Working in the field of Computer Vision for a decade, I have been using popular application frameworks to help me accomplish complex tasks, such as image processing, object tracking, face detection, and more.
Two of these tools are OpenCV and CUDA.
- OpenCV is a set of programming modules that help us accomplish complex real-time computer vision tasks easily. Due to its reliability and open-source philosophy, it’s been one of the most popular computer vision tools.
- CUDA is a parallel computing toolkit that allows us to use the power of an NVidia GPU to significantly accelerate the performance of our applications.
James Bowley has published a detailed performance comparison, where you can see the impact of CUDA on OpenCV.
However, the official OpenCV binaries do not include GPU support out-of-the-box. To harness the full power of your GPU, you’ll need to build the library yourself.
So, the following guide will show you how to compile OpenCV with CUDA support.
For your convenience, I have uploaded the latest stable compiled binaries.
Download OpenCV CUDA binariesRemember: if you download the binaries, share this post to help other developers, too!
Did you know?
As a professional software developer, I have been helping Fortune-500 companies and innovative startups create amazing Computer Vision applications. If you are looking to get your business to the next level, start a project with me.
System Requirements
Hardware
Ensure you have a computer with the following specifications:
- GPU with CUDA support (tested on Nvidia 1060)
- CPU Intel Core i7 (recommended)
Software
Download and install the following software:
- Windows 10 Operating System
- Visual Studio 2015 Community or Professional
- CUDA Toolkit 9.2
- CUDA Toolkit 9.2 Patch 1
- cuDNN 7.1.4
- CMake 3.10.1
Nvidia CUDA download page:
Nvidia cuDNN download page:
Build/Compile OpenCV
Build/Compile OpenCVThis is how to download and compile the latest version of OpenCV with CUDA support.
1) Clone OpenCV
Clone OpenCV to the desired location in your disk:
git clone https://github.com/opencv/opencv.gitIn my case:
C:\Users\Vangos\Projects\OpenCVAlso, switch to switch to the latest stable branch, e.g. “3.4”.
2) Generate the Visual Studio solution
Before going on, ensure that your Path Environment Variable includes the following path:
C:\Program Files (x86)\Windows Kits\8.1\bin\x64
In case the path is not included, add it manually.
Open CMD and type the following command. Replace PATH_TO_SOURCE with the path of the OpenCV installation (e.g. C:\Users\Vangos\Projects\OpenCV).
"C:\Program Files\CMake\bin\cmake.exe" -B"PATH_TO_SOURCE\build" -H"PATH_TO_SOURCE" -G"Visual Studio 14 2015 Win64" -DBUILD_opencv_world=ON -DWITH_CUDA=ON -DCUDA_FAST_MATH=ON -DWITH_CUBLAS=ON -DINSTALL_TESTS=ON -DINSTALL_C_EXAMPLES=ON -DBUILD_EXAMPLES=ONIn my case:
"C:\Program Files\CMake\bin\cmake.exe" -B"C:\Users\Vangos\Projects\OpenCV\build" -H"C:\Users\Vangos\Projects\OpenCV" -G"Visual Studio 14 2015 Win64" -DBUILD_opencv_world=ON -DWITH_CUDA=ON -DCUDA_FAST_MATH=ON -DWITH_CUBLAS=ON -DINSTALL_TESTS=ON -DINSTALL_C_EXAMPLES=ON -DBUILD_EXAMPLES=ON3) Compile!
Navigate to the PATH_TO_SOURCE folder and open the build directory. Find the OpenCV.sln file and open it with Visual Studio 2015.
In the Configuration Manager, select the following:
- Active solution configuration: Release
- Active solution platform: x64
In your Solution Explorer, find the project named INSTALL.
Right-click the INSTALL project and select “Build”.
Then, wait patiently while Visual Studio builds the project. Depending on your computer specs, it may take lots of minutes or even hours!
4) Verify the installation
When the compilation process is finished, the binaries should be under the build\install directory.
You can now follow the official OpenCV guide and integrate OpenCV with CUDA support in your own applications!
