Call a C++ function from Python and convert a OpenCV Mat to a Numpy array

Question

Background situation

I'm trying to use the OpenCV Stitching module via the Python bindings, but I'm getting an error:

import cv2
stitcher = cv2.createStitcher(False)

imageL = cv2.imread("imageL.jpg")
imageC = cv2.imread("imageC.jpg")
imageR = cv2.imread("imageR.jpg")

stitcher.stitch((imageL, imageC))

error: /home/user/OpenCV3.1.0/opencv/modules/python/src2/cv2.cpp:163: error: (-215) The data should normally be NULL! in function allocate

Similar people suffering this:

• https://stackoverflow.com/a/36646256/1253729
• How to stitch images from a UAV using opencv python with Stitcher class
• https://github.com/opencv/opencv/issues/6969

The problem at hand

So I decided to use a official C++ OpenCV stitching example and use Python to call it using Boost.Python. However, I'm still unable to figure out how to properly use Boost.Python + numpy-opencv-converter to handle the C++ Mat vs Numpy array conversion.

¿How do I call the numpy-opencv-converter? I've only got Boost.Python in place, and when running my python function to call the C++ file I got this (expected) outcome:

$ python python_caller.py 
Traceback (most recent call last):
  File "python_caller.py", line 10, in <module>
    visualize(A)
Boost.Python.ArgumentError: Python argument types in
    testing.visualize(numpy.ndarray)
did not match C++ signature:
    visualize(cv::Mat)

Thanks.

PD: I'm in Ubuntu 14.04, Python 2.7.4 using OpenCV 3.1.0 compiled from sources and inside a virtualenv.

---

These are the files I'm using.

testing.cpp:

#include <stdio.h>
#include <opencv2/opencv.hpp>
#include <boost/python.hpp>

using namespace cv;

int main(){}

Mat visualize(const cv::Mat input_image)
{
    cv::Mat image;
    image = input_image;

    namedWindow("Display Image", WINDOW_AUTOSIZE );
    imshow("Display Image", image);
    waitKey(0);

    return image;
}

using namespace boost::python;
BOOST_PYTHON_MODULE(testing) // file name
{   
    def("visualize", visualize); //function name
}

python_caller.py:

import cv2
import numpy as np
from testing import visualize

A = cv2.imread("imageL.jpg")

visualize(A)

Makefile:

CFLAGS=`pkg-config --cflags opencv`

LDFLAGS=`pkg-config --libs opencv`

testing.so: testing.o
    g++ -shared -Wl,--export-dynamic -o testing.so testing.o -L/usr/lib -lboost_python -L/usr/lib/python2.7/config -lpython2.7 -L/usr/lib/x86_64-linux-gnu/ -lopencv_calib3d -lopencv_contrib -lopencv_core -lopencv_features2d -lopencv_flann -lopencv_gpu -lopencv_highgui -lopencv_imgproc -lopencv_legacy -lopencv_ml -lopencv_objdetect -lopencv_ocl -lopencv_photo -lopencv_stitching -lopencv_superres -lopencv_ts -lopencv_video -lopencv_videostab 

testing.o: testing.cpp
    g++ -I/usr/include/python2.7 -I/usr/include -fPIC -c testing.cpp

Answer 1

You need to convert the Python NDArray <=> C++ cv::Mat. I can recommend this GitHub Repo. It contains an example that should fit to your needs. I am using the converter on Ubuntu 15.10 with Python 2.7/3.4 and OpenCV 3.1.