PDF version of this document

Part 2: Planned Experiments

Values in blue needs to be agreed upon

• We wish to take a registered set, perturb it, and then evaluate its model
• Given a perturbation method which behaves the way we expect,
  • Run a series of [10] instantiations. This means re-selecting random warps and averaging the results to get smoother curves.
    • Run a series of progressively increasing warps. A large enough series is required in order to sample the model quality curve. The more points, the smoother the curve; [7] point might be a reasonable number.
      • Try a variety of shuffle distances in the evaluation, e.g. Euclidean, 5 neighbours, 9, and 12 neighbours? (totalling in [4]shuffle radii)
      • Investigate the inclusion of a different number of modes in the evaluation. Will just [1] choice (say [5]principal modes) suffice?
• Compare results with overlap measures
• Open questions: which results to compare specifically? How can different results, corresponding to different parameters, be composed in a single figure?