ENH: Streaming-equivalence GTest for GaussianInterpolateImageFunction

[romangrothausmann]

Adds a GTest regression guard for issue #1011, which was fixed by #1202: ResampleImageFilter + GaussianInterpolateImageFunction must produce identical output whether or not its downstream is streamed.

Force-pushed to revive this dormant draft from 2019. Original three commits superseded by a single modern GTest TEST block; authorship preserved via Co-Authored-By: trailer.

Why the original draft no longer reproduced

The pipeline in the original draft fed the resampler an in-memory Image object directly. An in-memory image's BufferedRegion always equals its LargestPossibleRegion regardless of downstream RequestedRegion propagation, so the bug surface in GaussianInterpolateImageFunction::ComputeBoundingBox (which read from BufferedRegion pre-#1202) never triggered.

The new TEST inserts a CastImageFilter upstream of the resampler — a streaming-aware no-op that produces a chunked BufferedRegion per stream division. With the fix in place: 0 / 32 768 pixel diffs. With #1202 reverted locally: 28 672 / 32 768 pixels diverge, output contains NaN. Strong regression guard.

Test summary

Form	New TEST block appended to Modules/Filtering/ImageGrid/test/itkResampleImageFilterGTest.cxx (no new files)
Geometry	32³ float ramp, AffineTransform Scale(0.9), GaussianInterp σ=1, α=1
Pipeline	Image → CastImageFilter → ResampleImageFilter → StreamingImageFilter
Comparison	streamed (8 divisions) vs non-streamed (1 division), pixel-by-pixel EXPECT_FLOAT_EQ
Local timing	12 ms

[romangrothausmann]

Since all tests succeed, it seem that just using GaussianInterpolateImageFunction with a test based on itkResampleImageTest7 does not seem to trigger the problem of #1011. I also tried using 3D data instead of 2D, but then the test fails already without using GaussianInterpolateImageFunction probably due to the fixed tests concerning the chunk sizes. @thewtex How did you come up with these values:

ITK/Modules/Filtering/ImageGrid/test/itkResampleImageTest7.cxx

Lines 135 to 137 in 0b58825

	ITK_TEST_SET_GET_VALUE( 48, finalRequestedRegion.GetIndex(1) );
	ITK_TEST_SET_GET_VALUE( 60, finalRequestedRegion.GetSize(0) );
	ITK_TEST_SET_GET_VALUE( 12, finalRequestedRegion.GetSize(1) );

[dzenanz]

@romangrothausmann examining what the actual values are? The GetRadius patch has increased the region by 1 pixel on all sides.

[kwrobot-v1]

Errors:

• Failed to run the checks: service error.

[romangrothausmann]

examining what the actual values are?

@dzenanz What do you mean by that?
The increase in the reagion is clear but how to get those values for a 3D case as e.g. in: 3f5fd25
And why does the test succeed even for GaussianInterpolateImageFunction (where the radius should be larger)?

[dzenanz]

examining what the actual values are?

Run the program under debugger and examine the values.

GaussianInterpolateImageFunction's radius depends on the sigma parameter. I guess the default sigma results in radius of 1, which is the same as linear interpolation.

[dzenanz]

This PR should be revived after #1191 and #1213 are merged.

[dzenanz]

@romangrothausmann can you rebase this PR on current master?

[romangrothausmann]

That should be possible, I'll give it a try.

[romangrothausmann]

3D version without GaussianInterpolateImageFunction (26c91e7) now works for me locally. Let's see if it still does with GaussianInterpolateImageFunction.

[romangrothausmann]

Oddly, on my local PC ctest -R ResampleImageTest8 passes even though I would expect that it fails with only the changes of 452bba1 because I would expect that the 3rd dimension would demand changes to

ITK/Modules/Filtering/ImageGrid/test/itkResampleImageTest8.cxx

Lines 133 to 138 in 452bba1

	// Verify that we only requested a smaller chunk when streaming
	const ImageRegionType finalRequestedRegion( image->GetRequestedRegion() );
	ITK_TEST_SET_GET_VALUE( 0, finalRequestedRegion.GetIndex(0) );
	ITK_TEST_SET_GET_VALUE( 48, finalRequestedRegion.GetIndex(1) );
	ITK_TEST_SET_GET_VALUE( 60, finalRequestedRegion.GetSize(0) );
	ITK_TEST_SET_GET_VALUE( 12, finalRequestedRegion.GetSize(1) );

@dzenanz How to "Run the program under debugger" with ctest?

[dzenanz]

ctest -VV will print the exact command line for each test. That way, you can run that command line under debugger (e.g. copy command-line parameters into Visual Studio).

[stale]

This issue has been automatically marked as stale because it has not had recent activity. Thank you for your contributions.

[hjmjohnson]

Hi @romangrothausmann — I hope you don't mind, but I took the liberty of force-pushing a single commit to this branch to revive your dormant draft as a regression test for #1011.

The original three commits have been replaced (your authorship is preserved via Co-Authored-By: in the new commit's message). The new test is a TEST block appended to itkResampleImageFilterGTest.cxx, and inserts a CastImageFilter upstream of the resampler so the bug actually surfaces under streaming — with that change, locally reverting #1202's fix produces 28 672 / 32 768 NaN-or-divergent pixels, which gives us the strong regression guard you set out to add seven years ago.

cc @dzenanz — keeping this PR draft pending CI; happy to flip to ready-for-review once CI is green if you'd like.

If you'd prefer a different attribution arrangement or a separate PR from my fork, just say the word and I'll redo it.

[dzenanz]

Let's give @romangrothausmann a few days to review this.

[hjmjohnson]

Folded the two P2 style nits from greptile's review into the existing commit (autosquash, force-push to refresh CI). Both are pure stylistic improvements — no logic change, test still passes locally.

Greptile P2	Fix
Brace-aggregate { { 32, 32, 32 } } for SizeType	Replaced with ImageType::SizeType::Filled(32) (matches ITK convention for uniform sizes)
Magic literal 8 for stream divisions	Extracted as constexpr unsigned int kNumberOfStreamDivisions{ 8 }

@romangrothausmann — the patchset is still a single commit so the diff you'll see vs. upstream/main reflects the latest polish. No need for you to address greptile's findings yourself.

cc @dzenanz — taking the courtesy window into account; this just preempts the polish so Roman gets the final form on first read.