PDF version of this document

Part 1: Validation of Evaluation

• Comparing Euclidean, shuffles

  - Dataset comprises 104 brains, which have been registered in a group-wise fashion in VXL

  - Perturb the data uniformly and create a collection of perturbed image sets (repeat 10 times for better statistics)

  - The perturbed sets need to have pixel displacements which increase linearly between each given set and its predecessors

  - Apply symmetric shuffle distance with radius of size 1-5 (Euclidean and shuffles)

  - 1000 synthetic images to be generated

  - Derive Generalisability and Specificity with errors for all shuffle radii

  - Derive sensitivity

  - Repeat 5 time for all sets of perturbed images and take the averages

• Use the face data similarly. The Surrey face database contains ~60 instances.