![Page 1: Properties of Histograms and their Use for Recognition Stathis Hadjidemetriou, Michael Grossberg, Shree Nayar Department of Computer Science Columbia University](https://reader030.vdocuments.mx/reader030/viewer/2022032704/56649d6a5503460f94a48e47/html5/thumbnails/1.jpg)
Properties of Histograms and their Use for Recognition
Stathis Hadjidemetriou, Michael Grossberg,
Shree Nayar
Department of Computer Science
Columbia University
New York, NY 10027
![Page 2: Properties of Histograms and their Use for Recognition Stathis Hadjidemetriou, Michael Grossberg, Shree Nayar Department of Computer Science Columbia University](https://reader030.vdocuments.mx/reader030/viewer/2022032704/56649d6a5503460f94a48e47/html5/thumbnails/2.jpg)
Motivation
• Histogramming is a simple operation:
![Page 3: Properties of Histograms and their Use for Recognition Stathis Hadjidemetriou, Michael Grossberg, Shree Nayar Department of Computer Science Columbia University](https://reader030.vdocuments.mx/reader030/viewer/2022032704/56649d6a5503460f94a48e47/html5/thumbnails/3.jpg)
• Histograms have been used for:– Object recognition [Swain & Ballard 91, Stricker & Orengo 95]
– Indexing from visual databases [Bach et al, 96, Niblack et al 93, Zhang et al 95]
• Histogram advantages:– Efficient– Robust [Chatterjee, 96]
• Histogram limitation: – Do not represent spatial information
Motivation
![Page 4: Properties of Histograms and their Use for Recognition Stathis Hadjidemetriou, Michael Grossberg, Shree Nayar Department of Computer Science Columbia University](https://reader030.vdocuments.mx/reader030/viewer/2022032704/56649d6a5503460f94a48e47/html5/thumbnails/4.jpg)
Overview
•Image transformations that preserve the histogram
•Image structure through the multiresolution histogram
•Multiresolution histogram compared with other features
![Page 5: Properties of Histograms and their Use for Recognition Stathis Hadjidemetriou, Michael Grossberg, Shree Nayar Department of Computer Science Columbia University](https://reader030.vdocuments.mx/reader030/viewer/2022032704/56649d6a5503460f94a48e47/html5/thumbnails/5.jpg)
Invariance of Histogram with Discontinuous Transformations
Cut and rearrange regions
Shuffle pixels
![Page 6: Properties of Histograms and their Use for Recognition Stathis Hadjidemetriou, Michael Grossberg, Shree Nayar Department of Computer Science Columbia University](https://reader030.vdocuments.mx/reader030/viewer/2022032704/56649d6a5503460f94a48e47/html5/thumbnails/6.jpg)
Invariance of Histogram with Continuous Transformations
Rotation
Shear
![Page 7: Properties of Histograms and their Use for Recognition Stathis Hadjidemetriou, Michael Grossberg, Shree Nayar Department of Computer Science Columbia University](https://reader030.vdocuments.mx/reader030/viewer/2022032704/56649d6a5503460f94a48e47/html5/thumbnails/7.jpg)
What is the complete class of continuous transformations that
preserves the histogram?
