Opencv python build windows

Keep OpenCV Free

OpenCV is raising funds to keep the library free for everyone, and we need the support of the entire community to do it. Donate to OpenCV on Github to show your support.

• OpenCV on Wheels
  • Installation and Usage
• Frequently Asked Questions
• Documentation for opencv-python
  • CI build process
  • Manual builds
    • Manual debug builds
    • Source distributions
  • Licensing
  • Versioning
  • Releases
  • Development builds
  • Manylinux wheels
  • Supported Python versions
  • Backward compatibility

OpenCV on Wheels

Pre-built CPU-only OpenCV packages for Python.

Check the manual build section if you wish to compile the bindings from source to enable additional modules such as CUDA.

Installation and Usage

1. If you have previous/other manually installed (= not installed via pip) version of OpenCV installed (e.g. cv2 module in the root of Python’s site-packages), remove it before installation to avoid conflicts.

2. Make sure that your pip version is up-to-date (19.3 is the minimum supported version): pip install --upgrade pip. Check version with pip -V. For example Linux distributions ship usually with very old pip versions which cause a lot of unexpected problems especially with the manylinux format.

3. Select the correct package for your environment:

   There are four different packages (see options 1, 2, 3 and 4 below) and you should SELECT ONLY ONE OF THEM. Do not install multiple different packages in the same environment. There is no plugin architecture: all the packages use the same namespace (cv2). If you installed multiple different packages in the same environment, uninstall them all with pip uninstall and reinstall only one package.

   a. Packages for standard desktop environments (Windows, macOS, almost any GNU/Linux distribution)

   • Option 1 — Main modules package: pip install opencv-python
   • Option 2 — Full package (contains both main modules and contrib/extra modules): pip install opencv-contrib-python (check contrib/extra modules listing from OpenCV documentation)

   b. Packages for server (headless) environments (such as Docker, cloud environments etc.), no GUI library dependencies

   These packages are smaller than the two other packages above because they do not contain any GUI functionality (not compiled with Qt / other GUI components). This means that the packages avoid a heavy dependency chain to X11 libraries and you will have for example smaller Docker images as a result. You should always use these packages if you do not use cv2.imshow et al. or you are using some other package (such as PyQt) than OpenCV to create your GUI.

   • Option 3 — Headless main modules package: pip install opencv-python-headless
   • Option 4 — Headless full package (contains both main modules and contrib/extra modules): pip install opencv-contrib-python-headless (check contrib/extra modules listing from OpenCV documentation)

4. Import the package:

   import cv2

   All packages contain Haar cascade files. cv2.data.haarcascades can be used as a shortcut to the data folder. For example:

   cv2.CascadeClassifier(cv2.data.haarcascades + "haarcascade_frontalface_default.xml")

5. Read OpenCV documentation

6. Before opening a new issue, read the FAQ below and have a look at the other issues which are already open.

Frequently Asked Questions

Q: Do I need to install also OpenCV separately?

A: No, the packages are special wheel binary packages and they already contain statically built OpenCV binaries.

Q: Pip install fails with ModuleNotFoundError: No module named 'skbuild'?

Since opencv-python version 4.3.0.*, manylinux1 wheels were replaced by manylinux2014 wheels. If your pip is too old, it will try to use the new source distribution introduced in 4.3.0.38 to manually build OpenCV because it does not know how to install manylinux2014 wheels. However, source build will also fail because of too old pip because it does not understand build dependencies in pyproject.toml. To use the new manylinux2014 pre-built wheels (or to build from source), your pip version must be >= 19.3. Please upgrade pip with pip install --upgrade pip.

Q: Import fails on Windows: ImportError: DLL load failed: The specified module could not be found.?

A: If the import fails on Windows, make sure you have Visual C++ redistributable 2015 installed. If you are using older Windows version than Windows 10 and latest system updates are not installed, Universal C Runtime might be also required.

Windows N and KN editions do not include Media Feature Pack which is required by OpenCV. If you are using Windows N or KN edition, please install also Windows Media Feature Pack.

If you have Windows Server 2012+, media DLLs are probably missing too; please install the Feature called «Media Foundation» in the Server Manager. Beware, some posts advise to install «Windows Server Essentials Media Pack», but this one requires the «Windows Server Essentials Experience» role, and this role will deeply affect your Windows Server configuration (by enforcing active directory integration etc.); so just installing the «Media Foundation» should be a safer choice.

If the above does not help, check if you are using Anaconda. Old Anaconda versions have a bug which causes the error, see this issue for a manual fix.

If you still encounter the error after you have checked all the previous solutions, download Dependencies and open the cv2.pyd (located usually at C:\Users\username\AppData\Local\Programs\Python\PythonXX\Lib\site-packages\cv2) file with it to debug missing DLL issues.

