Both C++ and Python are excellent languages that complement each other in many ways. I have been working on Computer Vision and Document Analysis problem and I have had the need of offloading some performance critical code to C++ and expose it neatly to the other pieces, which in turn are in Python.

Here, I will take a simpler problem as a running example - matrix multiplication. We want to write a super-efficient C++ function that multiplies two matrices and expose that function to be callable from Python. The Python code should be able to pass in two numpy.ndarray objects to be multiplied and get back the result in a numpy.ndarray.

Note that the converter descibed here is available on GitHub.

We will call our module matrmul and the function mul. Let us see how the calls to this function should look like from Python:

import numpy
import matrmul # our module

a = numpy.array([[1., 2., 3.]])
b = numpy.array([[1.],
print(matrmul.mul(a, b)) # should print [[14.]]

With the interface in place, we now begin by starting a Boost Python module with the mul function. In matrmul.cpp, put

// matrmul.cpp
#include <iostream>
#include <opencv2/imgproc/imgproc.hpp>
#include <boost/python.hpp>
#include "conversion.h"

namespace py = boost::python;

mul(PyObject *left, PyObject *right)
    NDArrayConverter cvt;
    cv::Mat leftMat, rightMat;
    leftMat = cvt.toMat(left);
    rightMat = cvt.toMat(right);
    auto c1 = leftMat.cols, r2 = rightMat.rows;
    // Work only with 2-D matrices that can be legally multiplied.
    if (c1 != r2)
                        "Incompatible sizes for matrix multiplication.");
    cv::Mat result = leftMat * rightMat;

    PyObject* ret = cvt.toNDArray(result);

    return ret;

static void init()

    py::def("mul", mul);

As a quick tour of that code, the main multiplication routine mul takes two parameters of the type PyObject*. The PyObject* type represents the Python object type in the C API. Since Python is dynamically typed, the function has to figure out what is the actual type of the objects being passed. First we use the NDArrayConverter class to convert the two arguments(which we believe will NumPy arrays) to the OpenCV cv::Mat type. Then we check if the two matrices are multiplication-compatible. If not, a TypeError is thrown using the code

                "Incompatible sizes for matrix multiplication.");

py refers to the namespace boost::python, which provides many helpful, idiomatic C++ wrappers around the C/Python API. Then we multiply the two Mat objects and convert the result back to an ndarray using the NDArrayConverter::toNDArray() function the definition of which we shall see in a moment.

The part that exports this function mul as a function in Python is

    py::def("mul", mul);

So we use a macro defined by the Boost Python library that declares a Python module called matrmul. Within that declaration, we call init() that in turn initializes the Python runtime and the numpy C library. The latter is absolutely necessary, as without it, any calls to the numpy C API will cause a segmentation fault. Then, we use py::def() to define a module level function called mul (we could give any other name here - this is the name seen by Python code). The second argument to py:def is of course, our function.

Now, we need to define the actual conversion functions for

  • Numpy ndarray to cv::Mat conversion

  • cv::Mat to Numpy ndarray conversion

The code for doing these conversions has been copied(and modified slightly) from the OpenCV sources, and resides in conversion.h and conversion.cpp. The high level class exposing functions to do the conversions is called NDArrayConverter

class NDArrayConverter
    void init();
    cv::Mat toMat(const PyObject* o);
    PyObject* toNDArray(const cv::Mat& mat);

So, NDArrayConverter::toMat() takes a numpy ndarray, again as a PyObject*, tests whether it is a valid numpy array(needed, as a PyObject* can point to any Python object) and returns the equivalent cv::Mat.

NDArrayConverter::toNDArray() does the reverse - takes a reference to a cv::Mat and returns a PyObject* that represents the numpy array which can be returned to the Python runtime via Boost.

Boost Python can automatically convert native types (like strstd::string, Python long ⇋ C++ long, int, etc.), but we need to do the conversion ourselves when we have stuff like ndarrays that Boost Python does not know about. So the process of wrapping a function can be summed up as:

  • Write the C++ function to take normal native types if native types is all you expect to be passed to it from Python. Boost takes care of the plumbing and type checking.

  • If the function(in Python) takes anything other than native types, write the C++ function to take PyObject* - the generic Python object.

  • If using PyObject*, convert to whatever format is expected(or throw a TypeError if a malformed object was passed) and produce the result.

  • If the result is a standard C++ type that can be handled by Boost, no conversion is needed. Otherwise, convert that result to a PyObject* and return this.

  • Use BOOST_PYTHON_MODULE to expose your function to Python.

After the code is ready, we need to compile the code to a shared object file. The easiest way to do that would be to use GNU make. The Makefile should compile the conversion code(conversion.cpp), the module wrapper code(matrmul.cpp) and link them together. We also need to pass the Boost and OpenCV headers while compiling and must link against the Boost, Python and OpenCV libs. A working Makefile can be cloned and tweaked to specific needs. After compiling, one can start the python shell from the same directory as and test the multiplication routine:

In [1]: import numpy as np

In [2]: import matrmul

In [3]: a = np.array([[1., 2., 3.]])

In [4]: b = a.reshape(3, 1)

In [5]: b
array([[ 1.],
       [ 2.],
       [ 3.]])

In [6]:
Out[6]: array([[ 14.]])

In [7]: matrmul.mul(a, b)
Out[7]: array([[ 14.]])

Using the Makefile provided, you can also type make test to run the tests in, which verifies that our code is correct.

The converter code with extra examples is available on GitHub.