2011_08_03

Post date: Aug 3, 2011 8:47:10 PM

Morning half went to Lib check some new books.

Afternoon:

• Cleaned the ITSBN toolbox version 3.0.1 and assume that the code document explains the same thing :-P. Yeah, I really believe that they are >95% correspondent. I put the file on dropbox but did not make it available on the Internet yet.
• Also I cleaned ITSBN toolbox version 4.0.0 which supports the evidence in multiple scales, but I haven't compiled the cleaned file and put on the internet though.
• Discussed with Julius about the format of the data before putting to the DT code.
• The last thing for me today is GMM for image segmentation. At least I got the code run already with several features, e.g., gRGB, sCIELuv, sCIELab, sGray, Texton. What left is to clean the code and write the document and make the code available on the Internet.
• Tomorrow I should be able to overlaying the superpixels on the GMM segmentation results and use majority vote to assign class labels to each superpixel. Let's see how the results come out!
• At this point, I got an idea..why don't I just connect each pixel to its superpixel parent. This way, the framework would combine both pixel and superpixel together seamlessly. This can be comparable with "Associative Hierarchical CRFs for Object Class Image Segmentation", which is discriminative approach, but mine is generative approach.
  • It is worth to note that we don't need to use all the pixels in the image, we can just some of them obtained from sampling technique. For instance, you can just randomly pick 1/10 of all the pixels as to be children nodes of their corresponding parent. The network might look like the figure below.

• Another approach similar to the first one, but allow us to use the same model of ITSBN with evidence in all the scale is illustrated as follows: