CODEWITHSUNDEEP
pythonc++opencvnumpy
zhangxaochen
zhangxaochen

Reputation: 34017

OpenCV mask operation, elementwise assignment in c++

I'd like to assign each pixel of a mat matA to some value according to values of matB, my code is a nested for-loop:

clock_t begint=clock();
for(size_t i=0; i<depthImg.rows; i++){
    for(size_t j=0; j<depthImg.cols; j++){
        datatype px=depthImg.at<datatype>(i, j);
        if(px==0)
            depthImg.at<datatype>(i, j)=lastDepthImg.at<datatype>(i, j);
    }
}
cout<<"~~~~~~~~time: "<<clock()-begint<<endl;

and it costs about 40~70ms for a mat of size 640*480.

I could do this easily in python numpy using fancy indexing:

In [18]: b=np.vstack((np.ones(3), np.arange(3)))

In [19]: b
Out[19]: 
array([[ 1.,  1.,  1.],
       [ 0.,  1.,  2.]])

In [22]: a=np.vstack((np.arange(3), np.zeros(3)))

In [23]: a=np.tile(a, (320, 160))

In [24]: a.shape
Out[24]: (640, 480)

In [25]: b=np.tile(b, (320, 160))

In [26]: %timeit a[a==0]=b[a==0]
100 loops, best of 3: 2.81 ms per loop

and this is much faster than my hand writing for-loop.

So is there such operation in opencv c++ api?

Upvotes: 4

Views: 836

Answers (2)

Aurelius
Aurelius

Reputation: 11329

I am unable to replicate your timing results on my machine Your C++ code runs in under 1ms on my machine. However, whenever you have slow iteration, at<>() should be immediately suspect. OpenCV has a tutorial on iterating through images, which I recommend.

However, for the operation you describe, there is a better way. Mat::copyTo() allows masked operations:

lastDepthImg.copyTo(depthImg, depthImg == 0);

This is both faster (about 2x as fast) and far more readable than your nested-loop solution. In addition, it may benefit from hardware optimizations like SSE.

Upvotes: 4

Jaime
Jaime

Reputation: 67427

In your C++ code, at every pixel you are making a function call, and passing in two indices which are getting converted into a flat index doing something like i*depthImageCols + j.

My C++ skills are mostly lacking, but using this as a template, I guess you could try something like, which should get rid of most of that overhead:

MatIterator_<datatype> it1 = depthImg.begin<datatype>(),
                       it1_end = depthImg.end<datatype>();
MatConstIterator_<datatype> it2 = lastDepthImg.begin<datatype>();

for(; it1 != it1_end; ++it1, ++it2) {
    if (*it1 == 0) {
        *it1 = *it2;
    }
}

Upvotes: 1

Related Questions