Running OpenCV with Visual C++
When running your OpenCV projects using Visual Studio, you need to add the following information in your Project Properties window:
C/C++ → General → Additional Include Directories
C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v9.2\includeand
C:\Users\Vangos\Projects\OpenCV\build\install\includeLinker → General → Additional Library Directories
C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v9.2\lib\x64and
C:\Users\Vangos\Projects\OpenCV\build\install\x64\vc14\libLinker → General → Additional Dependencies
opencv_world400.liband
cudnn.libThis is it! You can now use the full power of a GPU-accelerated OpenCV build.
Again, if you just want the precompiled binaries, I have uploaded them for you. Before downloading, please do me a favor and share this post on social media. This way, you are going to help other developers access great resources.
Download OpenCV CUDA binariesSo, did you find this tutorial helpful? How are you going to use OpenCV with GPU CUDA support? Let me know in the comments below!
Before you go…
As a professional software developer, I have been helping Fortune-500 companies and innovative startups create amazing Computer Vision applications. If you are looking to get your business to the next level, start a project with me.
hi,
could you update your binaries from opencv400 and cuda10,
thanks
opencv4.0.1 rather? I meant the latest.
The tutorial helpful,
I’m going to use for OpenCv(contrib->face) LBPH face recognition.
Sure, I’m planning to update the blog post with the latest release in the next weeks!
hi,
I successfully compiled(core2duo processor) with 0 failed,
opencv_401_with_cuda BUT the file is too large, I mean Opencv_World401.dll = almost 1gb(916mb)
I include this “modules”,
cudev, cudaarithm, cudabgsegm, cudacodec, cudafeatures2d, cudafilters, cudaimgproc, cudalegacy, cudaobjdetect, cudaoptflow, cudastereo, cudawarping,
bioinspired,
face,
dnn_objdetect,
surface_matching,
superres
I have to use cudaobjdetect.hpp for “cuda::haarcascade…”(for face recognition with cuda enabled) that’s why I include “cudaobjdetect”
did I do something wrong? I have no viruses here! or that was it?
This are all the modules I had included:
bioinspired, calib3d, core, cudaarithm, cudabgsegm, cudacodec, cudafeatures2d, cudafilters, cudaimgproc, cudalegacy, cudaobjdetect, cudaoptflow, cudastereo, cudawarping, cudev, dnn, dnn_objdetect, face, features2d, flann, gapi, highgui, imgcodecs, imgproc, java_bindings_generator, ml, objdetect, photo, python_bindings_generator, shape, stitching, superres, surface_matching, video, videoio, world
Hello! The file size is normal. Usually, the file size is bigger than 800MB, so I would consider that normal.
hi,
my problem is building opencv 3.0.0++ or 4.0.0++ with cuda in 32 bit x86, I tried cuda toolkit 6.5.19 32 bit in windows 7 32 bit system, but it wouldn’t work.
any ideas how to build opencv with cuda in 32 bit,
here are the results that I have from cmake 3.13.2,
OpenNI2: YES (ver 2.2.0, build 33)
NVIDIA CUDA: YES (ver 6.5, CUFFT CUBLAS NVCUVID)
NVIDIA GPU arch: 20 30 35 37 50 52 60 61
NVIDIA PTX archs:
I hope you could also include the latest possible opencv and cuda in 32bit binaries with this modules, poor I have only core2duo processor.
here are I want to build in x64(possible) and x86(impossible += ERRORS coming out):
cudev, cudalegacy, cudaarithm, cudaimgproc, cudaobjdetect, cudawarping,
bioinspired, dnn_objdetect, face, stitching, tracking, plot, superres, surface_matching, shape,
calib3d, core, dnn, features2d, flann, highgui, imgcodecs, imgproc, ml, objdetect, photo, video, videoio, videostab,
world
thanks you
Building for 32-bit architecture is has limited support. Not sure how that could be accomplished.
Hi,
I have manage to build OpenCv_World345 and Cuda in 32 Bit in Windows 7 32 bit with MsVs 2013
and Yes I’ve noticed the file size is normal, but I only pick up some modules that are important I think and are listed from my previous reply.
I also included OpenNi2 in-case I would like to explore it.
