How to appropriately destroy mxArray created by ocvMxArray​FromImage_​{DataType} functions

Dear all, I am trying to integrate Matlab code into C++ program with OpenCV. I found the program will gradually increase memory usage.

After some working around, I believe the problem has something to do with destroying mxArray. I wrote a simple test program with an infinite loop that keeps create mxArray and destroy it, as shown below

 cv::Mat imageMat = cv::imread(imagePath);
 while(true)
 {
   mxArray* imageMXA = ocvMxArrayFromImage_uint8(imageMat);
   mxDestroyArray(imageMXA);
 }

this program will gradually increase memory usage.

However, another test program that creates mxArray with mxCreateNumericArray does not have such problem. the code is shown below

 while(true)
 {
   mxArray* imageMXA = mxCreateNumericArray(dimensionNumber, dimensions, mxDOUBLE_CLASS,  mxREAL);
   mxDestroyArray(imageMXA);
 }

It seems that there is something wrong in the code that destroy the mxArray created by ocvMxArrayFromImage_{DataType} functions. My question is, how should I appropriately destroy such mxArray?

[James Tursa]

You would probably need to see the code for ocvMxArrayFromImage_uint8 to answer your question, and you don't have that. However, it looks like it is a simple nD transpose and reordering operation, so it wouldn't be too difficult to write your own conversion routine if needed.

[Eddie Chiou]

Hi James,

Thanks for the suggestion.

I tried a manual conversion in c++ and the memory usage is stable so far. However, the speed is about 30 times slower. Is there any suggestion to efficiently convert the image?

Currently I assign the matrix element by element within 3 for-loop. Program will crash using MxArray of opencvmex, another library interfacing OpenCV and Matlab.

[Eddie Chiou]

hmm... I tried another way to do the conversion. I first transpose the cv::Mat, split into 3 channels and then copy data into mxArray. This has similar speed as ocvMxArrayFromImage_uint8 and has no memory problem. Though the first time executing this function is very slow and I don't know why.

[James Tursa]

I can't offer any advice unless I see your code.

[Eddie Chiou]

mxArray * ConvertMat(const cv::Mat & image)
{
  cv::Mat imageT = image.t();

  std::vector<cv::Mat> splitChannel;
  cv::split(imageT, splitChannel);

  size_t rows = image.rows;
  size_t cols = image.cols;
  size_t channels = image.channels();
  mxArray *pMXA = NULL;
  UINT8 *input = NULL;

  {
    const int dimensionNumber = 3;
    mwSize dimension[dimensionNumber] = { rows, cols, channels };
    pMXA = mxCreateNumericArray(dimensionNumber, dimension, mxUINT8_CLASS, mxREAL);
    input = (UINT8 *)mxGetData(pMXA);
    size_t step = rows * cols;

    for(int channelIndex = 0; channelIndex < channels; channelIndex++)
    {
      memcpy(input + step * channelIndex, splitChannel[channels - channelIndex - 1].data, step);
      
    }  
  }//3 channels

  return pMXA;
}

[James Tursa]

Is channels always either going to be 1 or 3?

[Eddie Chiou]

In general case, yes. Both this function and ocvMxArrayFromImage_uint8 take about 0.001 sec converting 640x480 image. This function does not increase memory usage after destroy the mxArray. It takes ~0.25 sec the first time executed but this is not a big problem.