Q: I have some other import errors?

A: Make sure you have removed old manual installations of OpenCV Python bindings (cv2.so or cv2.pyd in site-packages).

Q: Function foo() or method bar() returns wrong result, throws exception or crashes interpreter. What should I do?

A: The repository contains only OpenCV-Python package build scripts, but not OpenCV itself. Python bindings for OpenCV are developed in official OpenCV repository and it’s the best place to report issues. Also please check OpenCV wiki and the official OpenCV forum before file new bugs.

Q: Why the packages do not include non-free algorithms?

A: Non-free algorithms such as SURF are not included in these packages because they are patented / non-free and therefore cannot be distributed as built binaries. Note that SIFT is included in the builds due to patent expiration since OpenCV versions 4.3.0 and 3.4.10. See this issue for more info: https://github.com/skvark/opencv-python/issues/126

Q: Why the package and import are different (opencv-python vs. cv2)?

A: It’s easier for users to understand opencv-python than cv2 and it makes it easier to find the package with search engines. cv2 (old interface in old OpenCV versions was named as cv) is the name that OpenCV developers chose when they created the binding generators. This is kept as the import name to be consistent with different kind of tutorials around the internet. Changing the import name or behaviour would be also confusing to experienced users who are accustomed to the import cv2.

Documentation for opencv-python

The aim of this repository is to provide means to package each new OpenCV release for the most used Python versions and platforms.

CI build process

The project is structured like a normal Python package with a standard setup.py file.
The build process for a single entry in the build matrices is as follows (see for example .github/workflows/build_wheels_linux.yml file):

1. In Linux and MacOS build: get OpenCV’s optional C dependencies that we compile against

2. Checkout repository and submodules

   • OpenCV is included as submodule and the version is updated
     manually by maintainers when a new OpenCV release has been made
   • Contrib modules are also included as a submodule

3. Find OpenCV version from the sources

4. Build OpenCV

   • tests are disabled, otherwise build time increases too much
   • there are 4 build matrix entries for each build combination: with and without contrib modules, with and without GUI (headless)
   • Linux builds run in manylinux Docker containers (CentOS 5)
   • source distributions are separate entries in the build matrix

5. Rearrange OpenCV’s build result, add our custom files and generate wheel

6. Linux and macOS wheels are transformed with auditwheel and delocate, correspondingly

7. Install the generated wheel

8. Test that Python can import the library and run some sanity checks

9. Use twine to upload the generated wheel to PyPI (only in release builds)

Steps 1—4 are handled by pip wheel.

The build can be customized with environment variables. In addition to any variables that OpenCV’s build accepts, we recognize:

• CI_BUILD. Set to 1 to emulate the CI environment build behaviour. Used only in CI builds to force certain build flags on in setup.py. Do not use this unless you know what you are doing.
• ENABLE_CONTRIB and ENABLE_HEADLESS. Set to 1 to build the contrib and/or headless version
• ENABLE_JAVA, Set to 1 to enable the Java client build. This is disabled by default.
• CMAKE_ARGS. Additional arguments for OpenCV’s CMake invocation. You can use this to make a custom build.

See the next section for more info about manual builds outside the CI environment.

Manual builds

If some dependency is not enabled in the pre-built wheels, you can also run the build locally to create a custom wheel.

1. Clone this repository: git clone --recursive https://github.com/opencv/opencv-python.git
2. cd opencv-python
   • you can use git to checkout some other version of OpenCV in the opencv and opencv_contrib submodules if needed
3. Add custom Cmake flags if needed, for example: export CMAKE_ARGS="-DSOME_FLAG=ON -DSOME_OTHER_FLAG=OFF" (in Windows you need to set environment variables differently depending on Command Line or PowerShell)
4. Select the package flavor which you wish to build with ENABLE_CONTRIB and ENABLE_HEADLESS: i.e. export ENABLE_CONTRIB=1 if you wish to build opencv-contrib-python
5. Run pip wheel . --verbose. NOTE: make sure you have the latest pip version, the pip wheel command replaces the old python setup.py bdist_wheel command which does not support pyproject.toml.
   • this might take anything from 5 minutes to over 2 hours depending on your hardware
6. Pip will print fresh wheel location at the end of build procedure. If you use old approach with setup.py file wheel package will be placed in dist folder. Package is ready and you can do with that whatever you wish.
   • Optional: on Linux use some of the manylinux images as a build hosts if maximum portability is needed and run auditwheel for the wheel after build
   • Optional: on macOS use delocate (same as auditwheel but for macOS) for better portability

Manual debug builds

In order to build opencv-python in an unoptimized debug build, you need to side-step the normal process a bit.

