Please note that more complex scenes cause longer
processing times in all cases (indeed, every graph tends to
be an enlarged or shrinked versions of another) but higher
492
values for N increase significantly the average processing
times (dashed lines). Surprisingly enough, the prediction
error will not decrease by considering a larger number of
inputs to the RBF network. Figure 3 shows that prediction
errors are higher in more complex scenes (around frames 30
and 80-90), but no relationship seems to hold between the
prediction error level and the value of N
This can be explained considering that only a small number
of significant behaviour prototypes exist, and adding new
ones does not necessarily mean adding fresh information
[2]. Although too small values of N may lead back to
linear-like prediction, it is needless to increase N beyond
some units. Optimal value for N, however, increases when
we consider predictions further in the future. For our
experiments we set N=8 for next-frame (k+1) forecasting,
N=9 for k+2-frame forecasting, N=10 for k+3-frame
forecasting.
With all the parameters set using the above criteria, we get
average processing times around 14 minutes per map for
the synthesis phase and around 48 seconds for the forecast
phase.