![Page 8: Properties of Histograms and their Use for Recognition Stathis Hadjidemetriou, Michael Grossberg, Shree Nayar Department of Computer Science Columbia University](https://reader030.vdocuments.mx/reader030/viewer/2022032704/56649d6a5503460f94a48e47/html5/thumbnails/8.jpg)
Model for Image
Continuous domain Image: Map from continuous domain to intensities
![Page 9: Properties of Histograms and their Use for Recognition Stathis Hadjidemetriou, Michael Grossberg, Shree Nayar Department of Computer Science Columbia University](https://reader030.vdocuments.mx/reader030/viewer/2022032704/56649d6a5503460f94a48e47/html5/thumbnails/9.jpg)
Model for Histogram
U
Histogram count for bin U
≡Area bounded by level sets U
U
![Page 10: Properties of Histograms and their Use for Recognition Stathis Hadjidemetriou, Michael Grossberg, Shree Nayar Department of Computer Science Columbia University](https://reader030.vdocuments.mx/reader030/viewer/2022032704/56649d6a5503460f94a48e47/html5/thumbnails/10.jpg)
•Vector fields, X, morph images [Spivak, 65]:
Continuous Image Transformations
X
![Page 11: Properties of Histograms and their Use for Recognition Stathis Hadjidemetriou, Michael Grossberg, Shree Nayar Department of Computer Science Columbia University](https://reader030.vdocuments.mx/reader030/viewer/2022032704/56649d6a5503460f94a48e47/html5/thumbnails/11.jpg)
Gradient Transformations
FX
Original
5.122 )( yxF
40
)(sin
22 yxF
xyF
yxF 2
20
)(sin.
.20
)(sin
yx
yxF
![Page 12: Properties of Histograms and their Use for Recognition Stathis Hadjidemetriou, Michael Grossberg, Shree Nayar Department of Computer Science Columbia University](https://reader030.vdocuments.mx/reader030/viewer/2022032704/56649d6a5503460f94a48e47/html5/thumbnails/12.jpg)
Histograms of Gradient Transformations
![Page 13: Properties of Histograms and their Use for Recognition Stathis Hadjidemetriou, Michael Grossberg, Shree Nayar Department of Computer Science Columbia University](https://reader030.vdocuments.mx/reader030/viewer/2022032704/56649d6a5503460f94a48e47/html5/thumbnails/13.jpg)
Condition 1: Histogram Preservation and Local Area
Histogram preservedLocal area preserved
[Hadjidemetriou et al, 01]
T
……
……
T
![Page 14: Properties of Histograms and their Use for Recognition Stathis Hadjidemetriou, Michael Grossberg, Shree Nayar Department of Computer Science Columbia University](https://reader030.vdocuments.mx/reader030/viewer/2022032704/56649d6a5503460f94a48e47/html5/thumbnails/14.jpg)
Condition 2: Local Area Preservation and Divergence
•Divergence is rate of area change per unit area
Local area preserveddivergence is zero [Arnold, 89]
Small region
![Page 15: Properties of Histograms and their Use for Recognition Stathis Hadjidemetriou, Michael Grossberg, Shree Nayar Department of Computer Science Columbia University](https://reader030.vdocuments.mx/reader030/viewer/2022032704/56649d6a5503460f94a48e47/html5/thumbnails/15.jpg)
•Fields along isovalue contours of an energy function F
Isovalue contours
)sin(xyF
Hamiltonian Fields
•Flow of incompressible fluids [Arnold, 89]
Hamiltonian flow
![Page 16: Properties of Histograms and their Use for Recognition Stathis Hadjidemetriou, Michael Grossberg, Shree Nayar Department of Computer Science Columbia University](https://reader030.vdocuments.mx/reader030/viewer/2022032704/56649d6a5503460f94a48e47/html5/thumbnails/16.jpg)
Computing Hamiltonian Fields
22 yxF Gradient of
1. Compute gradient of F 2. Rotate gradient pointwise 900
jx
Fi
y
FFR
)(90
22 yxF Hamiltonian of
090R