1. Install the packages scikit-build and numpy via pip.
2. Run the command python setup.py bdist_wheel --build-type=Debug.
3. Install the generated wheel file in the dist/ folder with pip install dist/wheelname.whl.

If you would like the build produce all compiler commands, then the following combination of flags and environment variables has been tested to work on Linux:

export CMAKE_ARGS='-DCMAKE_VERBOSE_MAKEFILE=ON'
export VERBOSE=1

python3 setup.py bdist_wheel --build-type=Debug

See this issue for more discussion: https://github.com/opencv/opencv-python/issues/424

Source distributions

Since OpenCV version 4.3.0, also source distributions are provided in PyPI. This means that if your system is not compatible with any of the wheels in PyPI, pip will attempt to build OpenCV from sources. If you need a OpenCV version which is not available in PyPI as a source distribution, please follow the manual build guidance above instead of this one.

You can also force pip to build the wheels from the source distribution. Some examples:

• pip install --no-binary opencv-python opencv-python
• pip install --no-binary :all: opencv-python

If you need contrib modules or headless version, just change the package name (step 4 in the previous section is not needed). However, any additional CMake flags can be provided via environment variables as described in step 3 of the manual build section. If none are provided, OpenCV’s CMake scripts will attempt to find and enable any suitable dependencies. Headless distributions have hard coded CMake flags which disable all possible GUI dependencies.

On slow systems such as Raspberry Pi the full build may take several hours. On a 8-core Ryzen 7 3700X the build takes about 6 minutes.

Licensing

Opencv-python package (scripts in this repository) is available under MIT license.

OpenCV itself is available under Apache 2 license.

Third party package licenses are at LICENSE-3RD-PARTY.txt.

All wheels ship with FFmpeg licensed under the LGPLv2.1.

Non-headless Linux wheels ship with Qt 5 licensed under the LGPLv3.

The packages include also other binaries. Full list of licenses can be found from LICENSE-3RD-PARTY.txt.

Versioning

find_version.py script searches for the version information from OpenCV sources and appends also a revision number specific to this repository to the version string. It saves the version information to version.py file under cv2 in addition to some other flags.

Releases

A release is made and uploaded to PyPI when a new tag is pushed to master branch. These tags differentiate packages (this repo might have modifications but OpenCV version stays same) and should be incremented sequentially. In practice, release version numbers look like this:

cv_major.cv_minor.cv_revision.package_revision e.g. 3.1.0.0

The master branch follows OpenCV master branch releases. 3.4 branch follows OpenCV 3.4 bugfix releases.

Development builds

Every commit to the master branch of this repo will be built. Possible build artifacts use local version identifiers:

cv_major.cv_minor.cv_revision+git_hash_of_this_repo e.g. 3.1.0+14a8d39

These artifacts can’t be and will not be uploaded to PyPI.

Manylinux wheels

Linux wheels are built using manylinux2014. These wheels should work out of the box for most of the distros (which use GNU C standard library) out there since they are built against an old version of glibc.

The default manylinux2014 images have been extended with some OpenCV dependencies. See Docker folder for more info.

Supported Python versions

Python 3.x compatible pre-built wheels are provided for the officially supported Python versions (not in EOL):

• 3.7
• 3.8
• 3.9
• 3.10
• 3.11
• 3.12

Backward compatibility

Starting from 4.2.0 and 3.4.9 builds the macOS Travis build environment was updated to XCode 9.4. The change effectively dropped support for older than 10.13 macOS versions.

Starting from 4.3.0 and 3.4.10 builds the Linux build environment was updated from manylinux1 to manylinux2014. This dropped support for old Linux distributions.

Starting from version 4.7.0 the Mac OS GitHub Actions build environment was update to version 11. Mac OS 10.x support deprecated. See https://github.com/actions/runner-images/issues/5583

Starting from version 4.9.0 the Mac OS GitHub Actions build environment was update to version 12. Mac OS 10.x support deprecated by Brew and most of used packages.

Last Updated :
07 Aug, 2024

Prerequisite: Python Language Introduction   OpenCV is the huge open-source library for computer vision, machine learning, and image processing and now it plays a major role in real-time operation which is very important in today’s systems. By using it, one can process images and videos to identify objects, faces, or even the handwriting of a human. When it integrated with various libraries, such as Numpy, python is capable of processing the OpenCV array structure for analysis. To Identify image patterns and its various features we use vector space and perform mathematical operations on these features. To install OpenCV, one must have Python and PIP, preinstalled on their system. To check if your system already contains Python, go through the following instructions: Open the Command line(search for cmd in the Run dialog( + R). Now run the following command:

python --version

If Python is already installed, it will generate a message with the Python version available.

