updated levmar library

[nobody]

Last December, an updated levmar was released (2.2), with change that could be very interesting for hugin:

 - Fixed an error in the computation of the damping factor of levmar_dif() in lm_core.c. Thanks to John Ervin for reporting.

This change is interesting, because levmar_dif is used for photometric optimisation by hugin.

I took the liberty to try hugin with this newer version of the levmar library, and on my test panorama it took slightly longer but reached the same end result.

This is of course not 0.7 material, but perhaps something for early 0.7.1.

[nobody (nobody-users)]

The file hugin-update-levmar-2.2.diff was added: patch updating foreign/levmar from 2.1.3 to 2.2

[nobody (nobody-users)]

Logged In: NO

Sorry, I notice now that forgot to state my e-mail address:

<email address hidden>

[rew (r-e-wolff)]

By now 2.5 seesm to be current levmar version (since dec 2009).

It seems Levmar is carried along in source form with the hugin source as opposed to external linking.

[tmodes (tmodes)]

Updated internal levmar library to 2.6 in default branch.

[Stefan Peter (s-peter-deactivatedaccount)]

Hi Thomas

Hugin 7031:461bf6f58d2e fails linking libhuginbase:
...
Linking CXX shared library libhuginbase.so
cd /«BUILDDIR»/hugin-2015.1.0+hg7031+dfsg/obj-x86_64-linux-gnu/src/hugin_base && /usr/bin/cmake -E cmake_link_script CMakeFiles/huginbase.dir/link.txt --verbose=1
/usr/bin/x86_64-linux-gnu-g++ -fPIC -g -O2 -fstack-protector --param=ssp-buffer-size=4 -Wformat -Werror=format-security -Wall -D_FORTIFY_SOURCE=2 --std=c++11 -Wl,-Bsymbolic-functions -Wl,-z,relro -Wl,--as-needed -Wl,-z,defs -fopenmp -shared -Wl,-soname,libhuginbase.so.0.0 -o libhuginbase.so.0.0 CMakeFiles/huginbase.dir/algorithms/nona/NonaFileStitcher.cpp.o CMakeFiles/huginbase.dir/algorithms/basic/CalculateCPStatistics.cpp.o CMakeFiles/huginbase.dir/algorithms/basic/CalculateMeanExposure.cpp.o CMakeFiles/huginbase.dir/algorithms/basic/CalculateOptimalScale.cpp.o CMakeFiles/huginbase.dir/algorithms/basic/CalculateOptimalROI.cpp.o CMakeFiles/huginbase.dir/algorithms/basic/RotatePanorama.cpp.o CMakeFiles/huginbase.dir/algorithms/basic/TranslatePanorama.cpp.o CMakeFiles/huginbase.dir/algorithms/basic/StraightenPanorama.cpp.o CMakeFiles/huginbase.dir/algorithms/basic/CalculateOverlap.cpp.o CMakeFiles/huginbase.dir/algorithms/basic/LayerStacks.cpp.o CMakeFiles/huginbase.dir/algorithms/nona/CalculateFOV.cpp.o CMakeFiles/huginbase.dir/algorithms/nona/CenterHorizontally.cpp.o CMakeFiles/huginbase.dir/algorithms/nona/FitPanorama.cpp.o CMakeFiles/huginbase.dir/algorithms/nona/ComputeImageROI.cpp.o CMakeFiles/huginbase.dir/algorithms/optimizer/ImageGraph.cpp.o CMakeFiles/huginbase.dir/algorithms/optimizer/PhotometricOptimizer.cpp.o CMakeFiles/huginbase.dir/algorithms/optimizer/PTOptimizer.cpp.o CMakeFiles/huginbase.dir/algorithms/point_sampler/PointSampler.cpp.o CMakeFiles/huginbase.dir/algorithms/control_points/CleanCP.cpp.o CMakeFiles/huginbase.dir/appbase/ProgressDisplay.cpp.o CMakeFiles/huginbase.dir/huginapp/CachedImageRemapper.cpp.o CMakeFiles/huginbase.dir/huginapp/ImageCache.cpp.o CMakeFiles/huginbase.dir/hugin_math/eig_jacobi.cpp.o CMakeFiles/huginbase.dir/hugin_math/graph.cpp.o CMakeFiles/huginbase.dir/hugin_math/Matrix3.cpp.o CMakeFiles/huginbase.dir/hugin_math/Vector3.cpp.o CMakeFiles/huginbase.dir/hugin_utils/alphanum.cpp.o CMakeFiles/huginbase.dir/hugin_utils/utils.cpp.o CMakeFiles/huginbase.dir/hugin_utils/platform.cpp.o CMakeFiles/huginbase.dir/lensdb/LensDB.cpp.o CMakeFiles/huginbase.dir/nona/SpaceTransform.cpp.o CMakeFiles/huginbase.dir/nona/Stitcher1.cpp.o CMakeFiles/huginbase.dir/nona/Stitcher2.cpp.o CMakeFiles/huginbase.dir/nona/Stitcher3.cpp.o CMakeFiles/huginbase.dir/nona/Stitcher4.cpp.o CMakeFiles/huginbase.dir/nona/Stitcher.cpp.o CMakeFiles/huginbase.dir/nona/StitcherOptions.cpp.o CMakeFiles/huginbase.dir/panodata/ControlPoint.cpp.o CMakeFiles/huginbase.dir/panodata/Lens.cpp.o CMakeFiles/huginbase.dir/panodata/Mask.cpp.o CMakeFiles/huginbase.dir/panodata/Panorama.cpp.o CMakeFiles/huginbase.dir/panodata/PanoramaOptions.cpp.o CMakeFiles/huginbase.dir/panodata/PanoramaVariable.cpp.o CMakeFiles/huginbase.dir/panodata/PTScriptParsing.cpp.o CMakeFiles/huginbase.dir/panodata/SrcPanoImage.cpp.o CMakeFiles/huginbase.dir/panodata/ImageVariableGroup.cpp.o CMakeFiles/hu...

