[digikam] [Bug 349601] New: Compiling digikam with opencv 3.0.0 gives crash but not with opencv 2.x

[Detlef]

https://bugs.kde.org/show_bug.cgi?id=349601

            Bug ID: 349601
           Summary: Compiling digikam with opencv 3.0.0 gives crash but
                    not with opencv 2.x
           Product: digikam
           Version: 4.9.0
          Platform: MacPorts Packages
                OS: OS X
            Status: UNCONFIRMED
          Severity: normal
          Priority: NOR
         Component: libkface
          Assignee: [hidden email]
          Reporter: [hidden email]

Installing digikam via MacPort is building a binary very well, but now has
MacPort an upgrade of opencv from 2.x to 3.x and now it is impossible to
compile digikam. Every time there come an error for an not existing library
"opencv/legacy/compat.hpp" not found. deactivating opencv 3.x and activating
opencv 2.x and then compiling digikam works fine.

Probing the running of digikam with binaries of a package for version 4.10.x
and 4.11.x gives a crash while starting the digikam.app.

Reproducible: Always

Steps to Reproduce:
1. port Activate opencv @3.0.0
2. port clean digikam; port build digikam
3. Error while compiling
4. port activate opencv 2.x
5. port clean digikam; port install digikam
6. Now exist a running digikam

Actual Results:  
I can't use opencv 3.x

Expected Results:  
A running digikam while opencv is upgraded

--
You are receiving this mail because:
You are the assignee for the bug.
_______________________________________________
Digikam-devel mailing list
[hidden email]
https://mail.kde.org/mailman/listinfo/digikam-devel

[Gilles Caulier-4]

https://bugs.kde.org/show_bug.cgi?id=349601

Gilles Caulier <[hidden email]> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
          Component|libkface                    |Portability
                 CC|                            |[hidden email]

--
You are receiving this mail because:
You are the assignee for the bug.
_______________________________________________
Digikam-devel mailing list
[hidden email]
https://mail.kde.org/mailman/listinfo/digikam-devel

[Gilles Caulier-4]

https://bugs.kde.org/show_bug.cgi?id=349601

Gilles Caulier <[hidden email]> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
            Summary|Compiling digikam with      |digikam compile fine with
                   |opencv 3.0.0 gives crash    |OpenCV 2.X but not with 3.x
                   |but not with opencv 2.x     |

--
You are receiving this mail because:
You are the assignee for the bug.
_______________________________________________
Digikam-devel mailing list
[hidden email]
https://mail.kde.org/mailman/listinfo/digikam-devel

[demm]

https://bugs.kde.org/show_bug.cgi?id=349601

--- Comment #1 from [hidden email] ---
Same error if trying to compile on Linux with OpenCV 3.0.0

--
You are receiving this mail because:
You are the assignee for the bug.
_______________________________________________
Digikam-devel mailing list
[hidden email]
https://mail.kde.org/mailman/listinfo/digikam-devel

[Jozef Mlich]

https://bugs.kde.org/show_bug.cgi?id=349601

Jozef Mlich <[hidden email]> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |[hidden email]

--- Comment #2 from Jozef Mlich <[hidden email]> ---
Created attachment 93709
  --> https://bugs.kde.org/attachment.cgi?id=93709&action=edit
incomplete patch

According to OpenCV Transition Guide [1] was CV_INIT_ALGORITHM macro removed. I
have started writing of patch for that. It is not finished, but it may help
someone.

[1] http://docs.opencv.org/master/db/dfa/tutorial_transition_guide.html

--
You are receiving this mail because:
You are the assignee for the bug.
_______________________________________________
Digikam-devel mailing list
[hidden email]
https://mail.kde.org/mailman/listinfo/digikam-devel

[Nirvandil]

https://bugs.kde.org/show_bug.cgi?id=349601

Nirvandil <[hidden email]> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |[hidden email]

--- Comment #3 from Nirvandil <[hidden email]> ---
Same problem in Gentoo ~amd64, when compiling libkface-4.12.0 with opencv-3.0.0
I've got error:
/var/tmp/portage/media-libs/libkface-4.12.0/work/libkface-4.12.0_build/libkface/libopencv.h:55:37:
fatal error: opencv2/core/internal.hpp: No such file or directory
 #include <opencv2/core/internal.hpp>

