PDF version of this document

Evaluating Registration

The method was used to learn about registration algorithms that often lack benchmark tools. Crum et al. [17] have devised validation methods based on overlap measures of markup. These can measure the quality of registration, however they require ground-truth annotation.

Fig. 8 shows an appearance model which was built automatically using group-wise registration. Group-wise registration account for the entire set when registering, whereas pair-wise register with respect to a single reference. The initialisation algorithm is an information-theoretic group-wise algorithm that is described in [16].

Fig. 8. Appearance model which was built automatically by group-wise registration. First mode is shown, $\pm2.5$ standard deviations.

It can be inferred from the results (Fig. 9) that group-wise registration outperforms pair-wise registration. This means that a group-wise registration leads to better model and is hence better representative of the group.

  

Fig. 9. Registration evaluation which compares 3 different registration algorithms. Specificity is shown on the left and generalisation ability on the right. Values are the mean over a wide range of modes in the model.

It was ensured that the comparison is not dependent upon the process where syntheses are extracted from the model. Since a finite number of modes is selected in synthesis, the range of 2-20 modes was investigated and the figure shows the average. Nonetheless, the properties of the bar charts did remain consistent throughout.