WIP: Replace "vnl_sample.h" with <random> in ConnectedComponent tests

[N-Dekker]

• Another follow-up to pull request ENH: Replace "vnl/vnl_sample.h" with <random>, in Core tests #5611

[dzenanz]

Looks good on a glance.

[N-Dekker]

• The first of the two commits (on ConnectedComponent tests) has test failures that I find hard to fix. So I branched off the other one (on Voronoi), which should be easier to get passed: pull request ENH: Replace "vnl_sample.h" with <random> in VoronoiSegmentation tests #5624

[N-Dekker]

There appeared test failures at https://open.cdash.org/viewTest.php?onlyfailed&buildid=10817138 on the following four tests:

1. itkConnectedComponentImageFilterTest
2. itkConnectedComponentImageFilterTest2
3. itkConnectedComponentImageFilterTest3
4. itkConnectedComponentImageFilterTestRGB

I'm sorry, I don't really know how to fix them, but I guess we would have to change the ground truth images (somewhere in ${ITK_DATA_ROOT}, in order to make this work. For example, for itkConnectedComponentImageFilterTestRGB:

ITK/Modules/Segmentation/ConnectedComponents/test/CMakeLists.txt

Lines 44 to 55 in 6568876

	itk_add_test(
	NAME itkConnectedComponentImageFilterTestRGB
	COMMAND
	ITKConnectedComponentsTestDriver
	--compare
	DATA{${ITK_DATA_ROOT}/Baseline/BasicFilters/ConnectedComponentImageFilterTest.png,:}
	${ITK_TEST_OUTPUT_DIR}/ConnectedComponentImageFilterTestRGB.png
	itkConnectedComponentImageFilterTestRGB
	DATA{${ITK_DATA_ROOT}/Input/cthead1.png}
	${ITK_TEST_OUTPUT_DIR}/ConnectedComponentImageFilterTestRGB.png
	130
	145

Does anyone have suggestion?

[hjmjohnson]

@N-Dekker does {} initialize the seed to a constant value?

I think

constexpr auto random_seed = 1031571;
std::mt19937 randomNumberEngine{random_seed};

would make this easier to interpret.

[N-Dekker]

The default constructor of std::mt19937 always uses the same seed: default_seed = 5489. That's different from 1031571, obviously. So we'll see if using 1031571 will make the CI happy...

[hjmjohnson]

That's great to know, but not obvious. I would have guessed the default was to use an unknown changing seed unless explicitly specified. I do think that setting the seed explicitly (with what ever value you want, even 5489) would be more meaningful.

[N-Dekker]

When the C++ standard specifies that the default-constructor of std::mt19937 always uses the same seed, default_seed, I think we should be allowed to just trust the C++ standard.

We could make it explicit:

std::mt19937 randomNumberEngine(std::mt19937::default_seed);

And then add to our unit tests:

static_assert(std::mt19937::default_seed == 5489, "Check on any platform!");

But that all seems overdone to me. If we want to have non-deterministic random numbers, we should use std::random_device.