Figure 6(d) and Figure 8(a) show that the same face can be detected multiple times by a correctly
trained cascade. This should come as no surprise as the positive examples (see Figure 3(a)) do
allow a certain flexibility in pose and expression. On the contrary, many false positives do not enjoy
this stability, despite the fact that taken out of context, some of them do look like a face (also see
Figure 5). This observation lends support to the following detection confidence based heuristics for
further reducing false positives and cleaning up the detected result