Discussion on the exception handling. It is a serious problem which we have to
think about. Probably a short package has to be designed using the standard 
C++ exception mechanism.

The granularity of the classes should be optimized. One should split a class 
to separate classes only if you use them independently. A copy-paste pattern 
is a serious candidate for a function. A big function can be done more 
modular by adding private functions to the class if it is really needed.
Not all the nouns in the subject dictionary become classes in the 
implementation. 

The seeding in the parallel tracking is very similar to the one described by
Sylwester. The difference is that more criteria are used, not only the number
of points, but chi2 too. The approach of Sylwester uses more seeding passes.

For the moment we don't have any procedure to decide whether two close tracks
are in fact the same track. It is a difficult problem (a kind of art).

Marian don't put any error information in the clusters for the moment, but 
only information about the shape.

We have to see where do we recuperate the tracks, we expect this occurs 
at the outer corners of the sectors.

The seeding in the new approach takes 1s for 30000 simulated particles
(N^2 problem).