ML::EM with multiple Features

asked 2016-06-28 10:48:14 -0500


I have a float[33] sized feature vector derived from pcl (Point Cloud Library). I would like to perform a Gaussian Mixture Model on this and found that OpenCV has this implemented in ML::EM.

I use OpenCV 3.0. Now unfortunately the EM model needs a one channel matrix. Does this mean I can only calculate a one dimensional Model? I have 33 features, not only one.

Here is my class where I convert my pcl feature PointCloud to OpenCV:

FPFHtoCV::FPFHtoCV(pcl::PointCloud<pcl::FPFHSignature33>::Ptr fpfhs)
   cv::Mat openCVPointCloud( fpfhs->points.size(), 1, CV_32FC(33));
    for(int i=0; i < fpfhs->points.size();i++)
       pcl::FPFHSignature33 p = fpfhs->;
       for(size_t j = 0; j < 33; j++)
 <cv::Vec<float,33> >(1,0)[j] = p.histogram[j];
   _mat = openCVPointCloud;

And here how I try to run the GMM on the OpenCV matrix.

    FPFHtoCV ftc(fpfhs);
    cv::Mat mat = ftc.mat();
    cv::Ptr<cv::ml::EM> source_model = cv::ml::EM::create();
    cv::Mat logs;
    cv::Mat labels;
    cv::Mat probs;

It does not work of course, as there is more than one channel.

edit retag flag offensive close merge delete



what about reshaping your Mat, so it's [33, rows] , and single channel ?

why did you set it up like this (with only 1 col, but 33 channels) in the 1st place ?

berak gravatar imageberak ( 2016-06-29 05:57:13 -0500 )edit

@berak I consider this 33 featuers per point similar to something like rgb. While it is not color information it is addtional non spatial information. As OpenCV is for picture anaylsis I just assumend setting up mat[33,row s] with 1 channel would result in 33xrows 1 dimensinoal observations put into the EM algorithm. Will try though as suggested.

simpletree gravatar imagesimpletree ( 2016-06-29 07:08:16 -0500 )edit