Once a model has been built or loaded, an instance of the model can be located in the selected Tv image using the search function, which executes a downhill simplex optimisation to search for the model in the image which minimises the selected Cost function. Executing this routine again continues the search from the last solution.
Generally when automating the final system there is no harm in executing the search process twice to try to ensure convergence. Searching for objects through temporal sequences of images generally benefits from starting each search from the previous solution. Coarse to fine searches for both snakes and deformable templates and any permutation can be supported using the macro facility and combinations of trained models. Multiple objects or disjoint or hinged structures should be located using separate trained models for each section.
This option makes the selection between snake localisation and localisation based on grey-level profiles. Simple snake location operates via optimisation of an image potential using the specified shape parameters only. When specifying this option the local image potential is optimised along a line tangential to the control points. When locating a grey-level template, the local template profile is acquired tangential to the local boundary. If this is not the case check the order in which the boundary points were specified for the model (see above).
Use of mean squared error (mse) or sum of absolute differences (a robust statistic).