Out of curiosity, we thought it might be interesting to compare the behaviour of the CyberGlove to that of the PowerGlove. Unfortunately, however, there was insufficient data for a complete comparison; so the results from a single person using the PowerGlove are compared with the results of a single person using the CyberGlove.
As before, 5 instances total were collected for each of the variational handshapes, both with the CyberGlove and the PowerGlove. One out of the 5 was used for testing.
The results are shown in table A.4.
Table A.4: Results of comparison between PowerGlove and CyberGlove on
handshape recognition
The results here show that the CyberGlove has less than half the error rate as the PowerGlove with C4.5 and one tenth or less when using IBL1. IBL performs worse when used with the PowerGlove, because it appears that C4.5 is capable of forming better concept boundaries with the small collection of attributes than IBL is. For the large number of attributes available with the CyberGlove, IBL performs much better.
The result with the CyberGlove with additional features using IBL1 seems a little alarming to say the least. It means we can recognise handshapes with an accuracy of 98.4 per cent, considerably better than before with the full test. This result seemed extremely suspect, and the test was repeated on all the other data-sets to verify that this was no coincidence, since before, a random one of the seven available subjects was chosen.
The results are shown in table A.5.
Table A.5: Results of CyberGlove data for each of the data-sets
It appears that the above was just a a ``fluke'', but it does illustrate some aspects of the variability. It must be realised that this result is not impossible, considering that the IBL algorithm is being used. What happened before is that we lumped all of the data-sets together, regardless or not of whether they were from that person or not. This thus introduces points into the data-set that did not originally belong to that person, but also to how someone else did the handshapes. Once we only include examples that belong only to that person, it now is only possible to match that person's version of the handshape and not someone else's false handshape.
Thus the results obtained above are only valid if the system is only trained on one person.