EECS556_code

Introduction

There are three parts in code/ folder:

1. CODE/TOOLS

• Geometric Context/ and multipleSegmentations/
  Hoiem’s segmentation training and testing codes
• segment/
  Felzenzswalb’s oversegmentation tool
• trainDataGeneration/
  Our own training data aruto-extraction tool given ModelNumber_TissueType.mnc.gz from EECS556/BrainModels

2. DATASET

• Brainweb_MRIDatabase/ MRI dataset
• MSRC_ObjCategImageDatabase_v2/ MSRC dataset (Hoiem used)

3. RESULTS
   trained classifiers; test results (accuracy and labeled test images);

• mri_results/ results for MRI dataset
• msrc_results/ results for MSRC dataset

#Usage

1. In Matlab, navigate to EECS556/code; Run

    >> addpath(genpath(‘.’))
   

2. Manually unzip the ModelNumber_TissueType.mnc.gz into Brainweb_MRIDatabase/WebDownloads, then run

    >> get_data_ready
   

which will automatically generate original images in .ppm format, and ground truth labels train_data.mat in folder Brainweb_MRIDatabase/Images.
3. Run

    >> mriTrain
classifiers are learned and stored in `mri_results/train_results`

4. run

    >> mriTestScript
   

   test images will be loaded and labeld based on learned classifiers.
   Test results are stored in mri_results/test_results

#Parameters Parameters that can be tuned:
ncv: nvc-fold cross-validation
nsegments: list of number of regions for multiple hypothesis
train: training images indicies
test: test iamges indicies
trainind1: training images for edge classifer
sigma: Used to smooth the input image before segmenting it (0.1)
k: Value for the threshold function (100)
min: Minimum component size enforced by post-processing (5)
currently we use
ncv=1: no cross validation trainind1=half of the training set

#Experiments

1. Cross-tests
   We have four patterns of bias field (p1, p2, p3, p4), and we form train-test pair between any two of clean, p1, p2, p3, p4, mixed. (mixed is the dataset that contains all the images biased by every pattern).

2. Feature ablation study
   We have four kinds of features intensity, texture, shape, location. And we turn on either feature and compare the performance with the one that utilizes all of them.

3. Number of multiple hypothesis
   We can try different nsegments, i.e.

• nsegments = [different integers from small to large to inf]
• nsegments = [different integers from large to inf]
• nsegments = [several same integers1, several same integers2, ..., inf]

#Troubleshooting

1. Complain about treevalc, try navigate to GeometricContext/src/boosting and then

    >> mex treevalc
   

2. Complain about image file not exist, such as

    File "./tmpimsp424511.ppm" does not exist.
   
    Error in processSuperpixelImage (line 24)
    	im = imread(fn{f});
   
    Error in msCreateSuperpixels (line 12)
    imsegs = processSuperpixelImage(outfn);
   

   try

    cd EECS556_code/segment/
    rm segment
    make
   

   and test whether segment works by

    ./segment 0.1 100 5 test_in.ppm test_out.ppm
   

#Git Clone Repository Instructions If you have not generated any SSH key for github on your machine, please do so following this tutorial

1. Create a new folder for the repository you are going to clone, and navigate into the folder, eg

    $  mkdir EECS556_code
    $  cd EECS556_code/
    $  git init
   

2. Set up git

    $  git remote add origin _SSH_
    $  git pull origin master
   

   where _SSH_ is the SSH clone URL shown on the github repository webpage (near the bottom of the right column, above "Download ZIP").

3. Push to git After you commit any changes, push to git

    $  git push origin master