![Page 17: Properties of Histograms and their Use for Recognition Stathis Hadjidemetriou, Michael Grossberg, Shree Nayar Department of Computer Science Columbia University](https://reader030.vdocuments.mx/reader030/viewer/2022032704/56649d6a5503460f94a48e47/html5/thumbnails/17.jpg)
Transformations preserve histogram of all images corresponding field is Hamiltonian
[Hadjidemetriou et al, CVPR, 00, Hadjidemetriou et al, IJCV, 01]
Theorem
Condition 3: Divergence and Hamiltonian Fields
Divergence of field is zero Hamiltonian field [Arnold, 89]
0
jx
Fi
y
Fdiv
![Page 18: Properties of Histograms and their Use for Recognition Stathis Hadjidemetriou, Michael Grossberg, Shree Nayar Department of Computer Science Columbia University](https://reader030.vdocuments.mx/reader030/viewer/2022032704/56649d6a5503460f94a48e47/html5/thumbnails/18.jpg)
Examples of Hamiltonian TransformationsLinear: Translations, rotations, shears
Original 3xF
7.022 )( yxF 5.122 )( yxF 10
)(sin
yxF
yxF 2
![Page 19: Properties of Histograms and their Use for Recognition Stathis Hadjidemetriou, Michael Grossberg, Shree Nayar Department of Computer Science Columbia University](https://reader030.vdocuments.mx/reader030/viewer/2022032704/56649d6a5503460f94a48e47/html5/thumbnails/19.jpg)
Examples of Hamiltonian Transformations
4
)(sin
4
)(sin
yxyxF
40
)(sin
22 yxF
65
coshyx
F
)(3 2244 yxyxF
![Page 20: Properties of Histograms and their Use for Recognition Stathis Hadjidemetriou, Michael Grossberg, Shree Nayar Department of Computer Science Columbia University](https://reader030.vdocuments.mx/reader030/viewer/2022032704/56649d6a5503460f94a48e47/html5/thumbnails/20.jpg)
Border Preserving Hamiltonian Transformations
0border
F
~
FFF win0|
,0|
borderwin
borderwin
F
F
2)2/(
,
2
1wx
xwin eF
0 w
h
![Page 21: Properties of Histograms and their Use for Recognition Stathis Hadjidemetriou, Michael Grossberg, Shree Nayar Department of Computer Science Columbia University](https://reader030.vdocuments.mx/reader030/viewer/2022032704/56649d6a5503460f94a48e47/html5/thumbnails/21.jpg)
Examples of Windowed Hamiltonian Transformations
yxF ~
2~
yF 3~
xF 24~
3xxF )3(
324
24~
yy
xxF
xyF ~
10
)(sin
~ yxF
22
~
yxF 5.122~
yxF
20
sin
.20
sin~
yx
yxF
![Page 22: Properties of Histograms and their Use for Recognition Stathis Hadjidemetriou, Michael Grossberg, Shree Nayar Department of Computer Science Columbia University](https://reader030.vdocuments.mx/reader030/viewer/2022032704/56649d6a5503460f94a48e47/html5/thumbnails/22.jpg)
Examples of Windowed Hamiltonian Transformations
Identical histograms:
4
sin
.4
sin~
yx
yxF
65
cosh~ yxF
yxF 2
~
![Page 23: Properties of Histograms and their Use for Recognition Stathis Hadjidemetriou, Michael Grossberg, Shree Nayar Department of Computer Science Columbia University](https://reader030.vdocuments.mx/reader030/viewer/2022032704/56649d6a5503460f94a48e47/html5/thumbnails/23.jpg)
Weak Perspective Projection
•Depth (z) causes scaling
cos2
2
z
fmw
[Hadjidemetriou et al, 01]
•Planar object tilt causes shearing and scaling
![Page 24: Properties of Histograms and their Use for Recognition Stathis Hadjidemetriou, Michael Grossberg, Shree Nayar Department of Computer Science Columbia University](https://reader030.vdocuments.mx/reader030/viewer/2022032704/56649d6a5503460f94a48e47/html5/thumbnails/24.jpg)
The Hamiltonian transformations is the complete class of continuous image transformations that preserves the
histogram
![Page 25: Properties of Histograms and their Use for Recognition Stathis Hadjidemetriou, Michael Grossberg, Shree Nayar Department of Computer Science Columbia University](https://reader030.vdocuments.mx/reader030/viewer/2022032704/56649d6a5503460f94a48e47/html5/thumbnails/25.jpg)
How can spatial information be embedded into the histogram?
