partitions et structures hiérarchiques - greyc€¦ · partitions et structures hi´erarchiques...
TRANSCRIPT
![Page 1: Partitions et structures hiérarchiques - GREYC€¦ · Partitions et structures hi´erarchiques Luc Brun Groupe de Recherche en Informatique, Image, Automatique et Instrumentation](https://reader035.vdocuments.site/reader035/viewer/2022070613/5b976d9b09d3f206218c8c4d/html5/thumbnails/1.jpg)
PresentationIntroduction
Regular PyramidsIrregular PyramidsSome applications
Partitions et structures hierarchiques
Luc Brun
Groupe de Recherche en Informatique,Image, Automatique et Instrumentation
de Caen (GREYC)
1 / 51
![Page 2: Partitions et structures hiérarchiques - GREYC€¦ · Partitions et structures hi´erarchiques Luc Brun Groupe de Recherche en Informatique, Image, Automatique et Instrumentation](https://reader035.vdocuments.site/reader035/viewer/2022070613/5b976d9b09d3f206218c8c4d/html5/thumbnails/2.jpg)
PresentationIntroduction
Regular PyramidsIrregular PyramidsSome applications
Plan
Introduction
Regular PyramidsTree PyramidsMatrix Pyramids
Irregular PyramidsHierarchical encodingStructural properties within pyramidsCombinatorial Pyramids
Some applications
2 / 51
![Page 3: Partitions et structures hiérarchiques - GREYC€¦ · Partitions et structures hi´erarchiques Luc Brun Groupe de Recherche en Informatique, Image, Automatique et Instrumentation](https://reader035.vdocuments.site/reader035/viewer/2022070613/5b976d9b09d3f206218c8c4d/html5/thumbnails/3.jpg)
PresentationIntroduction
Regular PyramidsIrregular PyramidsSome applications
Introduction
◮ Defining a partition involves a choice :
◮ All usefull information must be in the provided partition yq q◮ Provides not one but a full stack of partitions successively
reduced. yq q
3 / 51
![Page 4: Partitions et structures hiérarchiques - GREYC€¦ · Partitions et structures hi´erarchiques Luc Brun Groupe de Recherche en Informatique, Image, Automatique et Instrumentation](https://reader035.vdocuments.site/reader035/viewer/2022070613/5b976d9b09d3f206218c8c4d/html5/thumbnails/4.jpg)
PresentationIntroduction
Regular PyramidsIrregular PyramidsSome applications
Tree PyramidsMatrix Pyramids
T-pyramids (quadtrees)
◮ Top-down construction of the partition by a recursivedecomposition into squares.
◮ Avantages :◮ Efficient Acces to some geometrical information.
◮ Drawbacks (see M-pyramids below)
4 / 51
![Page 5: Partitions et structures hiérarchiques - GREYC€¦ · Partitions et structures hi´erarchiques Luc Brun Groupe de Recherche en Informatique, Image, Automatique et Instrumentation](https://reader035.vdocuments.site/reader035/viewer/2022070613/5b976d9b09d3f206218c8c4d/html5/thumbnails/5.jpg)
PresentationIntroduction
Regular PyramidsIrregular PyramidsSome applications
Tree PyramidsMatrix Pyramids
Matrix Pyramids
◮ Stack of images with progressively reduced resolution.
bb bbr b b bb b bb b bb b br r rbbb
bbb
bbb
bbb
rrr
bbb
bbb
bbb
bbb
bbb
bbb
bbb
bbb
bbb
bbb
bbb
bbb
rrr
rrr
rrr
bb bbr bb bbrbb bbr bb bbr rb bbr b
A 2 × 2/4 pyramid
◮ At each step the pixels of the image above (�) may beassociated to pointels (•)of the image below.
5 / 51
![Page 6: Partitions et structures hiérarchiques - GREYC€¦ · Partitions et structures hi´erarchiques Luc Brun Groupe de Recherche en Informatique, Image, Automatique et Instrumentation](https://reader035.vdocuments.site/reader035/viewer/2022070613/5b976d9b09d3f206218c8c4d/html5/thumbnails/6.jpg)
PresentationIntroduction
Regular PyramidsIrregular PyramidsSome applications
Tree PyramidsMatrix Pyramids
Definition (1/2)
bb bbr bb bbrbb bbr bb bbr rb bbr b
◮ A N × N/q M-pyramid is defined by :◮ A reduction window N × N corresponds to a connected set of
pixels used to compute the value of a pixel in image above.The function applied to compute this value is called areduction function.
◮ A pixel is the father of all the pixels belonging to itsreduction window.
◮ Any pixel within a reduction window is the child of at leastone father (see below).
◮ The Reduction factor q encodes the ratio between the sizes oftwo consecutive images. This ratio is fixed along the pyramid.
6 / 51
![Page 7: Partitions et structures hiérarchiques - GREYC€¦ · Partitions et structures hi´erarchiques Luc Brun Groupe de Recherche en Informatique, Image, Automatique et Instrumentation](https://reader035.vdocuments.site/reader035/viewer/2022070613/5b976d9b09d3f206218c8c4d/html5/thumbnails/7.jpg)
PresentationIntroduction
Regular PyramidsIrregular PyramidsSome applications
Tree PyramidsMatrix Pyramids
Definition (2/2)
◮ The Receptive field is defined as the transitive closure of thefather/child relationship.
bb bbr b b bb b bb b bb b br r rbrr
brr
brr
brr
rrr
brr
bbb
bbb
brr
bbb
bbb
brr
bbb
bbb
brr
bbb
bbb
rrr
rrr
rrr
bb bbr bb bbrrr rrr bb bbr rb bbr b
Receptive field
7 / 51
![Page 8: Partitions et structures hiérarchiques - GREYC€¦ · Partitions et structures hi´erarchiques Luc Brun Groupe de Recherche en Informatique, Image, Automatique et Instrumentation](https://reader035.vdocuments.site/reader035/viewer/2022070613/5b976d9b09d3f206218c8c4d/html5/thumbnails/8.jpg)
PresentationIntroduction
Regular PyramidsIrregular PyramidsSome applications
Tree PyramidsMatrix Pyramids
Different types of M-Pyramids◮ If N × N/q < 1, the pyramid is named a
non-overlapping holed pyramid(a). Some pixels have nofathers (e.g., the center pixel in Fig. (a)).
◮ If N × N/q = 1, the pyramid is called anon overlapping pyramid without hole(b). Each pixel in thereduction window has exactly one father.
◮ If N × N/q > 1, the pyramid is named anOverlapping pyramid (c). Each pixel has severalpotential parents.
bb bbr bb bbrbb bbr bb bbr bb bbr bb bbr
bb bbr bb bbrbb r
bbbbrrbbbb r bb
a) 2 × 2/5b)
2 × 2/4c)
2 × 2/28 / 51
![Page 9: Partitions et structures hiérarchiques - GREYC€¦ · Partitions et structures hi´erarchiques Luc Brun Groupe de Recherche en Informatique, Image, Automatique et Instrumentation](https://reader035.vdocuments.site/reader035/viewer/2022070613/5b976d9b09d3f206218c8c4d/html5/thumbnails/9.jpg)
PresentationIntroduction
Regular PyramidsIrregular PyramidsSome applications
Tree PyramidsMatrix Pyramids
Example : A 2 × 2/4 non overlapping pyramid
9 / 51
![Page 10: Partitions et structures hiérarchiques - GREYC€¦ · Partitions et structures hi´erarchiques Luc Brun Groupe de Recherche en Informatique, Image, Automatique et Instrumentation](https://reader035.vdocuments.site/reader035/viewer/2022070613/5b976d9b09d3f206218c8c4d/html5/thumbnails/10.jpg)
PresentationIntroduction
Regular PyramidsIrregular PyramidsSome applications
Tree PyramidsMatrix Pyramids
Overlapping pyramids
◮ q inner childs,
◮ N2 − q outerchilds
0 1 2 3 4 5
0
1
2
3
45
◮ NxN/q > 1 : Each pixel contributes toseveral father ⇒ Each pixel has severalpotential father
0 1 2 3 4 5012
3
45
10 / 51
![Page 11: Partitions et structures hiérarchiques - GREYC€¦ · Partitions et structures hi´erarchiques Luc Brun Groupe de Recherche en Informatique, Image, Automatique et Instrumentation](https://reader035.vdocuments.site/reader035/viewer/2022070613/5b976d9b09d3f206218c8c4d/html5/thumbnails/11.jpg)
PresentationIntroduction
Regular PyramidsIrregular PyramidsSome applications
Tree PyramidsMatrix Pyramids
Segmentation algorithm◮ Main notations :
◮ Legitimate father : Closest father (strongest link w)◮ P ′ : son of P, Po : legitimate father of P.◮ Root : Link(P,Legitimate(P))<threshold)
◮ Algorithm :◮ From Bottom to Top
◮ Compute values and
v(P) =
P
P′ v(P ′)a(P ′)w(P, P ′)P
P′ w(P, P ′)a(P ′)a(P) =
X
P′
a(P ′)w(P, P ′
P
P′o w(P, P ′o)
◮ update links untill no change occur.
w(P, Po) =
(v(P) − v(Po))−2
P
Po ′(v(Po ′) − v(Po))−2
◮ From Top to Bottom◮ Select roots◮ Link non roots to legitimate fathers
11 / 51
![Page 12: Partitions et structures hiérarchiques - GREYC€¦ · Partitions et structures hi´erarchiques Luc Brun Groupe de Recherche en Informatique, Image, Automatique et Instrumentation](https://reader035.vdocuments.site/reader035/viewer/2022070613/5b976d9b09d3f206218c8c4d/html5/thumbnails/12.jpg)
PresentationIntroduction
Regular PyramidsIrregular PyramidsSome applications
Tree PyramidsMatrix Pyramids
Examples
12 / 51
![Page 13: Partitions et structures hiérarchiques - GREYC€¦ · Partitions et structures hi´erarchiques Luc Brun Groupe de Recherche en Informatique, Image, Automatique et Instrumentation](https://reader035.vdocuments.site/reader035/viewer/2022070613/5b976d9b09d3f206218c8c4d/html5/thumbnails/13.jpg)
PresentationIntroduction
Regular PyramidsIrregular PyramidsSome applications
Tree PyramidsMatrix Pyramids
Advantages of Regular pyramids
◮ reduce the influence of noise
◮ makes the processes independent of the resolution
◮ convert global features to local ones
◮ reduce computational cost
◮ Analysis at low cost using low resolution images.
