[Insight-developers] Anisotropic level sets tests

[Luca Antiga]

Hi Bill,
  last night I investigated the tests that give the largest differences,
ThresholdSegmentationLevelSetGrayMatterTest
and
ThresholdSegmentationLevelSetWhiteMatterTest
(the other tests fail for very minor interpolation differences, which
are expected since the fudge factor affected zero level-set  
interpolation).

Considering ThresholdSegmentationLevelSetWhiteMatterTest, the "with
fudge factor" segmentation result is half the brain, and the "without  
fudge
factor" is the whole brain.

Stopping condition for both tests was:
Max. no. iterations: 1200
Max. RMS error: 0.02

At the end of the run, these were the parameters:

With fudge factor:
No. elpased iterations: 403
RMS change: 0.0173078

Without fudge factor:
No. elpased iterations: 269
RMS change: 0.0181523

Therefore, turns out that the difference is the RMS change, which in the
"without fudge factor" case doesn't fall below 0.02 when half the  
brain is
segmented, so level sets evolve until the whole brain is segmented.

While on one hand this explains the issue without much voodoo, I'm
worried that behavior of user code that relies on RMS changes as  
stopping
condition will change with the new modification.

I don't know, but at this point keeping the fudge factor as an option
in SparseFieldLevelSetImageFilter is probably a safer bet.

Best

Luca



On Mar 5, 2008, at 7:00 PM, Bill Lorensen wrote:

> Luca,
>
> I looked at the new results and they certainly look reasonable. I
> would say that if you are satisfied that the differences are not bugs,
> go ahead and check in the new baselines.
>
> Thanks for all of your hard work,
>
> Bill
>
> On Wed, Mar 5, 2008 at 9:39 AM, Luca Antiga <luca.antiga at gmail.com>  
> wrote:
>> Dear all,
>>  today I worked at making all tests work with the new modifications
>> to level sets for support of anisotropic spacings. It looks like
>> we're close.
>> Today 5 tests were failing. Now, after temporarily reintroducing the
>> fudge factor (and fixing a bug), and in the absence of major
>> surprises on the dashboard, all tests should pass (except for
>> itkVectorThresholdSegmentationLevelSetImageFilterTest, which already
>> has a baseline image for the fudge factor-free level set).
>> If there aren't objections, tomorrow I'll permanently remove the
>> fudge factor (Insight/Code/BasicFilters/
>> itkSparseFieldLevelSetImageFilter.txx:991) and provide new baseline
>> images for the following tests
>>
>> CannySegmentationLevelSetImageFilterTest
>> LaplacianSegmentationLevelSetImageFilterTest
>> ThresholdSegmentationLevelSetGrayMatterTest
>> ThresholdSegmentationLevelSetVentricleTest
>> ThresholdSegmentationLevelSetWhiteMatterTest
>>
>> Alternatively, we could back up from our decisions and provide a
>> runtime flag to activate/deactivate the fudge factor. This way we
>> wouldn't need to provide new baseline images. In my opinion this is a
>> worse solution, but it's doable.
>>
>> Any opinion on this?
>>
>> Luca
>>
>>
>> _______________________________________________
>> Insight-developers mailing list
>> Insight-developers at itk.org
>> http://www.itk.org/mailman/listinfo/insight-developers
>>