--
You are receiving this mail because:
You are the assignee for the bug.
_______________________________________________
Digikam-devel mailing list
[hidden email]
https://mail.kde.org/mailman/listinfo/digikam-devel

[Adam Stylinski]

https://bugs.kde.org/show_bug.cgi?id=349601

[hidden email] changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |[hidden email]

--- Comment #4 from [hidden email] ---
Created attachment 94088
  --> https://bugs.kde.org/attachment.cgi?id=94088&action=edit
completed patch

If anyone's interested, I may have this completed?  It compiles, I haven't
tested it just yet, though.  The CV_ALGORITHM_INIT macro is now gone, but I'm
not sure if/where I'd need to call the new factory they are providing in
facerec.hpp (OpenCV's semantics for defining an algorithm are kind of confusing
in C++).  

I've attached the updated patch if anyone's interested.  It's against the 14.12
branch in git, but it will likely work in later ones as well.

--
You are receiving this mail because:
You are the assignee for the bug.
_______________________________________________
Digikam-devel mailing list
[hidden email]
https://mail.kde.org/mailman/listinfo/digikam-devel

[Adam Stylinski]

https://bugs.kde.org/show_bug.cgi?id=349601

--- Comment #5 from [hidden email] ---
Just finally got around to testing this (digikam is using the legacy API for
some reason in imgqsort, the change is extremely minor to make it compile and
work).  Libkface is not working properly, it's claiming the model hasn't been
trained a lot and eventually it crashes from all the openCV exceptions.  I
suspect that we'll need to find the analog to CV_ALGORITHM_INIT.

--
You are receiving this mail because:
You are the assignee for the bug.
_______________________________________________
Digikam-devel mailing list
[hidden email]
https://mail.kde.org/mailman/listinfo/digikam-devel

[Viktor Jancik]

https://bugs.kde.org/show_bug.cgi?id=349601

Viktor Jancik <[hidden email]> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |[hidden email]

--- Comment #6 from Viktor Jancik <[hidden email]> ---
Created attachment 94096
  --> https://bugs.kde.org/attachment.cgi?id=94096&action=edit
libkface-15.04.2-opencv3.patch

I found an additional error,

The haarcascade files are located in /usr/share/OpenCV/haarcascades, which can
be seen by running:
rpm -ql opencv

If you have opencv version 3 installed.

However the original ${OpenCV_DIR} points to the base of the OpenCV library
directory.
As described here: http://code.opencv.org/attachments/196/FindOpenCV.cmake
Specifically /usr/lib64/OpenCV

Instead, {OpenCV_SHARED} should be used which points to /usr/share/OpenCV as
described here:
https://github.com/Itseez/opencv/blob/master/cmake/OpenCVConfig.cmake

The values of these variables can be verified after building by looking at
CMake Cache.txt

Additionally I changed the minimum required version to 3.0.0 because I am not
sure if the changes also build for OpenCV version 2.4, this needs to be
checked.

And lastly I included face.hpp instead of face/facerec.hpp but they include
each other so it's the same.


The successful build with these changes can be seen here:
https://copr.fedoraproject.org/coprs/vjancik/opencv3-deps/build/109623/

--
You are receiving this mail because:
You are the assignee for the bug.
_______________________________________________
Digikam-devel mailing list
[hidden email]
https://mail.kde.org/mailman/listinfo/digikam-devel

[Gilles Caulier-4]

https://bugs.kde.org/show_bug.cgi?id=349601

--- Comment #7 from Gilles Caulier <[hidden email]> ---
In the pass, i accidentally checkout OpenCV from master which correspond to
current 3.0 (while 2.x was stable release).

I discovered that CV_ALGORITHM_INIT disappear. Thinking about working with 2.x,
i tried to fix the problem. The only solution was to import this macro into
libkface as well.

Of course, when i discovered that i don't work with right OpenCV  branch i
cancelled my work as well. Sound like this must be re-iterate if we want to not
re-write whole implementation about face management from libkface.

Gilles Caulier

--
You are receiving this mail because:
You are the assignee for the bug.
_______________________________________________
Digikam-devel mailing list
[hidden email]
https://mail.kde.org/mailman/listinfo/digikam-devel

[Viktor Jancik]

https://bugs.kde.org/show_bug.cgi?id=349601

--- Comment #8 from Viktor Jancik <[hidden email]> ---
Created attachment 94097
  --> https://bugs.kde.org/attachment.cgi?id=94097&action=edit
digikam-4.12.0-opencv3.patch

This patch doesn't contain fixes for libkface directly, because digikam is
configured to build with the system version of libkface, so instead I created a
separate patch that fixes the system version of libkface.

