PDF version of this document

Application to Model Evaluation

We used our new method to evaluate three different models built using an enlarged set of the brain data containing 104 affine aligned images (ground truth was not required for this experiment). It has been shown previously [14,11] that an appearance model can be built by registering each image in a set (pairwise) to a reference image. In [16] we argued that a 'groupwise' approach that took proper account of the whole group might be expected to perform better. We built three models, one using the pairwise approach, and two variants of our groupwise approach. The results, including different numbers of modes in the models, are shown in Figure 7 and demonstrate a clear advantage in terms of both Specificity and Generalisation for both groupwise methods over the pairwise approach. It was not possible to discriminate between the two groupwise methods.

Figure 7: Specificity and Generalisation of the three automatic model construction approaches

[width = 0.98 ]../Graphics/reg_results.png