Why is the value of cluster center and index keeps varying for an image in kmeans?

Dear Team,

     Wish you all a Happy New Year 2014:).

Problem statement: I'm working on image retrieval using SURF features. After extracting the features from image, i tried using kmeans algorithm. I gave the below command for getting its index and cluster centers.

code

nWords = 5;

 [idx,centers] = kmeans([features]',nWords);

When i try giving the above command in command window, value of idx and centers keeps varying for an image (gray scale image, size( 192 x 144)) whose extracted features is of size 64 x 10 single.Could you please let me know why the value of idx and centers are changing ?

Cheers,

Malini

Answer 1

[Walter Roberson]

kmeans by default uses random initialization. You can alter that by using the 'start' option; you could give a specific matrix for example.

[Malini]

Thank you sir for your immediate response. Could you kindly explain more on type of specific matrix to be used. As I tried giving one dimensional and two dimensional matrix but it is throwing me error.

code

sample = [0.7 0.2;0.3 0.4];

[Qidx,Qcenters]=kmeans(Qfeatures,5,'Start',sample);

Error using kmeans (line 200)

The 'Start' matrix must have K rows.

Also could you please explain the difference between use of fuzzy c means and kmeans with respect to random initialization.Sorry for the inconvenience. kindly request you to guide me.

Cheers,

Malini

[Walter Roberson]

As you want 5 clusters, you must have 5 rows in your "sample" matrix. The rows must have size(Qfeatures,2) columns.

I do not know at the moment how fcm() initializes the clusters.