'slic' crash with spacing

[Borda]

sigma division does not seems correct since sigma is a number and spacing is array_like

to Reproduce:

import numpy as np
from skimage import segmentation, io

image = io.imread('drosophila.png')

image = image[:,:,1]
image = image / float(np.max(image))
print image.shape
seg = segmentation.slic(image, n_segments=500, compactness=0.2, multichannel=False, spacing=(1.,1.)) 

[jni]

@Borda, the error message is obscure, I'll grant, =), but the docstring does specify that spacing must be length 3!

We don't currently support spacing in 2D images because it's rare to have different voxel spacing in just 2D images. (I have never come across this.) Do you have a use-case for 2D spacing? If so, I can add support for that. If not, I'll add a ValueError that explains what went wrong better than the current message.

Thanks for the report!

[Borda]

Well, I need segment 3D image but I wanted to start with simpler case (2D) a and give a try to the parameters I will need for 3D such as spacing... btw, what is the explanation of these numbers, 1 mean that distance between neighboring positions is 1? Thanks

[jni]

@Borda the spacing is meant to be relative. (5, 1, 1) means that the distance between adjacent voxels in the leading dimension (e.g. planes) is 5 times greater than the distance between voxels in the row / column dimensions.

[stefanv]

@jni How about a sentence in the user guide summarizing your comments here? We can perhaps also show how to address @Borda's user case (perhaps it requires adding a single dimension to make a 2D slice 3D).

[JDWarner]

I'd argue that we should be as general as possible when we can be. 2D anisotropic is a strange use case, but I think it's entirely possible to relax the length-3 requirement in favor of an iterable of length image.ndim. In fact, I'd recommend any algorithm which accepts a spacing kwarg should behave in that manner.

[Borda]

I can provide even code for 3D with interactive visualization in iPython if you like:

import libtiff
from skimage import segmentation

im = libtiff.TiffFile('drosophila.tiff') .get_tiff_array() 
img = np.empty(im.shape)
for i in range(img.shape[0]) :
    img[i,:,:] = np.array(im[i])
img = img / float(np.max(img))

seg = segmentation.slic(img, n_segments=5000, compactness=0.2, multichannel=False, spacing=(5.,1.,1.))

def showImageSegment(z):
    plt.figure(figsize=(12,8))
    _= plt.subplot(1,2,1), plt.imshow(img[z], cmap=plt.cm.gray), plt.axis('off')
    _= plt.subplot(1,2,2), plt.imshow(seg[z]), plt.axis('off')
    display()

interact( showImageSegment, z=widgets.IntSliderWidget(min=0,max=im.shape[0]-1,step=1,value=im.shape[0]/2) )

[jni]

@JDWarner I agree, but it's a bit finicky to deal with all the possible cases (multichannel = True/False, sigma being a scalar or an array-like). A quick glance at the code didn't suggest an immediate solution. PRs welcome, as always! ;)

Happy to keep this open as a reminder to do that?

[stefanv]

/ping @jni @JDWarner

Can we close this with an addition to the docstring, or is something more involved needed?

[emmanuelle]

A case for 2-D anisotropic images is the case of spatio-temporal diagrams (used for example in chaos and non-linear systems).

[Borda]

any update here? :)

[JDWarner]

Stand by my earlier comment that we should generalize spacing, here and elsewhere, to match the spatial dimensionality of the image.

I won't be able to dive into the code this week as I am on 12 hour overnight call.

[hmaarrfk]

i gave the code a quick scan, i think the real challenge is coming up with tests. PRs would help :D.

[jni]

I agree with @JDWarner, just haven't had the bandwidth to deal with this.

@Borda 👋 😊 maybe you'd like to tackle this? imho the right way to do it would be to normalise both spacing and sigma to be the same size array early on in the function.

[Borda]

@jni not sure what you meant by "tackle this"? 😃

[rfezzani]

• spacing can be a 1-element array like, or a 3-element array like, it crashes otherwise.
• sigma can only be a scalar, a 1-element array like or a 3-element array like.

We need to start a deprecation cycle if we want to Change this behavior as it will introduce breaking changes.