![Page 26: Properties of Histograms and their Use for Recognition Stathis Hadjidemetriou, Michael Grossberg, Shree Nayar Department of Computer Science Columbia University](https://reader030.vdocuments.mx/reader030/viewer/2022032704/56649d6a5503460f94a48e47/html5/thumbnails/26.jpg)
Previous work on Features combining the Histogram with Spatial Information
•Local statistics:−Local histograms [Hsu et al, 95, Smith & Chang, 96, Koenderink and Doorn, 99, Griffin, 97]
−Intensity patterns [Haralick,79, Huang et al, 97]
•One histogram: −Derivative filters [Schiele and Crowley, 00, Mel 97]
−Gaussian filter [Lee and Dickinson, 94]
•Many techniques are ad-hoc or not complete
![Page 27: Properties of Histograms and their Use for Recognition Stathis Hadjidemetriou, Michael Grossberg, Shree Nayar Department of Computer Science Columbia University](https://reader030.vdocuments.mx/reader030/viewer/2022032704/56649d6a5503460f94a48e47/html5/thumbnails/27.jpg)
Multiresolution HistogramG(l2)
![Page 28: Properties of Histograms and their Use for Recognition Stathis Hadjidemetriou, Michael Grossberg, Shree Nayar Department of Computer Science Columbia University](https://reader030.vdocuments.mx/reader030/viewer/2022032704/56649d6a5503460f94a48e47/html5/thumbnails/28.jpg)
Limitations of HistogramsDatabase of synthetic images with identical histograms
[Hadjidemetriou et al, 01]
![Page 29: Properties of Histograms and their Use for Recognition Stathis Hadjidemetriou, Michael Grossberg, Shree Nayar Department of Computer Science Columbia University](https://reader030.vdocuments.mx/reader030/viewer/2022032704/56649d6a5503460f94a48e47/html5/thumbnails/29.jpg)
Matching with Multiresolution Histograms
Match under Gaussian noise of st.dev. 15 graylevels:
![Page 30: Properties of Histograms and their Use for Recognition Stathis Hadjidemetriou, Michael Grossberg, Shree Nayar Department of Computer Science Columbia University](https://reader030.vdocuments.mx/reader030/viewer/2022032704/56649d6a5503460f94a48e47/html5/thumbnails/30.jpg)
Matching with Multiresolution Histograms Match under Gaussian noise of st.dev. 15 graylevels:
![Page 31: Properties of Histograms and their Use for Recognition Stathis Hadjidemetriou, Michael Grossberg, Shree Nayar Department of Computer Science Columbia University](https://reader030.vdocuments.mx/reader030/viewer/2022032704/56649d6a5503460f94a48e47/html5/thumbnails/31.jpg)
How is Image Structure Encoded in the Multiresolution Histogram?
?
Image structure
Differences of histogramsdl
l))(*(d GLh
LImage
h(L*G(l))Multiresolution histogram
![Page 32: Properties of Histograms and their Use for Recognition Stathis Hadjidemetriou, Michael Grossberg, Shree Nayar Department of Computer Science Columbia University](https://reader030.vdocuments.mx/reader030/viewer/2022032704/56649d6a5503460f94a48e47/html5/thumbnails/32.jpg)
Histogram Change with Resolution and Spatial Information
•Bin j: dl
lhdj
))(*( GL
Spatial information
•Averages of bins:
where Pj are proportionality factorsdl
ldh jm
jj
))(*(
1
GLP
ill-conditioned
well-conditioned
![Page 33: Properties of Histograms and their Use for Recognition Stathis Hadjidemetriou, Michael Grossberg, Shree Nayar Department of Computer Science Columbia University](https://reader030.vdocuments.mx/reader030/viewer/2022032704/56649d6a5503460f94a48e47/html5/thumbnails/33.jpg)
Histogram Change with Resolution and Fisher Information Measures
qJ= Generalized Fisher
information measures of order q [Stam, 59, Plastino et al, 97]
≡
)(LqJ L is the image
xdq
D
22
LLL
=
D is the image domain
dl
ldh jm
jqj
))(*(
1
GLP
Averages
![Page 34: Properties of Histograms and their Use for Recognition Stathis Hadjidemetriou, Michael Grossberg, Shree Nayar Department of Computer Science Columbia University](https://reader030.vdocuments.mx/reader030/viewer/2022032704/56649d6a5503460f94a48e47/html5/thumbnails/34.jpg)
Image Structure Through Fisher Information Measures
?