13 / 51
![Page 14: Partitions et structures hiérarchiques - GREYC€¦ · Partitions et structures hi´erarchiques Luc Brun Groupe de Recherche en Informatique, Image, Automatique et Instrumentation](https://reader035.vdocuments.site/reader035/viewer/2022070613/5b976d9b09d3f206218c8c4d/html5/thumbnails/14.jpg)
PresentationIntroduction
Regular PyramidsIrregular PyramidsSome applications
Tree PyramidsMatrix Pyramids
Drawbacks (1/2)
◮ Shift - Scale-Rotation variant
◮ Preservation of the connectivity is not guaranteed
◮ Limited number of regions at a given level0 31 2 4 5 6 7
1
3
4
5
6
7
0
2
with a 4 × 4/4 pyramid :
{
can be described only at level 3only 4 pixels left at level 3
14 / 51
![Page 15: Partitions et structures hiérarchiques - GREYC€¦ · Partitions et structures hi´erarchiques Luc Brun Groupe de Recherche en Informatique, Image, Automatique et Instrumentation](https://reader035.vdocuments.site/reader035/viewer/2022070613/5b976d9b09d3f206218c8c4d/html5/thumbnails/15.jpg)
PresentationIntroduction
Regular PyramidsIrregular PyramidsSome applications
Tree PyramidsMatrix Pyramids
Drawbacks (2/2)
◮ Difficulties to encode elongated objects
012
3
0 1 2 3
−→
01
0 1
−→
012
3
0 1 2 3
15 / 51
![Page 16: Partitions et structures hiérarchiques - GREYC€¦ · Partitions et structures hi´erarchiques Luc Brun Groupe de Recherche en Informatique, Image, Automatique et Instrumentation](https://reader035.vdocuments.site/reader035/viewer/2022070613/5b976d9b09d3f206218c8c4d/html5/thumbnails/16.jpg)
PresentationIntroduction
Regular PyramidsIrregular PyramidsSome applications
Hierarchical encodingStructural properties within pyramidsCombinatorial Pyramids
Irregular Pyramids◮ Can wee keep all the advantages of regular pyramids while
overcoming their drawbacks ?◮ Yes, using irregular pyramids :
◮ Stack of graphs (G0,G1, . . . ,Gn) successively reduced.◮ G0 : encodes the initial grid or an initial segmentation.
G0
G1
G2
G3
◮ G0, . . . ,Gn are only final results.16 / 51
![Page 17: Partitions et structures hiérarchiques - GREYC€¦ · Partitions et structures hi´erarchiques Luc Brun Groupe de Recherche en Informatique, Image, Automatique et Instrumentation](https://reader035.vdocuments.site/reader035/viewer/2022070613/5b976d9b09d3f206218c8c4d/html5/thumbnails/17.jpg)
PresentationIntroduction
Regular PyramidsIrregular PyramidsSome applications
Hierarchical encodingStructural properties within pyramidsCombinatorial Pyramids
Reduction window
◮ v ∈ Vi comes from the merge of a connected set of vertice inGi−1.
RWi (v) = {v1, . . . , vn} ⊂ Vi−1
◮ vj ∈ RWi (v) is a son of v ,◮ v is the father of all vj ∈ RWj(v).
17 / 51
![Page 18: Partitions et structures hiérarchiques - GREYC€¦ · Partitions et structures hi´erarchiques Luc Brun Groupe de Recherche en Informatique, Image, Automatique et Instrumentation](https://reader035.vdocuments.site/reader035/viewer/2022070613/5b976d9b09d3f206218c8c4d/html5/thumbnails/18.jpg)
PresentationIntroduction
Regular PyramidsIrregular PyramidsSome applications
Hierarchical encodingStructural properties within pyramidsCombinatorial Pyramids
Receptive field◮ Receptive field : transitive closure of the father/child
relationship.
RFi (v) =⋃
v ′∈RWi (v)
RWi−1(v′) ⊂ V0
◮ w ∈ RFi (v) is a descendant of v ,◮ v is an ancestor of w .
18 / 51
![Page 19: Partitions et structures hiérarchiques - GREYC€¦ · Partitions et structures hi´erarchiques Luc Brun Groupe de Recherche en Informatique, Image, Automatique et Instrumentation](https://reader035.vdocuments.site/reader035/viewer/2022070613/5b976d9b09d3f206218c8c4d/html5/thumbnails/19.jpg)
PresentationIntroduction
Regular PyramidsIrregular PyramidsSome applications
Hierarchical encodingStructural properties within pyramidsCombinatorial Pyramids
Research fields
◮ Pyramid construction schemes (verticaldefinition)
◮ sequential methods,◮ parallel methods.
◮ kernel method [Meer 89],◮ Data driven decimation [Jolion 2001],◮ decimation by maximal matching
[Haximusa et al. 2005].
◮ Encoding of partitions (horizontaldefinition)
◮ simple graphs,◮ dual graphs,◮ combinatorial maps
19 / 51
![Page 20: Partitions et structures hiérarchiques - GREYC€¦ · Partitions et structures hi´erarchiques Luc Brun Groupe de Recherche en Informatique, Image, Automatique et Instrumentation](https://reader035.vdocuments.site/reader035/viewer/2022070613/5b976d9b09d3f206218c8c4d/html5/thumbnails/20.jpg)
PresentationIntroduction
Regular PyramidsIrregular PyramidsSome applications
Hierarchical encodingStructural properties within pyramidsCombinatorial Pyramids
Construction schemes of the pyramid◮ sequential methods :
◮ sort the edges of the graphs◮ Union-find
◮ parallel method :◮ Define parallel merge operations◮ each step builds a new graph Gi+1 from Gi .◮ |Gi+1| is a fixed ratio of |Gi |.
|Gi+1| ≈ q|Gi | with q < 1 : reduction factor
◮ p the parallelism is a constraint for segmentation algorithms◮
yq q: “forces” a fixed amount of fusions at each step
|Gi+1| ≈ q|Gi |
◮yq q: bounds the number of graphs we have to build/store
P = (G0 . . . , Gn) with n = logr (|G0|)
20 / 51
![Page 21: Partitions et structures hiérarchiques - GREYC€¦ · Partitions et structures hi´erarchiques Luc Brun Groupe de Recherche en Informatique, Image, Automatique et Instrumentation](https://reader035.vdocuments.site/reader035/viewer/2022070613/5b976d9b09d3f206218c8c4d/html5/thumbnails/21.jpg)
PresentationIntroduction
Regular PyramidsIrregular PyramidsSome applications
Hierarchical encodingStructural properties within pyramidsCombinatorial Pyramids
Parallel construction schemes
◮ A set of independant processes merge vertices in parallel
◮ Problem : How to insure that : Vi
Vi−1/ 1
2
◮ computational time◮ storage memory.
21 / 51
![Page 22: Partitions et structures hiérarchiques - GREYC€¦ · Partitions et structures hi´erarchiques Luc Brun Groupe de Recherche en Informatique, Image, Automatique et Instrumentation](https://reader035.vdocuments.site/reader035/viewer/2022070613/5b976d9b09d3f206218c8c4d/html5/thumbnails/22.jpg)
PresentationIntroduction
Regular PyramidsIrregular PyramidsSome applications
Hierarchical encodingStructural properties within pyramidsCombinatorial Pyramids
Kernel methods
◮ Introduced by Meer in 1989.
◮ We build a set of surviving vertice which will correspond tothe vertice of Vi+1.
