Fix glass_blur for float dtype

[i-aki-y]

I found a bug about the GlassBlur (not GausssianBlur!) transform.
In this PR, I tried to fix the problem.
I would appreciate it if this PR is reviewed.

Problem

The GlassBlur transform returns wrong values when an input image's dtype is not uint8. This is caused by misimplementation of dtype conversion in the glass_blur function.

The following code is a reproducible example.
Note, the result of float32 case is expected to be non-zero.

import numpy as np
import albumentations

x_uint8 = np.zeros((5, 5)).astype(np.uint8)
x_uint8[2, 2] = 255

x_float32 = np.zeros((5, 5)).astype(np.float32)
x_float32[2, 2] = 1.0

glassblur = albumentations.augmentations.transforms.GlassBlur(always_apply=True, max_delta=1)

np.random.seed(13)
print(glassblur(image=x_uint8)["image"])

# [[ 6 10 10  3  1]
#  [14 23 18  7  5]
#  [25 36 20 11 14]
#  [14 22 14  9 10]
#  [ 6 11 14  7  5]]

np.random.seed(13)
print(glassblur(image=x_float32)["image"] * 255)

# [[0. 0. 0. 0. 0.]
# [0. 0. 0. 0. 0.]
# [0. 0. 0. 0. 0.]
# [0. 0. 0. 0. 0.]
# [0. 0. 0. 0. 0.]]

Fix

I change the following two lines.

1. The output of the cv2.GaussianBlur should be multiplied by 255 (max uint8) instead of coef (max of input image dtype).

+ x = np.uint8(cv2.GaussianBlur(np.array(img) / coef, sigmaX=sigma, ksize=(0, 0)) * coef)
- x = np.uint8(cv2.GaussianBlur(np.array(img) / coef, sigmaX=sigma, ksize=(0, 0)) * 255)

2. The transformed x values whose dtype is uint8 should be divided by 255 instead of coef before passed to the cv2.GaussianBlur.

+ return np.clip(cv2.GaussianBlur(x / coef, sigmaX=sigma, ksize=(0, 0)), 0, 1) * coef
- return np.clip(cv2.GaussianBlur(x / 255, sigmaX=sigma, ksize=(0, 0)), 0, 1) * coef

After the fix, I got the following results.

import numpy as np
import albumentations

x_uint8 = np.zeros((5, 5)).astype(np.uint8)
x_uint8[2, 2] = 255

x_float32 = np.zeros((5, 5)).astype(np.float32)
x_float32[2, 2] = 1.0

glassblur = albumentations.augmentations.transforms.GlassBlur(always_apply=True, max_delta=1)

np.random.seed(13)
print(glassblur(image=x_uint8)["image"])

# [[ 6 10 10  3  1]
#  [14 23 18  7  5]
#  [25 36 20 11 14]
#  [14 22 14  9 10]
#  [ 6 11 14  7  5]]


np.random.seed(13)
print(glassblur(image=x_float32)["image"] * 255)

# [[ 6.2885346 10.978549  10.776732   3.9743576  1.4602503]
#  [14.964603  23.983921  18.80205    7.6914196  5.5994816]
#  [25.221247  36.9323    20.978346  11.711028  14.315237 ]
#  [14.775828  22.008005  14.221998   9.483424  10.28976  ]
#  [ 6.3716593 11.8987875 14.053254   7.8504953  5.2329097]]

[Dipet]

Thank you very much! These are important fixes.
But could you a little bit rewrite your PR?

[Dipet]

Better to simple remove conversion into uint8 and convert to image data type

x = cv2.GaussianBlur(np.array(img) / coef, sigmaX=sigma, ksize=(0, 0)) * coef
x = x.astype(img.dtype)

[Dipet]

@creafz @BloodAxe @ternaus Here is a little bit strange implementation. Maybe we will remove normalization before applying blur? After these changes transform will work fine, but the result may differ from this implementation by 2 abs(old_result - new_result).max() <= 2

[creafz]

@Dipet, I agree with removing this normalization. I assume that it was used in the original implementation of glass_blur (https://github.com/hendrycks/robustness/blob/master/ImageNet-C/create_c/make_imagenet_c.py#L300) because skimage works mostly with float images that have values in the range [0.0, 1.0].

[i-aki-y]

@Dipet Thank you for your quick reply.

But could you a little bit rewrite your PR?

Of course! Now, I updated my PR.

@creafz @BloodAxe @ternaus Here is a little bit strange implementation. Maybe we will remove normalization before applying blur? After these changes transform will work fine, but the result may differ from this implementation by 2 abs(old_result - new_result).max() <= 2

I modified the test case to allow for the difference between uint8 and float at most 2.

If any fix is needed about this issue, please let me know.

[Dipet]

Can you remove normalization by coef?
With these changes we also don't need normalize x in the and and clip it

[i-aki-y]

@Dipet OK, I see your point.

Can I ask for some advice about the test case?
In the first PR version, I added a test case to check the transformed values of uint8, float32 images are approximately the same.
My intention against the test case is to confirm some dtype operations are done appropriately.
Thanks to your review, now the code has no explicit dtype operations, so I think the test case becomes less effective.
Is it better to remove my test case?
If there is any idea about this issue’s test case, let me know?

[i-aki-y]

Fixed.

[Dipet]

I think it is good that we have this test.
In my opinion all transforms in the library must have similar tests to check that transforms work the same for uint8 and float datatypes.

[i-aki-y]

OK, I see. Thanks for the comment.