Can we build opencv and opencv_contrib separately in 2 steps?

[jiapei100]

System information (version)

• OpenCV => :3.2.0-dev:
• Operating System / Platform => :Ubuntu 16.04.1:
• Compiler => :GCC 5.4.0:

Detailed description

Trying to build opencv together with opencv_contrib, with configuration:
-DOPENCV_EXTRA_MODULES_PATH=xxxxxxxxxx

However, it always reports some Error message as what I posted here:
opencv/opencv#7939

Just wonder if there is a way to build opencv and opencv_contrib separately? which seems to be more reasonable???

Thank you
Pei

[alalek]

No, you can't.

[jiapei100]

@alalek
However, for cnn_3dobj and dnn:
( refer to https://github.com/opencv/opencv_contrib/blob/master/modules/README.md )

cnn_3dobj: Deep Object Recognition and Pose -- Uses Caffe Deep Neural Net library to build, train and test a CNN model of visual object recognition and pose.
dnn: Deep Neural Networks (DNNs) -- This module can read in image recogniton networks trained in the Caffe neural netowrk library and run them efficiently on CPU.

Caffe requires opencv, while opencv_contrib is based on caffe.... This is a circle, isn't it?

Pei

[alalek]

It is resolved. But version should be the same to avoid conflicts.
You can build opencv without opencv_contrib, but you can't build opencv_contrib without opencv sources (yet).

BTW, gcc uses gcc to compile itself.

[jiapei100]

@alalek
Does that mean: the best way to build a full package of opencv, including both opencv and opencv_contrib, is to build in the following sequence:

1. opencv
2. caffe
3. opencv_contrib
   ?

Thank you
Pei

[alalek]

Probably yes, but it depends on details.

BTW, Caffe's dependency on OpenCV is optional, you need to check USE_OPENCV option.

[jiapei100]

@alalek

Just tested again.

opencv with Cuda Enabled successfully installed
caffe with Cuda Enabled successfully installed

opencv_contrib with Cuda Enabled, still the same issue as opencv/opencv#7939 .
I even turn BUILD_opencv_cnn_3dobj to OFF , because for cnn_3dobj, I obtained the following error message:
fatal error: opencv2/viz/vizcore.hpp

After I set all CUDA related options to OFF, including:
WITH_CUBLAS=OFF
WITH_CUDA=OFF
WITH_CUFFT=OFF
etc.
both opencv and opencv_contrib are NOW successfully built...

1. A phrasal conclusion NOW can be made as:
   There must be a conflict between CUDA 8.0.44 and opencv_contrib (git master)

2. A further conclusion can be made from my side:
   It has NEVER been a compile problem. It has ALWAYS been a link problem.

[ 27%] Linking CXX shared library ../../lib/libopencv_cudawarping.so
cd ....../opencv/build/modules/cudawarping && /usr/bin/cmake -E cmake_link_script CMakeFiles/opencv_cudawarping.dir/link.txt --verbose=1
/usr/bin/c++  -fPIC    -fsigned-char -W -Wall -Werror=return-type -Werror=non-virtual-dtor -Werror=address -Werror=sequence-point -Wformat -Werror=format-security   -Winit-self -Wpointer-arith  -Wsign-promo -Wno-narrowing -Wno-delete-non-virtual-dtor -Wno-comment -fdiagnostics-show-option -Wno-long-long -pthread -fno-omit-frame-pointer -ffast-math -msse -msse2 -mno-avx -msse3 -mno-ssse3 -mno-sse4.1 -mno-sse4.2 -pg -g -fvisibility=hidden -fvisibility-inlines-hidden -fopenmp -Wno-undef -Wno-missing-declarations -Wno-shadow -O3 -DNDEBUG  -DNDEBUG  -shared -Wl,-soname,libopencv_cudawarping.so.3.2 -o ../../lib/libopencv_cudawarping.so.3.2.0 CMakeFiles/opencv_cudawarping.dir/src/resize.cpp.o CMakeFiles/opencv_cudawarping.dir/src/remap.cpp.o CMakeFiles/opencv_cudawarping.dir/src/pyramids.cpp.o CMakeFiles/opencv_cudawarping.dir/src/warp.cpp.o CMakeFiles/cuda_compile.dir/src/cuda/cuda_compile_generated_warp.cu.o CMakeFiles/cuda_compile.dir/src/cuda/cuda_compile_generated_pyr_down.cu.o CMakeFiles/cuda_compile.dir/src/cuda/cuda_compile_generated_resize.cu.o CMakeFiles/cuda_compile.dir/src/cuda/cuda_compile_generated_pyr_up.cu.o CMakeFiles/cuda_compile.dir/src/cuda/cuda_compile_generated_remap.cu.o  -L/usr/local/cuda/lib64 ../../lib/libopencv_imgproc.so.3.2.0 -ldl -lm -lpthread -lrt -lGLU -lGL /usr/lib/x86_64-linux-gnu/libtbb.so ../../3rdparty/ippicv/ippicv_lnx/lib/intel64/libippicv.a -lopenblas -lcudart -lnppc -lnppi -lnpps -lcublas -lcufft -L/usr/local/cuda/lib64 -lcudart -lnppc -lnppi -lnpps ../../lib/libopencv_core.so.3.2.0 ../../lib/libopencv_cudev.so.3.2.0 -lcublas -lcufft -Wl,-rpath,/usr/local/cuda/lib64:....../opencv/build/lib: 
cd ....../opencv/build/modules/cudawarping && /usr/bin/cmake -E cmake_symlink_library ../../lib/libopencv_cudawarping.so.3.2.0 ../../lib/libopencv_cudawarping.so.3.2 ../../lib/libopencv_cudawarping.so
make[2]: Leaving directory '....../opencv/build'
[ 27%] Built target opencv_cudawarping
make[1]: Leaving directory '....../opencv/build'
Makefile:163: recipe for target 'all' failed
make: *** [all] Error 2

Any further suggestions???

Cheers
Pei

[alalek]

Check files owner in your build directory - it may be a problem if you run sudo make install or similar.

[jiapei100]

@alalek

No, I DONOT use sudo make install
Most of the time, I use cmake in this way:

ccmake ../
make -j8
sudo checkinstall ...

Further suggestions?

[alalek]

sudo breaks access right of your build sub-folders, so regular make command may fail.

I doesn't see error messages from your build log, so there is no other suggestions from me.

[jiapei100]

@alalek Problem solved on one of my desktops...

I believe it has something to do with VTK installed previously, which is based on a previous OpenCV. After upgrading to a new OpenCV, the dependencies messed up somewhere...

Thanks anyway
And, happy NEW year.

[jiapei100]

In addition, I did notice that cnn_3dobj relies on viz, which relies on VTK, which is supposed to be installed after OpenCV has been installed already... This seems to be a loop..

modules/viz/CMakeLists.txt:2:  ocv_module_disable(viz)
modules/viz/CMakeLists.txt:7:ocv_define_module(viz opencv_core ${VTK_LIBRARIES} WRAP python)

So, turn OFF cnn_3dobj seems to be a must...

Cheers
Pei