◮ Vi+1 must satisfy two constraints :
External stability :
∀v ∈ Vi − Vi+1 ∃v ′ ∈ Vi+1 : (v , v ′) ∈ Ei
Each non surviving vertex is adjacent to at leasta surviving one
Internal stability :∀(v , v ′) ∈ V 2
i+1 (v , v ′) ∈ Ei
Two adjacent vertice cannot both survive
22 / 51
![Page 23: Partitions et structures hiérarchiques - GREYC€¦ · Partitions et structures hi´erarchiques Luc Brun Groupe de Recherche en Informatique, Image, Automatique et Instrumentation](https://reader035.vdocuments.site/reader035/viewer/2022070613/5b976d9b09d3f206218c8c4d/html5/thumbnails/23.jpg)
PresentationIntroduction
Regular PyramidsIrregular PyramidsSome applications
Hierarchical encodingStructural properties within pyramidsCombinatorial Pyramids
Kernel construction scheme : selection of surviving vertice
i ii ii
◮ Three variables :pi = true if vi survivesqi = true if vi may become a surviving vertex (he is
candidate).xi value of the vertex (function or random variable)
p(1)i = xi = maxj∈V (vi ){xj}
q(1)i =
∧
j∈V (vi )pj
(1)
p(k+1)i = p
(k)i ∨ (q
(k)i ∧ xi = maxj∈V (vi ){q
(k)j xj})
q(k+1)i =
∧
j∈V (vi )pj
(k+1)23 / 51
![Page 24: Partitions et structures hiérarchiques - GREYC€¦ · Partitions et structures hi´erarchiques Luc Brun Groupe de Recherche en Informatique, Image, Automatique et Instrumentation](https://reader035.vdocuments.site/reader035/viewer/2022070613/5b976d9b09d3f206218c8c4d/html5/thumbnails/24.jpg)
PresentationIntroduction
Regular PyramidsIrregular PyramidsSome applications
Hierarchical encodingStructural properties within pyramidsCombinatorial Pyramids
Kernel construction scheme : selection of surviving vertice
i ii9 7 6 8 9ii
◮ Three variables :pi = true if vi survivesqi = true if vi may become a surviving vertex (he is
candidate).xi value of the vertex (function or random variable)
p(1)i = xi = maxj∈V (vi ){xj}
q(1)i =
∧
j∈V (vi )pj
(1)
p(k+1)i = p
(k)i ∨ (q
(k)i ∧ xi = maxj∈V (vi ){q
(k)j xj})
q(k+1)i =
∧
j∈V (vi )pj
(k+1)23 / 51
![Page 25: Partitions et structures hiérarchiques - GREYC€¦ · Partitions et structures hi´erarchiques Luc Brun Groupe de Recherche en Informatique, Image, Automatique et Instrumentation](https://reader035.vdocuments.site/reader035/viewer/2022070613/5b976d9b09d3f206218c8c4d/html5/thumbnails/25.jpg)
PresentationIntroduction
Regular PyramidsIrregular PyramidsSome applications
Hierarchical encodingStructural properties within pyramidsCombinatorial Pyramids
Kernel construction scheme : selection of surviving vertice
i ii9 7 6 8 9i iii
◮ Three variables :pi = true if vi survivesqi = true if vi may become a surviving vertex (he is
candidate).xi value of the vertex (function or random variable)
p(1)i = xi = maxj∈V (vi ){xj}
q(1)i =
∧
j∈V (vi )pj
(1)
p(k+1)i = p
(k)i ∨ (q
(k)i ∧ xi = maxj∈V (vi ){q
(k)j xj})
q(k+1)i =
∧
j∈V (vi )pj
(k+1)23 / 51
![Page 26: Partitions et structures hiérarchiques - GREYC€¦ · Partitions et structures hi´erarchiques Luc Brun Groupe de Recherche en Informatique, Image, Automatique et Instrumentation](https://reader035.vdocuments.site/reader035/viewer/2022070613/5b976d9b09d3f206218c8c4d/html5/thumbnails/26.jpg)
PresentationIntroduction
Regular PyramidsIrregular PyramidsSome applications
Hierarchical encodingStructural properties within pyramidsCombinatorial Pyramids
Kernel construction scheme : selection of surviving vertice
9 7 6 8 9i iii k
◮ Three variables :pi = true if vi survivesqi = true if vi may become a surviving vertex (he is
candidate).xi value of the vertex (function or random variable)
p(1)i = xi = maxj∈V (vi ){xj}
q(1)i =
∧
j∈V (vi )pj
(1)
p(k+1)i = p
(k)i ∨ (q
(k)i ∧ xi = maxj∈V (vi ){q
(k)j xj})
q(k+1)i =
∧
j∈V (vi )pj
(k+1)23 / 51
![Page 27: Partitions et structures hiérarchiques - GREYC€¦ · Partitions et structures hi´erarchiques Luc Brun Groupe de Recherche en Informatique, Image, Automatique et Instrumentation](https://reader035.vdocuments.site/reader035/viewer/2022070613/5b976d9b09d3f206218c8c4d/html5/thumbnails/27.jpg)
PresentationIntroduction
Regular PyramidsIrregular PyramidsSome applications
Hierarchical encodingStructural properties within pyramidsCombinatorial Pyramids
Kernel construction scheme : father/child relationships
ii
i
ii
ii
i
iii
1
5
10
11
6
20
9
6
15
11
3
9
17
7
10
21i
ii
ii
i
◮ link each non surviving vertex to one of its survivingneighbour ⇒ definition of the edges
◮ merge non surviving vertice to surviving ones along theselected edges(merge in simple graphs).
24 / 51
![Page 28: Partitions et structures hiérarchiques - GREYC€¦ · Partitions et structures hi´erarchiques Luc Brun Groupe de Recherche en Informatique, Image, Automatique et Instrumentation](https://reader035.vdocuments.site/reader035/viewer/2022070613/5b976d9b09d3f206218c8c4d/html5/thumbnails/28.jpg)
PresentationIntroduction
Regular PyramidsIrregular PyramidsSome applications
Hierarchical encodingStructural properties within pyramidsCombinatorial Pyramids
Kernel construction scheme : father/child relationships
ii
i
ii
ii
i
iii
1
5
10
11
6
20
9
6
15
11
3
9
17
7
10
21i
ii
ii
i
◮ link each non surviving vertex to one of its survivingneighbour ⇒ definition of the edges
◮ merge non surviving vertice to surviving ones along theselected edges(merge in simple graphs).
24 / 51
![Page 29: Partitions et structures hiérarchiques - GREYC€¦ · Partitions et structures hi´erarchiques Luc Brun Groupe de Recherche en Informatique, Image, Automatique et Instrumentation](https://reader035.vdocuments.site/reader035/viewer/2022070613/5b976d9b09d3f206218c8c4d/html5/thumbnails/29.jpg)
PresentationIntroduction
Regular PyramidsIrregular PyramidsSome applications
Hierarchical encodingStructural properties within pyramidsCombinatorial Pyramids
Kernel construction scheme : father/child relationships
ii
i
ii
ii
i
iii
1
5
10
11
6
20
9
6
15
11
3
9
17
7
10
21i
ii
ii
i i
ii
ii
i
���
��
�� @@
@@
@@
1
3
20
17
2111
◮ link each non surviving vertex to one of its survivingneighbour ⇒ definition of the edges
◮ merge non surviving vertice to surviving ones along theselected edges(merge in simple graphs).
24 / 51
![Page 30: Partitions et structures hiérarchiques - GREYC€¦ · Partitions et structures hi´erarchiques Luc Brun Groupe de Recherche en Informatique, Image, Automatique et Instrumentation](https://reader035.vdocuments.site/reader035/viewer/2022070613/5b976d9b09d3f206218c8c4d/html5/thumbnails/30.jpg)
PresentationIntroduction
Regular PyramidsIrregular PyramidsSome applications
Hierarchical encodingStructural properties within pyramidsCombinatorial Pyramids
Maximal matching : Motivations
◮ Method introduced by Haximusa & Kropatsch ≈ 2005
◮ within the kernel construction scheme the probability that avertex survives decreases with its degree.
◮ The mean degree of vertices increases within the pyramid.
◮ ⇒ The ratio Vi
Vi−1computed by the kernel method decreases
according to the level
◮ Increases the computational time, even on parallel processors.◮ Useless graph storage.
25 / 51
![Page 31: Partitions et structures hiérarchiques - GREYC€¦ · Partitions et structures hi´erarchiques Luc Brun Groupe de Recherche en Informatique, Image, Automatique et Instrumentation](https://reader035.vdocuments.site/reader035/viewer/2022070613/5b976d9b09d3f206218c8c4d/html5/thumbnails/31.jpg)
PresentationIntroduction
Regular PyramidsIrregular PyramidsSome applications
Hierarchical encodingStructural properties within pyramidsCombinatorial Pyramids
Maximal matching
◮ Define a maximal matching C (kernel of G ′ = (E , E ′))◮ (e, e′) ∈ E ′ iff e and e′ are incident to a same vertex.
◮ Complete the matching C to C+
◮ Remove edges from C+ in order to obtain trees of depth 1.
◮ Merge vertice adjacent along selected edges.
yyyy
yyyy
yyyy
yyyy
◮ A set C ⊂ E is said to be a matching ofG = (V , E ) if none of the edges of C areadjacent to a same vertex.
◮ A matching is said to be maximal if theaddition of any edge breaks the matchingproperty.
◮ A matching is said to be maximum if nolarger matching may be found.
26 / 51
![Page 32: Partitions et structures hiérarchiques - GREYC€¦ · Partitions et structures hi´erarchiques Luc Brun Groupe de Recherche en Informatique, Image, Automatique et Instrumentation](https://reader035.vdocuments.site/reader035/viewer/2022070613/5b976d9b09d3f206218c8c4d/html5/thumbnails/32.jpg)
PresentationIntroduction
Regular PyramidsIrregular PyramidsSome applications
Hierarchical encodingStructural properties within pyramidsCombinatorial Pyramids
Maximal matching
◮ Complete the matching C to C+
◮ Remove edges from C+ in order to obtain trees of depth 1.