Картинка с сайта: www.geeksforgeeks.org

If Python is not present, go through How to install Python on Windows? and follow the instructions provided.   PIP is a package management system used to install and manage software packages/libraries written in Python. These files are stored in a large “on-line repository” termed as Python Package Index (PyPI). To check if PIP is already installed on your system, just go to the command line and execute the following command:

pip -V

Картинка с сайта: www.geeksforgeeks.org

If PIP is not present, go through How to install PIP on Windows? and follow the instructions provided.

Downloading and Installing OpenCV:

OpenCV can be directly downloaded and installed with the use of pip (package manager). To install OpenCV, just go to the command-line and type the following command:

pip install opencv-python

Beginning with the installation:

To check if OpenCV is correctly installed, just run the following commands to perform a version check:

python
>>>import cv2
>>>print(cv2.__version__)

Картинка с сайта: www.geeksforgeeks.org

Introduction and Overview

This document contains instructions for installing and migrating to the latest release of OpenCV (version 3) and the Python bindings. In addition to some API changes, there are also changes to the Python interface (e.g. removal of cv2.cv) that may require changes in existing code to work with the new version of the cv2 module. After installation, see the next page, Migration & Changes, for details regarding the changes to the module and for help migrating existing Python code to the new version.

This page covers installing OpenCV 3 on Windows (using pre-built binaries) and Linux (compiled from source), including the Python interface (the cv2 module). OpenCV can be downloaded from the official OpenCV website. Note that this guide is written based on OpenCV version 3.1.0. After installation, it is recommended that you can check the version of OpenCV that Python is using:

import cv2
print cv2.__version__

# Should print 3.0.0-rc1 or newer.

Note that although OpenCV 3 is the latest release, by convention, the module is still named cv2.

Installing on Windows (pre-built binaries)

Using pre-built binaries is the quickest way to get a Python OpenCV environment up and running on Windows. Currently, only the Python 2 version of the cv2 module is built and included in the latest Windows release. Support for Python 3 (as well as adding other non-standard features/modules), requires compiling from source — see the official OpenCV documentation for details.

Downloading OpenCV and Python

To begin, download OpenCV for Windows (version 3.1, opencv-3.1.0.exe), and extract it to a directory of your choice. The Windows build includes both a 32-bit and 64-bit module for Python 2.7.

Before continuing, ensure that you have a working Python 2.7 installation, which can be downloaded from the Python website. This guide was tested using Python 2.7.9 (x86 installer / x64 installer). When installing, it is recommended that you allow the installer to add Python to your PATH environment variable, so you can run python and pip from a command prompt.

Installing Module Dependencies

The Python OpenCV cv2 module requires both the NumPy and SciPy stack. To get this, first ensure that you have a working python environment, and have pip installed (if not, use the links above, and ensure the Python folder is in your PATH variable).

Christoph Gohlke currently provides some (unofficial) pre-built Python packages for Windows, including NumPy and SciPy. Download the latest stable versions of NumPy and SciPy, and install them by calling pip install [module].whl from a command prompt. Note that you should download the version corresponding to your Python environment (2.7 in this case) and word length. For example, on 32-bit systems/interpreters:

pip install numpy-1.9.2+mkl-cp27-none-win32.whl
pip install scipy-0.15.1-cp27-none-win32.whl

And for 64-bit systems/interpreters:

pip install numpy-1.9.2+mkl-cp27-none-win_amd64.whl
pip install scipy-0.15.1-cp27-none-win_amd64.whl

After installing, you should be able to run import numpy and import scipy from a Python interpreter to verify that the modules were installed correctly. You can check what versions of NumPy and SciPy are installed from numpy.__version__ and scipy.__version__, respectively.

Installing Python-OpenCV Module

Lastly, we need to copy the OpenCV module into the local Python packages. In the files extracted from opencv-3.0.0-rc1.exe, go to the folder opencv\build\python\2.7\, and open either the x86\ (for 32-bit) or x64\ (for 64-bit) folder. In each, there will be a cv2.pyd file.

Copy cv2.pyd directly into the Lib\site-packages\ directory of your Python installation. For most users, this will be found at:

C:\Python27\Lib\site-packages 

Lastly, ensure that the OpenCV .dll files are somewhere in your system’s %PATH%, especially the opencv_ffmpeg310.dll (or opencv_ffmpeg310_64.dll for 64-bit systems) file required for loading video files.

Continue on to the Verifying Installation section to ensure everything was installed correctly, and the new version of OpenCV is being used. Also be sure to check out the next page, Migration & Changes, for details about the changes to the module and updating existing code.