Despite that fix, digikam still wouldn't build against OpenCV 3 and had
numerous errors.

cvCreateMat returns cvMat but variable d is of type Mat, a 1 to 1 replacement
was possible with the Mat constructor.
In core/libs/database/imgqsort/imgqsort.cpp vector containers were missing the
std:: namespace prefix. How did it build before?
In
extra/kipi-plugins/removeredeyes/detection/locators/haarclassifier/haarclassifierlocator.cpp
I had to add a definition of cvFillImage because it was removed from OpenCV 3.
The definition is exactly the same as it was in OpenCV 2.4.11.
I had to conditionally disable the building of ipptest because it used a
function that no longer exists in OpenCV 3 (cvGetModuleInfo) and I couldn't
find / create a replacement. As a consequence, I am not sure if the tests pass
properly, because I am not sure if this is a good way to disable a test.
And lastly, I had to modify the includes in 2 libopencv.h files and modify the
OPENCV_TEST_VERSION macros as they didn't work as intended for versions 2.3.0
up to 2.4.X Yes, there are two libopencv.h files.

--
You are receiving this mail because:
You are the assignee for the bug.
_______________________________________________
Digikam-devel mailing list
[hidden email]
https://mail.kde.org/mailman/listinfo/digikam-devel

[Viktor Jancik]

https://bugs.kde.org/show_bug.cgi?id=349601

--- Comment #9 from Viktor Jancik <[hidden email]> ---
Successful build with patched digikam and patched libkface can be seen here:
https://copr.fedoraproject.org/coprs/vjancik/opencv3-deps/build/109686/

--
You are receiving this mail because:
You are the assignee for the bug.
_______________________________________________
Digikam-devel mailing list
[hidden email]
https://mail.kde.org/mailman/listinfo/digikam-devel

[Gilles Caulier-4]

https://bugs.kde.org/show_bug.cgi?id=349601

--- Comment #10 from Gilles Caulier <[hidden email]> ---
I don't checked your patch yet, but did you take a care about to wrap current
implementation with a relevant pre-processor macro depending of OpenCV version
?

Typically, OpenCV 2 -> 3 transition will take a time and we can be necessary to
support both versions at the same time for a while.

Another solution is only to dedicate your patch to next digiKam 5.0.0 release
(frameworks branches) and to only support OpenCV 3 with this version. digiKam
4.x will only support OpenCV 2. As 5.0.0 is planed next year, this can be
acceptable.

Note that first solution will increase complexity of code.

Your viewpoint ?

Gilles Caulier

--
You are receiving this mail because:
You are the assignee for the bug.
_______________________________________________
Digikam-devel mailing list
[hidden email]
https://mail.kde.org/mailman/listinfo/digikam-devel

[Adam Stylinski]

https://bugs.kde.org/show_bug.cgi?id=349601

--- Comment #11 from [hidden email] ---
I see the first solution is worth the added complexity.  The changes are
relatively minor at the API level and the core algorithms generally behave the
same.  I think the missing haarcascades may have been why my version when
compiled would complain about lack of training data.  I'll try Viktor's patches
when I get home.

--
You are receiving this mail because:
You are the assignee for the bug.
_______________________________________________
Digikam-devel mailing list
[hidden email]
https://mail.kde.org/mailman/listinfo/digikam-devel

[Viktor Jancik]

https://bugs.kde.org/show_bug.cgi?id=349601

--- Comment #12 from Viktor Jancik <[hidden email]> ---
Yes, I believe the digikam changes will work for both OpenCV 2.4 and 3. You can
see, that most if not all of my modifications are preceded with a OpenCV
version conditional.

I didn't try to build it against 2.4, so there might be some slight error.

As for the libkface patch, I am fairly positive it won't build against OpenCV
2.4. There was no obvious infrastructure in place for such branching, and I
didn't want to commit work to making it build with 2.4, since I am not too
familiar with the project or the coding guidelines, which means it would have
to be reworked anyways.

