OpenCv create 3 channel Mat from continuous data array

Question

I'd like to create an OpenCV 3-channel Mat using data allocated elsewhere where the pixels of each channel are together, unlike the data for an OpenCV Mat where the data from different channels is interleaved.

Mat outputMat = Mat(dimY, dimX, CV_8UC3, rawData); 
// This works only if rawData interleaves channel data like an OpenCv Mat

Is there a way to create an OpenCV Mat without having to resort to the below solution of splitting channels from a temporary Mat and copying the right channel data to the corresponding location?

void createMat(unsigned char *rawData, unsigned int dimX, unsigned int dimY)
{
    Mat outputMat = Mat(dimY, dimX, CV_8UC3);

    // use outputMat to draw some stuff

    Mat channelR = Mat(dimY, dimX, CV_8UC1, rawData);
    Mat channelG = Mat(dimY, dimX, CV_8UC1, rawData + dimX * dimY);
    Mat channelB = Mat(dimY, dimX, CV_8UC1, rawData + 2 * dimX * dimY);

    std::vector<Mat> channels(3);
    split(outputMat, channels);

    channels[2].copyTo(channelR);
    channels[1].copyTo(channelG);
    channels[0].copyTo(channelB);
}

I need to do this operation frequently, so I was wondering if there is a solution that doesn't involve calling the split() and copyTo() functions each time.

Thanks!

Scevo

How about merge(channels,outputMat)? – Greenling

I already tried that, but it doesn't seem to work. The correct data doesn't go to the right channel. It is still interleaved. – Scevo

Your question is not clear to me, Do you want to get RGB channels in same manner from an OpenCV mat ? – Let

Not exactly. I wish to create an OpenCV Mat with external data where the RGB channels are not interleaved, but rather each channel is stored one after the other. – Scevo

Your statement is ambiguous to me. What does "not interleaved" refer to, the channels of the Mat, the channels of the external data, or both? – Infringement

I added an illustration showing how the different channels are defined for a 2x2 Mat in both cases. Hope it's clear now. – Scevo

Answer 1

You can avoid split and copyTo by using merge directly.

Mat createMat(unsigned char *rawData, unsigned int dimX, unsigned int dimY)
{
    // No need to allocate outputMat here
    Mat outputMat;

    // Build headers on your raw data
    Mat channelR(dimY, dimX, CV_8UC1, rawData);
    Mat channelG(dimY, dimX, CV_8UC1, rawData + dimX * dimY);
    Mat channelB(dimY, dimX, CV_8UC1, rawData + 2 * dimX * dimY);

    // Invert channels, 
    // don't copy data, just the matrix headers
    std::vector<Mat> channels{ channelB, channelG, channelR };

    // Create the output matrix
    merge(channels, outputMat);

    return outputMat;
}

---

I tested a few other approaches, but they result to be slower. Just for the records, I thought this would be faster, but the transpose is really heavy:

Mat outputMat(3, dimY*dimX, CV_8UC1, rawData);
Mat tmp = outputMat.t();
outputMat = tmp.reshape(3, dimY);
cvtColor(outputMat, outputMat, COLOR_RGB2BGR);

Entente

merge(channels, outputMat); doesn't seem to use the same data as rawData. For example, setting outputMat.at<Vec3b>(0, 0) = Vec3b(128, 128, 128); doesn't reflect the values at the locations rawData[0], rawData[dimX * dimY] and rawData[2 * dimX * dimY]. I checked, and it looks like outputMat.ptr<unsigned char>(0) != rawData – Scevo

1) the changes to outputMat are not reflected back in rawData. The data are deep copied while using merge. Since you're using a different layout than OpenCV, if you want to use your data you need to deep copy the data. 2) Obviously the data pointed by outputMat are different from rawData... you need to interleave your original data to use them in a Mat. – Entente

That's what I feared. For my application, I'm using OpenCV to draw some stuff and display in another application which allocates rawData. So, seems like whenever I change the content of the outputMat I need to do a deep copy for the changes to reflect in rawData. If I have a high refresh rate, there's a lot of overhead to manually interleave the data. If I understood it correctly, looks like there is no solution where I could do this just once and the subsequent changes in my OpenCV outputMat would reflect automatically in rawData. Thanks for the reply. @Entente – Scevo

Exactly. Some solutions to overcome the refresh would be to: 1) change the data layout your other application, 2) don't use OpenCV and make your own Matrix structure and the needed functions. If you just want to draw stuff, this shouldn't be difficult. 3) Use a library with the same data layout – Entente

What about YUV420? The split method seems to return just one channel with the initial Mat. – Micheal