RandomFog behaves differently in version 2.0.4 compared to previous versions

[huuquan1994]

Describe the bug

The result of RandomFog in version 1.4.6 is completely different from the newest version 2.0.4

To Reproduce

Steps to reproduce the behavior:

1. Environment (e.g., OS, Python version, Albumentations version, etc.)
   OS: MacOS Sonoma 14.2 (23C64)
   Python: 3.10.16
   Albumentations version: 1.4.6 and 2.0.4

2. Sample code that produces the bug.

import albumentations as A
from PIL import Image
import numpy as np

# read image
img_path = 'path/to/image'
pil_image = Image.open(img_path).convert('RGB')

# apply data aug
# for version 1.4.6
_augs_fnc = A.RandomFog(alpha_coef=0.08, fog_coef_lower=0.6, fog_coef_upper=1.0, p=1.0)

# for version 2.0.4
# _augs_fnc = A.RandomFog(alpha_coef=0.08, fog_coef_range=(0.6, 1.0), p=1.0,)

img_np = np.array(pil_image)
augmented = _augs_fnc(image=img_np)
aug_image = Image.fromarray(augmented['image'])

3. Any error messages or incorrect outputs.
   None

Expected behavior

I expect both results should look the same or very similar

Actual behavior

The two results are completely different and RandomFog in version 2.0.4 are not really foggy

Screenshots

Load image

Version 1.4.6

Version 2.0.4

[ternaus]

I changed a few things under the good there, but it should not create such a difference.

Will take a look.

[ternaus]

Reverted the algorithm back. Should work correctly now

[huuquan1994]

@ternaus
Thanks for the quick response!
I've checked the latest version. The result looks a bit better now but not quite close to version 1.4.6

If we decrease the alpha_coef, it's similar to Gaussian Blur but if we increase it, it shows noticeable "foggy circles" which looks unnatural compared to version 1.4.6.

Could you take a look again?

[ternaus]

Ok, will take a look