PDF version of this entire document
At the moment there are 3 similarity measures implemented: one checks the absolute differences (squared if that is deemed required) applying the residual, one adds the residual to the PCA model and assesses the changes this imposes upon the model, and the third (which is a slow one to run) is a family of methods that warp any residual corresponding to a pair into EDM space and then checks the differences, assuming there is nothing suspicious in the data which made the models (Mian's group too had problems with this). The PCA modes that we currently have have been improved considerably by reworking two phases; one is preparing the data for alignment with ICP and one filters the data for subsequent stages that compare pairs.
