Input matrix to opencv kmeans clustering - opencv

Question

This question is specific to opencv:
The kmeans example given in the opencv documentation has a 2-channel matrix - one channel for each dimension of the feature vector. But, some of the other example seem to say that it should be a one channel matrix with features along the columns with one row for each sample. Which of these is right?
if I have a 5 dimensional feature vector, what should be the input matrix that I use:
This one:
cv::Mat inputSamples(numSamples, 1, CV32FC(numFeatures))
or this one:
cv::Mat inputSamples(numSamples, numFeatures, CV_32F)

Answer 1

The correct answer is cv::Mat inputSamples(numSamples, numFeatures, CV_32F).
The OpenCV Documentation about kmeans says:
samples – Floating-point matrix of input samples, one row per sample
So it is not a Floating-point vector of n-Dimensional floats as in the other option. Which examples suggested such a behaviour?
Here is also a small example by me that shows how kmeans can be used. It clusters the pixels of an image and displays the result:
#include "opencv2/imgproc/imgproc.hpp"
#include "opencv2/highgui/highgui.hpp"
using namespace cv;
int main( int argc, char** argv )
{
Mat src = imread( argv[1], 1 );
Mat samples(src.rows * src.cols, 3, CV_32F);
for( int y = 0; y < src.rows; y++ )
for( int x = 0; x < src.cols; x++ )
for( int z = 0; z < 3; z++)
samples.at<float>(y + x*src.rows, z) = src.at<Vec3b>(y,x)[z];
int clusterCount = 15;
Mat labels;
int attempts = 5;
Mat centers;
kmeans(samples, clusterCount, labels, TermCriteria(CV_TERMCRIT_ITER|CV_TERMCRIT_EPS, 10000, 0.0001), attempts, KMEANS_PP_CENTERS, centers );
Mat new_image( src.size(), src.type() );
for( int y = 0; y < src.rows; y++ )
for( int x = 0; x < src.cols; x++ )
{
int cluster_idx = labels.at<int>(y + x*src.rows,0);
new_image.at<Vec3b>(y,x)[0] = centers.at<float>(cluster_idx, 0);
new_image.at<Vec3b>(y,x)[1] = centers.at<float>(cluster_idx, 1);
new_image.at<Vec3b>(y,x)[2] = centers.at<float>(cluster_idx, 2);
}
imshow( "clustered image", new_image );
waitKey( 0 );
}

Answer 2

As alternative to reshaping the input matrix manually, you can use OpenCV reshape function to achieve similar result with less code. Here is my working implementation of reducing colors count with K-Means method (in Java):
private final static int MAX_ITER = 10;
private final static int CLUSTERS = 16;
public static Mat colorMapKMeans(Mat img, int K, int maxIterations) {
Mat m = img.reshape(1, img.rows() * img.cols());
m.convertTo(m, CvType.CV_32F);
Mat bestLabels = new Mat(m.rows(), 1, CvType.CV_8U);
Mat centroids = new Mat(K, 1, CvType.CV_32F);
Core.kmeans(m, K, bestLabels,
new TermCriteria(TermCriteria.COUNT | TermCriteria.EPS, maxIterations, 1E-5),
1, Core.KMEANS_RANDOM_CENTERS, centroids);
List<Integer> idx = new ArrayList<>(m.rows());
Converters.Mat_to_vector_int(bestLabels, idx);
Mat imgMapped = new Mat(m.size(), m.type());
for(int i = 0; i < idx.size(); i++) {
Mat row = imgMapped.row(i);
centroids.row(idx.get(i)).copyTo(row);
}
return imgMapped.reshape(3, img.rows());
}
public static void main(String[] args) {
System.loadLibrary(Core.NATIVE_LIBRARY_NAME);
Highgui.imwrite("result.png",
colorMapKMeans(Highgui.imread(args[0], Highgui.CV_LOAD_IMAGE_COLOR),
CLUSTERS, MAX_ITER));
}
OpenCV reads image into 2 dimensional, 3 channel matrix. First call to reshape - img.reshape(1, img.rows() * img.cols()); - essentially unrolls 3 channels into columns. In resulting matrix one row corresponds to one pixel of the input image, and 3 columns corresponds to RGB components.
After K-Means algorithm finished its work, and color mapping has been applied, we call reshape again - imgMapped.reshape(3, img.rows()), but now rolling columns back into channels, and reducing row numbers to the original image row number, thus getting back the original matrix format, but only with reduced colors.