A Generic Approach To Coregistration

1
A Generic Approach ToCo-registration

• Rik Fransens
• Christoph Strecha
• Geert Caenen

2
Co-registration
3
Mathematical Framework

• Image
• Raster Image
• Transformation (Groups)
• Similarity Measures

4
Mathematical Framework

• Image I (x)
• Raster Image
• Transformation (Groups)
• Similarity Measures

I (x,y)
x
y
5
Mathematical Framework

• Image I (x)
• Raster Image I i1 int
• Transformation (Groups)
• Similarity Measures

I (x,y)
x
y
6
Mathematical Framework

• Image I (x)
• Raster Image I i1 int
• Transformation (Groups) T(xp)
• Similarity Measures

7
Mathematical Framework

• Image I (x)
• Raster Image I i1 int
• Transformation (Groups) T(xp)
• Similarity Measures C(X,Y)

Image 2
Image 2
Image 2
Xi aYi b
p(Xi,Yi)
Xi Yi
Image 1
Image 1
Image 1
8
Mathematical Framework

• Raster Images I1, I2 (resp. I1,I2)
• Transformation type T(xp)
• Similarity Measure C(X,Y)
• F(p) C( I1, I2(T(xip)i )
• MAX F(p)
• dF/dp dC/dI2 dI2/dx dx/dp

9
Implementation
Library with Plug Play-support
10
Implementation

• Optimisation
• Multi-resolution
• Deterministic / stochastic
• Dense vs. subsampled

I1
I1(1)
I1(2)
T
T(2)
T(1)
I2
I2(1)
I2(2)
11
Implementation

• Optimisation
• Multi-resolution
• Deterministic / stochastic
• Dense vs. subsampled

p (pF)
?F(p)
p
p
12
Implementation

• Optimisation
• Multi-resolution
• Deterministic / stochastic
• Dense vs. subsampled

SSD ? Xi-Yi2
c ? Xi-Yi2
All pixels
Fewer pixels
13
Experiments

• Ground-truth data

image
spectral noise
2?
pgt minp xdist,i-T(xip)2
14
Experiments

• Quality of the result

15
Experiments

• Quality of the result

16
Experiments

• Quality of the result

17
Experiments

• Influence of Geometric Distortion

18
Experiments

• Influence of Spectral Noise

19
Experiments

• Optimisation Speed

20
Experiments

• Real Experiments image map registration

21
Experiments

• Real Experiments optical flow

22
Experiments

• Real Experiments linear spline

23
Experiments

• Real Experiments linear spline

24
Experiments

• Real Experiments change detection

SSD 1 - exp -yi-xi2/(2s2) NCC 1 -
exp -yi-axi-b2/(2s2) MI max
1-p(yixi) , 1-p(xiyi)
25
(No Transcript)
26
(No Transcript)
27
Summary Discussion

• Mathematical Framework generic
• Implementation modular, object-oriented
• Similarity Measures
• Transformations
• Optimisation
• Experiments
• Ground-truth data subpixel convergence!
• Photometric distortion deteriorates results
• MI converges less accurately
• Parametrisation of transformations is an issue.
  (scaling)
• Real Data
• Change detection

28
Summary Discussion

• Future work
• Framework Deterministic ? Probabilistic
• Simultaneous Outlier Detection and registration
  (robust)
• Links co-registration with MVstereo
• Prior knowledge (intelligence)
• Spatial coherence of outliers using MRFs
• Vector Raster registration (!)
• Expand Transformation Library
• Speed / Initialisation
• Biased subsampling ? Feature based
  co-registration
• Initial registration
• Parallel processing

29
Summary Discussion
Input 3 images                               
                              
                               Resultsvisibili
ty maps                               
                               depth and ideal
image