MedianBlur randomly fails with OpenCV error

This issue is a bit mysterious from my point of view, as it’s kind of complicated to reproduce it a single try.

In [1]: from albumentations import MedianBlur
   ...: import numpy as np
   ...:
   ...: success, fail = 0, 0
   ...: for _ in range(100):
   ...:     img = np.random.rand(100, 100)
   ...:     try:
   ...:         img = MedianBlur()(image=img).get('image')
   ...:         success += 1
   ...:     except Exception:
   ...:         fail += 1
   ...: print(success, fail)
   ...:
OpenCV Error: Assertion failed (src.depth() == CV_8U && (cn == 1 || cn == 3 || cn == 4)) in medianBlur, file /Users/travis/build/skvark/opencv-python/opencv/modules/imgproc/src/smooth.cpp, line 3538
OpenCV Error: Unsupported format or combination of formats () in medianBlur, file /Users/travis/build/skvark/opencv-python/opencv/modules/imgproc/src/smooth.cpp, line 3529
OpenCV Error: Unsupported format or combination of formats () in medianBlur, file 

// some errors removed from the log

48 52

I’ve reproduced this stochastic issue on two different computers, both using opencv-python==3.3.0.10.

[creafz]

I think there are 2 reasons why those errors appear:

1. np.random.rand generates ndarrays with the float64 data type. This data type is not supported by cv2.medianBlur which is used internally in MedianBlur.
2. If the input type is not np.uint8, the only allowed ksize values for cv2.medianBlur are 3 and 5. Link to the OpenCV docs

This code that uses float32 inputs and ksize values 3 and 5 work without errors for me:

from albumentations import MedianBlur
import numpy as np
success, fail = 0, 0
for _ in range(100):
    img = np.random.rand(100, 100)
    img = img.astype('float32') # Cast a float64 image to float32
    try:
        # Use blur_limit 5 instead of default 7. So this augmentation will
        # apply `cv2.medianBlur` using ksize=3 or ksize=5.
        img = MedianBlur(blur_limit=5)(image=img).get('image')
        success += 1
    except Exception:
        fail += 1
print(success, fail)

I am afraid as long as we are using OpenCV functions there is nothing we can do to support float64 inputs and ksize values larger than 5.

Maybe we should check the data type of input images in all our augmentations and raise an exception if the input data type is not supported, so users won’t get those cryptic OpenCV errors? @albu @ternaus what do you think?

[arsenyinfo]

@creafz may be, it can be replaced with cv2.filter2D with a manually created filter?