--
You are receiving this mail because:
You are the assignee for the bug.
_______________________________________________
Digikam-devel mailing list
[hidden email]
https://mail.kde.org/mailman/listinfo/digikam-devel

[Adam Stylinski]

https://bugs.kde.org/show_bug.cgi?id=349601

--- Comment #13 from [hidden email] ---
It may require a lot of preprocessor statements but making libkface work with
both 3.x and 2.x is doable.  Many distributions are including 3, and 3 is
supposed be mainly a progression of the version 2 API (unlike OpenCV's earlier
version jumps).  Maybe the solution to being 2.4 compatible is writing a
separate source file for the FaceRecognizer and let CMake conditionally compile
it.

--
You are receiving this mail because:
You are the assignee for the bug.
_______________________________________________
Digikam-devel mailing list
[hidden email]
https://mail.kde.org/mailman/listinfo/digikam-devel

[Adam Stylinski]

https://bugs.kde.org/show_bug.cgi?id=349601

--- Comment #14 from Adam Stylinski <[hidden email]> ---
Viktor, just applied your patches, still having the same problem I was before.
After doing the face detect, libkface prints the following error:

OpenCV Error: Bad argument (This LBPH model is not computed yet. Did you call
the train method?) in predict, file
/scratchdir/tmp/portage/media-libs/libkface-4.12.0-r2/work/digikam-4.12.0/extra/libkface/libkface/recognition-opencv-lbph/facerec_borrowed.cpp,
line 384
digikam(17371)/KFACE: cv::Exception:
/scratchdir/tmp/portage/media-libs/libkface-4.12.0-r2/work/digikam-4.12.0/extra/libkface/libkface/recognition-opencv-lbph/facerec_borrowed.cpp:384:
error: (-5) This LBPH model is not computed yet. Did you call the train method?
in function predict

--
You are receiving this mail because:
You are the assignee for the bug.
_______________________________________________
Digikam-devel mailing list
[hidden email]
https://mail.kde.org/mailman/listinfo/digikam-devel

[Adam Stylinski]

https://bugs.kde.org/show_bug.cgi?id=349601

--- Comment #15 from Adam Stylinski <[hidden email]> ---
Hmm that error could have been innocuous (based on some searches with Google -
some more training data may have been needed to be supplied), but I get a very
real hard crash if I let it run long enough, with a double free:

