SUN_source_code_v2/code/vlfeat-0.9.5/docsrc/quickshift.html

<!DOCTYPE group PUBLIC "-//W3C//DTD XHTML 1.0 Transitional//EN" "http://www.w3.org/TR/xhtml1/DTD/xhtml1-transitional.dtd">

<group>

<p><b>Quick shift</b> is a mode seeking algorithm (like mean shift) which
instead of iteratively shifting each point towards a local mean instead forms a tree of links to the nearest neighbor which increases the density. For a more in-depth description of the algorithm, see our
<a href="%pathto:root;api/quickshift_8h.html">API reference for quick shift</a></p>

<ul>
 <li><a href="%pathto:tut.qs.superpixels;">Using quick shift to find superpixels</a></li>
 <li><a href="%pathto:tut.qs.multi;">Multiple segmentations</a></li>
</ul>
 
<!-- ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ -->
<h1 id="tut.qs.superpixels">Using quick shift to find superpixels</h1>
<!-- ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ -->

<p>This demo shows quick shift in a simple superpixelization problem where we seek to segment this image:</p>

<div class="figure">
  <image src="%pathto:root;demo/quickshift_image.jpg"/>
  <div class="caption">The image we wish to segment</div>
</div>

<p>As a feature vector, we choose the LAB colorspace representation of the image augmented with the x,y location of the pixel. <code>vl_quickseg</code> is a convenient wrapper function which takes care of the transformation of the image to LAB and performs segmentation, making our job as easy as:</p>

<pre>
ratio = 0.5;
kernelsize = 2;
Iseg = vl_quickseg(I, ratio, kernelsize, maxdist);
</pre>

<p>where <code>ratio</code> is the tradeoff between color importance and spatial importance (larger values give more importance to color), <code>kernelsize</code> is the size of the kernel used to estimate the density, and <code>maxdist</code> is the maximum distance between points in the feature space that may be linked if the density is increased.</p>

<div class="figure">
<image src="%pathto:root;demo/quickshift_qseg_1.jpg"/>
<image src="%pathto:root;demo/quickshift_qseg_2.jpg"/>
<div class="caption">The effect of <code>maxdist</code> on the superpixelization. As we increase <code>maxdist</code>, superpixels become larger and larger since we can link less similar points. Top: <code>maxdist=10</code>. Bottom: <code>maxdist=20</code>.</div>
</div>

<!-- ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ -->
<h1 id="tut.qs.multi">Multiple segmentations</h1>
<!-- ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ -->

<p>Quick shift arranges all of the data points into a tree where parents in the tree are the nearest neighbors in the feature space which increase the estimate of the density. By imposing a limit on the distance between nearest neighbors (<code>maxdist</code>), we decrease the amount of computation required to search for the nearest neighbors. However, we also break our tree into a forest, because local modes of the density will now have no neighbor which is close enough in the feature space to form a link.</p>

<p>In the previous section, we created a superpixel segmentation by taking each of the trees in this forest as a distinct cluster. However, since <code>maxdist</code> simply prevents new links from forming, the segmentation formed by every <code>dist &lt; maxdist</code> is contained in the result. <code>vl_quickvis</code> lets us visualize this by running quick shift once and forming multiple segmentations by cutting links in the tree which are smaller and smaller.</p>

<pre>
maxdist = 50;
ndists = 10;
Iedge = vl_quickvis(I, ratio, kernelsize, maxdist, ndists)
imagesc(Iedge);
axis equal off tight;
colormap gray;
</pre>

<div class="figure">
<image src="%pathto:root;demo/quickshift_qvis.jpg"/>
<div class="caption">A visualization of multiple <code>maxdist</code> thresholds on a single image. Here, boundaries are colored by the largest <code>maxdist</code> where the boundary is preserved.</div>
</div>

</group>