Installing on Linux (compiling from source)

Although this guide is written for Ubuntu 12.04/14.04 and other variants (e.g. Xubuntu/Lubuntu/Kubuntu), the process should largely be the same on other versions, as well as similar Debian-like Linux distributions. Pay attention to the output of each command to ensure everything worked correctly, and if there are any issues during the build process, see the bottom of this document for possible mitigations

Downloading OpenCV

To begin, download OpenCV for Linux (version 3.1.0, opencv-3.1.0.zip), and extract it to a directory of your choice (e.g. ~/opencv-src). Create a build folder inside the folder where the archive was extracted (the directory containing the CMakeLists.txt file), and open a terminal session there. For example:

# Assuming the files were extracted to ~/opencv-src/...
cd ~/opencv-src
mkdir build
cd build

Execute all of following commands from the build sub-folder itself, so the compiled files will be placed there.

Installing Build Dependencies

To compile OpenCV, we must ensure that the required dependencies are available, including the build tools themselves. We can get the required ones using apt on Ubuntu, but first, ensure the package list is up-to-date by running apt-get update. Next, execute the following commands to get the required packages (see below for a one-line list of all):

sudo apt-get install cmake build-essential pkg-config

sudo apt-get install libgtk2.0-dev libtbb-dev

sudo apt-get install python-dev python-numpy python-scipy

sudo apt-get install libjasper-dev libjpeg-dev libpng-dev libtiff-dev 

sudo apt-get install libavcodec-dev libavutil-dev libavformat-dev libswscale-dev

sudo apt-get install libdc1394-22-dev libv4l-dev

In single line:

sudo apt-get install cmake build-essential pkg-config libgtk2.0-dev libtbb-dev python-dev python-numpy python-scipy libjasper-dev libjpeg-dev libpng-dev libtiff-dev libavcodec-dev libavutil-dev libavformat-dev libswscale-dev libdc1394-22-dev libv4l-dev

Note that for additional OpenCV modules/features (e.g. GPU/CUDA support, or Python 3 module), you will need to download the respective development/SDK packages for those libraries as well. The dependencies listed above only cover building OpenCV itself and the Python 2.7 cv2 module.

Compiling and Installing OpenCV

Now that we have the required build dependencies, run cmake (again, in the build/ directory we created) to generate the Makefile to build OpenCV:

cmake -D CMAKE_BUILD_TYPE=Release -D CMAKE_INSTALL_PREFIX=/usr/local ..

If there are any errors, ensure that you downloaded all the required packages — the output should help track down what is missing. To ensure the Python module will be built, you should see python2 in the list of configured modules after running cmake:

[...]

--     Linker flags (Release):
--     Linker flags (Debug):
--     Precompiled headers:      YES
--
--   OpenCV modules:
--     To be built:              hal core flann imgproc ml [...] python2

[...]

If you do not see the python2 module listed, in the «To be built» list, check that you have the proper Python development packages installed, remove all files from the build/ folder, and try running the cmake command again.

Now, we can build OpenCV (and the Python module) using make, and install it to our system:

make
sudo make install

Ensure that the build was successful after calling make, and check the output before installing. If you run into build issues/errors, again ensure that you have all the required dependencies and header files on your system. If there are actual build issues with OpenCV itself, see the Linux Build Issues section below for some possible workarounds.

When installing OpenCV 3.0-rc1 on Ubuntu 12.04, I ran into build errors regarding some missing codec #define entries. As mentioned, the steps to do this are detailed below should you run into the same problem (the problem should be solved in version 3.1.0 and above). Ubuntu 14.04 and above should build successfully without requiring any modifications.

If the build was successful, but you can’t import cv2 from a Python shell after running make install, you can install the module manually by copying the cv2.so file we just built in the build/lib/ folder to /usr/local/lib/python2.7/dist-packages/. From the build/ folder, this can be done by:

sudo cp lib/cv2.so /usr/local/lib/python2.7/dist-packages/

After this step, the cv2 module is installed, and can now be imported by your Python environment. Continue on to the Verifying Installation section to ensure everything was installed correctly, and more importantly, that the correct version of OpenCV is being used by Python. You can also check out the next page, Migration & Changes, for details about changes to the cv2 module, and what changes need to be made to existing code to run with the updated module.

Verifying Installation

As mentioned previously, you can verify that the cv2 module was installed correctly by executing the following in a Python shell:

import cv2
print cv2.__version__

If the correct version string is printed (3.0.0-rc1 or newer if you used a more recent version), everything is installed correctly at this point!

Note that there are some major changes to the cv2 module hierarchy itself, and Python programs written with OpenCV 2.4.x in mind may not work properly anymore. See the next section, Migration & Changes for details about the changes, and how to modify programs to deal with the changes.

