threshold_li gets stuck in infinite loop due to value oscillation

[janniswi00]

Description:

When applying skimage.filters.threshold_li to a specific float32 image, the algorithm enters an infinite loop. The threshold value oscillates between two close values and never converges.

This happens in a scientific image processing pipeline, even without saving/loading from file. I saved the problematic image directly to verify this behavior.

Expected Behavior: threshold_li should converge to a stable threshold value or raise an error if convergence is not possible.

Actual Behavior: The threshold alternates indefinitely between two very close float values (0.018132983 ↔ 0.018132977) without satisfying the convergence condition.

This behavior also appears in scikit-image version 0.25.2. In specific cases this problem can be solved by manually setting the tolerance.

density_map_bug.zip

Way to reproduce:

import numpy as np
from skimage.filters import threshold_li
import zipfile

with zipfile.ZipFile("density_map_bug.zip", 'r') as zip_ref:
    zip_ref.extractall()

density_map = np.load("density_map_bug.npy")
threshold_li(density_map, iter_callback=print) 

Version information:

3.9.21 (main, Dec 11 2024, 16:24:11) 
[GCC 11.2.0]
Linux-6.8.0-58-generic-x86_64-with-glibc2.39
scikit-image version: 0.24.0
numpy version: 1.26.4

[jni]

Yikes. That's amazing. 😂 Thanks for the report @janniswi00!

My first instinct is that we could check against seen values, or maybe even just the most recent value, and then terminate.

I don't know whether this is true but my intuition is that you couldn't have longer loops than two values, so checking against the last seen value is probably sufficient.

[janniswi00]

Yes, my intuition is the same, there shouldn't be a loop with more than two values.

[lagru]

Hey @janniswi00 thanks for the report. Could I maybe ask you to simplify to a reproducing snippet or attach the image in a nonbinary format instead? I feel somewhat uncomfortable downloading and opening an "untrusted" ZIP.

Ideally, this would be a few lines of codes that reproduce the buggy scenario. We'd have to do that anyway to add a regression test.

You could also try a simple numpy.savetxt or if that doesn't work a maybe a GIF?

import imageio.v3 as iio  # Dependency of skimage
iio.imwrite("density_map_bug.gif", image)

[janniswi00]

Hi @lagru, I can upload the image as txt.gz file. The txt-file is too big and the bug is not reproducible with GIF.

density_map.txt.gz

import numpy as np
import skimage
dm = np.loadtxt("density_map.txt.gz", dtype='float32')
skimage.filters.threshold_li(dm, iter_callback=print)