Image structure
Differences of histogramsdl
l))(*(d GLh
LImage
h(L*G(l))Multiresolution histogram
Fisher information measures (Analysis)
P
Jq
![Page 35: Properties of Histograms and their Use for Recognition Stathis Hadjidemetriou, Michael Grossberg, Shree Nayar Department of Computer Science Columbia University](https://reader030.vdocuments.mx/reader030/viewer/2022032704/56649d6a5503460f94a48e47/html5/thumbnails/35.jpg)
Shape Boundary and Multiresolution Histogram15.0)( yxR LSuperquadrics:
=0.56
Histogram change with l is higher for complex boundary
=1.00 =1.48 =2.00
=6.67
![Page 36: Properties of Histograms and their Use for Recognition Stathis Hadjidemetriou, Michael Grossberg, Shree Nayar Department of Computer Science Columbia University](https://reader030.vdocuments.mx/reader030/viewer/2022032704/56649d6a5503460f94a48e47/html5/thumbnails/36.jpg)
Texel Repetition and Multiresolution Histogram
,, pTiling
Histogram change with l is proportional to number of texels (analytically)
![Page 37: Properties of Histograms and their Use for Recognition Stathis Hadjidemetriou, Michael Grossberg, Shree Nayar Department of Computer Science Columbia University](https://reader030.vdocuments.mx/reader030/viewer/2022032704/56649d6a5503460f94a48e47/html5/thumbnails/37.jpg)
Texel Placement and Multiresolution Histogram
Std. dev. of perturbation
Histogram change with l decreases with randomness
![Page 38: Properties of Histograms and their Use for Recognition Stathis Hadjidemetriou, Michael Grossberg, Shree Nayar Department of Computer Science Columbia University](https://reader030.vdocuments.mx/reader030/viewer/2022032704/56649d6a5503460f94a48e47/html5/thumbnails/38.jpg)
Matching Algorithm for Multiresolution Histograms
Burt-Adelson image pyramid
Cumulative histograms
L1 norm
Differences of histograms betweenconsecutive image resolutions
Concatenate to form feature vector
![Page 39: Properties of Histograms and their Use for Recognition Stathis Hadjidemetriou, Michael Grossberg, Shree Nayar Department of Computer Science Columbia University](https://reader030.vdocuments.mx/reader030/viewer/2022032704/56649d6a5503460f94a48e47/html5/thumbnails/39.jpg)
Histogram Parameters
•Bin width
•Smoothing to avoid aliasing
•Normalization:−Image size−Histogram size
179x179
89x89
44x44
5x5
……
![Page 40: Properties of Histograms and their Use for Recognition Stathis Hadjidemetriou, Michael Grossberg, Shree Nayar Department of Computer Science Columbia University](https://reader030.vdocuments.mx/reader030/viewer/2022032704/56649d6a5503460f94a48e47/html5/thumbnails/40.jpg)
Database of Synthetic Images
108 images with identical histograms [Hadjidemetriou et al, 01]
![Page 41: Properties of Histograms and their Use for Recognition Stathis Hadjidemetriou, Michael Grossberg, Shree Nayar Department of Computer Science Columbia University](https://reader030.vdocuments.mx/reader030/viewer/2022032704/56649d6a5503460f94a48e47/html5/thumbnails/41.jpg)
Sensitivity of Matching for Synthetic Images
![Page 42: Properties of Histograms and their Use for Recognition Stathis Hadjidemetriou, Michael Grossberg, Shree Nayar Department of Computer Science Columbia University](https://reader030.vdocuments.mx/reader030/viewer/2022032704/56649d6a5503460f94a48e47/html5/thumbnails/42.jpg)
