Memory error performing a binary_erosion

[plooney]

Using the following image

http://ubuntuone.com/0unwk6BVwTn5q9zGJR54vq

It's a pickled numpy array. When I use values on the order of a few hundred I get a memory error

[jni]

@plooney please include a minimal code snipped to reproduce the error, and the exact error printout you are getting.

[JDWarner]

I was going to play with this but it appears the file is Pickle protocol 3 thus Python 3.x only, whereas my environments are standardized to Python 2.x.

For a platform- and Python version-agnostic way to share NumPy arrays in the future, consider the *.npy format available through the functions np.save and np.load.

[plooney]

I have output the binary image as @JDWarner suggests. New file here

http://ubuntuone.com/5rToLHpQdlXzHdfOOhD1OK

Minimal code is

im = np.load(file_above)
binary_erosion(im,square(300))

Error is

Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "/usr/lib/python3/dist-packages/scipy/ndimage/morphology.py", line 404, in binary_erosion
output, border_value, origin, 0, brute_force)
  File "/usr/lib/python3/dist-packages/scipy/ndimage/morphology.py", line 260, in _binary_erosion
border_value, origin, invert, cit, 0)
MemoryError

[plooney]

It looks like the error is in scipy.

Using

scipy.ndimage.morphology.binary_erosion

binary_erosion(im,square(300))

I get the same error.

Should the following be equal

binary_erosion(im,square(50),iterations = 6)

and

binary_erosion(im,square(300))

?

The second crashes but the first is ok.

[jni]

Hey @plooney,

Sorry that we have taken so long to get back to you. It looks like your issue got lost in the stack. It also looks like UbuntuOne got shut down, so if you want me to have another look at this you'll have to share your file somewhere else!

As to your questions, you (usually) want an odd side length for your structuring elements, so that they are symmetric about a central pixel. So, binary_erosion(im, square(51), iterations=6) is indeed equal to binary_erosion(im, square(301)). But binary_erosion(im, square(50), iterations=6) will be approximately equal to binary_erosion(im, square(295)), but not exactly because the first call will result in six asymmetrical operations.

How big is your image? I'm not sure why it should result in a memory error...

[jni]

@plooney Could you post your test image somewhere?

[soupault]

Ok. I guess this is the case...
I've taken random large image from the web (http://www.hawaiisummersalsafest.com/wp-content/uploads/2015/11/free-b.jpg (3750x5000x3)), then:

import numpy as np
import skimage
from skimage import io, color, morphology

img = io.imread('free-b.jpg')

img_gray = color.rgb2gray(img)
img_bin = img_gray.copy()
img_bin[img_bin < 0.5] = 0
img_bin[img_bin >= 0.5] = 1

tmp = morphology.binary_erosion(img_bin, np.ones((300, 300), dtype=np.bool))

---------------------------------------------------------------------------
MemoryError                               Traceback (most recent call last)
<ipython-input-20-55b1b7efe510> in <module>()
----> 1 tmp = morphology.binary_erosion(img_bin, np.ones((300, 300), dtype=np.bool))

/usr/lib/python3.5/site-packages/skimage/morphology/misc.py in func_out(image, selem, *args, **kwargs)
35         if selem is None:
36             selem = _default_selem(image.ndim)
---> 37         return func(image, selem=selem, *args, **kwargs)
38 
39     return func_out

/usr/lib/python3.5/site-packages/skimage/morphology/binary.py in binary_erosion(image, selem, out)
40     if out is None:
41         out = np.empty(image.shape, dtype=np.bool)
---> 42     nd.binary_erosion(image, structure=selem, output=out)
43     return out
44 

/usr/lib/python3.5/site-packages/scipy/ndimage/morphology.py in binary_erosion(input, structure, iterations, mask, output, border_value, origin, brute_force)
373     """
    374     return _binary_erosion(input, structure, iterations, mask,
--> 375                            output, border_value, origin, 0, brute_force)
    376 
    377 

/usr/lib/python3.5/site-packages/scipy/ndimage/morphology.py in _binary_erosion(input, structure, iterations, mask, output, border_value, origin, invert, brute_force)
    244     if iterations == 1:
    245         _nd_image.binary_erosion(input, structure, mask, output,
--> 246                                      border_value, origin, invert, cit, 0)
    247         return return_value
    248     elif cit and not brute_force:

MemoryError: 

I suppose, the erosion kernel is too large.

I'd suggest to close the issue and discuss this behaviour with numpy/scipy devs.

[soupault]

@ev-br Hi! Do you think if it is worth reporting to scipy?

[ev-br]

Reporting it on the scipy tracker does no harm, I suppose. The blocker would be that someone actually looks into the ndimage sources and comes up with a fix. Is that you maybe? :-)

[soupault]

Closing as not a bug of skimage.
For further discussion and proposals, please, use scipy/scipy#6042 and scipy/scipy#4878 .