Actually I’m using it for C++Builder with face recognition(DNN+LBP) in a latest OpenCv,
thank you so much for this tutorial,
Blaise
p.s
it takes hours to build, lol
That’s great to hear! It’s definitely a pain, but worths it — given the performance advantages 🙂
Hi,
I created a very basic project file from sourceforge.net,
here for the curiosity of others how did I do it or used it, that includes cuda version with Opencv _World345
https://sourceforge.net/projects/opencv-world345-cbuilder-proj/
thank you again
blaise
Awesome! Thank you very much for sharing with me!
U can help to you this post
https://habr.com/ru/sandbox/145056/
Thanks a lot. Could you please update the cuda binaries for opencv 4.0.1? I tried but got many strange errors|:
Indeed, I’m planning to update the blog post with the latest OpenCV version in the next weeks!
Very good post! I am curious running Cuda-enabled OpenCV in Qt5. I am working on Qt project with OpenCV3.4.5 in windows, but encountered many compiling errors in building my Qt project. You may check my post at Stack Overflow (https://stackoverflow.com/questions/54130898/cuda-related-compling-errors-in-qt5-with-opencv3-4-5) for these errors. I wonder if you have some ideas about that. Thanks!
Thanks for this work!
I’d love to be able to use CUDA support on OpenCV 4.0… (which is recently out…)
Is there a reason to stick with the 3.4 series?
Hi Ken. The article was written before version 4.0 was out. I’ll publish a new article for OpenCV 4.0 + CUDA.
Have you compiled CUDA with the latest version of OpenCV?
Hi Vangos, thanks a lot for this informative article!
However, I tried opencv_traincascade from your download-binaries. It seems not to use the GPU. GPU-Shark shows no activity, training seems to run via CPU. Did I miss something?
I got the answer here: http://answers.opencv.org/question/179831/how-opencv_traincascade-can-using-gpu/
CUDA is not supported in opencv_traincascade
Nice post. But, are you sure you are building OpenCV with CUDA Toolkit 9.1? When I run my code using your binaries, I got error “…because cudaart64_92.dll was not found…” and another *_92.dll error. I believe that “92” means “CUDA Toolkit 9.2”.
Or, I dont understand naming convention correctly? Thank you.
Hello Wawan. Thank you for your message. Originally, the post (and binaries) was written for CUDA 9.1 and was later updated for CUDA 9.2. I assume that if you hard-refresh the page, you’ll be able to see the updated 9.2 images.
Vangos, do you have prebuilt OpenCV 4.0.0/1 with CUDA 10.1/2 uploaded ?
I am having nightmare of a time to get them to build.
Let us know, if you have fought and won the battles we are about to fight.
Hi NitinH
here I found prebuilt opencv 4.0.0 with cuda 10.1
but I have problem with compiling , got errors related to wrapper.h.
I read that it is maybe a bug in opencv4 and has fixed in opencv4.1 but not sure.
sorry here is the link
https://jamesbowley.co.uk/downloads/#OpenCV4.0.0
Hi, bro :
I followed this tutorial but when I built INSTALL in openCV.snl I found this message 24>C:/opencv4/sources/opencv/modules/core/include\opencv2/core/cuda.hpp(723): warning : field of class type without a DLL interface used in a class with a DLL interface many times…
I install with opencv_contrib and cmake but I couldn’t solve my problem, in addition to this , I couldn’t find opencv_world401.lib manually in my opencv folder … Please Help me !
Hello. You can download the prebuilt binaries directly. Not sure what’s causing this error.
Hi, thank you for this post. I’ve been trying to figure this out for a while.
But I think you skipped over an entire step where you clone the contrib modules from the github. Otherwise, some of the options you mentioned in your notes don’t make sense, such as the cudafastmath.
Also, where do I put the cudnn?
Hi, thank you for this post.
I want to ask if I still need to install the CUDA toolkit if I am using the pre-build binaries? I haven’t installed it and I get cudart64_92.dll missing. I am new to CUDA and sorry if my question was dumb.
great very helpful rare to find tuts on building opencv with cuda
One suggestion please add to the post that cuDNN archive need to be extracted inside cuda toolkit installed dir