Multivariate gaussian probability errors

Question

Sri 2016-10-17

0
链接

此问题的直接链接

https://ww2.mathworks.cn/matlabcentral/answers/307609-multivariate-gaussian-probability-errors

I am using multivariate Gaussian model to identify a yellow ball. I am given some training images with yellow balls. I have used these pixels to get the parameters of the model.

But when I am observing the probability values of the image with this model, I get the values of very low order (10^-8) and strangely I am finding the probability of foreground is lower than the background.

The code to get parameters are

mu = sum(Samples(:,:))/N;
Samples = double(Samples);
mu1 = repmat(mu, N, 1);
Sigma = ((Samples-mu1)'*(Samples-mu1))/N;

where Samples is N*3 matrix with RGB values. Using the following code, I am trying to get probability values

clc
clear al
close all
imagepath = './train';
D = 3;
load('parameters.mat',  'mu');
load('parameters.mat',  'Sigma');
dSigma = det(Sigma);
const = 1/(((2*pi)^(D/2))*((dSigma)^(1/2)));
thre = 1e-8;
for k=1:1
    samples = [];
    % Load image
    I = imread(sprintf('%s/%03d.png',imagepath,k));
      % You may consider other color space than RGB
      R = I(:,:,1);
      G = I(:,:,2);
      B = I(:,:,3);
      R = R(:);
      B = B(:);
      G = G(:);
      samples = [samples; [R G B]];
      samples = double(samples);
      m = length(R);
      p = zeros(m, 1);
      for i = 1:m
          p(i) = const*exp(-0.5*(samples(i)-mu)*pinv(Sigma)*(samples(i)-mu)');
      end
  end