*** Error in `digikam': free(): invalid next size (normal): 0x00007fe0b84a6bf0
***
======= Backtrace: =========
/lib64/libc.so.6(+0x7237c)[0x7fe0fb95337c]
/lib64/libc.so.6(+0x779ae)[0x7fe0fb9589ae]
/lib64/libc.so.6(+0x7818b)[0x7fe0fb95918b]
/usr/lib64/libopencv_imgproc.so.3.0(_ZN2cv5CannyERKNS_11_InputArrayERKNS_12_OutputArrayEddib+0x5ca)[0x7fe0ffab408a]
/usr/lib64/libopencv_imgproc.so.3.0(cvCanny+0x19b)[0x7fe0ffab5ecb]
/usr/lib64/libopencv_objdetect.so.3.0(_Z25cvHaarDetectObjectsForROCPKvP23CvHaarClassifierCascadeP12CvMemStorageRSt6vectorIiSaIiEERS5_IdSaIdEEdii6CvSizeSC_b+0x215f)[0x7fe0f7e1a29f]
/usr/lib64/libopencv_objdetect.so.3.0(+0x40b90)[0x7fe0f7e33b90]
/usr/lib64/libopencv_objdetect.so.3.0(+0x41082)[0x7fe0f7e34082]
/usr/lib64/libopencv_objdetect.so.3.0(+0x2cf3b)[0x7fe0f7e1ff3b]
/usr/lib64/libopencv_objdetect.so.3.0(_ZN2cv17CascadeClassifier16detectMultiScaleERKNS_11_InputArrayERSt6vectorINS_5Rect_IiEESaIS6_EEdiiNS_5Size_IiEESB_+0xc6)[0x7fe0f7e32fb6]
/usr/lib64/libkface.so.3(+0x1b742)[0x7fe0ffe96742]
/usr/lib64/libkface.so.3(+0x1e029)[0x7fe0ffe99029]
/usr/lib64/libkface.so.3(_ZN10KFaceIface12FaceDetector11detectFacesERK6QImageRK5QSize+0x9c)[0x7fe0ffe89d0c]
digikam[0x7a8a60]
digikam[0x7a8e12]
/usr/lib64/qt4/libQtCore.so.4(_ZN7QObject5eventEP6QEvent+0x25e)[0x7fe0fc3e632e]
/usr/lib64/qt4/libQtGui.so.4(_ZN19QApplicationPrivate13notify_helperEP7QObjectP6QEvent+0x8c)[0x7fe0fcb4effc]
/usr/lib64/qt4/libQtGui.so.4(_ZN12QApplication6notifyEP7QObjectP6QEvent+0x1c1)[0x7fe0fcb57d21]
/usr/lib64/libkdeui.so.5(_ZN12KApplication6notifyEP7QObjectP6QEvent+0x33)[0x7fe0fe173b43]
/usr/lib64/qt4/libQtCore.so.4(_ZN16QCoreApplication14notifyInternalEP7QObjectP6QEvent+0x75)[0x7fe0fc3c70d5]
/usr/lib64/qt4/libQtCore.so.4(_ZN23QCoreApplicationPrivate16sendPostedEventsEP7QObjectiP11QThreadData+0x1ea)[0x7fe0fc3cbd2a]
/usr/lib64/qt4/libQtCore.so.4(+0x1ea0ae)[0x7fe0fc3ff0ae]
/usr/lib64/libglib-2.0.so.0(g_main_context_dispatch+0x27c)[0x7fe0f4c2610c]
/usr/lib64/libglib-2.0.so.0(+0x54398)[0x7fe0f4c26398]
/usr/lib64/libglib-2.0.so.0(g_main_context_iteration+0x2c)[0x7fe0f4c2643c]
/usr/lib64/qt4/libQtCore.so.4(_ZN20QEventDispatcherGlib13processEventsE6QFlagsIN10QEventLoop17ProcessEventsFlagEE+0x7e)[0x7fe0fc3fe01e]
/usr/lib64/qt4/libQtCore.so.4(_ZN10QEventLoop13processEventsE6QFlagsINS_17ProcessEventsFlagEE+0x38)[0x7fe0fc3c50b8]
/usr/lib64/qt4/libQtCore.so.4(_ZN10QEventLoop4execE6QFlagsINS_17ProcessEventsFlagEE+0x185)[0x7fe0fc3c53d5]
/usr/lib64/libdigikamcore.so.4.12.0(+0x3a0a58)[0x7fe101569a58]
/usr/lib64/qt4/libQtCore.so.4(+0x7389b)[0x7fe0fc28889b]
/usr/lib64/qt4/libQtCore.so.4(+0x83327)[0x7fe0fc298327]
/lib64/libpthread.so.0(+0x7304)[0x7fe0fa7a4304]
/lib64/libc.so.6(clone+0x6d)[0x7fe0fb9c8c1d]

--
You are receiving this mail because:
You are the assignee for the bug.
_______________________________________________
Digikam-devel mailing list
[hidden email]
https://mail.kde.org/mailman/listinfo/digikam-devel

[Viktor Jancik]

https://bugs.kde.org/show_bug.cgi?id=349601

--- Comment #16 from Viktor Jancik <[hidden email]> ---
That error message gets created if the member variable m_histograms of the
class LBPHFaceRecognizer which is of type std::vector<cv::Mat> is empty.
Unless the getter/setter methods we created are not working (which they have no
reason not to) I have no idea why that vector is empty.

Could you try building and running the tests of the patched libkface?
Separately from digikam? A few of them use the train method, and should fail if
we broke it.

--
You are receiving this mail because:
You are the assignee for the bug.
_______________________________________________
Digikam-devel mailing list
[hidden email]
https://mail.kde.org/mailman/listinfo/digikam-devel

[Adam Stylinski]

https://bugs.kde.org/show_bug.cgi?id=349601

--- Comment #17 from Adam Stylinski <[hidden email]> ---
My feature vectors are also being serialized into a MySQL database for my
backend instead of the default sqlite3, so that could also have something to do
with it.  I'll try your tests but if your unit tests work I assume mine
probably will too.  The issues I'm seeing admittedly could have been there all
along, I haven't heavily used this feature.  The face detect seems to be
working right now, I manually gave it a ton of training data so I'm trying the
facial recognition right now.  I did the detect and recognize option when it
was churning forever (and eventually crashed).

--
You are receiving this mail because:
You are the assignee for the bug.
_______________________________________________
Digikam-devel mailing list
[hidden email]
https://mail.kde.org/mailman/listinfo/digikam-devel