Ideas on initializing ITSBN with GMM-class-majority on each superpixel

Post date: Aug 10, 2011 5:52:23 PM

I plan to experiment how well the superpixels and the GMM fits together

1. Run GMM (or VBGMM) on an image to get the pixel-level segmentation
2. Superpixelize the image, and overlay the superpixel on the GMM result obtained from 1
3. For a superpixel, assign the winner class from majority vote to every pixel in the superpixel.
4. Compare the superpixel result with the one from the pixel-level GMM

How can we use this to initialize the ITSBN?

Use the pixel-level GMM MAP class to initialize the superpixel-level GMM.