/SVM_ML

Linear SVM pipeline in MATLAB

Primary LanguageMatlab

SVM Pipeline in MATLAB

Dependencies

  • MATLAB (Tested in version 2013a)
  • LIBLINEAR (version 1.94)

Quick start

  • Compile LIBLINEAR MATLAB wrapper

    1. Download LIBLINEAR code
    2. Follow the FAQ and update linear.cpp file, change
    int check_probability_model(const struct model *model_)
{
    return (model_->param.solver_type==L2R_LR ||

    into

    int check_probability_model(const struct model *model_)
{
    return 1;

    This will allow probabilistic output for Linear SVM model.

    1. Specify MATLABDIR in matlab/Makefile, then make to compile .mex files. They will be used during training and testing.

  • Convert DTFV ASCII files into .mat binary files

    1. Prepare the feature list of DTFVs in the following format:
    [number of features listed]
[groundtruth label] [truncated path to the features]
...
[groundtruth label] [truncated path to the features]

    If you have DTFV features for HVC123456.mp4 stored in /temp/HVC123456.mp4.hog.fv.txt, /temp/HVC123456.mp4.hof.fv.txt, etc. The truncated path is just /temp/HVC123456.mp4 (i.e. excluding mode type and fv.txt)

    1. Call
    function convert_mat(featList, outputFile)

    featList is what you created in the previous step, outputFile is something like '/temp/train.mat'.

  • Training and testing with linear SVM

    1. Prepare training and testing .mat binary files
    2. Call
    function svm_pipeline(trainMat, testMat, eventID, saveDir)

    trainMat and testMat are the paths to the .mat files. eventID is an integer identifier for event classes. saveDir is the path to store models and result files.
    You can do training and testing separately by commenting out

    train_svm(trainMat, eventID, modelPath);

    or

    [confs, ap] = test_svm(testMat, eventID, modelPath);

    in this function.

  • Understanding the output formats

    1. After training, a model model.[eventID].[modeType].mat is stored in binary format
    2. After testing, a result file result.[eventID] is stored in ASCII. The i-th line gives the event confidence value (from 0 to 1) of the i-th item in testing list.

TODO

  1. Weird behavior under 10EX setting

Contact

chensun@usc.edu