Progress Report

July 5th, 2005

Overview

• Model Evaluation
  • bending energy used in analysis
  • forcing of re-sampling in perturbation
  • input data for experiments soon to arrive
• Returning to analysis of automatic model construction
  • new resource site and experiments log
  • preliminary experiments
  • towards propagation of segmentation (labels) using NRR

Discussed or Agreed Upon

• Plot mean bending energy versus degree of perturbation
• Apply a shift (translation) of 0.5 pixels to all warps
• Wait for image data, then run evaluation experiments
• Improved Euclidean distance to be viewed as a possible extension
• Reconsider segmentation joint with NRR

Image Re-sampling in Perturbation

• Re-sampling anomaly resolved
• Basic translation is being forced
• A fixed half-a-pixel offset to affect interpolation
• Blurring effect in images under all circumstances

Forcing Data Re-sampling

On the left: original data grid (no perturbation)
On the right: the same data re-sampled (translation of 0.5 pixels)

Bending Energy Investigated

• Computing mean bending energy in warped images
• 4 knot-points to get results quickly
• Averaged over 10 instantiations
• Warp magnitude sampled at points: [0,0.01, 0.05, 0.1, 0.2, 0.4]
• Corresponding mean bending energy: [0,6.2935e-12, 1.2062e-10, 7.4499e-10, 1.6296e-09, 1.8232e-08]

Bending Energy versus Perturbation

The increase in mean bending energy against mean of distribution of warp magnitude. These warps (CPS) make up perturbations.

Bending Energy and Perturbation: Side Notes

• Dots in the plot make the sample points more prominent
• Assumed that:
  • only a few points are needed
  • the general behaviour needs analysing
  • appears to be exponential
• Error bars can be added if required
• The errors are not of any particular interest

Perturbation Experiments and Ground Truth

• Experiments are ready to run once data becomes available
• Changes made to the code caused it to become unusable for the time being
• Waiting until everything changes to the new convention
• Also an urgency to obtain and visualise 3-D models (to show in IPMI)

Returning to NRR, Models and Segmentation

• Site with news section to keep track of progress
• Progress reports will still be delivered as presentations
• Migration and re-use of code from AART, which combined NRR and models in 1-D
• Project now titled MARS (Models of Appearance, Registration and Segmentation) to reflect on its goals
• Experiments log gets generated automatically

Current Work

• Investigate possible contributions:
  • what does segmentation have to offer to the process of automatic model-building?
  • can models and NRR to improve and aid segmentation?
  • can robustness to noise be archived by making use of the entire group?

Starting Point

• Can output video of models automatically constructed using NRR
• This video shows a bump model in motion:
  • the first mode incorporates bump position
  • second mode is primarily associated with height
• The model is built automatically using fast (coarse-level) non-rigid registration in 1-D

Model of Bump Data

A model built from bump data within seconds

Data Visualisation

Bump data (vectors) shown as meshes

Reconstruction from Model

The original data and its reconstruction from an automatically-built model

Reconstruction Error

Residual error calculated using mean of squared differences

The Next Goal

Illustrating the propagation of segmentation (labels) using NRR (or model construction which is related)

Similar Work

• UCL initiated similar work to be presented in the Oxford Plenary (July 21^st)
• Title found in the programme sent yesterday
• Talk involves registration, NRR and segmentation

Summary

• Evaluation experiments:
  • Expect no Specificity 'dip' at start of data degradation
  • Bending energy grows exponentially
• Segmentation and registration:
  • Improve segmentation by exploiting automatic model-building
  • will attempt to propagate labels using NRR techniques