ShapeLabelMapFilter floating point exception

Fijoy_Vadakkumpadan · November 5, 2018, 10:47pm

Hello,

I’m getting a floating point exception at line #364 of:

InsightSoftwareConsortium/ITK/blob/master/Modules/Filtering/LabelMap/include/itkShapeLabelMapFilter.hxx

/*=========================================================================
 *
 *  Copyright Insight Software Consortium
 *
 *  Licensed under the Apache License, Version 2.0 (the "License");
 *  you may not use this file except in compliance with the License.
 *  You may obtain a copy of the License at
 *
 *         http://www.apache.org/licenses/LICENSE-2.0.txt
 *
 *  Unless required by applicable law or agreed to in writing, software
 *  distributed under the License is distributed on an "AS IS" BASIS,
 *  WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 *  See the License for the specific language governing permissions and
 *  limitations under the License.
 *
 *=========================================================================*/
#ifndef itkShapeLabelMapFilter_hxx
#define itkShapeLabelMapFilter_hxx

This file has been truncated. show original

crashing the application. The reason seems to be numerical precision issues in principal moments calculation, resulting in a call to std::sqrt() with a negative value. In my test specifically, the principalMoments array at that line contains {-2.2737367544323206e-13, 4.5474735088646412e-13, 2524.329986572267}.

Perhaps the filter could use a more robust algorithm? Alternatively, that crashing line needs to be further guarded with a check for negative values?

FYI, I’m using ITK version 4.10.1. Just by visually comparing the source above with the one in 4.10.1, I do see that there have been some updates to the moments calculation, but the underlying numerical calculations look similar.

Thanks,
–Fijoy

blowekamp · November 7, 2018, 8:11pm

I agree this is a bug.
Can you please create an issue on Github with this information?

Thanks.

blowekamp · November 13, 2018, 8:42pm

Please find a patch to prevent the sqrt of a negative number here:

Fijoy_Vadakkumpadan · February 5, 2019, 8:26pm

Hi Bradley,

I now have a test case that hits arithmetic exception at line 393 of that same file, which does “edet = std::pow(edet, 1.0 / ImageDimension);”. The exception occurs because edet is negative, which in turn is due to numerical instability. The std::pow() cannot handle negative bases when the exponent is fractional. Can we update the file to do a check on that line as well?

Thanks,
–Fijoy

blowekamp · February 5, 2019, 9:17pm

It would be really great if we can add this case into the regression testing. Can you reduce the problem down to a simple binary image with the problematic label as non-zero and share it?

Thanks,
Brad

Fijoy_Vadakkumpadan · February 5, 2019, 10:14pm

I’ll try to find some time for making a shareable test case. The images are private and complex, so can’t promise anything at the moment.

blowekamp · February 21, 2019, 3:22pm

Any progress and producing a small test case fo this problem?

Fijoy_Vadakkumpadan · February 21, 2019, 5:19pm

I tried but couldn’t isolate the one connected component that’s causing the issue for making a separate shareable image. My test image has a very large number of components, and I can’t share it as is because of privacy. I can confirm that the issue went away once I inserted the check to my copy of that ITK class. Do you think you can add the check without my specific test case since it needs to be there anyway given the documentation for std::pow()?

blowekamp · February 22, 2019, 2:49pm

Perhaps you could try looking over all the components in the image then extract the component, then run this filter on just the one component?