Progress Report
October 25
th
, 2005
Overview
MIAS-IRC abstract submission: sensitivity, error propagation, values aggregated...
ISBI submission (progress dependent on abstract above)
Automatic model building in faces
Perturbation method revised (landmarks vs. dense correspondence)
Evaluation of face models: new experiments and results
Ways to proceed and forthcoming deadlines
MIAS-IRC Submission
Draft continuously modified
Latest version sent by E-mail
Includes a technical discussion of overlap
Requires a further overview for confirmation
Opportunity for improvements
Text/structure changes in alignment with a foreseeable ISBI submission
MIAS-IRC Submission - Assorted Points
The plan was to write up a 2-page paper for the 'internal' congregation and then expand it to fit 4 pages (ISBI)
The themes are merely identical, much like the flow of arguments
We can be more succinct by including just Tanimoto, which we have shown to be more sensitive than Dice
Remember to be self-critical of the methods
Model quality versus registration correctness - worth investigating on its own right
Generalisation - Final Plot
Generalisation is shown against the degree of misregistration
Overlap - Final Plot
Tanimoto overlap is shown against the degree of misregistration
Specificity - Final Plot
Specificity is shown against the degree of misregistration
Generalisation Sensitivity
Example sensitivity plot with errors propagated (Generalisation ability)
Specificity Sensitivity
Example sensitivity plot with errors propagated (Specificity)
Overlaid Plots of Sensitivity
Sensitivity is shown for each of the assessement methods
See
Finalised large figure
Aggregated Sensitivities
Aggregating the curves to derive a measure of sensitivity that is independent of the degree of displacement
Sensitivity - Final Plot
A comparison between the sensitivities of different assessment measures (grouped by method)
ISBI Submission Plan
Worth reviewing again and maybe composing text already, making use of the MIAS-IRC abstract
Deadline extended: November 15
th
ISBI Structure Plan (
PDF
)
Recent Work on Face Models
Work on some face datasets and automatic model construction
Aim is to register a set of faces (in practice, a single individual) and then evaluate the resulting model/s
We have a robust and well-understood perturbation framework
Can alternatively use the manually annotated data (difficult due to diversity) and show that we are able to properly evaluate models
Face Model Evaluation - Time Limitations
Need to re-build VXL in order to make use of piece-wise affine registration in the context of faces
To remain realistic and conform with deadlines, might make use our perturbation framework and show that we can assess face models
The gist:
take a set of manually annotated faces
deform each face image and measure mean pixel displacement, bending energy, etc.
show that our 'face evaluation' tool can detect slight difference in terms of model quality
Example faces
Example face image from the dataset, which includes 68 such images with fixed dimensions (140x150 pixels)
Face Models
Face Models with varying amounts of deformation applied to the training data
Face Models and Perturbed Sets
Face models as sigma of perturbation increases (smaller example)
Synthetic images
A synthetic image drawn from the appearance model (near the mean)
Shuffle Distance in Faces
Example of the effect of shuffle distance in the context of faces
Example Shuffle Distance
More detailed example of the shuffle distance
Face Model Generalisation
Generalisation of faces models where correspondence is progressively eroded
Face Model Specificity
Specificity of faces models where correspondence is progressively eroded
Normalisation Revisited
The normalisation framework once envisioned
Thesis Structural Outline
Has not been reviewed since last changes had been made
Could benefit from further extension and embedment of experiments as placeholders
Thesis Structure Plan (
PDF
)
Upcoming Deadlines and Submission Possibilities
ISBI 2006 - validation/assessment
Journal paper on registration - Possibly TMI
Possibly: CVPR 2006 on face model evaluation
Possibly: ICPR 2006 (to be determined, perhaps redundant)
Summary
Evaluating brains
Reasoning about the model evaluation framework
Sensitivities (conduction a comparison)
Flaws and areas to further explore
Evaluating faces
Reasoning about the framework
Consider NRR in faces
Comparison between facial AAM's
Present and Future
ISBI - extend MIAS-IRC abstract
CVPR - Work on faces, possibly automatic model building
Thesis progress
Possibility of normalisation, investigation of model-NRR relationship
Possibly: Exploration of other datasets