How to apply K means in a mask of an image instead the whole one

edit

A quick and dirty solution (maybe s.o. else knows sth more elegant):

// assume you loaded img as color-image
// and that mask is 0 in the areas which you don't want
std::vector<cv::Vec3b> points;
for( int y = 0; y < img.rows; y++){
    for( int x = 0; x < img.rows; x++){
        if ( mask(y,x) != 0 ) 
            points.push_back(img.at<cv::Vec3b>(y,x));
    }   
}

// now run kmeans using the points as input-data

as soon as you have the labels you can go through your image as above and set the image according to the label.

It looks like you execute kmeans on the whole image. Take a look at: Interactive foreground extraction using grab cut algortihm

Then you can use the result image as mask for src.copyTo(dst, mask); and then execute kmeans. This should work, because copied pixels are those wich are unequal to zero in mask image. But I didn't test it.