AdaptiveThreshold with non-square block sizes

edit

@berak why can we change int blockSize in Size s in adaptiveThreshold? In source code, adaptiveThreshod uses boxFilter and gaussianBlur with a parameter of size type

hehe, i did not dare to look there, but now i will....

@LBerger, that might indeed be possible, but there are some limitations, like only 3,5,7,9,11 can be used, so the most "horizontal" kernel would be 11x3

where did you find limit ? gaussianBlur and boxFilter

minimum is here

i'm probably wrong about the maximum size, but there are hardcoded opencl kernels for certain sizes only

Thanks a lot for your help. I never heard of the Range feature before. However, it behaves different from my Python code and the adaptiveThreshold seems to ignore the range boundaries.

Example:

unsigned int nHeight = oImgInput.size().height;
for (int nY = 0; nY < nHeight; ++nY)
{
    const auto& oLine = oImgInput(cv::Range(nY, nY + 1), cv::Range::all());
    const auto& oLineResult = oImgOutput(cv::Range(nY, nY + 1), cv::Range::all());
    cv::adaptiveThreshold(oLine, oLineResult, 255, cv::ADAPTIVE_THRESH_MEAN_C, cv::THRESH_BINARY, 9, 0);
}

Explanation: The loop applies the adaptiveThreshold to each line and stores the result to oImgOutput. But the result image is exactly the same as if I apply the threshold to the entire image. Is the code wrong?