3D Face--> Results

In this section we present the dimensions selected to form the combined fishersurface
systems (Figure 7) and the error rates obtained from a range of tests sets, making a
comparison to optimum individual systems in Figure 8.



We see that systems with lower EERs generally make the most contribution to the combined system, as would be expected. However, it is also interesting to note that even systems with particularly high EERs do contain some dimensions that make a positive contribution, although this is much more prominent for the cosine distance, showing that this metric is more suited to combing multiple surface spaces.
Having selected and combined the range of dimensions shown in Figure 7, we now apply these ombined systems to test sets A and B using both the cosine and Euclidean distance metric. We also perform an evaluation on the union of test sets A and B: an experiment analogous to training on a database (or gallery set) of known people, which are then compared with newly acquired (unseen) images.

Figure 8 shows the error curves obtained when optimum individual fishersurface systems and combined systems are applied to test set A (used to construct the combination), test set B (the unseen test set) and the full test set (all surfaces from sets A and B), using the cosine and Euclidean distance metrics. We see that the combined systems produce lower error rates than the optimum individual systems for all six
experiments. As would be expected, the lowest error rates are achieved when tested on the surfaces used to construct the combination (7.2% and 12.8% EER respectively).

However an improvement is also seen when applied to the unseen test set B, from 11.5% and 17.3% using the best single systems to 9.3% and 16.3% EER for the combined systems. Performing the evaluation on the larger set, providing 1,079,715 verification operations (completed in 14 minutes 23 seconds on a Pentium III 1.2GHz processor, providing a verification rate of 1251 per second), the error drops slightly to 8.2% and 14.4% EER, showing that a small improvement is introduced if some test data is available for training, as well as suggesting that the method scales well, considering the
large increase in verification operations.