Use CV2 on Headless Ubuntu

Method 1: use imgcat instead of cv2.imshow

• light version without imshow pip install opencv-python-headless
• use iTerm + imgcat https://iterm2.com/documentation-images.html(means installing shell integration of iTerm)
• pip install imgcat on remote server https://pypi.org/project/imgcat/

Method 2: use cv2 or PIL over SSH X11

• if want to use imshow, install regular opencv-python, set '/etc/ssh/sshd_config' to 'X11Forwarding yes', and connect with 'ssh -X', some authorization and dependencies may need.

Compilation

OpenCV Build on Linux

https://wiki.aleen42.com/qa/opencvonlinux.html

https://www.marearts.com/0-OpenCV-Build-Ubuntu-20-04-OpenCV-4-5-2-CUDA-11-2-4c876d664e84442e82ebec13c23586b3

• pre

  # check nvidia version and ubuntu version
  nvidia-smi
  nvcc --version
  lsb_release -a
  
  # download source code and create new dir
  mkdir opencv_with_cuda && cd opencv_with_cuda
  wget -O opencv.zip <https://github.com/opencv/opencv/archive/master.zip>
  wget -O opencv_contrib.zip <https://github.com/opencv/opencv_contrib/archive/master.zip>
  unzip opencv.zip
  unzip opencv_contrib.zip
  rm opencv.zip
  rm opencv_contrib.zip
  mkdir build && cd build
  
  # python env
  conda activate python3.8
  pip show numpy
  pip install --upgrade numpy
  

OpenCV with GPU for Python on Windows

https://www.youtube.com/watch?v=HsuKxjQhFU0&t=515s

https://docs.opencv.org/master/db/d05/tutorial_config_reference.html

1. install NVIDIA CUDA, OpenCV code, CMake as for C++

2. install Anaconda and set PATH correctly

3. download OpenCV code，(opencv, opencv_contrib)

4. CMake Configure:

   1. select the OpenCV source code and target build path, click Configure select os platform, start to build

      1. if 网络不佳，自行查看下载失败的包，去github下载好，修改相应.cmake 文件里的地址为file:///...。或者把raw.githubusersourcecontent 改为镜像地址，参考：https://www.codenong.com/cs107007432/

   2. upgrade NumPy version to newest

   3. check all python paths including the NumPy path，configure again

      

   4. uncheck OPENCV_GENERATE_SETUPVARS

   5. check these:

      1. WHITH_CUDA
      2. OPENCV_DNN_CUDA
      3. ENABLE_FAST_MATH
      4. BUILD_opencv_world

   6. OPENCV_EXTRA_MODULES_PATH = path/to/opencv_contrib/modules

   7. configure again

   8. check CUDA_FAST_MATH

   9. choose arch version （ https://en.wikipedia.org/wiki/CUDA, torch.cuda.get_arch_list()）, and release config, click ‘Configure’ again

   10. click ‘Generate'

   11. install in prompt

       cmake --build "D:\mnvai-alg\cv2_with_gpu\build_win_py" --target INSTALL --config Release
       

5. Verify:

   import cv2
   from cv2 import cuda
   
   cuda.printCudaDeviceInfo(0)