$next$ $up$ $previous$
: 5 Summary and Conclusions : A Framework for Evaluation : 3.3 Normalisation

4 Results

Experiments... then results...

Face landmark points have been perturbed by noise. The magnitude of noise is defined by a standard distribution and the value sigma of that distribution varies.

$\includegraphics[% scale=0.36]{EPS/Shuffle_Difference_Image/shuffle_distance_example.eps}$

Figure x: Example of shuffle distance...

$\includegraphics[% scale=0.3]{EPS/faces_models_with_sigma.eps}$

Figure x: The 3 Principal modes of variation of a correct face model (left) and the corresponding model whose landmarks have been perturbed by Gaussian noise of 3 standard deviations.

$\includegraphics[% scale=0.3]{EPS/faces_sigma_15032005_gen_processed.eps}$ $\includegraphics[% scale=0.3]{EPS/faces_sigma_15032005_spec_processed.eps}$

Figure x:

$\includegraphics[% scale=0.37]{EPS/Shuffle_Difference_Image/face_shuffle_3.eps}$ $\includegraphics[% scale=0.3]{EPS/shuffle3x3_power1.eps}$

Figure x: Two face images and the resulting shuffle difference image with neighbourhood of size 3 $\times$ 3 shuffle Figure x:Shuffle difference image of the brain 3 $\times$ 3.

$\begin{figure}\par \end{figure}$

To put the algorithm in a challenging position, images were warped while landmark points remained the same. This obtained fuzzier models, but models that appear merely identical. To control the amount of noise, a progressively-increasing number of clamped-plate splines (REF) was applied to all images. All previous warped remained unchanged since the previous passes for stability, but since warps can improve models as well as degrading them, monotonous were not expected. The trend, however, was expected to show the movement of the images around the points resulted in worse models, as one would expect.

In the case of the brains, an even greater amount of warps was applied to see the effect in a larger scale. The data used was MR-..... brains obtained from.... aligned affinely and sliced....

$\includegraphics[% scale=0.4]{EPS/gen_modes_processed.eps}$ $\includegraphics[% scale=0.4]{EPS/spec_modes_processed.eps}$

Figure x:

$next$ $up$ $previous$
: 5 Summary and Conclusions : A Framework for Evaluation : 3.3 Normalisation

Roy Schestowitz 平成17年6月23日