Barcelona June 2013 1
• The hierarchical cuts theory
• Climbing energies and optimal cuts
Jean Serra B. Ravi Kiran
LIMG ESIEE Université Paris-Est
UPC / Pompeu Fabra Barcelona, June 13 2013
Barcelona June 2013 2
The goal
Barcelona June 2013 3
Unicity problem
Barcelona June 2013 4
How to get out ?
Barcelona June 2013 5
Plan
1. Hierarchies
2. Singular energies and lattices
3. Optimal cuts and hierarchical increasingness
4. Compositions of energy by sums and by suprema
5. Climbing families of energies.
Barcelona June 2013 6
Hierarchy, or pyramid, of partitions
Barcelona June 2013 7
Hierarchy, or pyramid, of partitions
Barcelona June 2013 8
Representation of a hierarchy
a d g
b e h
c f i
A B C E
a b c d e f g h i
A B C
E
S C
S C
Barcelona June 2013 9
Energy and pyramid
Barcelona June 2013 10
Plan
1. Hierarchies
2. Singular energies and lattices
3. Optimal cuts and hierarchical increasingness
4. Compositions of energy by sums and by suprema
5. Climbing families of energies.
Barcelona June 2013 11
Energetic ordering on cuts
Barcelona June 2013 12
Energetic ordering and singular energy
….. ….. > or
<
Barcelona June 2013 13
Energetic Lattice
Barcelona June 2013 14
Energetic Lattice
Barcelona June 2013 15
Plan
1. Hierarchies
2. Singular energies and lattices
3. Optimal cuts and hierarchical increasingness
4. Compositions of energy by sums and by suprema
5. Climbing families of energies.
Barcelona June 2013 16
Hierarchical increasingness
Barcelona June 2013 17
Climbing energies
Barcelona June 2013 18
Hierarchical increasingness
Barcelona June 2013 19
Algorithms
Barcelona June 2013 20
Plan
1. Hierarchies
2. Singular energies and lattices
3. Optimal cuts and hierarchical increasingness
4. Compositions of energy by sums and by suprema
5. Climbing families of energies.
21 Kiran-Serra, Barcelona, June 2013
How to construct a climbing energy?
T3
T2
T1
S
Barcelona June 2013 22
How to construct a climbing energy?
Barcelona June 2013 23
Sum generated energies (Salembier-Guigues)
3 3 2 4 3 4 12 4 5
10 2 5 4 12 8
9 5 20
35
Barcelona June 2013 24
Sum generated energies (Salembier-Guigues)
3 3 2 4 3 4 12 4 5
10 2 5 4 12 8
9 5 20
35
Barcelona June 2013 25
An example : Mumford-Shah
Barcelona June 2013 26
T3
T2
T1
S
Optimal Cut: Luminance
Barcelona June 2013 27
• c
• cp
• cd
• g • r
• b
Luminance axis
Chromatic
plane
Luminance-Chrominance
Barcelona June 2013 28
Optimal Cut: Luminance
Barcelona June 2013 29
T3
T2
T1
S
Another example: color and texture
Barcelona June 2013 30
S
Partition with min variation
in component sizes
Another example: color and texture
Barcelona June 2013 31
Another example: color and texture
Barcelona June 2013 32
Another example: color and texture
Barcelona June 2013 33
Composition of additive energies
Barcelona June 2013 34
Barcelona June 2013 35
Additive energies and graph-cuts
1 1 4 2 3
3 3
8 T
S
Barcelona June 2013 36
Min-cut versus optimal cut
1 1 4 2 3
3 3
8 T
S
0 1 4 2 3
2 3
7
S
T
Barcelona June 2013 37
Min-cut versus optimal cut
1 1 4 2 3
3 3
8 T
S
0 0 4 2 3
1 3
6
S
T
Barcelona June 2013 38
Min-cut versus optimal cut
1 1 4 2 3
3 3
8 T
S
0 0 4 0 2
1 0
3
S
T
Barcelona June 2013 39
Min-cut versus optimal cut
3 3 2 4 3 4 12 4 5
10 2 5 4 12 8
22 5 21
35
Suprema generated energies (Soille-Grazzini)
Barcelona June 2013 40
3 3 2 4 3 4 12 4 5
10 2 5 4 12 8
22 5 21
35
Suprema generated energies (Soille-Grazzini)
Barcelona June 2013 41
V-generated energies (Akçay-Aksoy)
Barcelona June 2013 42
3 3 2 4 3 4 21 4 5
10 2 5 4 21 3
9 5 20
15
V-generated energies (Akçay-Aksoy)
