Object that is passed to function is changed, although no pointer is passed

[donelron]

Hey there,

I am posting this simplified piece of code that is a bit confusing for me. There are two functions that I call. One shows the expected results but the result of the other one rather puzzles me.

//#define defineVecTyp Vec3f
#define defineVecTyp float


template <typename vecTyp>
vecTyp buildLaplacianPyramid(cv::Mat inputmat) {
    vecTyp lapPyr;

    GaussianBlur(inputmat,inputmat,Size(GaussSize,GaussSize),0.0,0.0);
	code ...

    return lapPyr;
}


int sum1(int int1, int int2)
{
    int1=int1*2;
    int2=int2*2;

    int sum=int1+int2;
    return sum;
}


int main()
{

    int val1=1;
    int val2=10;
    int sumOfVals;

    cout << "val1= " << val1 << " ## val2= " << val2 << endl;

    sumOfVals=sum1(val1,val2);
    cout << "val1= " << val1 << " ## val2= " << val2 << endl;
    cout << "sumOfVals= " << sumOfVals << endl;


    cv::Mat M1;
    M1=imread("../image1.jpeg");
    imshow("M1",M1);


    vector<Mat_<defineVecTyp> > LaplacianPyramid;
    LaplacianPyramid = buildLaplacianPyramid< vector<Mat_<defineVecTyp> > >(M1);
    imshow("M1, after buildLaplacianPyramid",M1);


    return 0;
}

Calling the function sum1 does not change the values stored in the variables val1 and val2. The output of the program is as follows:
val1= 1 ## val2= 10 // before the call of function sum1
val1= 1 ## val2= 10 // after the call of function sum1
sumOfVals= 22

This is quite obvious and as expected and I just pasted this piece of code as an example for better clarification.


However, if I call the function buildLaplacianPyramid and apply a function for Gaussian Blurring, this also effects the cv::Mat passed to the function. The line
imshow("M1, after buildLaplacianPyramid",M1);
therefore shows an image that is blurred. Since I am not passing a pointer to the cv::Mat I do not understand why this should be happening.
I was assuming that there would be a copy of the cv::Mat M1 to be used within the function. Therefore I was expecting the cv::Mat M1 to retain its original value. I was expecting that all changes applied to cv::Mat inputmat within the function would not have any influence on the cv::Mat M1. Just like in my other example with the sum.

Does anybody have an idea how to explain this behaviour?!
I am grateful for any hints.
Thanks

[JLBorges]

> Since I am not passing a pointer to the cv::Mat I do not understand why this should be happening.

The copy constructor of cv::Mat does not do a deep copy.

To get value semantics, clone the object.

buildLaplacianPyramid< ... >( M1.clone() ) ;

[Thumper]

JLBorges wrote:

The copy constructor of cv::Mat does not do a deep copy

I.e., cv::Mat behaves a similar to std::shared_ptr. The matrix data isn't copied by assignment, instead, it gives a pointer to the existing data and increments a reference counter. This is done to protect against extreme memory usage which, without this type of feature, would be really easy to overlook in programs using OpenCV.

[donelron]

Thanks for the quick response!
Now that you mention it, it really makes a lot of sense to NOT copy the complete data.
Basically it would not have been a big problem to work around it, but I was not sure if the problem was due to some kind of mistake of mine. Now I know...