Linux Build Issues

On some systems you may run into issues when compiling OpenCV itself, depending on what packages are available. If the proper header files are available, but build issues still arise, try to see if there are any workarounds below for the problem.

Undefined AV_CODEC_ID_... in ffmpeg_codecs.hpp

If you run into an issue where build errors are caused by certain codecs not being defined, you can download a more recent version of ffmpeg_codecs.hpp from here, and replace it in the source code you downloaded/extracted in the modules/videoio/src/ folder.

Alternatively, you can declare them manually by editing the ffmpeg_codecs.hpp file itself by adding the missing codec entry for H263I, and renaming WNV to WNV1. You can see what changes have to be made by viewing this commit from the OpenCV GitHub project.

Once the file is updated, you can call make again from the build/ folder, and OpenCV should finish compiling without error now.

Время на прочтение7 мин

Количество просмотров41K

Введение

В данной статье речь пойдет о сборке и установке OpenCV 4 для C/C++, Python 2 и Python 3 из исходных файлов с дополнительными модулями CUDA 10 на ОС Windows.

Я постарался включить сюда все тонкости и нюансы, с которыми можно столкнуться в ходе установки, и про которые не написано в официальном мануале.

Картинка с сайта: habr.com

Сборка тестировалась для:

• Windows 8.1 + Visual Studio 2017 + Python 2/3 + CUDA 10.0 + GeForce 840m
• Windows 10 + Visual Studio 2019 + Python 2/3 + CUDA 10.0 + GeForce GTX 1060

Внимание! Сборка не подойдет для версии OpenCV 4.0.1 и/или CUDA ниже 10-й версии. CUDA 9 и ниже поддерживается OpenCV 3.

Что потребуется для установки

В моей сборке использовались следующие инструменты:

1. CMake 3.15
2. MS Visual Studio 2019 64-бит + средства CMake С++ для Windows
3. Python 3.7.3 64-бит + NumPy 64-бит
4. Python 2.7.16 64-бит + NumPy 64-бит
5. CUDA 10.0
6. CuDNN 7.6.2
7. OpenCV 4.1.1 и OpenCV-contrib-4.1.1

Установка

Так как установка производится через консольные команды, то следует внимательно и аккуратно выполнять все шаги. Также, при необходимости, меняйте установочные пути на свои.
Для начала необходимо установить требуемое ПО, причем Visual Studio должна быть установлена до CUDA:

• CMake (версия >= 3.9.1)
• MS Visual Studio
• Python 3.7 (дистрибутив Anaconda3)
• CUDA 10.0
• CuDNN 7.6.2

После установки всех компонент убедитесь, что пути для CMake, Visual Studio, Python, CUDA, CuDNN прописаны в переменных PATH, PYTHONPATH, CUDA_PATH и cudnn соответственно.

Далее загружаем архивы исходников opencv-4.1.1 и opencv-contrib-4.1.1 в желаемое место (в моем случае это C:\OpenCV\).

git clone https://github.com/opencv/opencv.git -b "4.1.1"
git clone https://github.com/opencv/opencv_contrib.git -b "4.1.1"

Картинка с сайта: habr.com

Создаем папку build/ внутри opencv-4.1.1.

Картинка с сайта: habr.com

Далее сгенерируем файлы сборки с помощью cmake. Мы будем использовать консольный вариант cmake, так как cmake-gui путает типы некоторых переменных (например, OPENCV_PYTHON3_VERSION) и, как следствие, неправильно генерирует файлы.

Открываем консоль по пути C:\OpenCV\ и прописываем переменные.

set "opencvSource=opencv-4.1.1"
set "opencvExtraModules=opencv_contrib-4.1.1/modules"
set "opencvBuild=%opencvSource%\build"
set "compiler=Visual Studio 16 2019"
set "buildType=Release"

Примечание. Для Visual Studio 2017 генератор пишется как «Visual Studio 15 2017 Win64» и без флага -A.

