During the construction of our tree, the MDF's are computed locally. For each subregion , we obtain DKL projection matrices and and mean vector based on the training samples within , where is the projection matrix to the MEF space and is the projection matrix to the MDF space as defined previously. The leaves of the partition tree correspond to the regions which contain the training samples from a single class. The approximator uses the following decision rule to classify the query fovea vector to the class of a leaf cell.
Since each local cell has its own DKL projection, in order to logically compare between two different cells, we use a measurement called Mixture Distance ( ).
Intuitively, what is being measured can be seen in Fig. 10. In Fig. 10, the original image space is a 3D space, the MEF space is a 2D subspace, and the MDF space is 1D subspace since two classes are well separated along the first MDF vector. The first term under the radical indicates the distance of the original vector from the population which indicates how well the MEF subspace represents the query vector . This term is necessary since it is entirely possible that a query vector that is miles away from a particular subregion's MEF subspace would project very near to the region's center. The second term indicates the distance between the MDF components of the query vector and the MDF components of the center vector in the original image space.
Figure 10:
Illustration of components in the Mixture Distance in a 3D original space.