'subpixel' boundaries at the edges of an image

[firerain42]

Description

When using the function find_boundaries with mode='subpixel' the behavior at the edges of the image seems strange to me.

From the documentation it is not clear to me what the intended behavior should be. Without an option to explicitly specify the behavior at the edges of the image, I would expect a constant input image to have no boundary at all or maybe straight lines at the edges of the image. However, for constant images with a non-zero value the actual image alternates between True and False. If the image is completely zero, there is no such effect.

Way to reproduce

import skimage.segmentation
import numpy as np
skimage.segmentation.find_boundaries(np.ones((3,3), dtype=int), mode='subpixel')

returns

array([[False,  True, False,  True, False],
       [ True, False, False, False,  True],
       [False, False, False, False, False],
       [ True, False, False, False,  True],
       [False,  True, False,  True, False]])

I would expect the same output as np.zeros((5,5), dtype=bool).

Version information

from __future__ import print_function
import sys; print(sys.version)
import platform; print(platform.platform())
import skimage; print("scikit-image version: {}".format(skimage.__version__))
import numpy; print("numpy version: {}".format(numpy.__version__))

3.7.10 | packaged by conda-forge | (default, Feb 19 2021, 16:07:37)
[GCC 9.3.0]
Linux-5.4.0-74-generic-x86_64-with-debian-bullseye-sid
scikit-image version: 0.18.1
numpy version: 1.20.3

[mkcor]

Hello @firerain42,

Thanks for reporting! I agree with you, the current behaviour is not consistent.

Would you like to try and investigate by yourself?

scikit-image/skimage/segmentation/boundaries.py

Line 11 in a1c5846

def _find_boundaries_subpixel(label_img):

You could run this function line by line, starting with label_img = np.ones((3, 3), dtype=int) and comparing with label_img = np.zeros((3, 3), dtype=int) to see what happens. Let me know, thanks!

[firerain42]

Hello @mkcor,

I think the problem is the padding in

scikit-image/skimage/segmentation/boundaries.py

Lines 37 to 39 in 23fd40b

	windows = view_as_windows(np.pad(label_img_expanded, 1,
	mode='constant', constant_values=0),
	(3,) * ndim)

which introduces 0 as an image value. Replacing it with np.pad(label_img_expanded, 1, mode='edge') should do the trick. I would have sent a pull request, but I couldn't get the tests running on my PC (and don't have the time to find the cause).