◮ Merge vertice adjacent along selected edges.
yyyy
yyyy
yyyy
yyyy
yyyy
yyyy
yyyy
yyyy
26 / 51
![Page 33: Partitions et structures hiérarchiques - GREYC€¦ · Partitions et structures hi´erarchiques Luc Brun Groupe de Recherche en Informatique, Image, Automatique et Instrumentation](https://reader035.vdocuments.site/reader035/viewer/2022070613/5b976d9b09d3f206218c8c4d/html5/thumbnails/33.jpg)
PresentationIntroduction
Regular PyramidsIrregular PyramidsSome applications
Hierarchical encodingStructural properties within pyramidsCombinatorial Pyramids
Maximal matching
◮ Complete the matching C to C+
◮ Remove edges from C+ in order to obtain trees of depth 1.
◮ Merge vertice adjacent along selected edges.
yyyy
yyyy
yyyy
yyyy
yyyy
yyyy
yyyy
yyyy
yyyy
yyyy
yyyy
yyyy
26 / 51
![Page 34: Partitions et structures hiérarchiques - GREYC€¦ · Partitions et structures hi´erarchiques Luc Brun Groupe de Recherche en Informatique, Image, Automatique et Instrumentation](https://reader035.vdocuments.site/reader035/viewer/2022070613/5b976d9b09d3f206218c8c4d/html5/thumbnails/34.jpg)
PresentationIntroduction
Regular PyramidsIrregular PyramidsSome applications
Hierarchical encodingStructural properties within pyramidsCombinatorial Pyramids
Data driven decimation
i9 6 9i i8i7i i
◮ perform on iteration of the kernel computation,
◮ attach each non surviving vertex to a surviving one
◮ merge vertices
◮ continue on the reduced graph
◮ Method introduced by Jolion ≈ 2001
27 / 51
![Page 35: Partitions et structures hiérarchiques - GREYC€¦ · Partitions et structures hi´erarchiques Luc Brun Groupe de Recherche en Informatique, Image, Automatique et Instrumentation](https://reader035.vdocuments.site/reader035/viewer/2022070613/5b976d9b09d3f206218c8c4d/html5/thumbnails/35.jpg)
PresentationIntroduction
Regular PyramidsIrregular PyramidsSome applications
Hierarchical encodingStructural properties within pyramidsCombinatorial Pyramids
Data driven decimation
i9 6 9i i8i7i ii iii
◮ perform on iteration of the kernel computation,
◮ attach each non surviving vertex to a surviving one
◮ merge vertices
◮ continue on the reduced graph
◮ Method introduced by Jolion ≈ 2001
27 / 51
![Page 36: Partitions et structures hiérarchiques - GREYC€¦ · Partitions et structures hi´erarchiques Luc Brun Groupe de Recherche en Informatique, Image, Automatique et Instrumentation](https://reader035.vdocuments.site/reader035/viewer/2022070613/5b976d9b09d3f206218c8c4d/html5/thumbnails/36.jpg)
PresentationIntroduction
Regular PyramidsIrregular PyramidsSome applications
Hierarchical encodingStructural properties within pyramidsCombinatorial Pyramids
Data driven decimation
i9 6 9i i8i7i ii iii
◮ perform on iteration of the kernel computation,
◮ attach each non surviving vertex to a surviving one
◮ merge vertices
◮ continue on the reduced graph
◮ Method introduced by Jolion ≈ 2001
27 / 51
![Page 37: Partitions et structures hiérarchiques - GREYC€¦ · Partitions et structures hi´erarchiques Luc Brun Groupe de Recherche en Informatique, Image, Automatique et Instrumentation](https://reader035.vdocuments.site/reader035/viewer/2022070613/5b976d9b09d3f206218c8c4d/html5/thumbnails/37.jpg)
PresentationIntroduction
Regular PyramidsIrregular PyramidsSome applications
Hierarchical encodingStructural properties within pyramidsCombinatorial Pyramids
Data driven decimation
i9 6 9i i
◮ perform on iteration of the kernel computation,
◮ attach each non surviving vertex to a surviving one
◮ merge vertices
◮ continue on the reduced graph
◮ Method introduced by Jolion ≈ 2001
27 / 51
![Page 38: Partitions et structures hiérarchiques - GREYC€¦ · Partitions et structures hi´erarchiques Luc Brun Groupe de Recherche en Informatique, Image, Automatique et Instrumentation](https://reader035.vdocuments.site/reader035/viewer/2022070613/5b976d9b09d3f206218c8c4d/html5/thumbnails/38.jpg)
PresentationIntroduction
Regular PyramidsIrregular PyramidsSome applications
Hierarchical encodingStructural properties within pyramidsCombinatorial Pyramids
Data driven decimation
i9 6 9i ii ii
◮ perform on iteration of the kernel computation,
◮ attach each non surviving vertex to a surviving one
◮ merge vertices
◮ continue on the reduced graph
◮ Method introduced by Jolion ≈ 2001
27 / 51
![Page 39: Partitions et structures hiérarchiques - GREYC€¦ · Partitions et structures hi´erarchiques Luc Brun Groupe de Recherche en Informatique, Image, Automatique et Instrumentation](https://reader035.vdocuments.site/reader035/viewer/2022070613/5b976d9b09d3f206218c8c4d/html5/thumbnails/39.jpg)
PresentationIntroduction
Regular PyramidsIrregular PyramidsSome applications
Hierarchical encodingStructural properties within pyramidsCombinatorial Pyramids
Data driven decimation : conclusion
◮ only one step of the kernel computation is performed◮ “Corresponds” to a model of the behavior of our brain,◮ allows to avoid (in some cases) wrong merge operations.
ii
i
ii
ii
i
iii
1
5
10
11
6
20
9
6
15
11
3
9
17
7
10
21i
ii
ii
i
28 / 51
![Page 40: Partitions et structures hiérarchiques - GREYC€¦ · Partitions et structures hi´erarchiques Luc Brun Groupe de Recherche en Informatique, Image, Automatique et Instrumentation](https://reader035.vdocuments.site/reader035/viewer/2022070613/5b976d9b09d3f206218c8c4d/html5/thumbnails/40.jpg)
PresentationIntroduction
Regular PyramidsIrregular PyramidsSome applications
Hierarchical encodingStructural properties within pyramidsCombinatorial Pyramids
Data driven decimation and maximal matching
◮ Method introduced by Pruvot & Brun◮ The maximal matching is defined as a MIS on the graph
G = (E , E ′).
1. Value each edge as a merging cost,2. Perform only one iteration of the maximal matching algorithm3. One edge is selected if it is locally minimal (the two regions
like each other more than any of their neighbour).
yyyy
yyyy
yyyy
yyyy
1 2 3 1
1
5
10 20 30 10
15 22 34 1111 8 14
6 7 5
5 3
13 17
29 / 51
![Page 41: Partitions et structures hiérarchiques - GREYC€¦ · Partitions et structures hi´erarchiques Luc Brun Groupe de Recherche en Informatique, Image, Automatique et Instrumentation](https://reader035.vdocuments.site/reader035/viewer/2022070613/5b976d9b09d3f206218c8c4d/html5/thumbnails/41.jpg)
PresentationIntroduction
Regular PyramidsIrregular PyramidsSome applications
Hierarchical encodingStructural properties within pyramidsCombinatorial Pyramids
Simple graph pyramids
Gi → Gi+1 by a merge operation between vertice
◮ Construction scheme :1. Define Ki ⊂ Ei (c.f. previous slides)2. Contract Ki
3. Remove any loops and double edges
yyyy
yyyy
yyyy
yyyy
−→Ki
y ��y �
�
� �
y yy
y� �
Gi Contraction Removal
30 / 51
![Page 42: Partitions et structures hiérarchiques - GREYC€¦ · Partitions et structures hi´erarchiques Luc Brun Groupe de Recherche en Informatique, Image, Automatique et Instrumentation](https://reader035.vdocuments.site/reader035/viewer/2022070613/5b976d9b09d3f206218c8c4d/html5/thumbnails/42.jpg)
PresentationIntroduction
Regular PyramidsIrregular PyramidsSome applications
Hierarchical encodingStructural properties within pyramidsCombinatorial Pyramids
Construction scheme of dual graph pyramids
◮ Gi = G0
◮ Define a set Ki of edges to be contracted
yyyy
yyyy
yyyy
yyyy
31 / 51
![Page 43: Partitions et structures hiérarchiques - GREYC€¦ · Partitions et structures hi´erarchiques Luc Brun Groupe de Recherche en Informatique, Image, Automatique et Instrumentation](https://reader035.vdocuments.site/reader035/viewer/2022070613/5b976d9b09d3f206218c8c4d/html5/thumbnails/43.jpg)
PresentationIntroduction
Regular PyramidsIrregular PyramidsSome applications
Hierarchical encodingStructural properties within pyramidsCombinatorial Pyramids
Construction scheme of dual graph pyramids
◮ Gi = G0
◮ Define a set Ki of edges to be contracted◮ Ki must be a forest of Gi (we do not contract loops) called a
contraction kernel
yyyy
yyyy
yyyy
yyyy
31 / 51
![Page 44: Partitions et structures hiérarchiques - GREYC€¦ · Partitions et structures hi´erarchiques Luc Brun Groupe de Recherche en Informatique, Image, Automatique et Instrumentation](https://reader035.vdocuments.site/reader035/viewer/2022070613/5b976d9b09d3f206218c8c4d/html5/thumbnails/44.jpg)
PresentationIntroduction
Regular PyramidsIrregular PyramidsSome applications
Hierarchical encodingStructural properties within pyramidsCombinatorial Pyramids
Construction scheme of dual graph pyramids
◮ Gi = G0
◮ Define a set Ki of edges to be contracted
◮ Contract Ki within Gi → Gi+1,
◮ Define a set Ki+1 of edges to remove
◮ Contract Ki+1 within Gi+1 → Gi+2.
