This code would require porting to opencv4

[hyunlee1o]

A lot of the changes are easy to make but others are not that easy.
Could you look into that?

[lejeunel]

I started this port to opencv4. It comes with a boost-python wrapper.

https://github.com/lejeunel/gPb-GSoC

Keep in mind though that it produces malloc/seg-fault errors... which I'm still trying to solve.

[hyunlee1o]

I can try to help, the main issues are found in the watershed file.
Because it need some refactoring to some malloc yes.

[lejeunel]

The watershed is ok for me.
The malloc issue seems to be in function contour_side. I'll let you know.

[hyunlee1o]

Also i got errors compiling numpy-opencv-converter

Could NOT find PythonLibs: Found unsuitable version "2.7.16", but required
  is at least "3" (found /usr/lib/x86_64-linux-gnu/libpython2.7.so)

[lejeunel]

Two options:

• Add a cmake flag to omit the python wrapper
• Install python and boost

[hyunlee1o]

I used this command

cmake .. -D CMAKE_BUILD_TYPE=RELEASE \
    -DPYTHON_LIBRARY=/usr/lib/x86_64-linux-gnu/libpython3.7m.so \
    -DPYTHON_INCLUDE_DIR=/usr/include/python3.7/ \
    -DPYTHON_EXECUTABLE=/usr/bin/python3.7

I had python and boost installed but it didn't work properly because of the numpy-opencv-converter CMakeLists.txt

Now it compiled without error.
Also mind to add the
cmake_minimum_required(VERSION 3.13)
To your CMakeLists.txt because it pops an error.

[lejeunel]

I'm not sure if I actually found the actual seg-fault culprit, but I got around it with shared pointers :x
See my last commit and let me know.

[hyunlee1o]

eh, well, first i need to know how to use it X)
I did a pull.

[hyunlee1o]

Ok, it works but it's very slow wow.

PYTHON TYPE CONVERTERS exported
size(in1): [640 x 512]
size(in2): [640 x 512]
mPb computation commencing ...
 --->  computing texton ...
 --->  computing bg cga cgb tg ...
sPb computation commencing ...
gPb computation commencing ...
`fish: Job 2, 'python3  $argv' terminated by signal SIGSEGV (Address boundary error)`

I have seen it fails when i use a starred operation. I can only use it one by one.

[hyunlee1o]

Also it's SUPER slow because EVERY SINGLE pixel access is made with img.at instead of a pointer.

[lejeunel]

indeed this code is pretty bad.
Regarding crashes: It seems that there's a size limit to images. For now, I resize them to 300x300 and resize as post-processing.
Regarding performance: Are you willing to make a PR?

Here a reference for iterating on pixels:
https://docs.opencv.org/master/d3/d63/classcv_1_1Mat.html#a952ef1a85d70a510240cb645a90efc0d

[hyunlee1o]

I have done some work with opencv C++ int the past but a reminder is a pretty good option, also that im working rn 100% in python.

Also i get random crashes.
I am trying with images 640 x 512 and i got crashes in 2 out of 10.

I will review the package and maybe i will do some PR