
• How you installed albumentations: pip (on top of a conda environment)

[Andredance]

I think it is related to #850.