Life after first submission

Post date: Oct 5, 2011 8:23:45 PM

Things to show in the dissertation

  1. The structure prior is LESS important when the superpixel size gets BIGGER. [Experiment is here]
  2. Comparison of superpixel generators: Mori's and UCM.
  3. How to pick the scale of the DDT; What is the reason to pick 10, 50, 120, 300?
  4. Improve the feature extraction for DDT. Currently we use only CIELuv + position.
  5. Compare all the architectures of DDT
  6. Compare them with others.
  7. Use MSRC and BSDS500 dataset in addition to BSDS300

Dissertation-oriented list of work

  1. How can we define the scale of superpixel
    1. Using entropy? which is equivalent to the # of boundary pixel/# of whole pixel.
  2. Superpixel quality.
    1. In one superpixel algorithm, plot the accuracy vs # of superpixel used vs the entropy
    2. Make sure they are on the same scale!!!
  3. Importance of tree-structure prior
    1. meDDT vs GMiND
    2. leDDT vs GMiND
  4. Importance of the evidence in all the scale
    1. TSBN vs meTSBN
    2. leDDT vs meDDT
  5. Feature selection for image segmentation when using superpixel and generative model GMM
    1. Dimensionality reduction, PCA, SNE,
    2. Experiment on superGMM
  6. Variant architectures of DDT
    1. leDDT
    2. meDDT
    3. coDDT