Why eigenvector & eigenvalue in LDA become zero?

Question

I'd like to implement fast PLDA (Probabilistic Linear Discriminant Analysis) in OpenCV. in this, LINK fast PLDA have been implemented in Matlab and Python. One of the parts of PLDA is LDA. I've written following code for implementing LDA in OpenCV:

int LDA_dim = 120;  // Load data  FileStorage fs("newStorageFile.yml", FileStorage::READ);  // Read data  Mat train_data, train_labels;  fs["train_data"] >> train_data; fs["train_labels"] >> train_labels;  // LDA  if (LDA_dim > 0) {     LDA lda(LDA_dim);     lda.compute(train_data, train_labels);          // compute eigenvectors      Mat eigenvectors = lda.eigenvectors(); } 

I've converted database that was introduced in above link from .mat to .yml. The result is newStorageFile.yml that I've uploaded here. train_data have 650 rows and 600 cols and train_labels have 650 rows and 1 cols. I don't know why eigenvectors and eigenvalue become zero!!? PLZ help me to fix this code.

It's better to bring the code that convert data from .mat to .yml :

function matlab2opencv( variable, fileName, flag)  [rows cols] = size(variable);  % Beware of Matlab's linear indexing variable = variable';  % Write mode as default if ( ~exist('flag','var') )     flag = 'w';  end  if ( ~exist(fileName,'file') || flag == 'w' )     % New file or write mode specified      file = fopen( fileName, 'w');     fprintf( file, '%%YAML:1.0\n'); else     % Append mode     file = fopen( fileName, 'a'); end  % Write variable header fprintf( file, '    %s: !!opencv-matrix\n', inputname(1)); fprintf( file, '        rows: %d\n', rows); fprintf( file, '        cols: %d\n', cols); fprintf( file, '        dt: f\n'); fprintf( file, '        data: [ ');  % Write variable data for i=1:rows*cols     fprintf( file, '%.6f', variable(i));     if (i == rows*cols), break, end     fprintf( file, ', ');     if mod(i+1,4) == 0         fprintf( file, '\n            ');     end end  fprintf( file, ']\n');  fclose(file); 

Edit 1 ) I've tried LDA with some sample that myself generate:

Mat train_data = (Mat_<double>(3, 3) << 25, 45, 44, 403, 607, 494, 2900, 5900, 2200);     Mat train_labels = (Mat_<int>(3, 1) << 1, 2, 3 );      LDA lda(LDA_dim);      lda.compute(train_data, train_labels);          // compute eigenvectors     Mat_<double> eigenvectors = lda.eigenvectors();     Mat_<double> eigenvalues = lda.eigenvalues();     cout << eigenvectors << endl << eigenvalues; 

but I've to got same result: eigenvalue and eigenvector become zero:

Answer 1

It is because of the floating point imprecision that the eigen values get close to zero.