Barcelona June 2013 43
3 3 2 4 3 4 21 4 5
10 2 5 4 21 3
9 5 20
15
Barcelona June 2013 44
Infima generated energies Ground truth Evaluation
Barcelona June 2013 45
Composition of V-generated energies
Plan
1. Hierarchies
2. Singular energies and lattices
3. Optimal cuts and hierarchical increasingness
4. Compositions of energy by sums and by suprema
5. Climbing families of energies.
Barcelona June 2013 46
Climbing families of energies
Barcelona June 2013 47
Barcelona June 2013 50
Another Example
Barcelona June 2013 51
Another Example
Barcelona June 2013 52
Another Example
• We replaced the numerical approach
by the lattice one
which adds a local meaning to the global energy w, (similar to
the uniform convergence versus the simple one).
energy +
singularity
Lattice of the cuts
optimal partition
energy optimal
partition
Barcelona June 2013 53
Conclusions
• We replaced the variational approach by the axiomatics
which allows the fast computation
• We introduced the climbing families of energies
Which results in
Singular and h-increasing energy = climbing energy
Climbing families
of energies
Hierarchies of
optimal partitions
climbing energy optimal cut in one pass
Barcelona June 2013 54
Conclusions
…. A study by my student Barcelona June 2013 55
…. A study by my student, and me. Barcelona June 2013 56
Thanks
For your
attention !
Barcelona June 2013 57
Barcelona June 2013 1
II- Ground truth energies and Saliency Transform
B. Ravi Kiran Jean Serra
LIMG ESIEE Université Paris-Est
Barcelona June 2013 2
UPC / Pompeu Fabra Barcelona, June 13 2013
Problem context
Barcelona June 2013 3
A Problem: Inputs
Barcelona June 2013 4
Ground truth: Evaluation of Hierarchies
Barcelona June 2013 5
Problems
1. Given a hierarchy H and ground truth partition G find the partition in H closest to G.
1. Closest from H -> G
2. Closest from G -> H
2. Compare any hierarchy H with multiple ground truth partitions of the same image
3. Compare any two hierarchies H1, H2, with respect to a common ground truth partition G
6 Barcelona June 2013
Hausdorff distance and associated problems
Y
X
i.e. smallest disc dilation of X that contains Y and of X to contain Y
- Global Measure - Large variations when object are asymmetric w.r.t each other
7 Barcelona June 2013
Local Hausdorff distances
• Local measures: Each class S in H is assigned 2 radii: ,
• Both are h-increasing energies
• Local optimization to obtain a globally optimal solution
8
minimum radius of dilation of ground truth contour that covers the contour of S.
