DDT version 6

Post date: Sep 2, 2011 6:05:54 PM

Basically, DDT is a new name for ITSBN. In this post we illustrate the performance of DDT version 6 with the following parameters:

• There are totally 6 levels (root included), whose number of superpixels in each level are approximately 1, 20, 60, 80, 200 and 1500 respectively. Note that the leaf level of DDT version 6 are pixel-level node 1% randomly sampled from pixels in each superpixel. This is the main focus of the version 6.
• Some levels are tied together for inference stability:
  • tie_level_table = [1 1
  • 2 2
  • 3 3
  • 4 4
  • 5 4
  • 6 4];
  • Each tied level has its own number of classes and the DDT parameters
  • tie_parameter_table = [1 7 1
  • 2 7 1
  • 3 5 1
  • 4 3 1];
• Number of maximum iterations is 15

There are several questions I try to answer in this experiment

1. Quality of DDTv6 vs superGMM
   1. We found that DDTv6 performs competitively to superGMM with the same initialization.
2. Does the quality of segmentation depend on the log-likelihood?
   1. The segmentation quality strictly depends on the log-likelihood in the very beginning of the iteration, say 1-5, however, after iteration 10, the log-likelihood still keeps increasing whereas the segmentation seems to remain the same. So, I think setting the maximum number of iterations to 10 is sufficient to get high-quality segmentation.
3. What is the optimal number of iteration?
   1. Around 10
4. Run-time vs # of nodes
   1. Please refer to the end of this post.

Sample results are shown in the slides below