Progress Report
February 22nd, 2005
Overview
Face models
Facial appearance models
Evaluation and perturbation
Results
Model Evaluation
Number of modes (in brain)
Future paths
Summary
Masked Faces
The background is omitted from all training images
The training set becomes easier to compare
Faces: Models Sigma
Models can then be built from a set
Faces: Models Sigma - Ctd.
Perturbation of the landmarks affects model quality
Perturbation with
Sigma=3
Masked Faces - Synthesis
Some of the synthetic images from the model:
Images still need to be re-scaled to overlap
Face Model Evaluation
Experiments investigate landmark noise
Sigma ranging from 0 to 5 inclusive
Distance measures based on 1000 syntheses
Specificity was expected to rise as a function of sigma
Face Model Evaluation: Results
Generalisability
Face Model Evaluation: Results
Specificity
Face Model Evaluation: Results
Example distance matrix
Brain Model Evaluation: Explanation
We investigate the effect of the number of modes
Modes which are used to evaluate a model
Synthetic data is generated from the appearance model
A finite number of principal modes is selected
Brain Model Evaluation: Explanation
Investigate the effect of the number of modes
Referring to modes which are used to evaluate a model
Synthetic data is generated from the appearance model
A finite number of principal modes is selected
Brain Model Evaluation: Figures
If all modes are altered in generation simultaneously...
...distance measures become less meaningful
Hence more synthetic data is required
Brain Model Evaluation: Experiments
3 models were used to learn the effect of the number of modes:
Model I: built from the results of pair-wise registration
Model II: built using the initialisation algorithm
Model III: built using group-wise registration
Brain Model Evaluation: Results
Generalisability
Brain Model Evaluation: Results
Specificity
Brain Model Evaluation: Conclusions
Generalisability on a steady decrease
Group-wise outperforms pair-wise regardless of number of modes
Complex methods surpass the basic initialisation algorithm
Specificity of group-wise almost independent of number of modes
Specificity of pair-wise degrades approximately linearly
Brain Model Evaluation: Future
Normalisation in evaluation:
Set size
Other factors
Extension of training set using synthetic data
Larger experiments
MICCAI / BMVC
Deadline is mid-April for both
Wiki
:
A Wiki can be set up
Used for collaborated editing of content
Web-based and database driven
Easy rollback, backup, '
diff
'
Possible Paper Structure [1]
Introduction
Pair-wise versus group-wise
Automatic model-building
Model evaluation
Proof of group-wise being most desirable
Possible Paper Structure [2]
Automatic Model-Building
Explanation of the algorithm/s
Hopefully tetrahedronisation
If 3-D results are available:
Pick a slice from the volume to evaluate the 3-D model
This will provide an approximation
Initialisation algorithm in 3-D
Possible Paper Structure [3]
Model Evaluation
Explanation of the methods:
Shuffle distance
Normalisation, scaling issues
Deriving Spec. and Gen. in relation to work on shape
Possible Paper Structure [4]
Brain
Sigma of noise versus Spec. and Gen.
Spec. and Gen. versus the number of modes
Spec. (maybe also Gen.) for different registration methods
Possible Paper Structure [5]
Faces
Show faces from the model
Show noise being applied to face mark-up and the resulting model/s
Experiments with noise - an explanation
Specificity and Generalisability versus sigma
Possible Paper Structure [6]
Conclusions, Summary, Discussion, etc.
Repeating the main points from the sections above
Future work:
Full 3-D (if not already done)
Larger training set (possibility of creating pseudo-training data)
Further investigation of registration methods
Investigation of evaluation parameters