minimum radius of dilation of the contour of S to cover GT within S. Barcelona June 2013
Example
Barcelona June 2013 9
Energy at various levels of H
10 Barcelona June 2013
Energy at various levels of H
11 Barcelona June 2013
Problems
1. Given a hierarchy H and ground truth partition G find the partition in H closest to G.
1. Closest from H -> G
2. Closest from G -> H
2. Compare any hierarchy H with multiple ground truth partitions of the same image
3. Compare any two hierarchies H1, H2, with respect to a common ground truth partition G
12 Barcelona June 2013
Optimal Cuts
Initial Image
GT2 13 Barcelona June 2013
Optimal Cuts
Initial Image
GT7 14 Barcelona June 2013
Problems
1. Given a hierarchy H and ground truth partition G find the partition in H closest to G.
1. Closest from H -> G
2. Closest from G -> H
2. Compare any hierarchy H with multiple ground truth partitions of the same image
3. Compare any two hierarchies H1, H2, with respect to a common ground truth partition G
15 Barcelona June 2013
Composition of ground truths:
GT5_1 GT5_2 GT5 = GT5_1 + GT5_2
The distance function of the union(sum) is the inf of the distance functions 16 Barcelona June 2013
Composition of ground truths:
GT5_1 GT5_2 GT5 = GT5_1 + GT5_2
17 Barcelona June 2013
Problems
1. Given a hierarchy H and ground truth partition G find the partition in H closest to G.
1. Closest from H -> G
2. Closest from G -> H
2. Compare any hierarchy H with multiple ground truth partitions of the same image
3. Compare any two hierarchies H1, H2, with respect to a common ground truth partition G
18 Barcelona June 2013
Global Precision-Recall Energies
Local dissimilarity measure Counterpart Global similarity measures
19
The two half distances yield two local and then two global energies: • Precision (P) : How close is on average the ground truth to the class (G->S) • Recall (R) : How close is on average the Class contour to the Ground truth (S->G)
Barcelona June 2013
Comparing Hierarchies (saliencies) with Precision-recall similarity measures
UCM Cousty (floodings of watershed)
UCM random hierarchy Cousty random hierarchy
GT6 GT5 GT4 GT3 GT2 GT7 GT1 20 Barcelona June 2013
Comparing Hierarchies (saliencies) with Precision-recall similarity measures
Image 25098
UCM UCM random
Cousty Cousty random
Precision energy
4.4 0.27 0.13 0.09
Recall energy
3.9 0.28 0.16 0.10
Integrals from PR equations expressed per 1000 pixels in the image
21 Barcelona June 2013
Problem context
Barcelona June 2013 22
Ground truth: Evaluation of Hierarchies
Barcelona June 2013 23
A problem: Transforming hierarchies
Barcelona June 2013 24
Representations of Hierarchies: Saliency function
Barcelona June 2013 25
Saliency Ultrametric contour Map (UCM)
Introducing an external function
Barcelona June 2013 26
Introducing an external function
Barcelona June 2013 27
Introducing an external function
Barcelona June 2013 28
Binary Class Opening
Barcelona June 2013 29
Numerical (grayscale) class opening
Barcelona June 2013 30
Example: Class Opening
Barcelona June 2013 31
Properties of class opening I
Barcelona June 2013 32
Properties of class opening II
Barcelona June 2013 33
Saliency degeneracy
Barcelona June 2013 34
1 2 3 4 5 6 7 8 9 100
2
4
6
8
10
12
14
16
18
20
linear
power law x > 1 (compression)
power law x 0 < < 1(expansion)
Logarithmic
Evaluating Hierarchies w.r.t Ground Truth
Barcelona June 2013 35
Evaluating Hierarchies w.r.t Ground Truth
Barcelona June 2013 36
Barcelona June 2013 37
Composing two external functions
Barcelona June 2013 38
Composing two external functions
Barcelona June 2013 39
Generating Random Hierarchies
Barcelona June 2013 40
Hierarchy Fusion: Matching hierarchies
Barcelona June 2013 41
Hierarchy Fusion: Matching hierarchies
Barcelona June 2013 42
Hierarchy Fusion: Matching hierarchies
Barcelona June 2013 43
Hierarchy Fusion: Matching hierarchies
Barcelona June 2013 44
Conclusion
Barcelona June 2013 45
http://www.esiee.fr/~kiranr/HierarchEvalGT.html
Future work
Barcelona June 2013 46
47 Barcelona June 2013