yyyy
yyyy
yyyy
yyyy
y ��y �
�
� �
y
31 / 51
![Page 45: Partitions et structures hiérarchiques - GREYC€¦ · Partitions et structures hi´erarchiques Luc Brun Groupe de Recherche en Informatique, Image, Automatique et Instrumentation](https://reader035.vdocuments.site/reader035/viewer/2022070613/5b976d9b09d3f206218c8c4d/html5/thumbnails/45.jpg)
PresentationIntroduction
Regular PyramidsIrregular PyramidsSome applications
Hierarchical encodingStructural properties within pyramidsCombinatorial Pyramids
Construction scheme of dual graph pyramids
◮ Gi = G0
◮ Define a set Ki of edges to be contracted
◮ Contract Ki within Gi → Gi+1,
◮ Define a set Ki+1 of edges to remove◮ Ki+1 is a forest of Gi+1 called a Removal kernel◮ e ∈ Ki+1 ⇒ e is incident to f ∈ Vi+1, d◦(f ) ≤ 2.
◮ Contract Ki+1 within Gi+1 → Gi+2.
yyyy
yyyy
yyyy
yyyy
y ��y �
�
� �
y
31 / 51
![Page 46: Partitions et structures hiérarchiques - GREYC€¦ · Partitions et structures hi´erarchiques Luc Brun Groupe de Recherche en Informatique, Image, Automatique et Instrumentation](https://reader035.vdocuments.site/reader035/viewer/2022070613/5b976d9b09d3f206218c8c4d/html5/thumbnails/46.jpg)
PresentationIntroduction
Regular PyramidsIrregular PyramidsSome applications
Hierarchical encodingStructural properties within pyramidsCombinatorial Pyramids
Construction scheme of dual graph pyramids
◮ Gi = G0
◮ Define a set Ki of edges to be contracted
◮ Contract Ki within Gi → Gi+1,
◮ Define a set Ki+1 of edges to remove
◮ Contract Ki+1 within Gi+1 → Gi+2.
yyyy
yyyy
yyyy
yyyy
y ��y �
�
� �
y yy
y� �
31 / 51
![Page 47: Partitions et structures hiérarchiques - GREYC€¦ · Partitions et structures hi´erarchiques Luc Brun Groupe de Recherche en Informatique, Image, Automatique et Instrumentation](https://reader035.vdocuments.site/reader035/viewer/2022070613/5b976d9b09d3f206218c8c4d/html5/thumbnails/47.jpg)
PresentationIntroduction
Regular PyramidsIrregular PyramidsSome applications
Hierarchical encodingStructural properties within pyramidsCombinatorial Pyramids
Comparison of dual and simple graph pyramidalconstruction schemes
◮ Construction :Simple Pyr. Dual Graph Pyr.
Etap 1 Define K Define K
Etap 2 Merge according to K Contract K within G
Etap 3 Define K
Etap 4 Contract K within G
◮ Information associated to edges :◮ Simple graphs : adjacency between regions.◮ Dual graphs : boundary information
◮ Reduction window, receptive fields : same definitions in bothcases.
32 / 51
![Page 48: Partitions et structures hiérarchiques - GREYC€¦ · Partitions et structures hi´erarchiques Luc Brun Groupe de Recherche en Informatique, Image, Automatique et Instrumentation](https://reader035.vdocuments.site/reader035/viewer/2022070613/5b976d9b09d3f206218c8c4d/html5/thumbnails/48.jpg)
PresentationIntroduction
Regular PyramidsIrregular PyramidsSome applications
Hierarchical encodingStructural properties within pyramidsCombinatorial Pyramids
Comparison of dual and simple graph pyramidalconstruction schemes
◮ Construction :Simple Pyr. Dual Graph Pyr.
Etap 1 Define K Define K
Etap 2 Merge according to K Contract K within G
Etap 3 Define K
Etap 4 Contract K within G
◮ Information associated to edges :◮ Simple graphs : adjacency between regions.◮ Dual graphs : boundary information
◮ Reduction window, receptive fields : same definitions in bothcases.
32 / 51
![Page 49: Partitions et structures hiérarchiques - GREYC€¦ · Partitions et structures hi´erarchiques Luc Brun Groupe de Recherche en Informatique, Image, Automatique et Instrumentation](https://reader035.vdocuments.site/reader035/viewer/2022070613/5b976d9b09d3f206218c8c4d/html5/thumbnails/49.jpg)
PresentationIntroduction
Regular PyramidsIrregular PyramidsSome applications
Hierarchical encodingStructural properties within pyramidsCombinatorial Pyramids
Comparison of dual and simple graph pyramidalconstruction schemes
◮ Construction :Simple Pyr. Dual Graph Pyr.
Etap 1 Define K Define K
Etap 2 Merge according to K Contract K within G
Etap 3 Define K
Etap 4 Contract K within G
◮ Information associated to edges :◮ Simple graphs : adjacency between regions.◮ Dual graphs : boundary information
◮ Reduction window, receptive fields : same definitions in bothcases.
32 / 51
![Page 50: Partitions et structures hiérarchiques - GREYC€¦ · Partitions et structures hi´erarchiques Luc Brun Groupe de Recherche en Informatique, Image, Automatique et Instrumentation](https://reader035.vdocuments.site/reader035/viewer/2022070613/5b976d9b09d3f206218c8c4d/html5/thumbnails/50.jpg)
PresentationIntroduction
Regular PyramidsIrregular PyramidsSome applications
Hierarchical encodingStructural properties within pyramidsCombinatorial Pyramids
Combinatorial Pyramids : Construction
◮ Same contruction scheme than for the dual graph pyramids
1. Definition of a contraction kernel K of G ,2. Definition of two removal kernels K1 and K2 of G removing
respectively :◮ the empty loops (|ϕ∗(b)| = 1) and◮ the double edges (|ϕ∗(b)| = 2).
◮ P = (G0, . . . ,Gn) Gi is deduced from Gi−1 by a contraction ora removal kernel
Dn ⊂ Dn−1 ⊂ · · · ⊂ D0
33 / 51
![Page 51: Partitions et structures hiérarchiques - GREYC€¦ · Partitions et structures hi´erarchiques Luc Brun Groupe de Recherche en Informatique, Image, Automatique et Instrumentation](https://reader035.vdocuments.site/reader035/viewer/2022070613/5b976d9b09d3f206218c8c4d/html5/thumbnails/51.jpg)
PresentationIntroduction
Regular PyramidsIrregular PyramidsSome applications
Hierarchical encodingStructural properties within pyramidsCombinatorial Pyramids
Combinatorial Pyramids and graphs
◮ One single map Gi = (Di , σi , αi ) instead of two graphs(Gi , Gi ) at each level.
◮ Reduction window problem :◮ p a vertex of Gi is defined by a cycle σ∗
i (b), b ∈ Di
◮ ⇒ No explicit encoding of vertice.
◮ Two solutions :◮ create a labeling (explicit encoding) of vertice
yq q◮ Speak map language yq q
34 / 51
![Page 52: Partitions et structures hiérarchiques - GREYC€¦ · Partitions et structures hi´erarchiques Luc Brun Groupe de Recherche en Informatique, Image, Automatique et Instrumentation](https://reader035.vdocuments.site/reader035/viewer/2022070613/5b976d9b09d3f206218c8c4d/html5/thumbnails/52.jpg)
PresentationIntroduction
Regular PyramidsIrregular PyramidsSome applications
Hierarchical encodingStructural properties within pyramidsCombinatorial Pyramids
Combinatorial Pyramids and graphs
◮ One single map Gi = (Di , σi , αi ) instead of two graphs(Gi , Gi ) at each level.
◮ Reduction window problem :
RFi (v) = {v1, . . . , vn}
◮ p a vertex of Gi is defined by a cycle σ∗
i (b), b ∈ Di
◮ ⇒ No explicit encoding of vertice.
◮ Two solutions :◮ create a labeling (explicit encoding) of vertice yq q◮ Speak map language
yq q
34 / 51
![Page 53: Partitions et structures hiérarchiques - GREYC€¦ · Partitions et structures hi´erarchiques Luc Brun Groupe de Recherche en Informatique, Image, Automatique et Instrumentation](https://reader035.vdocuments.site/reader035/viewer/2022070613/5b976d9b09d3f206218c8c4d/html5/thumbnails/53.jpg)
PresentationIntroduction
Regular PyramidsIrregular PyramidsSome applications
Hierarchical encodingStructural properties within pyramidsCombinatorial Pyramids
Combinatorial Pyramids and graphs
◮ One single map Gi = (Di , σi , αi ) instead of two graphs(Gi , Gi ) at each level.
◮ Reduction window problem :◮ p a vertex of Gi is defined by a cycle σ∗
i (b), b ∈ Di
◮ ⇒ No explicit encoding of vertice.
