How to convert OpenCV Mat to Python ndarray?

Question

C++ / C#

Hello everyone, experts, in general, there is a function that I use through Python Boost, I need to transfer OpenCV Mat to Python, when I try to do it directly, boost gives TypeError: No to_python (by-value) converter found for C++ type: cv::Mat, Which makes sense since Python doesn't use ndarray for images.

#include <iostream>
#include <opencv2/highgui.hpp>
#include "opencv2/cudawarping.hpp"
#include <boost/python.hpp>

using namespace std;
using namespace cv;


Mat gpuResize(string path, int maxwidth, int height)
{   
    Mat inputCpu = imread(path, IMREAD_COLOR);
    cuda::GpuMat input(inputCpu);
    cuda::GpuMat output;
    cuda::resize(input, output, Size(maxwidth, height), 0, 0, INTER_LINEAR);
    
    Mat outputCpu;
    output.download(outputCpu);

    // PyObject* ret = toNDArray(outputCpu);

    return outputCpu;
}

char const* greet()
{
   return "hello, world";
}


BOOST_PYTHON_MODULE(cvcuda)
{
    using namespace boost::python;
    def("gpuResize",gpuResize);
}

from cvcuda import gpuResize
path = '/home/alenorze/Workspace/Current/hair/backend/test_data/0e6303c4-0b29-457e-ab74-ac2d1cfb1ecf.jpeg'
maxwidth = 650
height = 500
print(gpuResize(path, maxwidth, height))

I need a simple solution, all tens of thousands of string converters are outdated and hard to get started. Thank you in advance.

Answer 1

In general, there is one solution, and it is completely wrong, but logical. OpenCV does not have code for Python, it is completely in C++, there is a converter in python/src2/cv2.cpp that can convert a lot of things, including Mat, but since you need to have at least some experience with pluses to run it separately ( which I don't have).
I decided to modify the library itself and insert new functions into it in order to import directly through cv2. I found the imread function inside loadsave.cpp, added my function, declared it next to imread, and everything worked, no problems. But again, I do not recommend doing this, it's a sin.

Answer 2

At one time I made a computational module for python in C, transferred C arrays from the module to python and displayed them there using Matplotlib. Implemented the Python Buffer Protocol in the module . Further in python, I received such an object from the module and, using numpy.asarray(), distilled it into ndarray.
If Mat is also based on a regular array, then the Buffer Protocol can be implemented for it in a similar way.
I don't know how to do it with python boost, I used Python C API.
It is also possible to create numpy objects directly in the module, numpy has its own numpy C API. But I had no experience with this.
I don't have any easier options.