Results
Click for full size images
Comparing the two algorithms tested, AdaBoost and
AdaPlusBoost, it can be seen that they both achieve similar recognition rates. It should be noted, however, that
AdaPlusBoost uses far fewer classifiers (often a factor of
ten fewer than AdaBoost) to produce similar results. This
means that the system would be much more suited to realtime
applications if using the AdaPlusBoost classifiers, for
while they take longer to train (due to re-calculating the
strong classifier each iteration) they reduce the per frame
time required for detection. |
Known Signers
|
|
When working on purely known signers (those who appeared in the training data) little difference is made by using a threshold on the difference image fed into the integral volume, the advantages of the threshold are only fully noticeable when the system is tested on a mixture of both known and unknown signers. On this type of data the threshold reduces the disparity between different signers' skin colours and clothing colours making the system more robust to unseen signers. |
Mix of Known and Unknown signers
|
