This directory contains a simple implementation of the SEMANTIC PROFILES 
algorithm by Hofmanninger et al. When using this code, please cite the
following paper: 

@inproceedings{hofmanninger2015mapping,
  title={Mapping Visual Features to Semantic Profiles for Retrieval in Medical Imaging},
  author={Hofmanninger, Johannes and Langs, Georg},
  booktitle={Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition},
  pages={457--465},
  year={2015}
}

Look at toyExample.m for an exemplary application.

To generate semantic profiles you need to perform a training:

        % specification of learning parameters
        p.num_ferns = 1200;      % number of ferns to be generated (1200 default)
        p.ferns_depth = 8;       % depth of one fern (e.g. 2^8 partitions per fern) (8 default)
        p.sub_dims = 9;          % number of sub-dimensions used on each split (usually <12) (9 default)
        p.partitionRes = 5000;   % parameter K in the Paper.  (5000 default)
        p.classSmoothing = 20;   % parameter gamma in the paper (prevents overfitting) (20 default)

        % training of the model
        % trainingData is dxn
        % weakLabels is a binary matrix nxC where C is the number of classes
        % where 1 indicates weak membership of the class. (o 
        model = sptrainmodel(trainingData,weakLabels,p);

and use the model for novel records:

        profiles = spgetprofiles(testData,model);