View
234
Download
1
Tags:
Embed Size (px)
Citation preview
"Everything is related to everything else, but near things are more related to each other.”
-Waldo Tobler (1970)
First Law Of Geography:
Pyramids of FeaturesFor Categorization
Presented by Greg Griffin
Project Partner Will Coulter
Pyramids of FeaturesFor Categorization
Presented by Greg Griffin
Project Partner Will Coulter
Buckets of FeaturesFor Categorization
Presented by Greg Griffin
Project Partner Will Coulter
This talk is mostly about this paper:
With a little bit about benchmarking:
CALTECH 256
Images Features
€
rX 1K
r X Nx
€
rY 1K
r Y NY
€
K(X,Y )
A Number
Images Features
€
rX 1K
r X Nx
€
rY 1K
r Y NY
€
K(X,Y )
A Number
“How well dothey match”
“Weak Features”
“Weak Features”
(I think?)
“…points whose gradient magnitude in a given direction exceeds a minimum threshold.”
This is just their toy exampleThey use SIFT descriptors as “Strong Features”.
But you could use any features you want!
Images Features A Number
1 2 3 4 5 6 7 8
Images Features
X1 X5
Y1 Y5
Images Features
€
l D Il
Y1 Y5
X1 X5
Features
€
l D Il
€
K L (X,Y )
X1 X5
Y1 Y5
Features
€
l D Il
€
I l = min(HXl
i=1
D
∑ (i),HYl (i))
€
K L (X,Y )
Start By Matching Reds
Features
€
l D Il
0 1
€
I l = min(HXl
i=1
D
∑ (i),HYl (i))
€
K L (X,Y )
Start By Matching Reds
Features
€
l D Il
0 1 10
€
I l = min(HXl
i=1
D
∑ (i),HYl (i))
€
K L (X,Y )
Start By Matching Reds
10
10
Features
€
l D Il
0 1 10
1 4
€
I l = min(HXl
i=1
D
∑ (i),HYl (i))
€
K L (X,Y )
Start By Matching Reds
Features
€
l D Il
0 1 10
1 4
€
I l = min(HXl
i=1
D
∑ (i),HYl (i))
€
K L (X,Y )
Start By Matching Reds
6
7
2
1
1
2
1
0
Features
€
l D Il
0 1 10
1 4 8
€
I l = min(HXl
i=1
D
∑ (i),HYl (i))
€
K L (X,Y )
Start By Matching Reds
6
7
2
1
1
2
1
0
Features
€
l D Il
0 1 10
1 4 8
2 16
€
I l = min(HXl
i=1
D
∑ (i),HYl (i))
€
K L (X,Y )
Start By Matching Reds
Features
€
l D Il
0 1 10
1 4 8
2 16 6
€
I l = min(HXl
i=1
D
∑ (i),HYl (i))
€
K L (X,Y )
Start By Matching Reds
5
2
1
1
23
1 1 1
1
€
L = 2
2
€
l D Il
0 1 10
1 4 8
2 16 6
€
K L (X,Y )
Start By Matching Reds
€
κ L (X1,Y1) = IL +1
2L−ll = 0
L−1
∑ I l = 6 +8
21+
10
22
€
L = 2
€
I l = min(HXl
i=1
D
∑ (i),HYl (i))
€
l D Il
0 1 10
1 4 8
2 16 6
€
K L (X,Y )
Start By Matching Reds
€
κ L (X1,Y1) = IL +1
2L−ll = 0
L−1
∑ I l = 6 +8
21+
10
22
€
L = 2
€
I l = min(HXl
i=1
D
∑ (i),HYl (i))
A compact set of matches is preferable to widely dispersed matches
€
l D Il
0 1 10
1 4 8
2 16 6
€
K L (X,Y )
Start By Matching Reds
€
κ L (X1,Y1) = IL +1
2L−ll = 0
L−1
∑ I l = 6 +8
21+
10
22
€
L = 2
€
I l = min(HXl
i=1
D
∑ (i),HYl (i))
€
l D Il
0 1 10
1 4 8
2 16 6
€
K L (X,Y )
Start By Matching Reds
€
κ L (X1,Y1) = IL +1
2L−ll = 0
L−1
∑ I l = 6 +8
