The **Template Tool** has been designed for use in automatically
locating deformable shape in noisy (generally medical) data sets.
Using it well requires some familiarity with the data
and a little thought as to what is inteneded to be achieved.
Start by determining the best number of points
needed to accurately define the boundary of the object and whether
an inner contour is also needed. Then experiment with the data to
try to determine a reproducable way of positioning boundary points
as accurately as possible. draw up guidelines for features to look
out for and try to define the location of the boundary in a simple and
reproducable manner. If necessary images can be pre-processed by the
**Imcalc Tool** in order to assist with this process (eg:
enhancing edges etc.).

Once you have a good idea of what you need to do then proceed with the following steps;

- Mark up a series of example profiles which you feel span the space of possible shape and grey-level variation. The number you will need will depend upon the degree of variability in the data and cannot be specified a-priori, but even simple models will generally require tens of examples.
- Construct a list file for the data you wish to include in the model and train using the PCA model builder. Use a maximum number of parameters without global covariance at this stage.
- Test the performance of the model on unseen datasets varying the number of parameters used for the shape and grey level variation. Try to use a minimum of the number of shape parameters which accurately locates the boundary. Investigate the effects of robust optimisation and data normalisation to see if there are any particular advatages to using these options. Always take care that you use the same cost function options during search that were used during PCA analysis and never attempt to search with more parameters than were written out during the analysis process (you can however use fewer).
- Test on additional datasets and if the location process repeatedly fails for a particular image mark it up and add it to the dataset for retraining.

The posibilities for data pre-processing before building the deformable model
are limitless, but an attempt should be made to work with data-sets
which have the properties of uniform random errors (see the **Imcalc
Tool**) as this is what is most consistent with the statistical
assumptions behind greylevel modelling
and least-squares based location algorithms.

The resulting tracking variables of orientation, scale and principle eigen-modes of the shape model make a good starting point as a reduced representation of the data for any subsequent anaysis such as statistical pattern recognition (eg: classification). However, these techniques are not expected to form the basis of a generic object recognition system as there is no robust way of automatically selecting or building models for arbitrary image data sets.