ICCV 2007 tutorial Part III Messagepassing algorithms for
![ICCV 2007 tutorial Part III Message-passing algorithms for energy minimization Vladimir Kolmogorov University College ICCV 2007 tutorial Part III Message-passing algorithms for energy minimization Vladimir Kolmogorov University College](https://slidetodoc.com/presentation_image_h/fac31c50397b539d78a2c9afaa3d2f68/image-1.jpg)
![Message passing p q • Iteratively pass messages between nodes. . . • Message Message passing p q • Iteratively pass messages between nodes. . . • Message](https://slidetodoc.com/presentation_image_h/fac31c50397b539d78a2c9afaa3d2f68/image-2.jpg)
![Outline • Belief propagation – BP on a tree • Min-marginals – BP in Outline • Belief propagation – BP on a tree • Min-marginals – BP in](https://slidetodoc.com/presentation_image_h/fac31c50397b539d78a2c9afaa3d2f68/image-3.jpg)
![Belief propagation (BP) Belief propagation (BP)](https://slidetodoc.com/presentation_image_h/fac31c50397b539d78a2c9afaa3d2f68/image-4.jpg)
![BP on a tree [Pearl’ 88] leaf p q leaf r root • Dynamic BP on a tree [Pearl’ 88] leaf p q leaf r root • Dynamic](https://slidetodoc.com/presentation_image_h/fac31c50397b539d78a2c9afaa3d2f68/image-5.jpg)
![Inward pass (dynamic programming) p q r Inward pass (dynamic programming) p q r](https://slidetodoc.com/presentation_image_h/fac31c50397b539d78a2c9afaa3d2f68/image-6.jpg)
![Inward pass (dynamic programming) p q r Inward pass (dynamic programming) p q r](https://slidetodoc.com/presentation_image_h/fac31c50397b539d78a2c9afaa3d2f68/image-7.jpg)
![Inward pass (dynamic programming) p q r Inward pass (dynamic programming) p q r](https://slidetodoc.com/presentation_image_h/fac31c50397b539d78a2c9afaa3d2f68/image-8.jpg)
![Inward pass (dynamic programming) p q r Inward pass (dynamic programming) p q r](https://slidetodoc.com/presentation_image_h/fac31c50397b539d78a2c9afaa3d2f68/image-9.jpg)
![Inward pass (dynamic programming) p q r Inward pass (dynamic programming) p q r](https://slidetodoc.com/presentation_image_h/fac31c50397b539d78a2c9afaa3d2f68/image-10.jpg)
![Inward pass (dynamic programming) p q r Inward pass (dynamic programming) p q r](https://slidetodoc.com/presentation_image_h/fac31c50397b539d78a2c9afaa3d2f68/image-11.jpg)
![Inward pass (dynamic programming) p q r Inward pass (dynamic programming) p q r](https://slidetodoc.com/presentation_image_h/fac31c50397b539d78a2c9afaa3d2f68/image-12.jpg)
![Outward pass p q r Outward pass p q r](https://slidetodoc.com/presentation_image_h/fac31c50397b539d78a2c9afaa3d2f68/image-13.jpg)
![BP on a tree: min-marginals p Min-marginal for node q and label j: q BP on a tree: min-marginals p Min-marginal for node q and label j: q](https://slidetodoc.com/presentation_image_h/fac31c50397b539d78a2c9afaa3d2f68/image-14.jpg)
![BP in a general graph • Pass messages using same rules – Empirically often BP in a general graph • Pass messages using same rules – Empirically often](https://slidetodoc.com/presentation_image_h/fac31c50397b539d78a2c9afaa3d2f68/image-15.jpg)
![Distance transforms [Felzenszwalb & Huttenlocher’ 04] • Naïve implementation: O(K 2) • Often can Distance transforms [Felzenszwalb & Huttenlocher’ 04] • Naïve implementation: O(K 2) • Often can](https://slidetodoc.com/presentation_image_h/fac31c50397b539d78a2c9afaa3d2f68/image-16.jpg)
![Reparameterization Reparameterization](https://slidetodoc.com/presentation_image_h/fac31c50397b539d78a2c9afaa3d2f68/image-17.jpg)
![Energy function - visualization label 0 0 0 5 4 label 1 2 node Energy function - visualization label 0 0 0 5 4 label 1 2 node](https://slidetodoc.com/presentation_image_h/fac31c50397b539d78a2c9afaa3d2f68/image-18.jpg)
![Energy function - visualization label 0 0 0 5 4 label 1 2 node Energy function - visualization label 0 0 0 5 4 label 1 2 node](https://slidetodoc.com/presentation_image_h/fac31c50397b539d78a2c9afaa3d2f68/image-19.jpg)
