The features come from a predefined pool. They are constructed so as to
have sufficiently low density on generic background. Training for an object class,
involves taking a small number of examples at reference pose, i.e registered to the
reference grid, and identifying those features from the pool, which are frequent
at particular locations in the reference grid. These locations implicitly define the
global spatial relations. There is no special learning phase for these relations.
The object representation is simply the list of selected features with the selected
locations on the reference grid.