Database of Brodatz Textures91 images with identical equalized histograms: 13 textures
different rotations
![Page 43: Properties of Histograms and their Use for Recognition Stathis Hadjidemetriou, Michael Grossberg, Shree Nayar Department of Computer Science Columbia University](https://reader030.vdocuments.mx/reader030/viewer/2022032704/56649d6a5503460f94a48e47/html5/thumbnails/43.jpg)
Match Results for Brodatz Textures
Match under Gaussian noise of st.dev. 15 graylevels:
![Page 44: Properties of Histograms and their Use for Recognition Stathis Hadjidemetriou, Michael Grossberg, Shree Nayar Department of Computer Science Columbia University](https://reader030.vdocuments.mx/reader030/viewer/2022032704/56649d6a5503460f94a48e47/html5/thumbnails/44.jpg)
Sensitivity of Class Matching for Brodatz Textures
![Page 45: Properties of Histograms and their Use for Recognition Stathis Hadjidemetriou, Michael Grossberg, Shree Nayar Department of Computer Science Columbia University](https://reader030.vdocuments.mx/reader030/viewer/2022032704/56649d6a5503460f94a48e47/html5/thumbnails/45.jpg)
Database of CUReT Textures 8,046 images with identical equalized histograms : 61
materials under different illuminations [Dana et al, 99]
![Page 46: Properties of Histograms and their Use for Recognition Stathis Hadjidemetriou, Michael Grossberg, Shree Nayar Department of Computer Science Columbia University](https://reader030.vdocuments.mx/reader030/viewer/2022032704/56649d6a5503460f94a48e47/html5/thumbnails/46.jpg)
Match Results for CUReT Textures
Match under Gaussian noise of st.dev. 15 graylevels:
![Page 47: Properties of Histograms and their Use for Recognition Stathis Hadjidemetriou, Michael Grossberg, Shree Nayar Department of Computer Science Columbia University](https://reader030.vdocuments.mx/reader030/viewer/2022032704/56649d6a5503460f94a48e47/html5/thumbnails/47.jpg)
Match Results for CUReT Textures
Match under Gaussian noise of st.dev. 15 graylevels:
![Page 48: Properties of Histograms and their Use for Recognition Stathis Hadjidemetriou, Michael Grossberg, Shree Nayar Department of Computer Science Columbia University](https://reader030.vdocuments.mx/reader030/viewer/2022032704/56649d6a5503460f94a48e47/html5/thumbnails/48.jpg)
Sensitivity of Class Matching for CUReT Textures
100 randomly selected images per noise level
![Page 49: Properties of Histograms and their Use for Recognition Stathis Hadjidemetriou, Michael Grossberg, Shree Nayar Department of Computer Science Columbia University](https://reader030.vdocuments.mx/reader030/viewer/2022032704/56649d6a5503460f94a48e47/html5/thumbnails/49.jpg)
Embed spatial information into the histogram with the multiresolution histogram
![Page 50: Properties of Histograms and their Use for Recognition Stathis Hadjidemetriou, Michael Grossberg, Shree Nayar Department of Computer Science Columbia University](https://reader030.vdocuments.mx/reader030/viewer/2022032704/56649d6a5503460f94a48e47/html5/thumbnails/50.jpg)
How well does the multiresolution histogram perform compared to other image features?