◮ Two solutions :◮ create a labeling (explicit encoding) of vertice
yq q◮ Speak map language yq q
34 / 51
![Page 54: Partitions et structures hiérarchiques - GREYC€¦ · Partitions et structures hi´erarchiques Luc Brun Groupe de Recherche en Informatique, Image, Automatique et Instrumentation](https://reader035.vdocuments.site/reader035/viewer/2022070613/5b976d9b09d3f206218c8c4d/html5/thumbnails/54.jpg)
PresentationIntroduction
Regular PyramidsIrregular PyramidsSome applications
Hierarchical encodingStructural properties within pyramidsCombinatorial Pyramids
Combinatorial Pyramids and graphs
◮ One single map Gi = (Di , σi , αi ) instead of two graphs(Gi , Gi ) at each level.
◮ Reduction window problem :◮ p a vertex of Gi is defined by a cycle σ∗
i (b), b ∈ Di
◮ ⇒ No explicit encoding of vertice.
◮ Two solutions :◮ create a labeling (explicit encoding) of vertice
yq q◮ Speak map language yq q
34 / 51
![Page 55: Partitions et structures hiérarchiques - GREYC€¦ · Partitions et structures hi´erarchiques Luc Brun Groupe de Recherche en Informatique, Image, Automatique et Instrumentation](https://reader035.vdocuments.site/reader035/viewer/2022070613/5b976d9b09d3f206218c8c4d/html5/thumbnails/55.jpg)
PresentationIntroduction
Regular PyramidsIrregular PyramidsSome applications
Hierarchical encodingStructural properties within pyramidsCombinatorial Pyramids
Combinatorial Pyramids and graphs
◮ One single map Gi = (Di , σi , αi ) instead of two graphs(Gi , Gi ) at each level.
◮ Reduction window problem :◮ p a vertex of Gi is defined by a cycle σ∗
i (b), b ∈ Di
◮ ⇒ No explicit encoding of vertice.
◮ Two solutions :◮ create a labeling (explicit encoding) of vertice
yq q◮ Speak map language yq q
34 / 51
![Page 56: Partitions et structures hiérarchiques - GREYC€¦ · Partitions et structures hi´erarchiques Luc Brun Groupe de Recherche en Informatique, Image, Automatique et Instrumentation](https://reader035.vdocuments.site/reader035/viewer/2022070613/5b976d9b09d3f206218c8c4d/html5/thumbnails/56.jpg)
PresentationIntroduction
Regular PyramidsIrregular PyramidsSome applications
Hierarchical encodingStructural properties within pyramidsCombinatorial Pyramids
Connecting walks :
◮ Let b ∈ Di , CWi (b) : sequence of darts to traversee withinGi−1 in order to connect b to
◮ ϕi (b) if Ki−1 is a contraction kernel◮ σi (b) si Ki−1 is a removal kernel.
35 / 51
![Page 57: Partitions et structures hiérarchiques - GREYC€¦ · Partitions et structures hi´erarchiques Luc Brun Groupe de Recherche en Informatique, Image, Automatique et Instrumentation](https://reader035.vdocuments.site/reader035/viewer/2022070613/5b976d9b09d3f206218c8c4d/html5/thumbnails/57.jpg)
PresentationIntroduction
Regular PyramidsIrregular PyramidsSome applications
Hierarchical encodingStructural properties within pyramidsCombinatorial Pyramids
Connecting walks :◮ If Ki−1 is a contraction kernel :
w w w@@ 2⊲3⊳�
�4⊲
−4⊳
1⊲6△ �
�9⊳5⊲ 8⊲@@7
⊳w 8⊲�
�9⊳6
△
@@ 2⊲3⊳�
�4 ⊲
7▽
CWi (b) = bϕi−1(b) . . . ϕn−1i−1 (b);
with n = Min{k ∈ N∗ | ϕk
i−1(b) ∈ Di},CWi (−4) = −4.1.5.
◮ If Ki−1 is a removal kernel
w 8⊲��9⊳
6△
@@ 2⊲3⊳�
�4 ⊲
7▽
w 8⊲@@ 2⊲3⊳�
�4 ⊲
7▽35 / 51
![Page 58: Partitions et structures hiérarchiques - GREYC€¦ · Partitions et structures hi´erarchiques Luc Brun Groupe de Recherche en Informatique, Image, Automatique et Instrumentation](https://reader035.vdocuments.site/reader035/viewer/2022070613/5b976d9b09d3f206218c8c4d/html5/thumbnails/58.jpg)
PresentationIntroduction
Regular PyramidsIrregular PyramidsSome applications
Hierarchical encodingStructural properties within pyramidsCombinatorial Pyramids
Connecting walks :
◮ If Ki−1 is a removal kernel
w 8⊲��9⊳
6△
@@ 2⊲3⊳�
�4 ⊲
7▽
w 8⊲@@ 2⊲3⊳�
�4 ⊲
7▽
CWi (b) = b.σi−1(b) . . . σn−1i−1 (b)
with n = Min{k ∈ N∗ | σk
i−1(b) ∈ Di}.CWi (8) = 8.9.6
35 / 51
![Page 59: Partitions et structures hiérarchiques - GREYC€¦ · Partitions et structures hi´erarchiques Luc Brun Groupe de Recherche en Informatique, Image, Automatique et Instrumentation](https://reader035.vdocuments.site/reader035/viewer/2022070613/5b976d9b09d3f206218c8c4d/html5/thumbnails/59.jpg)
PresentationIntroduction
Regular PyramidsIrregular PyramidsSome applications
Hierarchical encodingStructural properties within pyramidsCombinatorial Pyramids
Connecting walks
◮ In short :CWi (b) = b.b1. . . . .bp
◮ {b1, . . . , bp} ⊂ Di−1
ϕi (b) = ϕi−1(bp) If Ki−1 is a contraction kernelσi (b) = σi−1(bp) If Ki−1 is a removal kernel
◮ the connecting walks allows to compute Gi from Gi−1 and Ki
36 / 51
![Page 60: Partitions et structures hiérarchiques - GREYC€¦ · Partitions et structures hi´erarchiques Luc Brun Groupe de Recherche en Informatique, Image, Automatique et Instrumentation](https://reader035.vdocuments.site/reader035/viewer/2022070613/5b976d9b09d3f206218c8c4d/html5/thumbnails/60.jpg)
PresentationIntroduction
Regular PyramidsIrregular PyramidsSome applications
Hierarchical encodingStructural properties within pyramidsCombinatorial Pyramids
Connecting dart sequences
◮ Receptive field : transitive closure of reduction windows.
CRi (v) =⋃
v ′∈FRi (v)
CRi−1(v′) ⊂ V0
◮ Connecting dart sequences : closure (concatenation) ofconnecting walks
SCi (b) = SCi−1(b1) . . . .SCi−1(bp) ⊂ D0
with CWi (b) = b1 . . . .bp.
◮ nb : This last formula is only valid when two consecutivekernels (Ki−1, Ki have a same type (both contraction orremoval) kernels.
37 / 51
![Page 61: Partitions et structures hiérarchiques - GREYC€¦ · Partitions et structures hi´erarchiques Luc Brun Groupe de Recherche en Informatique, Image, Automatique et Instrumentation](https://reader035.vdocuments.site/reader035/viewer/2022070613/5b976d9b09d3f206218c8c4d/html5/thumbnails/61.jpg)
PresentationIntroduction
Regular PyramidsIrregular PyramidsSome applications
Hierarchical encodingStructural properties within pyramidsCombinatorial Pyramids
Connecting dart sequences
◮ Let G0 = (D0, σ0, α0).
∀b ∈ D0 SC0(b) = b
◮ for all i in {1, . . . , n}
∀b ∈ Di SCi (b) = b1.SC ∗
i−1(αi−1(b1)) . . . .bp.SC ∗
i−1(αi−1(bp))
with◮ CWi (b) = b1 . . . .bp,◮ SC∗
i−1(bj) : SCi−1(bj) minus its first dart (bj).
◮ SCi (b) is defined in G0.
∀b ∈ D1 SC1(b) = CW1(b)
37 / 51
![Page 62: Partitions et structures hiérarchiques - GREYC€¦ · Partitions et structures hi´erarchiques Luc Brun Groupe de Recherche en Informatique, Image, Automatique et Instrumentation](https://reader035.vdocuments.site/reader035/viewer/2022070613/5b976d9b09d3f206218c8c4d/html5/thumbnails/62.jpg)
PresentationIntroduction
Regular PyramidsIrregular PyramidsSome applications
Hierarchical encodingStructural properties within pyramidsCombinatorial Pyramids
Connecting dart sequence : Properties
CW ∗
i (b) ⊂ Ki and SC ∗
i (b) ⊂i−1⊔
j=0
Kj
◮ b ∈ Di , SCi (b) = b.b1 . . . , bp, p > 1
◮ If Ki is a contraction kernel :
ϕi (b) =
{
ϕ0(bp) If bp is contractedσ0(bp) Si bp is removed.
◮ If Ki is a removal kernel :
σi (b) =
{
ϕ0(bp) If bp is contractedσ0(bp) If bp is removed.
