Progress Report
March 15th, 2005
Protein Gel Registration Experiments
Images are similar (in their general form)
They comprise many small separate elements
One-to-one correspondence exists
Discrepancy image is rough after a short registration
shown in the next slide
Protein Gel Registration Experiments - Ctd.
CPS warps are used to achieve this
Further efforts have been conceded at this stage
Model Evaluation Improvements
GUI
tools for syntheses generation
Generation was made about 5 times quicker
Gaussian distribution for choice of syntheses
An option for a flat distribution still exists
Many more syntheses attempted recently
10,000 at most, IPMI experiment included 1,000
This results in heavy calculations
1,040,000 'distances' for the brain experiments
VXL QT Tools
Synthesis from a model added to the
GUI
More options (customisations) are available
VXL QT Tools - Ctd.
There are more possible extensions like automated evaluation
Implementation of these would be effort-consuming
Model Evaluation Experiments
Further model evaluations
Similar experiments to ones in the IPMI submission
Gaussian distribution to draw syntheses
Much larger experiments (~7 days run-time for each)
Run in parallel (lowered
significantly
in scheduler)
Model Evaluation Experiments - Explanation
The following are a few brief notes
Recapitulate the problem being solved
Explain how the concepts were extended
Model Evaluation Experiments - Explanation - Ctd.
A collection of raw brains given
Data from Paul Bromiley: 104 brain slices, affinely-aligned
Model Evaluation Experiments - Explanation - Ctd.
Syntheses are produced from model
These tend to be smaller in size
A little fuzzy if instantiated after automatic model-building
36 hand-annotated brains available too
Model Evaluation Experiments - Framework
Model Evaluation Experiments - Distances
All possible training image-synthesis pairings are considered
Distance is calculated using the shuffle transform
A matrix of distances is obtained
Model Evaluation Experiments - Measures
Specificity and generalisability are derived directly from the matrix
Various shuffle radii/window sizes attempted
Model Evaluation Experiments
C0: Pairwise, 16 x 16 grid, residuals evaluated in reference frame
C1: Pairwise, 16 x 16 grid, residuals evaluated in training image frame
C2: Pairwise, 16 x 16 grid, ignoring low variance regions
Model Evaluation Experiments - Ctd.
C3: Pairwise, 16 x 16 grid, ignoring low variance regions, moving to strong edges
C4: Groupwise, (points as C3), (mean only) - i.e. 'Initialisation' algorithm
C5: Groupwise, (points as C3), including shape model [width=10]
Model Evaluation Experiments - Results
10,000 synthetic images
10 most principal modes of variation
All drawn from a normal distribution
Model Evaluation Experiments - Specificity
Some interesting behaviour to analyse
Model Evaluation Experiments - Generalisability
Relatively noisy, no valid explanation yet
Face Experiments
Set of face images obtained from Tim
68 faces from Germany; quite diverse
Many glasses and beards; pose and light variation
Images are marked up, but not aligned
Images were later rigidly aligned
Face model evaluation in progress
Faces - Examples
Below are two images from the set after rigid alignment under MATLAB
The black frames near the borders are due to alignment
Models Sigma
Models can then be built from the set
Models Sigma - Ctd.
Perturbation of the landmarks affects model quality
Perturbation with
Sigma=3
Evaluation Difficulties
Note the background artefacts in original images
These interfere with sensible distance (similarity) metrics
Masked Faces
The background is omitted from all images
This obtains a new model
The training set becomes more easily-comparable
Masked Faces - Synthesis
Some of the synthetic images from the model:
Similarity to images from the training set becomes clearer
Images still need to be re-scaled to overlap
Face Model Evaluation
Current experiments investigate landmark noise
Sigma ranging from 0 to 6 inclusive
Distance measures based on 1000 syntheses
Carole yet to give generalisability and specificity figures
Specificity is expected (or hoped) to rise as a function of sigma
Miscellany
MATLAB Meeting
MICCAI, BMVC, MIUA Deadlines
Recommendation:
Google Desktop