![Page 51: Properties of Histograms and their Use for Recognition Stathis Hadjidemetriou, Michael Grossberg, Shree Nayar Department of Computer Science Columbia University](https://reader030.vdocuments.mx/reader030/viewer/2022032704/56649d6a5503460f94a48e47/html5/thumbnails/51.jpg)
Comparison of Multiresolution Histogram with Other Features
•Multiresolution histogram:−Variable bin width−Histogram smoothing
•Fourier power spectrum annuli [Bajsky, 73]
•Gabor features [Farrokhnia & Jain, 91]
•Daubechies wavelet packets energies [Laine & Fan, 93]
•Auto-cooccurrence matrix [Haralick, 92]
•Markov random field parameters [Lee & Lee, 96]
![Page 52: Properties of Histograms and their Use for Recognition Stathis Hadjidemetriou, Michael Grossberg, Shree Nayar Department of Computer Science Columbia University](https://reader030.vdocuments.mx/reader030/viewer/2022032704/56649d6a5503460f94a48e47/html5/thumbnails/52.jpg)
Comparison of Effects of Transformations on the Features
Feature Translation RotationUniform Scaling
1Fourier power
spectrum annuliinvariant robust equivariant
2 Gabor features invariant sensitive equivariant
3Daubechies wavelet
energiessensitive sensitive sensitive
4Multiresolution
histogramsinvariant invariant equivariant
5Auto-cooccurrence
matrixinvariant robust equivariant
6Markov random field
parametersinvariant sensitive sensitive
![Page 53: Properties of Histograms and their Use for Recognition Stathis Hadjidemetriou, Michael Grossberg, Shree Nayar Department of Computer Science Columbia University](https://reader030.vdocuments.mx/reader030/viewer/2022032704/56649d6a5503460f94a48e47/html5/thumbnails/53.jpg)
Comparison of Class Matching Sensitivity of Features
Database of Brodatz textures
![Page 54: Properties of Histograms and their Use for Recognition Stathis Hadjidemetriou, Michael Grossberg, Shree Nayar Department of Computer Science Columbia University](https://reader030.vdocuments.mx/reader030/viewer/2022032704/56649d6a5503460f94a48e47/html5/thumbnails/54.jpg)
Comparison of Class Matching Sensitivity of Features
•Database ofCUReT textures•100 randomly selected images per noise level
![Page 55: Properties of Histograms and their Use for Recognition Stathis Hadjidemetriou, Michael Grossberg, Shree Nayar Department of Computer Science Columbia University](https://reader030.vdocuments.mx/reader030/viewer/2022032704/56649d6a5503460f94a48e47/html5/thumbnails/55.jpg)
Sensitivity of Features to Matching
FeatureGaussian
NoiseDatabase
size,# classesIlluminati-
on Parameter selection
Fourier power spectrum annuli
sensitive sensitive robust very sensitive
Gabor features robust robust robust sensitive
Daubechies wavelet energies
sensitive robust robust robust
Multiresolution histogram
robust robust robust robust
Auto-cooccurrence matrix
very sensitive very sensitive very sensitive very sensitive
Markov random field parameters
very sensitive very sensitive sensitive N/A
![Page 56: Properties of Histograms and their Use for Recognition Stathis Hadjidemetriou, Michael Grossberg, Shree Nayar Department of Computer Science Columbia University](https://reader030.vdocuments.mx/reader030/viewer/2022032704/56649d6a5503460f94a48e47/html5/thumbnails/56.jpg)
Comparison of Computation Costs of Features
1Markov random field
parametersO(n(2-1)2-(2-1)3/3)
2 Gabor features ( (logn+1)nlogn)
3Fourier power spectrum
annuliO(n3/2)
4 Auto-cooccurrence matrix O(n)
5 Wavelet packets energies O(nl)
6 Multiresolution histograms nn- number of pixels- window widthl- resolution levels
Decreasing cost
![Page 57: Properties of Histograms and their Use for Recognition Stathis Hadjidemetriou, Michael Grossberg, Shree Nayar Department of Computer Science Columbia University](https://reader030.vdocuments.mx/reader030/viewer/2022032704/56649d6a5503460f94a48e47/html5/thumbnails/57.jpg)
The multiresolution histogram compared to other image features is robust and efficient
