Tips on using hierarchicalClustering in OpenCV

Function hiererchicalClustering is an implementation of  algorithm:  “Scalable Recognition with a Vocabulary Tree”,  the method clusters the given feature vectors by constructing a hierarchical k-means tree and choosing a cut in the tree that minimizes the cluster’s variance. It returns the number of clusters found.

Some tips need to be known when using this function:

1.DO Check whether you have correctly included the header file  “opencv2/flann/flann.hpp”,  it’s not included in “opencv2/opencv.hpp”  of version  2.4.5, so we should manually modify the problem.

2. Note that the third parameter is const cvflann::KMeansIndexParams& params, not cv:flann. So define var. like this:

cvflann::KMeansIndexParams kmeans_param;

3. Make sure that the Mat & centers  in

template<typename Distance> int flann::hierarchicalClustering(const Mat& features, Mat& centers, const cvflann::KMeansIndexParams& params, Distance d=Distance())

 have the same rows as you desire to be.

For example, do this before clustering:

Mat centers(100,128,CV_32F);

4. Now you can successfully cluster data, two forms are available(assuming data type is float and using L2-Distance):

flann::hierarchicalClustering<float,float>(features, centers, kmeans_param);

or

flann::hierarchicalClustering<flann::L2<float>>(features, centers, kmeans_param);

it returns the number of clustering centers.

I have to emphasize that whatever the features type is, the centers type must be float according to dist.h:

template<typename T>
struct Accumulator { typedef T Type; };
template<>
struct Accumulator<unsigned char>  { typedef float Type; };
template<>
struct Accumulator<unsigned short> { typedef float Type; };
template<>
struct Accumulator<unsigned int> { typedef float Type; };
template<>
struct Accumulator<char>   { typedef float Type; };
template<>
struct Accumulator<short>  { typedef float Type; };
template<>
struct Accumulator<int> { typedef float Type; };

5. Don’t forget to  remove the last redundant rows because the centers are sometimes(maybe always) less than you desire. As the function returns the actual numbers of centers, it’s easy to delete undesired data .

Enjoy it!

Reference:

1. D. Nistér and H. Stewénius. Scalable recognition with a vocabulary tree. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), volume 2, pages 2161-2168, June 2006. oral presentation.

2. OpenCV online documentation-flann::hierarchicalClustering<Distance>.