[tmodes (tmodes)]

Hi Stefan,

strange. I tested on Linux before committing and it compiled fine on my machine. Also the original CMakeLists.txt from levmar does not contain the -fPIC switch which your compiler complains.
I added the switch again. It should now work again.

[jpe (jean-pierre-eckmann)]

thanks. works for me now

[Stefan Peter (s-peter-deactivatedaccount)]

Hi Thomas

The hugin nightlies of 2015.1.0.12f7dc74bbc2 compiled successfully, thank you very much for the prompt support.

Do you have any insight regarding the

#warning Linear inequalities constrained optimization requires LAPACK and was not compiled!

in the levmar build? I don't think that an additional lapack dependency would do any harm. But I don't know if this really is needed. And I'd need help adapting the build system (if this is needed) to honour the lapack dependency. Or does it already support this and we just missed to specify it?

With kind regards

Stefan

[tmodes (tmodes)]

Hi Stephan,

there are 2 type of warnings in levmar.

The first one is of type
#warning Linearly constrained optimization requires LAPACK and was not compiled!

levmar implement different unconstrained and constrained versions of the Levenberg-Marquard algorithm. Some of the contrained version need LAPACK to work. But Hugin is only using the unconstrained version. This type of warning can therefore ignored, at least for Hugin.

The second warning is
#warning LAPACK not available, LU will be used for matrix inversion when computing the covariance; this might be unstable at times

This warning is trigged in code which is used by Hugin. The warning states that LAPACK could help to calculate the covariance. But as far as I know Hugin is not using the returned covariance. So LAPACK could help to make it more robust. But until now I have not seen any issues when using levmar without LAPACK.

The LAPACK is already in CMake, but needs explicitly activated:
##
## LAPACK (optional, enable by -DENABLE_LAPACK=ON)
##
OPTION (ENABLE_LAPACK "Use LAPACK based solver in levmar" OFF)

[Stefan Peter (s-peter-deactivatedaccount)]

Hi Thomas

Thank you very much for your explanations. I will try to add the lapack
dependency to the Ubuntu packages I maintain and to the upstream Debian
packages as well. However, this effort will be concentrate on
hugin>2015.0.0 because I do not want to introduce a new dependency
(and, maybe, a new behavior and new problem spots) before the upcoming
2015.0.0 release.

Again, thank you very much for your advice.

With kind regards

Stefan

On 24.06.2015 20:10, tmodes wrote:
> Hi Stephan,
>
> there are 2 type of warnings in levmar.
>
> The first one is of type
> #warning Linearly constrained optimization requires LAPACK and was not compiled!
>
> levmar implement different unconstrained and constrained versions of the
> Levenberg-Marquard algorithm. Some of the contrained version need LAPACK
> to work. But Hugin is only using the unconstrained version. This type of
> warning can therefore ignored, at least for Hugin.
>
> The second warning is
> #warning LAPACK not available, LU will be used for matrix inversion when computing the covariance; this might be unstable at times
>
> This warning is trigged in code which is used by Hugin. The warning
> states that LAPACK could help to calculate the covariance. But as far as
> I know Hugin is not using the returned covariance. So LAPACK could help
> to make it more robust. But until now I have not seen any issues when
> using levmar without LAPACK.
>
> The LAPACK is already in CMake, but needs explicitly activated:
> ##
> ## LAPACK (optional, enable by -DENABLE_LAPACK=ON)
> ##
> OPTION (ENABLE_LAPACK "Use LAPACK based solver in levmar" OFF)
>

--
Any technology that does not appear magical is insufficiently advanced.
~ Gregory Benford

[tmodes (tmodes)]

Hi Stephan,

I committed changeset ad2dbc6c09bb. With this changeset the covariance matrix is not longer calculated, we don't need this matrix. So now the second warning can be ignored and compiling/linking against LAPACK should not make a difference for Hugin, because the codes which is optionally using LAPACK is not used by Hugin.

So I think it should not be necessary to pull in the LAPACK dependency.