Some ideas on ITSBN

Post date: Aug 4, 2011 12:52:36 AM

There are also some new ideas I want to try if time permits:

• Multiscale feature with ITSBN: Using real multiscale features with itsbn -- for instance using blurred version of image, using bigger kernel's size. So far we have done using the same scale features in multiple levels.
  • # of class determined using VBGMM: Initialize the number of class labels using VBGMM in every level with different scales of alpha parameter.
  • Ignore evidence in the top levels: Nodes at the top (e.g., top 3) levels don't need to have evidence because the evidence at the top level might not be valid. Remember that the evidence at the top levels are so blurred and might not be very informative. Also when we tie the top levels together, it might be the case that those evidence in the top levels might be an outlier respect to evidence in the lower layers, which case might affect the Gaussian components.
  • Structure obtained from dynamic tree: Make an experiment using 1 dimensional data to test the hypothesis that the structure obtained using mean-field dynamic tree is similar to that obtained from simple clustering i.e., k-mean, PRI.
  • Gaussian prior for structure: Another idea is to embed the Gaussian prior to the structure learning (in mean-field dynamic tree) and compare ITSBN vs ITSBN with deformable structure using Gaussian prior. I anticipate that the improvement might not be significant.
  • Build structure using PRI: Using PRI to learn the structure of the pixel-level image then run the itsbn on top of that. Compare the results with using TSBN and superpixel level.