PDF version of this document

Validity of the Method

To satisfy ourselves that the method behaves properly, we demonstrated that a hand-annotated model gets assigned fixed values for specificity and generalisation. Noise was then applied to the annotation, the model re-built and as a result of that noise, values of specificity and generalisability measures were negatively affected. Noise that was applied to the mark-up resulted a steady rise in these value (indicating exacerbation) so even less trivial experiments were embarked upon.

To established even more confidence in the evaluation criterion, subtle changes were made to the images rather than the mark-up. To do so, images were transformed locally while the landmark points upon them remained unchanged. This means that correspondences will be badly affected and lose their meaning. By applying a different number of warps at each stage (while keeping older warps in place), landmark points should usually be located at worse positions. Hence, the model representing all images under consideration is expected to be aggravated. Shown below in Fig. 6 (also confer Fig. 1 where the corresponding correct model is included) are models created from images that were subjected to a varying number of small, localised clamped-plate spline warps. It can be seen that the model becomes more fuzzy and less realistic as more warps are applied.

Fig. 6. The first mode of an appearance model of the brain whose training set was subjected to diffeomorphic warps - 15 warps for the model at the top and 30 for the one at the bottom. $\pm2.5$ standard deviations are shown.

  

Fig. 7. On the left: The effect of image perturbation on specificity; On the right: The effect of image perturbation on generalisation ability.