Can somebody explain why I cannot use rescale()
from here to rescale a segmentation mask? I tried the following:
>>> from skimage.transform import rescale as r >>> a = np.ones((8,8), dtype=np.uint8) >>> a array([[1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1]], dtype=uint8) >>> r(a, 0.5, order=0) array([[0.00392157, 0.00392157, 0.00392157, 0.00392157], [0.00392157, 0.00392157, 0.00392157, 0.00392157], [0.00392157, 0.00392157, 0.00392157, 0.00392157], [0.00392157, 0.00392157, 0.00392157, 0.00392157]])
I’d expect the result to contain only ones, but this does not happen. What is the correct way of downscaling the height and width of a segmentation mask by half?
Advertisement
Answer
I think we can classify this as a bug. The origin of your problem is the way scikit-image views data types, which is detailed in this page:
https://scikit-image.org/docs/dev/user_guide/data_types.html
In some sense, scikit-image sees those values (float 0.00392157) as “equivalent” to (uint8 1).
You can use the keyword argument preserve_range=True
in order to keep the scale as the input, but you will still get the incorrect dtype.
In [1]: import numpy as np In [2]: from skimage import transform In [3]: seg_mask = np.random.randint(0, 2**16, size=(16, 16)).astype(np.uint16) In [4]: transform.rescale(seg_mask, 0.5, order=0) Out[4]: array([[0.5222858 , 0.55330739, 0.76211185, 0.61666285, 0.45897612, 0.19795529, 0.52283513, 0.56203555], [0.57460899, 0.81582361, 0.61560998, 0.68281071, 0.3169604 , 0.19830625, 0.40459297, 0.48615244], [0.47759213, 0.62240024, 0.76609445, 0.333196 , 0.88154421, 0.43877317, 0.50019074, 0.65618372], [0.84681468, 0.58448157, 0.56137942, 0.2824445 , 0.76746777, 0.75156786, 0.47310597, 0.66036469], [0.18850996, 0.60334173, 0.25194171, 0.83747616, 0.56156252, 0.61664759, 0.80070192, 0.48683909], [0.29718471, 0.3195697 , 0.32796216, 0.58196384, 0.81673915, 0.50057221, 0.48458076, 0.27048142], [0.52285039, 0.35194934, 0.84243534, 0.68207828, 0.66150912, 0.46347753, 0.23147936, 0.84484627], [0.51114672, 0.35536736, 0.87023728, 0.44734874, 0.6835584 , 0.54543374, 0.43472953, 0.78928817]]) In [5]: transform.rescale(seg_mask, 0.5, order=0, preserve_range=True) Out[5]: array([[34228., 36261., 49945., 40413., 30079., 12973., 34264., 36833.], [37657., 53465., 40344., 44748., 20772., 12996., 26515., 31860.], [31299., 40789., 50206., 21836., 57772., 28755., 32780., 43003.], [55496., 38304., 36790., 18510., 50296., 49254., 31005., 43277.], [12354., 39540., 16511., 54884., 36802., 40412., 52474., 31905.], [19476., 20943., 21493., 38139., 53525., 32805., 31757., 17726.], [34265., 23065., 55209., 44700., 43352., 30374., 15170., 55367.], [33498., 23289., 57031., 29317., 44797., 35745., 28490., 51726.]])
I think for order=0 we should not change the dtype, and have made an issue here:
https://github.com/scikit-image/scikit-image/issues/5268
In the meantime, I hope you can use preserve_range=True
and .astype(int)
to get unstuck.