
Online Learning of Discriminative Patterns from Unlimited 
Sequences of Candidates 

Ilkka Autio and J.T. Lindgren

Package version v1.0 / 10.Aug.2006 

For License, see file COPYING


This source code package performs the experiments described
in our ICPR2006 paper with the same name.

run_tests.m            runs online accuracy experiments
run_xval.m             runs cross-validation experiments
algorithm_bayes.m      Algorithm 'Beta-Learn' of the paper
algorithm_heuristic.m  Algorithm 'Beta-Heuristic' of the paper
calc_stats.m           Computes the accuracies after experiments have been ran
calc_xval.m            Computes the xval after experiments have been ran
calc_plots.m           Computes the plots after experiments have been ran

Note that we do not provide the third party datasets for testing.
After you have obtained some dataset, the above scripts use these
in online fashion through the image filenames (i.e. one image 
is read and discarded at a time). The filenames must be 
provided in .mat files in datasets/ directory.

datasets/fn_negatives.mat  

This file must contain a cell array named 'negfn'. Each
array entry must be a full path to an image matlab can load.
These are the filenames to the 'no object' images (negative 
examples).

datasets/fn_'identifier'.mat

Similar but the cell array must be called 'posfn'. Identifier is
replaced with the actual dataset name, i.e. if you specify
"datasets = {'rothcars'}" in the testing script, it expects to
find "datasets/fn_rothcars.mat". These are the filenames
of the 'object present' images (positive examples).

A script 'create_datasets.m' shows how to create these 
files for your own image sets.

In case of bugs/problems/questions/ideas related to this code 
package, please send email to <jtlindgr@cs.helsinki.fi>.


 - J.L. 

