Because certain methods neglect available information

- NRR and Models
- Evaluation method
- Validation of the method
- Evaluation of NRR
- Comparison between NRR algorithms
- Summary and conclusions

- Align images using
- spatial transformations
- similarity measures

- Align images using
- spatial transformations
- similarity measures

- Take a data set

- Find correspondences in set
*Learn*how correspondences vary

- Take a data set
- Find correspondences in set

*Learn*how correspondences vary

- Take a data set
- Find correspondences in set
*Learn*how correspondences vary

- Need for an automatic approach

- Difficulties in 3-D
- Non-rigid registration to align data
- Produce deformation fields/grid

- Need for an automatic approach
- Difficulties in 3-D

Where is a corresponding point in the volume? Picture from Johan Montagnat, INRIA |

- Need for an automatic approach
- Difficulties in 3-D
- Non-rigid registration to align data

- Produce deformation fields/grid

- Need for an automatic approach
- Difficulties in 3-D
- Non-rigid registration to align data
- Produce deformation fields/grid

- Grids of deformation encapsulate variation

- Perform statistical analysis on grids
- Use model of variation for synthesis

- Grids of deformation encapsulate variation
- Perform statistical analysis on grids

- Use model of variation for synthesis

- Grids of deformation encapsulate variation
- Perform statistical analysis on grids
- Use model of variation for synthesis

First variation mode Second variation mode |

- Each registration results in a model
- Better registration » better model
- A good model is:
*specific*, i.e. instantiates only valid examples- capable of
*generalising*to new, unseen examples

- Specificity and Generalisation successfully exampled
- Optimal shape models (Davies
*et al.*2001)

*0* to *5* CPS warps perturbing the correct solution.

Shown is the first mode of the model, ±2.5 SD

Model of the registered images and synthesis from the model

A hyperspace representation where 'clouds' of images overlap

Calculating Specificity and Generalisation ability

- Distance naturally assumed Euclidean
- Shuffle distance performs better

As correspondences degrade, so does Generalisability (low values are good)

As correspondences degrade, so does Specificity

Investigate measures most *sensitive* to change

Shuffle distances covering a large region are sensitive to differences

The choice of shuffle distance radius becomes an efficiency vs. performance trade-off

- Registration builds models
*automatically* - Model from group-wise registration presented below

- The evaluation requires
**no ground truth**

Group-wise methods surpass pair-wise regardless of the expressiveness of the model used

- Each registration leads to a model
- Models can be evaluated
- Shuffle distance is used in evaluation
- Registration evaluated without ground truth

- Model construction and NRR are analogous
- NRR can be evaluated using its resulting model
- Models can be evaluated using a sparse distance map
- Shuffle distance is more robust than Euclidean
- Group-wise registration surpasses pair-wise