◮ May be extended to any j ≤ i → (G0, . . . ,Gn)
38 / 51
![Page 63: Partitions et structures hiérarchiques - GREYC€¦ · Partitions et structures hi´erarchiques Luc Brun Groupe de Recherche en Informatique, Image, Automatique et Instrumentation](https://reader035.vdocuments.site/reader035/viewer/2022070613/5b976d9b09d3f206218c8c4d/html5/thumbnails/63.jpg)
PresentationIntroduction
Regular PyramidsIrregular PyramidsSome applications
Hierarchical encodingStructural properties within pyramidsCombinatorial Pyramids
Connecting dart sequences : traversal
◮ Let SCi (b) = b.b1, . . . , bp, p > 1 :
b1 =
{
ϕ(b) If Ki is a contraction kernelσ(b) If Ki is a removal kernel
et
∀j ∈ {2, . . . , p} bj =
{
ϕ(bj−1) Si bj−1 is contractedσ(bj−1) Si bj−1 is supprimed.
◮ We need to know :◮ The type of Ki ,◮ The operation aplied to each dart.
39 / 51
![Page 64: Partitions et structures hiérarchiques - GREYC€¦ · Partitions et structures hi´erarchiques Luc Brun Groupe de Recherche en Informatique, Image, Automatique et Instrumentation](https://reader035.vdocuments.site/reader035/viewer/2022070613/5b976d9b09d3f206218c8c4d/html5/thumbnails/64.jpg)
PresentationIntroduction
Regular PyramidsIrregular PyramidsSome applications
Hierarchical encodingStructural properties within pyramidsCombinatorial Pyramids
Implicit encoding◮ Let the two functions :
◮ state
{1, . . . , n} → {0, 1, 2}
i 7→
0 If Ki cont. kernel1 If Ki rem. kernel (empty selft loops) ;2 If Ki rem. kernel (double edges)
In practice state(i) = i mod 3◮ level :
(
D0 → {1, . . . , n + 1}b 7→ max{i ∈ {1, . . . , n + 1} | b ∈ Di−1}.
◮ In terms of encoding :◮ state : arrray of bits of size n.◮ level :array of integers of size |D0|.
40 / 51
![Page 65: Partitions et structures hiérarchiques - GREYC€¦ · Partitions et structures hi´erarchiques Luc Brun Groupe de Recherche en Informatique, Image, Automatique et Instrumentation](https://reader035.vdocuments.site/reader035/viewer/2022070613/5b976d9b09d3f206218c8c4d/html5/thumbnails/65.jpg)
PresentationIntroduction
Regular PyramidsIrregular PyramidsSome applications
Hierarchical encodingStructural properties within pyramidsCombinatorial Pyramids
Implicit encoding : Definition
◮ Explicit encoding P = (G0, . . . ,Gn)
◮ Implicit encoding : P = (G0, state, level).◮ Any may Gi may be retreived from the implicit encoding◮ No more restriction on the number fo levels yq q yq q◮ Maximal compression :
◮ G0 : 4 connected grid → implicitly encoded◮ : state(i) : i mod 3◮ Encoding : P = (level)
◮ For practical reasons :◮ P = (G0, Gn, level) or◮ P = (Gn, level).
41 / 51
![Page 66: Partitions et structures hiérarchiques - GREYC€¦ · Partitions et structures hi´erarchiques Luc Brun Groupe de Recherche en Informatique, Image, Automatique et Instrumentation](https://reader035.vdocuments.site/reader035/viewer/2022070613/5b976d9b09d3f206218c8c4d/html5/thumbnails/66.jpg)
PresentationIntroduction
Regular PyramidsIrregular PyramidsSome applications
Hierarchical encodingStructural properties within pyramidsCombinatorial Pyramids
Implicit encoding
◮ Trversal of SCi (b) = b.b1 . . . .bp using stateand level :
b1 =
{
ϕ0(b) If state(i)=contractedσ0(b) If state(i)=removed
et
bj =
ϕ0(bj−1) If state(level(bj−1)) = Contractedσ0(bj−1) If state(level(bj−1)) = Removed
for j ∈ {2, . . . , p}
42 / 51
![Page 67: Partitions et structures hiérarchiques - GREYC€¦ · Partitions et structures hi´erarchiques Luc Brun Groupe de Recherche en Informatique, Image, Automatique et Instrumentation](https://reader035.vdocuments.site/reader035/viewer/2022070613/5b976d9b09d3f206218c8c4d/html5/thumbnails/67.jpg)
PresentationIntroduction
Regular PyramidsIrregular PyramidsSome applications
Hierarchical encodingStructural properties within pyramidsCombinatorial Pyramids
Conversion Implicit/Explicit
◮ Solution 1 : Compute successively each map G1, . . . ,Gnyq q .
◮ Solution 2 : study the structure of the connecting dartsequences
SCi (b) = b1.SC ∗
i−1(αi−1(b1)) . . . .bp.SC ∗
i−1(αi−1(bp))
◮ Computation of all maps (G0, . . . ,Gi ) in one traversal of theconnecting dart sequences of level i .
43 / 51
![Page 68: Partitions et structures hiérarchiques - GREYC€¦ · Partitions et structures hi´erarchiques Luc Brun Groupe de Recherche en Informatique, Image, Automatique et Instrumentation](https://reader035.vdocuments.site/reader035/viewer/2022070613/5b976d9b09d3f206218c8c4d/html5/thumbnails/68.jpg)
PresentationIntroduction
Regular PyramidsIrregular PyramidsSome applications
Hierarchical encodingStructural properties within pyramidsCombinatorial Pyramids
Conversion Implicit/Explicit
5
−11b
4
−5b1
1
6b2
σ0(−5)
1
−12b3
2
−9b4
σ1(−5)
1
−4b5
2
−8b6
3
−3b7
σ2(−5)
1
−7b8
1
1b9
2
8b10
1
10b11
4
5b12
σ3(−5)
1
−10b13
3
3b14
◮ If Dn = {b, αn(b)}, the whole pyramid is conmupted by onetraversal of SCn(b) et SCn(αn(b)).
∀i ∈ {0, . . . , n}⊔
b∈Di
SCi (b) = D0
43 / 51
![Page 69: Partitions et structures hiérarchiques - GREYC€¦ · Partitions et structures hi´erarchiques Luc Brun Groupe de Recherche en Informatique, Image, Automatique et Instrumentation](https://reader035.vdocuments.site/reader035/viewer/2022070613/5b976d9b09d3f206218c8c4d/html5/thumbnails/69.jpg)
PresentationIntroduction
Regular PyramidsIrregular PyramidsSome applications
Hierarchical encodingStructural properties within pyramidsCombinatorial Pyramids
Vertex receptive fields
◮ Can we obtain intermediate results between the darts and thewhole map ?
◮ Let σ∗
i (b) = (b1, . . . , bp). We want to connect bj toσi (bj) = bj+1 in G0.
RFi (b) =
{
SCi (b) If Ki is a removal kernelb.SC ∗
i (α(b)) If Ki is a contraction kernel.
◮ Receptive field of σ∗
i (b) :
Rσ∗i(b) =
⊙pj=1RFi (bj)
44 / 51
![Page 70: Partitions et structures hiérarchiques - GREYC€¦ · Partitions et structures hi´erarchiques Luc Brun Groupe de Recherche en Informatique, Image, Automatique et Instrumentation](https://reader035.vdocuments.site/reader035/viewer/2022070613/5b976d9b09d3f206218c8c4d/html5/thumbnails/70.jpg)
PresentationIntroduction
Regular PyramidsIrregular PyramidsSome applications
Hierarchical encodingStructural properties within pyramidsCombinatorial Pyramids
Embedding
◮ If G0 encode the 4 connected grid◮ σ∗
0 (b) corresponds to a pixel,◮ α∗
0(b) corresponds to a lignel,◮ b corresponds to an oriented lignel.
◮ ϕ∗
0(b) corresponds to a pointel.
