[Solved] Preliminary results and problems on ITSBN 6.0

Post date: Aug 31, 2011 2:59:05 AM

Good news and bad news.

One feature added in version 6.0 is the pixel-level evidence at the bottom most of the network (l=1). Pixels are sampled from each of the superpixels in level l=2. Originally this feature is implemented in the function fn_make_tree_structure by sampling the pixels at the end. However, the result on the pixel-level does not look good, so I changed the strategy by building the pixel level right after making superpixels. In fact, this enables more flexibility to the algorithm because from now on we can input any type of superpixel before making the tree, for instance, we can build a mask for TSBN pretty easily.

However, this does not seem to solve the problem of pixel partial sample; this requires more investigation.

Some preliminary results are shown here:

Results for each level

majority-assignment of the level l=1

description

using the number of classes C= [3 3 5 7] for each level from top to bottom. The leaf level is implemented by gridding the image into 20x20 small rectangles.

using the number of classes C= [3 3 5 7] for each level from top to bottom. The leaf level is implemented by gridding the image into 40x50 small rectangles.

using the number of classes C= [3 3 5 7] for each level from top to bottom. The leaf level is implemented by gridding the image into 20x20 small rectangles.However, in this case, 10% of the patches are removed, and the results are totally weird. I suspect the rearrange function. To be investigated!

In this case, only 1 patch is missing, but the result starts to look weird already.

Bug fixed!

After 6 hours of sleeping, I found the small, arrogant, cute, unforgivable bug in my function fn_make_evidence. I couldn't believe this took me 12 working hours to close the deal. Details of the bug can be found here.

Here are some results after the bug is fixed

Results for each level

majority-assignment of the level l=1

missing-patch gRGB image

description

The l-1 level is composed of 20x20 patches with 2 patches removed.

The l-1 level is composed of 2x200 sampled pixels from the above layer having 200 superpixels. The points in l-1 are very sparse, but still the majority-vote construction scheme gives pretty good result.

We use ITSBN version 6 to implement TSBN with evidence at all the levels. The TSBN here has unmber of nodes in each level from root to leaf: 1, 4, 16, 64, 256 and 1024 respectively.

Another TSBN using 1, 4, 16, 64, 256, 1024 and 4096 nodes in each level. This takes 5472 sec to run!!!

Run-time

Summary

  1. ITSBN version 6.0 can replace the results from version 4.0.0 and perhaps version 4.5.0 too.
  2. We can experiment on any shape of patch by simply build masks for it, which can be either superpixel or pixel level with no limitation of patch shape.