2×1=10

Don’t judge me, I know how this title sounds. The harsh reality is — as per writing of this post — that I always have a hard time getting CMake to recognize the right Python path when I’m using an Anaconda environment. In theory it’s just

git clone https://github.com/opencv/opencv.git opencv
mkdir opencv/build && cd opencv/build
cmake -DCMAKE_INSTALL_PREFIX=/usr/local  ..
make -j8
(sudo) make install

and there you go. It will happily pick up your environment’s Python installation and build that for you. Of course, it looks more along the lines of

cmake \
    -DCMAKE_INSTALL_PREFIX="/usr/local" \
    -DOPENCV_EXTRA_MODULES_PATH="../opencv_contrib/modules" \
    -DBUILD_DOCS=OFF \
    -DBUILD_TESTS=OFF \
    -DBUILD_EXAMPLES=OFF \
    -DBUILD_PERF_TESTS=OFF \
    -DBUILD_opencv_dnn=OFF \
    -DENABLE_FAST_MATH=ON \
    -DWITH_OPENMP=ON \
    -DWITH_TBB=ON \
    -DMKL_WITH_TBB=ON \
    -DMKL_WITH_OPENMP=ON \
    -DCMAKE_CXX_COMPILER="/usr/bin/g++-5" \
    -DCMAKE_C_COMPILER="/usr/bin/gcc-5" \
    -DCUDA_HOST_COMPILER="/usr/bin/gcc-5" \
    -DCUDA_FAST_MATH=ON \
    -DCUDA_ARCH_BIN="5.2" \
    -DWITH_CUBLAS=ON \
    ..

Whenever you then add

conda activate my-environment

to the mix though, all hell breaks loose and you end up with partially configured Python 2 and no Python 3 support at all. The trick seems to be not to rely on OpenCV’s standard Python configuration values

PYTHON3_LIBRARY
PYTHON3_EXECUTABLE
PYTHON3_INCLUDE_DIR
PYTHON3_INCLUDE_DIR2
PYTHON3_NUMPY_INCLUDE_DIRS

but rather to use the seemingly undocumented values

PYTHON3_LIBRARIES
PYTHON3_INCLUDE_PATH

as well, giving it the nice appearance of

cmake \
    -DCMAKE_BUILD_TYPE=RELEASE \
    -DCMAKE_INSTALL_PREFIX="/your/anaconda3" \
    -DOPENCV_EXTRA_MODULES_PATH="../opencv_contrib/modules" \
    -DBUILD_DOCS=OFF \
    -DBUILD_TESTS=OFF \
    -DBUILD_EXAMPLES=OFF \
    -DBUILD_PERF_TESTS=OFF \
    -DBUILD_opencv_dnn=ON \
    -DTINYDNN_USE_NNPACK=OFF \
    -DTINYDNN_USE_TBB=ON \
    -DTINYDNN_USE_OMP=ON \
    -DENABLE_FAST_MATH=ON \
    -DWITH_OPENMP=ON \
    -DWITH_TBB=ON \
    -DMKL_WITH_TBB=ON \
    -DMKL_WITH_OPENMP=ON \
    -DCMAKE_CXX_COMPILER="/usr/bin/g++-5" \
    -DCMAKE_C_COMPILER="/usr/bin/gcc-5" \
    -DCUDA_HOST_COMPILER="/usr/bin/gcc-5" \
    -DCUDA_FAST_MATH=ON \
    -DCUDA_ARCH_BIN="5.2" \
    -DWITH_CUBLAS=ON \
    -DBUILD_opencv_python2=OFF \
    -DPYTHON_EXECUTABLE="/your/anaconda3/bin/python3" \
    -DPYTHON_LIBRARY="/your/anaconda3/lib/libpython3.6m.so" \
    -DPYTHON3_LIBRARY="/your/anaconda3/lib/libpython3.6m.so" \
    -DPYTHON3_EXECUTABLE="/your/anaconda3/bin/python3" \
    -DPYTHON3_INCLUDE_DIR="/your/anaconda3/include/python3.6m" \
    -DPYTHON3_INCLUDE_DIR2="/your/anaconda3/include" \
    -DPYTHON3_NUMPY_INCLUDE_DIRS="/your/anaconda3/lib/python3.6/site-packages/numpy/core/include" \
    -DPYTHON3_INCLUDE_PATH="/your/anaconda3/include/python3.6m" \
    -DPYTHON3_LIBRARIES="/your/anaconda3/lib/libpython3.6m.so" \
    -DHDF5_C_LIBRARY_z="/your/anaconda3/lib/libz.so" \
    ..

And then it’ll do what you want it to do. Because I’m tired of retrying everytime, on github.com/sunsided/opencv-cmake is my repo with support for the OpenCV Extras Module and some documentation for this.

So, what paths go in there? Well, this blog post has a very flashy way of finding out.

For CMAKE_INSTALL_PREFIX you use

python3 -c "import sys; print(sys.prefix)"

so it will install OpenCV directly into your Anaconda installation.
For PYTHON3_EXECUTABLE you call

which python3

and PYTHON3_INCLUDE_DIR is given by

python3 -c "from distutils.sysconfig import get_python_inc; print(get_python_inc())"

Finally, a PYTHON3_PACKAGES_PATH can be found using

python3 -c "from distutils.sysconfig import get_python_lib; print(get_python_lib())"

which you a can use to determine the PYTHON3_NUMPY_INCLUDE_DIRS, although CMake strangely did that part right for me.