Skip to main content

Install OpenCV

To enable GPU support for OpenCV's DNN module, OpenCV must be built from source with CUDA support.

Clone the Repositories​

Clone both OpenCV and the extra modules (opencv_contrib):

git clone https://github.com/opencv/opencv.git
git clone https://github.com/opencv/opencv_contrib.git

Build from Source​

cd opencv
mkdir build
cd build
cmake -DCMAKE_BUILD_TYPE=RELEASE \
      -DCMAKE_INSTALL_PREFIX=/usr/local \
      -DCUDA_ARCH_BIN="7.5;8.9;12.0" \
      -DENABLE_CUDA_FIRST_CLASS_LANGUAGE=ON \
      -DOPENCV_DNN_CUDA=ON \
      -DWITH_CUBLAS=1 \
      -DWITH_CUDA=ON \
      -DWITH_CUDNN=ON \
      -DCUDA_FAST_MATH=1 \
      -DENABLE_FAST_MATH=1 \
      -DBUILD_opencv_python3=ON \
      -DPYTHON3_EXECUTABLE=$(which python3) \
      -DWITH_TBB=ON \
      -DWITH_GSTREAMER=ON \
      -DWITH_FFMPEG=ON \
      -DWITH_OPENMP=ON \
      -DBUILD_opencv_sfm=ON \
      -DBUILD_opencv_cudacodec=ON \
      -DBUILD_EXAMPLES=ON \
      -DINSTALL_PYTHON_EXAMPLES=ON \
      -DINSTALL_C_EXAMPLES=ON \
      -DOPENCV_ENABLE_NONFREE=ON \
      -DOPENCV_EXTRA_MODULES_PATH=/home/yi-chen/thirdparty/opencv_contrib/modules \
      ..
info

CUDA_ARCH_BIN depends on your NVIDIA GPU's compute capability. Check the CUDA GPU list on Wikipedia for the correct value for your card. For example:

  • RTX 3090 / A100 → 8.6 / 8.0
  • RTX 4090 → 8.9
  • RTX 5090 → 12.0

You can turn off flags unrelated to CUDA, but the CUDA-related flags must be enabled and CUDA_ARCH_BIN must be specified.

Then build and install:

cmake --build . -j$(nproc)
sudo cmake --install .
sudo ldconfig