Также можно явно указать питоновские библиотеки для python 2 и python 3 на случай, если сборщик не сможет найти их автоматически.

set "python2_executable=C:/Python27/python.exe"
set "python2_include_dir=C:/Python27/include"
set "python2_library=C:/Python27/libs/python27.lib"
set "python2_numpy_include_dirs=C:/Python27/Lib/site-packages/numpy/core/include"
set "python2_packages_path=C:/Python27/Lib/site-packages"

set "python3_executable=C:/Users/root/Anaconda3/python.exe"
set "python3_include_dir=C:/Users/root/Anaconda3/include"
set "python3_library=C:/Users/root/Anaconda3/libs/python37.lib"
set "python3_numpy_include_dirs=C:/Users/root/Anaconda3/lib/site-packages/numpy/core/include"
set "python3_packages_path=C:/Users/root/Anaconda3/Lib/site-packages"

Примечание. Обратите внимание на то, что библиотека NumPy должна быть той же разрядности, что и OpenCV. Проверить это легко:

import numpy.distutils.system_info as sysinfo
print(sysinfo.platform_bits)

Выполняем генерацию файлов сборки с помощью длинной команды ниже. При неудачной генерации или наличии ошибок после выполнения команды, повторную генерацию следует производить вычистив все файлы в build/ и .cache/.

cmake ^
-B"%opencvBuild%/" ^
-H"%opencvSource%/" ^
-G"%compiler%" ^
-Ax64 ^
-DCMAKE_BUILD_TYPE=%buildType% ^
-DBUILD_opencv_world=ON ^
-DINSTALL_TESTS=OFF ^
-DINSTALL_C_EXAMPLES=OFF ^
-DBUILD_EXAMPLES=OFF ^
-DOPENCV_EXTRA_MODULES_PATH="%opencvExtraModules%/" ^
-DBUILD_PROTOBUF=ON ^
-DBUILD_opencv_python_bindings_generator=ON ^
-DWITH_CUDA=ON ^
-DCUDA_FAST_MATH=ON ^
-DWITH_CUBLAS=ON ^
-DCUDA_ARCH_PTX=7.5 ^
-DBUILD_opencv_python2=ON ^
-DPYTHON2_EXECUTABLE="%python2_executable%" ^
-DOPENCV_PYTHON2_VERSION=2.7.16 ^
-DPYTHON2_INCLUDE_DIR="%python2_include_dir%" ^
-DPYTHON2_LIBRARY="%python2_library%" ^
-DPYTHON2_NUMPY_INCLUDE_DIRS="%python2_numpy_include_dirs%" ^
-DPYTHON2_PACKAGES_PATH="%python2_packages_path%" ^
-DBUILD_opencv_python3=ON ^
-DPYTHON3_EXECUTABLE="%python3_executable%" ^
-DOPENCV_PYTHON3_VERSION=3.7.3 ^
-DPYTHON3_INCLUDE_DIR="%python3_include_dir%" ^
-DPYTHON3_LIBRARY="%python3_library%" ^
-DPYTHON3_NUMPY_INCLUDE_DIRS="%python3_numpy_include_dirs%" ^
-DPYTHON3_PACKAGES_PATH="%python3_packages_path%"

Спустя примерно 10 минут в консоле должна появиться информация о сборке и завершающие строки «Configuring done» и «Generating done». Проверяем всю информацию, особенно разделы NVIDIA CUDA, Python 2, Python 3.

Картинка с сайта: habr.com

Далее собираем решение. На сборку может уйти несколько часов в зависимости от вашего процессора и версии Visual Studio.

cmake --build %opencvBuild% --target ALL_BUILD --config Release

Устанавливаем решение.

cmake --build %opencvBuild% --target INSTALL --config Release

После успешной установки создаем системную переменную OPENCV_DIR со значением C:\OpenCV\opencv-4.1.1\build\install\x64\vc15\bin и также добавляем ее в PATH.

Тест

Проверим работоспособность OpenCV с модулем CUDA на простом примере умножения матриц.

Пример на Python 3

import numpy as np
import cv2 as cv
import time

rand = np.random.random((1024, 1024)).astype(np.float32)
h_array1 = np.stack([rand, rand],axis=2)
h_array2 = h_array1
d_array1 = cv.cuda_GpuMat()
d_array2 = cv.cuda_GpuMat()
d_array1.upload(h_array1)
d_array2.upload(h_array2)

start = time.time()
cv.cuda.gemm(d_array1, d_array2, 1, None, 0, None, 1)
end = time.time()
print("Time elapsed:", end - start, "sec")

Вывод в консоли

Time elapsed: 0.3130002021789551

Пример на C++

#include <iostream>
#include <opencv2/opencv.hpp>
#include <opencv2/cudaarithm.hpp>

using namespace std;
using namespace cv;
using namespace cv::cuda;

int main() {

	Mat h_array1 = Mat::ones(1024, 1024, CV_32FC2);
	Mat h_array2 = Mat::ones(1024, 1024, CV_32FC2);
	Mat h_array3 = Mat::zeros(1024, 1024, CV_32FC2);
	Mat h_result;
	GpuMat d_array1, d_array2, d_array3, d_result;

	d_array1.upload(h_array1);
	d_array2.upload(h_array2);

	const clock_t begin_time = clock();
	cuda::gemm(d_array1, d_array2, 1.0, d_array3, 0.0, d_result);
	cout << "Time elapsed: " << (float(clock() - begin_time) / CLOCKS_PER_SEC) << " sec" << endl;

	d_result.download(h_result);

	return 0;
}

Вывод в консоли

Time elapsed: 0.354

Удаление

Чтобы удалить OpenCV, нужно выполнить команду.

cmake --build %opencvBuild% --target uninstall --config Release

и удалить системную переменную OPENCV_DIR и убрать путь до OpenCV из PATH.

