next up previous index
Next: Application in MRF Object Recognition Up: Theory of Parameter Estimation Previous: A Non-parametric Learning Algorithm

7.2.5

Reducing Search Space

The data set in (7.20) may be very large because there are a combinatorial number of possible configurations in . In principle, all should be considered. However, we assume that the configurations which are neighboring have the largest influence on the selection of . Define the neighborhood of as

 

where reads ``exists one and only one'' and is the set of admissible labels for every . The consists of all that differ from by one and only one component. This confinement reduces the search space to an enormous extent. After the configuration space is confined to , the set of training data is computed as

 

which is much smaller than the in Eq.(7.20).