7⊲
10⊲
8⊲
11⊲
9⊲
12⊲
1△ 2△
3△ 4△
5△ 6△
13⊳ 14⊳ 15⊳16△
17△
18△
21⊲
20⊲
19⊲
24▽
23▽
22▽ v
1⊲13△
24⊳7
▽ 1△
-1▽
13⊳24
▽
7⊲45 / 51
![Page 71: Partitions et structures hiérarchiques - GREYC€¦ · Partitions et structures hi´erarchiques Luc Brun Groupe de Recherche en Informatique, Image, Automatique et Instrumentation](https://reader035.vdocuments.site/reader035/viewer/2022070613/5b976d9b09d3f206218c8c4d/html5/thumbnails/71.jpg)
PresentationIntroduction
Regular PyramidsIrregular PyramidsSome applications
Hierarchical encodingStructural properties within pyramidsCombinatorial Pyramids
Embedding of receptive fieldsuuu
uuu
uuu
1⊲ -1⊳ 2⊲ -2⊳
3⊲ -3⊳ 4⊲ -4⊳
5⊲ -5⊳ 6⊲ -6⊳
7▽
-7△
10▽
-10△
8▽
-8△
11▽
-11△
9▽
-9△
12▽
-12△
13△
14△
15△
21▽ 20▽ 19▽
24⊳
23⊳
22⊳
16⊲
17⊲
18⊲
7⊲
10⊲
8⊲
11⊲
9⊲
12⊲
1△ 2△
3△ 4△
5△ 6△
13⊳ 14⊳ 15⊳16△
17△
18△21⊲
20⊲
19⊲
24▽
23▽
22▽
⊳23
⊲16
▽5
▽20
△-7 △-8
y
y
y 7⊲
8⊲
5△
16△
20⊲
23▽
46 / 51
![Page 72: Partitions et structures hiérarchiques - GREYC€¦ · Partitions et structures hi´erarchiques Luc Brun Groupe de Recherche en Informatique, Image, Automatique et Instrumentation](https://reader035.vdocuments.site/reader035/viewer/2022070613/5b976d9b09d3f206218c8c4d/html5/thumbnails/72.jpg)
PresentationIntroduction
Regular PyramidsIrregular PyramidsSome applications
Hierarchical encodingStructural properties within pyramidsCombinatorial Pyramids
Embedding of receptive fields
7⊲
10⊲8⊲
11⊲9⊲
12⊲
1△ 2△
3△ 4△
5△ 6△
13⊳ 14⊳ 15⊳16△
17△
18△
21⊲ 20⊲ 19⊲
24▽
23▽
22▽
⊳23
⊲16
▽5
▽ 20
△-7 △-8vvv 7⊲ 8⊲
5△
16△
20⊲
23▽
◮ σ∗
i (16) = (16, 7, 8)
Rσ∗i(16) = 16.15.−2.14.−1.13.24.7.1.8.2.9
46 / 51
![Page 73: Partitions et structures hiérarchiques - GREYC€¦ · Partitions et structures hi´erarchiques Luc Brun Groupe de Recherche en Informatique, Image, Automatique et Instrumentation](https://reader035.vdocuments.site/reader035/viewer/2022070613/5b976d9b09d3f206218c8c4d/html5/thumbnails/73.jpg)
PresentationIntroduction
Regular PyramidsIrregular PyramidsSome applications
Hierarchical encodingStructural properties within pyramidsCombinatorial Pyramids
Traversal of borders
◮ Traversal of darts encoding lignel borders :
∂Rσ∗i(b) = b1, . . . , bp avec
{
b1 = b
bj = ϕnj
0 (α0(bj))
where nj = Min{p ∈ N∗ |ϕp
0(α0(bj)) ∈ Di or double edge}.
◮ Rem : b double edge ⇔ level(b)mod3 = 2(cont., rem. emptyself-loops, rem. double edges).
◮ The border is included within the region ∂Rσ∗i(b) ⊂ R
σ∗i(b)
47 / 51
![Page 74: Partitions et structures hiérarchiques - GREYC€¦ · Partitions et structures hi´erarchiques Luc Brun Groupe de Recherche en Informatique, Image, Automatique et Instrumentation](https://reader035.vdocuments.site/reader035/viewer/2022070613/5b976d9b09d3f206218c8c4d/html5/thumbnails/74.jpg)
PresentationIntroduction
Regular PyramidsIrregular PyramidsSome applications
Hierarchical encodingStructural properties within pyramidsCombinatorial Pyramids
Traversal of borders
∂Rσ∗i(b) = b1, . . . , bp avec
{
b1 = b
bj = ϕnj
0 (α0(bj))
Rσ∗i(16) = 16.15.−2.14.−1.13.24.7.1.8.2.9
∂Rσ∗i(16) = 16.15.14.13.24.7.8.9
7⊲
10⊲
8⊲
11⊲
9⊲
12⊲
1△ 2△
3△ 4△
5△ 6△
13⊳ 14⊳ 15⊳16△
17△
18△
21⊲
20⊲
19⊲
24▽
23▽
22▽
47 / 51
![Page 75: Partitions et structures hiérarchiques - GREYC€¦ · Partitions et structures hi´erarchiques Luc Brun Groupe de Recherche en Informatique, Image, Automatique et Instrumentation](https://reader035.vdocuments.site/reader035/viewer/2022070613/5b976d9b09d3f206218c8c4d/html5/thumbnails/75.jpg)
PresentationIntroduction
Regular PyramidsIrregular PyramidsSome applications
Hierarchical encodingStructural properties within pyramidsCombinatorial Pyramids
Traversal of borders
◮ If we modify the operation of self loops removal :◮ SCi (16) = 16.15.−2.14.−1.13.24
→ ∂SCi (16) = 16.15.14.13.24◮ SCi (7) = 7.1 → ∂SCi (7) = 7◮ SCi (8) = 8.2.9 → ∂SCi (8) = 8.9
7⊲
10⊲
8⊲
11⊲
9⊲
12⊲
1△ 2△
3△ 4△
5△ 6△
13⊳ 14⊳ 15⊳16△
17△
18△
21⊲
20⊲
19⊲
24▽
23▽
22▽
47 / 51
![Page 76: Partitions et structures hiérarchiques - GREYC€¦ · Partitions et structures hi´erarchiques Luc Brun Groupe de Recherche en Informatique, Image, Automatique et Instrumentation](https://reader035.vdocuments.site/reader035/viewer/2022070613/5b976d9b09d3f206218c8c4d/html5/thumbnails/76.jpg)
PresentationIntroduction
Regular PyramidsIrregular PyramidsSome applications
Hierarchical encodingStructural properties within pyramidsCombinatorial Pyramids
Inside relationships
◮ Inside relationships are characterized by the loops
G = (D, σ, α) G = (D, ϕ, α)
2
6
1
-1
2
-2
-5
-4
4
6
-6
-3
3
-4
-64
-55
1-1
-2
5
3 -3
48 / 51
![Page 77: Partitions et structures hiérarchiques - GREYC€¦ · Partitions et structures hi´erarchiques Luc Brun Groupe de Recherche en Informatique, Image, Automatique et Instrumentation](https://reader035.vdocuments.site/reader035/viewer/2022070613/5b976d9b09d3f206218c8c4d/html5/thumbnails/77.jpg)
PresentationIntroduction
Regular PyramidsIrregular PyramidsSome applications
Hierarchical encodingStructural properties within pyramidsCombinatorial Pyramids
Inside relationships
◮ But the location of loops is ambigous.
2
-55
1-1
-2
2
-55
1-1
-2
σ = (2, 1,−2, 5)(−1)(−5)
48 / 51
![Page 78: Partitions et structures hiérarchiques - GREYC€¦ · Partitions et structures hi´erarchiques Luc Brun Groupe de Recherche en Informatique, Image, Automatique et Instrumentation](https://reader035.vdocuments.site/reader035/viewer/2022070613/5b976d9b09d3f206218c8c4d/html5/thumbnails/78.jpg)
PresentationIntroduction
Regular PyramidsIrregular PyramidsSome applications
Hierarchical encodingStructural properties within pyramidsCombinatorial Pyramids
Inside relationships
◮ Solution : Use the orientation
G = (D, σ, α) G = (D, ϕ, α)
2
6
1
-1
2
-2
-5
-4
4
6
-6
-3
3
-4
-64
-55
1-1
-2
5
3 -3
48 / 51
![Page 79: Partitions et structures hiérarchiques - GREYC€¦ · Partitions et structures hi´erarchiques Luc Brun Groupe de Recherche en Informatique, Image, Automatique et Instrumentation](https://reader035.vdocuments.site/reader035/viewer/2022070613/5b976d9b09d3f206218c8c4d/html5/thumbnails/79.jpg)
PresentationIntroduction
Regular PyramidsIrregular PyramidsSome applications
Hierarchical encodingStructural properties within pyramidsCombinatorial Pyramids
Complexities
◮ Coputation of one map : O (size of the borders),
◮ Computation of all maps : O(D0) ≈ O(|I |)
◮ Traversal of one border : O (size (in lignels) of the border)
◮ Inside relationships O(|σ∗
i (b)|).
49 / 51
![Page 80: Partitions et structures hiérarchiques - GREYC€¦ · Partitions et structures hi´erarchiques Luc Brun Groupe de Recherche en Informatique, Image, Automatique et Instrumentation](https://reader035.vdocuments.site/reader035/viewer/2022070613/5b976d9b09d3f206218c8c4d/html5/thumbnails/80.jpg)
PresentationIntroduction
Regular PyramidsIrregular PyramidsSome applications
Application 1 : hierarchical Watershed
Image LPE
◮ hierarchical Watershed :
1. Compute the watershed2. valuate the importance of each contour
2.1 minimal value of the gradient along the contour. . .
3. sort the contours :◮ remove the less significative ones.◮ examin the merge of regions according to the importance of
the contours.
50 / 51
![Page 81: Partitions et structures hiérarchiques - GREYC€¦ · Partitions et structures hi´erarchiques Luc Brun Groupe de Recherche en Informatique, Image, Automatique et Instrumentation](https://reader035.vdocuments.site/reader035/viewer/2022070613/5b976d9b09d3f206218c8c4d/html5/thumbnails/81.jpg)
PresentationIntroduction
Regular PyramidsIrregular PyramidsSome applications
Application 1 : hierarchical Watershed
Image LPE Dynamics
◮ Improvements :1. Taking into account the
evolution of the partition◮ implicit encoding
2. robust valuation of thecontours
◮ geometric embeddingof the darts Segmentation
50 / 51
![Page 82: Partitions et structures hiérarchiques - GREYC€¦ · Partitions et structures hi´erarchiques Luc Brun Groupe de Recherche en Informatique, Image, Automatique et Instrumentation](https://reader035.vdocuments.site/reader035/viewer/2022070613/5b976d9b09d3f206218c8c4d/html5/thumbnails/82.jpg)
PresentationIntroduction
Regular PyramidsIrregular PyramidsSome applications
Application 2 : Inside relationships
51 / 51