Evaluation of Models and Non-Rigid Registration

April 22^nd, 2005

Overview

• NRR and Models
• Evaluation method
• Method validation
• Evaluation of appearance models
• Comparison between NRR methods

Non-Rigid Registration «-» Modelling

• Models require correspondences
• Problems:
  • Correspondence in large training sets
  • Subjective, prone to error
  • Laborious, especially in 3-D
• NRR identifies correspondences automatically

Non-Rigid Registration «-» Modelling

• Good NRR results in a good model
• A good model is:
  • Compact
  • Specific
  • Capable of generalising to new, unseen examples

Evaluation of Appearance Models

• Model quality degrades when correspondences are lost

Evaluation of Appearance Models

0 to 5 large CPS warps applied to each training image

Evaluation of Appearance Models

• Model quality degrades when correspondences are lost
• Devise a measure which evaluates models

Evaluation of Appearance Models

• Model quality degrades when correspondences are lost
• Devise a measure which evaluates models
• Compare model examples against training set

Evaluation of Appearance Models

The model evaluation framework

Evaluation of Appearance Models

A hyperspace equivalent

Evaluation of Appearance Models

Calculating Specificity and Generalisation ability

Measuring Distance

Measuring Distance - Shuffle Distance

Validation of the Method

Validation of the Method

Evaluation of Registration

Registration Algorithms - Comparison

Summary

Conclusions