Let us consider the situation where an object consists of different types of
features such as points and lines. Obviously, a point in the scene
should not be matched to a line in an object model. This is a symbolic
constraint. In this case, the positivity condition of MRF in
Equ.(1.19) does not hold any more if the configuration
space is still defined as the simple product as
in Eq.(5.4) for a single MRF.
To overcome the limitation, we partition the whole set of labels to
a few admissible sets for different types of sites. This results in a
few coupled MRFs. These MRFs are coupled to each other via
inter-relations
(
). For example, the distance between a
point and a line can constrain the two different types of features.
Furthermore, they are also coupled via the label NULL
which is a
``wildcard'' compatible to all types of features.
If there are two different types of features, then can be
partitioned into two admissible sets, with each set consisting of
indices to a single type of features. In the most general case, each of
the m sites has its own set of labels
(
), each
being determined using the symbolic unary
constraints; and the label for site i assumes a value
where
. Then, the configuration space is defined
as
In this situation, the energy has the same form as usual and the
solution is still found by
. The only
difference is in the definition of the configuration space
in which
the solution is searched for.