1
$\begingroup$

Imagine we are clustering pixels in an image. If we use Mean Shift Clustering, at least in my understanding, we will embed each pixel into some dimensional space (intensity, rgb, texture, etc) and cluster them. This embedding process, however, inherently destroys the graph structure of the image, which could be quite useful: e.g. neighboring pixels tend to form same semantic cluster.

Is there a way I can preserve this graph structure when using Mean Shift Clustering? You can assume grid like graph structure like a 2D image.

Thanks in advance!

$\endgroup$
  • $\begingroup$ Simplest solution for this is to use quick-shift algorithm $\endgroup$ – foothill Dec 18 '16 at 20:33
1
$\begingroup$

The simplest thing to do is to add the position information as another channel as well, such that points spatially close to each other are likely to be in the same cluster.

You can try different kernels and weights for the position channel to get different results.

$\endgroup$
  • $\begingroup$ Position information such as pixel coordinates are weaker than connectivity information in graphs. $\endgroup$ – foothill Nov 15 '16 at 23:22

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.