21+
10
22
€
L = 2
€
κ L (X1,Y1) = IL +1
2L−ll = 0
L−1
∑ ( I l − I l +1) = 6 +8 − 6
2+
10 − 8
22= 7.5
€
I l = min(HXl
i=1
D
∑ (i),HYl (i))
€
l D Il
0 1 10
1 4 5.5
2 16 1.9
€
K L (X,Y )
€
κ L (X1,Y1) = IL +1
2L−ll = 0
L−1
∑ I l = 6 +8
21+
10
22
€
L = 2
€
I l = min(HXl
i=1
D
∑ (i),HYl (i))
€
κ L (X1,Y1) = IL +1
2L−ll = 0
L−1
∑ ( I l − I l +1) = 5.4 < 7.5
A Sanity Check:Features X vs. Purely Isotropic Y
€
l D Il
0 1 10
1 4 8
2 16 6
€
K L (X,Y )
Start By Matching Reds, Then The Blues, Then…
€
κ L (X1,Y1) = IL +1
2L−ll = 0
L−1
∑ I l = 6 +8
21+
10
22
€
L = 2
€
κ L (X1,Y1) = IL +1
2L−ll = 0
L−1
∑ ( I l − I l +1) = 6 +8 − 6
2+
10 − 8
22= 7.5
€
I l = min(HXl
i=1
D
∑ (i),HYl (i))
€
ΚL (X,Y ) = κ L (X m,Y m )m=1
M
∑ = 7.5 + ...
1 2 3 4 5 6 7 8
M = 8
Features
€
l D Il
0 1 10
€
I l = min(HXl
i=1
D
∑ (i),HYl (i))
Features
€
l D Il
0 1 10
1 4 8
€
I l = min(HXl
i=1
D
∑ (i),HYl (i))
Features
€
l D Il
0 1 10
1 4 8
2 16 6
€
I l = min(HXl
i=1
D
∑ (i),HYl (i))
€
κ L (X m,Y m ) = IL +1
2L−ll = 0
L−1
∑ ( I l − I l +1)
€
I l = min(HXl
i=1
D
∑ (i),HYl (i))
€
ΚL (X,Y ) =1
NxmNy
mκ L (X m,Y m )
m=1
M
∑
Foreach feature M=1…mForeach level = 0…L
Foreach cell i=1…D
€
l
Training Set
15Categories
100 Imagesper Category
Test Set
100-300 Imagesper Category
3.4
5.6
7.8
1.5
5.4
SL(X,Y)
office
store
coast
street
suburb
Confusion Matrix
Train on 100Test on 100-300
(per category)
Scene Database
Caltech 101
Hypothesis
Pyramid Matching works well when:
• Objects are aligned and localized– ie. certain Caltech 101 categories– biased by different values of Ntest?
• A few common features that define the category get randomly permuted through many positions, thanks to a large dataset– ie. scene database– now and then pyramid matching gets lucky
Test
How well will Pyramid matching work?
• Objects are not well aligned, or cluttered – Caltech 256 is more challenging in this respect
Scene Database
Caltech 101
Cluster Matching
k-means of SIFT positions
More Flexible Than A Grid?
k-means of SIFT positions + color
Position Invariance & Clutter
1. Any cluster can match any cluster2. Clusters respect duck / water boundaries
(sort of)
Tackles Alignment Problem
But… how to match efficiently?How many clusters? How big?
Summary
Spatial Pyramid Matching is efficient and handles a range of scales, but seems to be
sensitive to translation and clutter.
Cluster Matching has the potential to improve translational invariance and toleranceof clutter. But inefficient. Less principled: how
many clusters are optimal? How big should they be? No scale invariance.
Can we have the best of both worlds?
Try Sliding Spatial Pyramids?
Slide puzzle photo from:http://www.tenfootpolesoftware.com/products/slidepuzzle/