Exhaustive family of energies minimizable exactly by a graph cut
read more
Citations
Learning with Submodular Functions: A Convex Optimization Perspective
Learning with Submodular Functions: A Convex Optimization Perspective
Minimisation du risque empirique avec des fonctions de perte nonmodulaires
Efficient learning for discriminative segmentation with supermodular losses
An Efficient Decomposition Framework for Discriminative Segmentation with Supermodular Losses.
References
Fast approximate energy minimization via graph cuts
An experimental comparison of min-cut/max- flow algorithms for energy minimization in vision
Fast approximate energy minimization via graph cuts
An Experimental Comparison of Min-cut/Max-flow Algorithms for Energy Minimization in Vision
What energy functions can be minimized via graph cuts
Related Papers (5)
Frequently Asked Questions (12)
Q2. What is the simplest way to represent the energy of binary variables?
Since α1,1 has to be positive, any energy of binary variables which is in F0 has to satisfy that there exists a label ordering for which V 0i,j is submodular for all (i, j) simultaneously.
Q3. What is the utEnergie of the family F0?
utEnergies of the family F0 are written in (1) as linear combinations of quantities bki , b k i b k′ i and b k i b l j , which can be seen as vectors or matrices as functions of the subgraph states representing xi and xj .
Q4. What is the definition of a graph with a finite number of nodes?
If the authors restrict the study to the case of graphs with a finite number of nodes, then the number of possible cuts of a graph is finite, so the number of possible states of all variables (xi)16i6p has to be finite.
Q5. What is the simplest way to represent the canonized potentials?
If a subgraph contains only #Si = 1 node, then A1 =( −1 1 ) and since α1 ∈ R, all canonized potentials D0i of abinary variable are representable.
Q6. What is the complexity of the MRF energy?
The complexity of testing whether an MRF energy belongs to F0, and, in that case, of finding an ordering of labels (Ii)i so that it can be expressed as in Proposition 3, is linear as a function of the data size almost surely.
Q7. What is the smallest possible dimension of the space of canonized potentials?
The space of all canonized interaction matrices M (for any MRF, not necessarily in F0) has only one degree of freedom, so that any interaction matrix M writes necessarily as αW 1,1 for an α ∈ R.
Q8. What is the simplest way to represent the space of canonized potentials?
In the case of #Si = 2 nodes :A1i = −1 −1 1 1 , A2i = −1 1 −1 1 , A1,2i = −1 1 1 −1 and since the dimension of the space of canonized potentials is 3, with α1, α2 ∈ R and α1,2 ∈ R+, half of them are representable for any fixed interpretation.
Q9. What is the family of energies minimizable by graph cuts?
The exhaustive family F of all energies minimizable by a graph cut is the set of functions which either belong to the former family F0, either can be written as an infimum of an energy in F0 with respect to some of its variables, globally (i.e. over all possible labels of these variables : minxsupp ) or partially (i.e. over a subset of the possible labels of these variables : minxi∈Xri ⊂Xi ).
Q10. What is the canonical form of E?
The canonical form of E is then : D0i = qi∑ k=1 αki A k i + qi∑ k=1 qi∑ k′ = 1 k′ 6= k αk,k ′ i A k,k′ i V 0i,j = qi∑k=1 qj∑ l=1 αk,li,j W k,l i,j(4)where αki = 1 2 (wsource ki − wk sinki ) + 1 4 ∑ k′ (wk k ′ i − wk k ′ i ) + 1 4 ∑ j (wk k ′ i,j − wk k ′ i,j ) αk,k ′ i = 1 4 (wk k ′ i + w k k′ i )αk,li,j = 1 4 (wk li,j + w k l i,j )can be seen as parameters of the energy E, i.e. as degrees of freedom in the design of energies within F0.
Q11. What is the simplest way to write an energy in F0?
The writing of an energy in F0 as in Proposition 3 is almost surely unique, i.e. there exists a.s. only one unique valid interpretation (= label orderings), up to node numbering and global reversion of all orderings.
Q12. What is the matrices available in F0 for the interaction between a subgraph?
The matrices available in F0 for the interaction between a subgraph of 1 node and one of 2 nodes are :( −1 −1 1 1 1 1 −1 −1 ) , ( −1 1 −1 1 1 −1 1 −1 ) .