![Page 58: Properties of Histograms and their Use for Recognition Stathis Hadjidemetriou, Michael Grossberg, Shree Nayar Department of Computer Science Columbia University](https://reader030.vdocuments.mx/reader030/viewer/2022032704/56649d6a5503460f94a48e47/html5/thumbnails/58.jpg)
Summary and Discussion
•Hamiltonian transformations preserve features based on:−Histogram−Image topology
•Multiresolution histograms:−Embed spatial information
•Comparison of multiresolution histograms with other features:−Efficient and robust
![Page 59: Properties of Histograms and their Use for Recognition Stathis Hadjidemetriou, Michael Grossberg, Shree Nayar Department of Computer Science Columbia University](https://reader030.vdocuments.mx/reader030/viewer/2022032704/56649d6a5503460f94a48e47/html5/thumbnails/59.jpg)
Recognition of 3D Matte Polyhedral Objects
•Face histograms:–Magnitude scaled by tilt angle ()
–Intensity scaled by illumination (ai )
•In an object database find [Hadjidemetriou et al, 00]:–Object identity
–Pose ()–Illumination (ai )
•Total histogram: Sum of h(i) of visible faces
![Page 60: Properties of Histograms and their Use for Recognition Stathis Hadjidemetriou, Michael Grossberg, Shree Nayar Department of Computer Science Columbia University](https://reader030.vdocuments.mx/reader030/viewer/2022032704/56649d6a5503460f94a48e47/html5/thumbnails/60.jpg)
A Simple Experiment
Object 1: Object 2:
Object 3: Object 4:
Object Tests Rank=1 Rank=2
Total 40 38 2
![Page 61: Properties of Histograms and their Use for Recognition Stathis Hadjidemetriou, Michael Grossberg, Shree Nayar Department of Computer Science Columbia University](https://reader030.vdocuments.mx/reader030/viewer/2022032704/56649d6a5503460f94a48e47/html5/thumbnails/61.jpg)
Shape Elongation and Multiresolution Histogram
Elongation:
y
x
St. dev. along axes: x, y.
•Gaussian:Sides of base : rx, ry.
y
x
r
r
•Pyramid:
Elongation:
1
(analytically)
Histogram change with l
![Page 62: Properties of Histograms and their Use for Recognition Stathis Hadjidemetriou, Michael Grossberg, Shree Nayar Department of Computer Science Columbia University](https://reader030.vdocuments.mx/reader030/viewer/2022032704/56649d6a5503460f94a48e47/html5/thumbnails/62.jpg)
Are all image resolutions equally significant?
![Page 63: Properties of Histograms and their Use for Recognition Stathis Hadjidemetriou, Michael Grossberg, Shree Nayar Department of Computer Science Columbia University](https://reader030.vdocuments.mx/reader030/viewer/2022032704/56649d6a5503460f94a48e47/html5/thumbnails/63.jpg)
Resolution Selection with Entropy of Multiresolution Histograms
•Entropy-resolution plot [Hadjidemetriou
et al, ECCV, 02] :–Global–Non-monotonic
….
l
…………………………………
![Page 64: Properties of Histograms and their Use for Recognition Stathis Hadjidemetriou, Michael Grossberg, Shree Nayar Department of Computer Science Columbia University](https://reader030.vdocuments.mx/reader030/viewer/2022032704/56649d6a5503460f94a48e47/html5/thumbnails/64.jpg)
Examples of Entropy-Resolution Plots
![Page 65: Properties of Histograms and their Use for Recognition Stathis Hadjidemetriou, Michael Grossberg, Shree Nayar Department of Computer Science Columbia University](https://reader030.vdocuments.mx/reader030/viewer/2022032704/56649d6a5503460f94a48e47/html5/thumbnails/65.jpg)
The entropy of the multiresolution histogram can be used to detect significant image resolutions
![Page 66: Properties of Histograms and their Use for Recognition Stathis Hadjidemetriou, Michael Grossberg, Shree Nayar Department of Computer Science Columbia University](https://reader030.vdocuments.mx/reader030/viewer/2022032704/56649d6a5503460f94a48e47/html5/thumbnails/66.jpg)
Future Work
•Histogram preserving fields:−Transformations over limited regions−Sensitivity of features to image transformation
•Multiresolution histograms:−Color images−Rotational variance with elliptic Gaussians
•Resolution selection:−Preprocessing step−Non-monotonic features