5.3.1 Validation methodology

To check the validity of a method, they were tested on the data collected.

In total five datasets were collected. These are named after the person who produced them. Each was collected on a different day, and for practical reasons, not the same number of samples were obtained from each signer.

The datasets are:

adam

8 samples per sign.

andrew

8 samples per sign.

john

18 samples per sign.

stephen

16 samples per sign.

waleed

20 samples per sign.

In total 6 650 signs were collected.

It was expected that there would be significant variations between the efficacy of the features on the various signers. This is because just as accents exist in spoken languages, each signer has his own gestural ``accent''. Some signers are conservative in their movements, some are energetic. Any system will prove to be better able to cope with some signers than with others, just as handwriting recognition systems are better able to work with some people's handwriting than with others.

Each of the features was tested with the learning algorithms already discussed.

In each of these cases, five-fold cross validation was used. Five-fold cross-validation means that the sample set is divided into fifths. One fifth is used as a test set and the learner is trained on the other four fifths. This is repeated five times with a different fifth used for testing each time. The average error rate is then taken.