next up previous index
Next: How Minimal Configuration Changes Up: Application in MRF Object Recognition Previous: Posterior Energy

7.3.2

Energy in Linear Form

Parameters involved in are the noise variances and the prior penalties (n=1,2). Let the parameters be denoted uniformly by . For k=0,

and for ,

 

Note all . Let the different energy components be uniformly denoted by . For k=0,

 

is the number of NULL labels in f and

 

is the number of label pairs at least one of which is NULL . For , relates to the likelihood energy components; They measure how much the observations deviate from the should-be true values under f:

 

and

 

After some manipulation, the energy can be written as

where and are column vectors of components. Given an instance , the is a known vector of real numbers. The is the vector of unknown weights to be determined. The stability follows immediately as

where .

Some remarks on , and E are in order. Obviously, all and are non-negative. In the ideal case of exact (possibly partial) matching, all () are zeros because and are exactly the same and so are and . In the general case of inexact matching, the sum of the should be as small as possible for the minimal solution. The following are some properties of E:

The first property enables us to use the results we established on linear classifiers for learning correct and optimal . According to the second property, a larger relative to the rest makes the constraint and play a more important role. Useless and mis-leading constraints and should be weighted by 0. Using the third property, one can decrease values to increase the number of the NULL labels. This is because for the minimum energy matching, lower cost for NULL labels makes more sites be labeled NULL , which is equivalent to discarding more not-so-reliable non- NULL labels into the NULL bin.