One of the obvious ways to test how well the system works is to simply combine all the feature sets discussed above into one test.
In total, 119 features were used. The same techniques are employed for assessment -- 5-fold cross validation.
Table 6.1: Results of learning algorithms on all attributes combined
We can see, using the above, that altogether, GRASP can have an accuracy up to about 80 per cent. In all the large data sets, the error rate is approximately 80 per cent. The results of the andrew data-set indicate that perhaps that if an insufficient number of samples is provided, then the accuracy goes down rapidly. As previously mentioned, the adam dataset should not be taken as representative, because when the data was sampled, the words were not randomly sorted. This of course meant that the fatigue factor was removed. This also perhaps indicates the relative sensitivity of GRASP to intra-signer consistency. From the above, it would appear that GRASP is sensitive to variations in the way a signer makes a sign.
The effects of the number of samples will be explored in section 6.3.1.