cv::Point_<T> always gets converted to mxDOUBLE_CLASS mxArray

[amroamroamro]

In the current implementation of MxArray, when passing some of the templated OpenCV data types back to MATLAB from a MEX file, the conversion ignores the template type and creates a double array no matter what. Example:

create_point.cpp

#include "mexopencv.hpp"
using namespace std;
using namespace cv;
void mexFunction(int nlhs, mxArray *plhs[], int nrhs, const mxArray *prhs[])
{
    Point_<int> p(1,2);
    plhs[0] = MxArray(p);
}

MATLAB

>> p = create_point();
>> class(p)      % this is double

This implicit conversion can show up in unexpected places.. For example cv.approxPolyDP that takes a vector of int points, returns the approximated points as type double.

From what I can tell, this applies to Point_<T>, Point3_<T>, Size_<T>, Rect_<T>, Scalar_<T>, and possibly other places...

Do you think this should be fixed, or is this by design (possibly because some code depends on it somewhere else)?

Here is what I had in mind to fix the constructor:

template <typename T>
MxArray::MxArray(const cv::Point_<T>& p)
{
    p_ = mxCreateNumericMatrix(1, 2, ClassIDOf[DataType<T>::type], mxREAL);
    T *x = static_cast<T*>(mxGetData(p_));
    x[0] = p.x;
    x[1] = p.y;
}

where ClassIDOf is the same map defined in MxArray.cpp that translates opencv types to matlab types

[kyamagu]

Yes, this is by design. The reason is that the default numeric type is double in matlab, and also integer arithmetic in matlab sucks. In my opinion, returning integer types is benefitial only for large arrays, such as pixels. Other than that, it's probably better to return values as double.

[amroamroamro]

alright got it, thanks Kota