ferrubio / gmforslinr

Graphical Models for Semantic Localization in Robotics

BASIC STEPS TO REPRODUCE THE COMPARATIVE STUDY OF BAYESIAN NETWORKS

DOWNLOAD THIS TOOLBOX AND UNZIP IT INTO ANY FOLDER (CALLED $MATLAB_CODE HERE)

DOWNLOAD FREIBURG, LJUBLJANA AND SAARBRUCKEN SEQUENCES FROM COLD DATASET (http://www.cas.kth.se/COLD/downloads.php), INTO $MATLAB_CODE/dataset/COLD/ WITH THE NEXT STRUCTURE

  $MATLAB_CODE/datasets/COLD/Freiburg/seq1_cloudy1.tar
  ...
  $MATLAB_CODE/datasets/COLD/Freiburg/seq3_sunny3.tar
  $MATLAB_CODE/datasets/COLD/Ljubljana/seq1_cloudy1.tar
  ...
  $MATLAB_CODE/datasets/COLD/Ljubljana/seq2_sunny3.tar
  $MATLAB_CODE/datasets/COLD/Saarbrucken/seq1_cloudy1.tar
  ...
  $MATLAB_CODE/datasets/COLD/Saarbrucken/seq4_sunny3.tar

ALSO DOWNLOAD THE DUMBO AND MINNIE SEQUENCES FROM KTH-IDOL2 DATASET(http://www.cas.kth.se/IDOL/#Download) INTO $MATLAB_CODE/datasets/KTH-IDOL WITH THE NEXT STRUCTURE

  $MATLAB_CODE/datasets/KTH-IDOL/dum_cloudy1.tar
  ...
  $MATLAB_CODE/datasets/KTH-IDOL/dum_sunny4.tar
  $MATLAB_CODE/datasets/KTH-IDOL/min_cloudy1.tar
  ...
  $MATLAB_CODE/datasets/KTH-IDOL/min_sunny4.tar

TO CREATE THE FOLDER STRUCTURE THAT OUR FUNCTIONS NEED, TWO SHELL SCRIPT MUST BE LAUNCH FROM THE TERMINAL

  $MATLAB_CODE/COLDFolders.sh
  $MATLAB_CODE/IDOLFolders.sh

DOWNLOAD THE LAST VERSION OF BNT FOR MATLAB (https://code.google.com/p/bnt/) AND UNZIP IT IN $MATLAB_CODE FOLDER WITH THE NEXT STRUCTURE

  $MATLAB_CODE/bnt-master

DOWNLOAD THE LAST VERSION OF libsvm FOR MATLAB (http://www.csie.ntu.edu.tw/~cjlin/libsvm/) AND UNZIP IT IN $MATLAB_CODE FOLDER WITH THE NEXT STRUCTURE

  $MATLAB_CODE/libsvm

IF $LIBSVM_FOLDER IS DIFFERENT TO libsvm, RENAME IT.

BEFORE USING libsvm, IT MUST BE COMPILED. LAUNCH MATLAB AND ESTABLISH $MATLAB_CODE/libsvm/matlab AS CURRENT FOLDER, THEN TYPE

  >> make

IF THIS DOESN'T WORK READ THE README.TXT IN $MATLAB_CODE/$LIBSVM_FOLDER/matlab FOLDER FOR MORE INFORMATION.

DOWNLOAD THE PHOG CODE FOR MATLAB (http://www.robots.ox.ac.uk/~vgg/research/caltech/phog.html) AND UNZIP IT IN THE FOLDER

$MATLAB_CODE/Descriptors

OPEN MATLAB AND SET $MATLAB_CODE AS CURRENT FOLDER

LEARN DIFFERENTS TRAINING MODELS AND TEST THEM

+ The experiments can be reproduced with completeProcessCV function:
    completeProcessCV(0,0,1,360,5,5) learns and tests a Naive Bayes network with 
    continuous data from DUMBO dataset and store the accuracy and the confusion 
    matrix in an output folder. The first call to this function takes a long time,
    in order to extract all the features and store them in the correct directory 
    structure.

+ Prove different combinations for completeProcessCV. The variables and their 
  options are detailed in the comments.

THIS CODE INCLUDES FUNCTIONS TO EXTRACT PHOG DESCRIPTORS OF DIFFERENT IMAGES AND TO LEARN BAYESIAN NETWORKS WITH DIFFERENT ALGORITHMS. THE PREVIOUS STEPS SERVE TO REPRODUCE A SERIES OF EXPERIMENTS, BUT THE FUNCTIONS CAN BE TESTED EASILY BY FOLLOWING THE CODE FROM completeProcessCV.m IN THE CASE OF USING DIFFERENT DATA.