Perturbation Method
Notes for discussion
Problems Identified: Locality
Problems Identified: Re-sampling
Note:
Warp fields aggregation will resolve this issue
Problems Identified: Diffeomorphism
Diffeomorphism breaks with multi knot-point splines
Required Perturbation Properties: #1
Range of scales
Images are subjected to warps of varying intensity
Perturbation framework to allow intensity (magnitude) to increase by a fixed amount
Flexible in terms of the scales supported or else obscenely large deformations cannot be investigated.
To investigate:
at which point does the amount of perturbation become difficult to detect and quantify?
The effect of small warps
Required Perturbation Properties: #2
Diffeomorphism
No folding and tearing
Required Perturbation Properties: #3
Large number of perturbation 'sources'
To make the effects of perturbation less local and more global
Large number of random processes, e.g. warps
Spread around the image boundaries
Required Perturbation Properties: #4
No pixel condensation at images edges and corners
'Stuffing' of pixels
Tends to happen when there is not sufficient freedom for pixels to be moved outside the image boundaries
Required Perturbation Properties: #5
Stochastic
Perturbation needs to have a stochastic nature
Points needs to be displaced by a random unit
Displacement drawn from a normal distribution
Required Perturbation Properties: #6
Predictable
For any given point, the distribution of its displacements needs to be well-understood
Required Perturbation Properties: #7
Similar distributions across the entire image
Displacements affect all parts of the image similarly
May be difficult to assure
Required Perturbation Properties: #8
No re-sampling error
Images are warped (transformed), interpolated and re-sampled
Loss of detail, visible in the form of blurring
Good perturbation will have small errors
Errors add noise to the final results
Alternative Approach
Add buffers to images
Apply smaller localised warps
Single-point CPS