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](https://github.com/opencv/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: opencv#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):

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

1. 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

2. Find OpenCV version from the sources

3. 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

4. Rearrange OpenCV's build result, add our custom files and generate wheel

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

6. Install the generated wheel

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

8. 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. You'll have the wheel file in the dist folder 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: opencv#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):

• 3.6
• 3.7
• 3.8
• 3.9
• 3.10

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.