|
|
<!DOCTYPE group PUBLIC "-//W3C//DTD XHTML 1.0 Transitional//EN" "http://www.w3.org/TR/xhtml1/DTD/xhtml1-transitional.dtd"> |
|
|
|
|
|
<group> |
|
|
|
|
|
<p><b>VLFeat</b> includes a basic implementation of k-means clustering |
|
|
and hierarchical k-means clustering. They are designed to be |
|
|
lightweight in order to work on large datasets. In particular, they |
|
|
assume that the data are vectors of unsigned chars (one byte). While |
|
|
this is limiting for some application, it works well for clustering |
|
|
image descriptors, where very high precision is usually unnecessary. |
|
|
For more details, see the |
|
|
<a href="%pathto:root;api/ikmeans_8h.html">Integer k-means API |
|
|
reference</a>.</p> |
|
|
|
|
|
|
|
|
<ul> |
|
|
<li><a href="%pathto:tut.ikm.usage;">Usage</a></li> |
|
|
<li><a href="%pathto:tut.ikm.elkan;">Elkan</a></li> |
|
|
</ul> |
|
|
|
|
|
|
|
|
<h1 id="tut.ikm.usage">Usage</h1> |
|
|
|
|
|
|
|
|
<p>Integer k-means (IKM) is run by the command |
|
|
<code>vl_ikmeans</code>. In order to demonstrate the usage of this |
|
|
command, we sample 1000 random points in the <code>[0,255]^2</code> |
|
|
integer square and use <code>vl_ikmeans</code> to get k=3 |
|
|
clusters:</p> |
|
|
|
|
|
<pre> |
|
|
K = 3 ; |
|
|
data = uint8(rand(2,1000) * 255) ; |
|
|
[C,A] = vl_ikmeans(data,K) ; |
|
|
</pre> |
|
|
|
|
|
<p>The program returns both the cluster centers <code>C</code> and the |
|
|
data-to-cluster assignments <code>A</code>. By means of the cluster |
|
|
centers |
|
|
<code>C</code> we can project more data on the same clusters</p> |
|
|
|
|
|
<pre> |
|
|
datat = uint8(rand(2,10000) * 255) ; |
|
|
AT = vl_ikmeanspush(datat,C) ; |
|
|
</pre> |
|
|
|
|
|
<p>In order to visualize the results, we associate to each cluster a |
|
|
color and we plot the points:</p> |
|
|
|
|
|
<pre> |
|
|
cl = get(gca,'ColorOrder') ; |
|
|
ncl = size(cl,1) ; |
|
|
for k=1:K |
|
|
sel = find(A == k) ; |
|
|
selt = find(AT == k) ; |
|
|
plot(data(1,sel), data(2,sel), '.',... |
|
|
'Color',cl(mod(k,ncl)+1,:)) ; |
|
|
plot(datat(1,selt),datat(2,selt),'+',... |
|
|
'Color',cl(mod(k,ncl)+1,:)) ; |
|
|
end |
|
|
</pre> |
|
|
|
|
|
<div class="figure"> |
|
|
<image src="%pathto:root;demo/ikmeans_lloyd.jpg"/> |
|
|
<div class="caption"> |
|
|
<span class="content"> |
|
|
<b>Integer k-means.</b> We show clusters of 2-D points obtained by |
|
|
integer k-means. There are <code>k=3</code> clusters represented with |
|
|
different colors. The clusters have been estimated from 1000 points |
|
|
(displayed as dots). Then 10000 different points have been projected on |
|
|
the same clusters (displayed as crosses). The three big markers |
|
|
represent the cluster centers. |
|
|
</span> |
|
|
</div> |
|
|
</div> |
|
|
|
|
|
|
|
|
<h1 id="tut.ikm.elkan">Elkan</h1> |
|
|
|
|
|
|
|
|
<p>VLFeat supports two different implementations of k-means. While |
|
|
they produce identical output, the Elkan method requires fewer |
|
|
distance computations. The <code>method</code> parameters controls |
|
|
which method is used. Consider the case when <code>K=100</code> and our |
|
|
data is now 128 dimensional (e.g. SIFT descriptors):</p> |
|
|
|
|
|
<pre> |
|
|
K=100; |
|
|
data = uint8(rand(128,10000) * 255); |
|
|
tic; |
|
|
[C,A] = vl_ikmeans(data,K,'method', 'lloyd') ; % default |
|
|
t_lloyd = toc |
|
|
tic; |
|
|
[C,A] = vl_ikmeans(data,K,'method', 'elkan') ; |
|
|
t_elkan = toc |
|
|
|
|
|
t_lloyd = |
|
|
|
|
|
10.2884 |
|
|
|
|
|
t_elkan = |
|
|
|
|
|
5.1405 |
|
|
</pre> |
|
|
|
|
|
|
|
|
</group> |
|
|
|