Заключение

В статье мы рассмотрели установку OpenCV 4 для ОС Windows 10. Данный алгоритм тестировался на Windows 8.1 и Windows 10, но, в теории, может собираться и на Windows 7. За дополнительной информацией можно обратиться к списку источников ниже.

P.S. Напоследок сведем все установочные команды в один .bat файл для тех, кто хочет автоматизировать установку.

cd C:\OpenCV

git clone https://github.com/opencv/opencv_contrib.git -b "4.1.1"
git clone https://github.com/opencv/opencv.git -b "4.1.1"

ren opencv opencv-4.1.1
ren opencv_contrib-4.1.1

set "opencvSource=opencv-4.1.1"
set "opencvExtraModules=opencv_contrib-4.1.1/modules"
set "opencvBuild=%opencvSource%\build"
set "compiler=Visual Studio 16 2019"
set "buildType=Release"

set "python2_executable=C:/Python27/python.exe"
set "python2_include_dir=C:/Python27/include"
set "python2_library=C:/Python27/libs/python27.lib"
set "python2_numpy_include_dirs=C:/Python27/Lib/site-packages/numpy/core/include"
set "python2_packages_path=C:/Python27/Lib/site-packages"

set "python3_executable=C:/Users/root/Anaconda3/python.exe"
set "python3_include_dir=C:/Users/root/Anaconda3/include"
set "python3_library=C:/Users/root/Anaconda3/libs/python37.lib"
set "python3_numpy_include_dirs=C:/Users/root/Anaconda3/lib/site-packages/numpy/core/include"
set "python3_packages_path=C:/Users/root/Anaconda3/Lib/site-packages"

cmake ^
-B"%opencvBuild%/" ^
-H"%opencvSource%/" ^
-G"%compiler%" ^
-Ax64 ^
-DCMAKE_BUILD_TYPE=%buildType% ^
-DBUILD_opencv_world=ON ^
-DINSTALL_TESTS=OFF ^
-DINSTALL_C_EXAMPLES=OFF ^
-DBUILD_EXAMPLES=OFF ^
-DOPENCV_EXTRA_MODULES_PATH="%opencvExtraModules%/" ^
-DBUILD_PROTOBUF=ON ^
-DBUILD_opencv_python_bindings_generator=ON ^
-DWITH_CUDA=ON ^
-DCUDA_FAST_MATH=ON ^
-DWITH_CUBLAS=ON ^
-DCUDA_ARCH_PTX=7.5 ^
-DBUILD_opencv_python2=ON ^
-DPYTHON2_EXECUTABLE="%python2_executable%" ^
-DOPENCV_PYTHON2_VERSION=2.7.16 ^
-DPYTHON2_INCLUDE_DIR="%python2_include_dir%" ^
-DPYTHON2_LIBRARY="%python2_library%" ^
-DPYTHON2_NUMPY_INCLUDE_DIRS="%python2_numpy_include_dirs%" ^
-DPYTHON2_PACKAGES_PATH="%python2_packages_path%" ^
-DBUILD_opencv_python3=ON ^
-DPYTHON3_EXECUTABLE="%python3_executable%" ^
-DOPENCV_PYTHON3_VERSION=3.7.3 ^
-DPYTHON3_INCLUDE_DIR="%python3_include_dir%" ^
-DPYTHON3_LIBRARY="%python3_library%" ^
-DPYTHON3_NUMPY_INCLUDE_DIRS="%python3_numpy_include_dirs%" ^
-DPYTHON3_PACKAGES_PATH="%python3_packages_path%"

cmake --build %opencvBuild% --target ALL_BUILD --config Release
cmake --build %opencvBuild% --target INSTALL --config Release

Список источников

1. docs.opencv.org/4.1.1/d3/d52/tutorial_windows_install.html
2. www.learnopencv.com/install-opencv-4-on-windows
3. jamesbowley.co.uk/build-opencv-4-0-0-with-cuda-10-0-and-intel-mkl-tbb-in-windows
4. lightbuzz.com/opencv-cuda
5. Глава «Building libraries from Source» книги Hands-on GPU-Accelerated Computer Vision with OpenCV and CUDA