[ITK-dev] Regression from change with WindowConvergenceMonitoringFunction?

[Bradley Lowekamp]

Hello,

I am upgrading SimpleITK's superbuild to the latest ITK. I have encountered an odd regression..

Here is the failing test output [1] and here is the source [2], where the check at line 679 fails, because the optimization didn't move.

What is very odd is that after running a git bisect I found that this patch [3] was the root of the change in behavior:

commit 51c2ff58e04a25166b6aafc7d7590c2ae74f2ec6
Author: Nick Tustison <ntustison at gmail.com>
Date:   Tue Jun 30 21:29:07 2015 -0700

    BUG:  Set a default b-spline epsilon.
    
    The B-spline domain is defined on the closed-half interval
    [a,b) which presents difficulty when we define the image
    domain to be co-extensive with the B-spline domain.  Earlier
    attempts at calculating the B-spline domain didn't work as
    this error kept popping up.  Therefore we're defining a
    default B-spline epsilon which the user can change depending
    on usage.
    
    Change-Id: I64605557fc7131e148c725e0d1cce2e5aa84f31f

After grepping around I found that the BSplineScatteredDataPointSetToImageFilter is used by the WindowConvergenceMonitoringFunction. Which could cause this early convergence I was seeing.

Perhaps this changed epsilon default value does not not make sense for the ConvergenceMonitoringFunctions?

Thanks for the additional eyes and opinions on this!

Brad

[1] https://open.cdash.org/testDetails.php?test=380566217&build=4051508
[2] https://github.com/SimpleITK/SimpleITK/blob/master/Testing/Unit/sitkImageRegistrationMethodTests.cxx#L627-L683
[3] https://github.com/InsightSoftwareConsortium/ITK/commit/51c2ff58e04a25166b6aafc7d7590c2ae74f2ec6
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://public.kitware.com/pipermail/insight-developers/attachments/20151007/a9e434f3/attachment.html>