Progress Report

June 20th, 2005

Previously Suggested/Agreed Upon

• Fit curve to plot of warp magnitude versus pixel displacement
• Toward comprehensive overview on the method
  • plan required diagrams
  • add captions and headings where they ought to reside
  • add content to existing draft which incorporate NRR as well as model evaluation
  • yet to be passed around the Group and discussed
  • awaiting decision on the BMVC submission
• Consider Imperial College for porting of interfaces to suit NRR evaluation

Warps and Displacements

Linear relationship between warp magnitude and average displacement

Warps and Displacements

Linear relationship between warp magnitude and maximum displacement

Main Milestones

• Simultaneous perturbation of labels and images working for experiments
• Completing and testing the perturbation of labels and images
• Ensuring images and labels are deformed more properly than before
• Better understanding of the behaviour of pixel displacements

Implementation

• We have a set of m = 37 images with 8 corresponding labels for each
• The process involves the following steps:
  • Warp all images and labels to create n sets of n increasing perturbations
  • Build models for each set comprising m images
  • Evaluate the model

Example Warps #1

Original sample image

Warps and Displacements #2

image affected by warps with magnitude W_m = 0.02

Example Warps #3

image affected by warps with magnitude W_m = 0.05

Example Warps #4

image affected by warps with magnitude W_m = 0.1

Example Warps #5

image affected by warps with magnitude W_m = 0.3

Specificity

The value of model Specificity versus the mean pixel displacement

Generalisation

The value of model Generalisation versus the mean pixel displacement

Commentary on the Results

• The sampling of displacement values is more dense at the start
• Hence, curves will be smoother at more crucial points
• No error bars plotted, but these are significantly small
• Errors bars can be made tiny by running the evaluation for longer

Commentary on the Results - Ctd.

• 10 instantiations used to learn about mean and maximum pixel displacement
• 5 principal modes were used in the evaluation (future experiments to incorporate more, albeit experiments in BMVC submission suggest it results in insignificant gain)
• Shuffle: 13 neighbouring pixels (disc with diameter of 5 pixels)
• No Euclidean and various shuffle distances yet
• Corresponding labels are available for similar evaluation based ground truth

Extensions and Ways Forward

• Better registration with labels
• Repeating for stats
• More shuffle sizes
• Avoid going beyond 3-4 pixels of mean deformation
• Comparing against overlap-based measures by sending some new data to UCL

Extensions and Ways Forward - Ctd.

• Speed up the process of evaluation -- needs compiling on a networked, shared filestore
• Finer-level registration might to possibly make the 'dip' at the start vanish
• Repetitions of experiments with different instantiations to make the curve smoother

Related Documents

• Evaluation of AAM's and NRR (PDF)
• Pending Experiments (PDF, HTML)

Lateral Thoughts

• Can models of the labels can be of any use?
• Registrations are based on image data alone
• Perform a comparison between registration based on images and registration that depends on the ground truth
• Similar to the use of the labels to compute overlap