![Reparameterization 0 0 5 4 2 3 1 0 Reparameterization 0 0 5 4 2 3 1 0](https://slidetodoc.com/presentation_image_h/fac31c50397b539d78a2c9afaa3d2f68/image-20.jpg)
![Reparameterization 0 0 3 5 4 d 2 1 d 0 +d Reparameterization 0 0 3 5 4 d 2 1 d 0 +d](https://slidetodoc.com/presentation_image_h/fac31c50397b539d78a2c9afaa3d2f68/image-21.jpg)
![Reparameterization 0 0 4 -d 2 3 5 1 -d d • Definition. is Reparameterization 0 0 4 -d 2 3 5 1 -d d • Definition. is](https://slidetodoc.com/presentation_image_h/fac31c50397b539d78a2c9afaa3d2f68/image-22.jpg)
![BP as reparameterization [Wainwright et al. 04] • Messages define reparameterization: Mpq min-marginals (for BP as reparameterization [Wainwright et al. 04] • Messages define reparameterization: Mpq min-marginals (for](https://slidetodoc.com/presentation_image_h/fac31c50397b539d78a2c9afaa3d2f68/image-23.jpg)
![Tree-reweighted message passing (TRW) Tree-reweighted message passing (TRW)](https://slidetodoc.com/presentation_image_h/fac31c50397b539d78a2c9afaa3d2f68/image-24.jpg)
![Linear Programming relaxation • Energy minimization: NP-hard problem • Relax integrality constraint: xp {0, Linear Programming relaxation • Energy minimization: NP-hard problem • Relax integrality constraint: xp {0,](https://slidetodoc.com/presentation_image_h/fac31c50397b539d78a2c9afaa3d2f68/image-25.jpg)
![Convex combination of trees [Wainwright, Jaakkola, Willsky ’ 02] • Goal: compute minimum of Convex combination of trees [Wainwright, Jaakkola, Willsky ’ 02] • Goal: compute minimum of](https://slidetodoc.com/presentation_image_h/fac31c50397b539d78a2c9afaa3d2f68/image-26.jpg)
![Convex combination of trees (cont’d) graph maximize tree T’ lower bound on the energy Convex combination of trees (cont’d) graph maximize tree T’ lower bound on the energy](https://slidetodoc.com/presentation_image_h/fac31c50397b539d78a2c9afaa3d2f68/image-27.jpg)
![Maximizing lower bound • Subgradient methods – [Schlesinger&Giginyak’ 07], [Komodakis et al. ’ 07] Maximizing lower bound • Subgradient methods – [Schlesinger&Giginyak’ 07], [Komodakis et al. ’ 07]](https://slidetodoc.com/presentation_image_h/fac31c50397b539d78a2c9afaa3d2f68/image-28.jpg)
![TRW algorithms • Two reparameterization operations: – Ordinary BP on trees – Node averaging TRW algorithms • Two reparameterization operations: – Ordinary BP on trees – Node averaging](https://slidetodoc.com/presentation_image_h/fac31c50397b539d78a2c9afaa3d2f68/image-29.jpg)
![TRW algorithms • Two reparameterization operations: – Ordinary BP on trees – Node averaging TRW algorithms • Two reparameterization operations: – Ordinary BP on trees – Node averaging](https://slidetodoc.com/presentation_image_h/fac31c50397b539d78a2c9afaa3d2f68/image-30.jpg)
![TRW algorithms • Two reparameterization operations: – Ordinary BP on trees – Node averaging TRW algorithms • Two reparameterization operations: – Ordinary BP on trees – Node averaging](https://slidetodoc.com/presentation_image_h/fac31c50397b539d78a2c9afaa3d2f68/image-31.jpg)
![TRW algorithms • Order of operations? – Affects performance dramatically • Algorithms: – [Wainwright TRW algorithms • Order of operations? – Affects performance dramatically • Algorithms: – [Wainwright](https://slidetodoc.com/presentation_image_h/fac31c50397b539d78a2c9afaa3d2f68/image-32.jpg)
![TRW algorithm of Wainwright et al. with tree-based updates (TRW-T) Run BP on all TRW algorithm of Wainwright et al. with tree-based updates (TRW-T) Run BP on all](https://slidetodoc.com/presentation_image_h/fac31c50397b539d78a2c9afaa3d2f68/image-33.jpg)
![Sequential TRW algorithm (TRW-S) [Kolmogorov’ 05] Pick node p Run BP on all trees Sequential TRW algorithm (TRW-S) [Kolmogorov’ 05] Pick node p Run BP on all trees](https://slidetodoc.com/presentation_image_h/fac31c50397b539d78a2c9afaa3d2f68/image-34.jpg)
![Main property of TRW-S • Theorem: lower bound never decreases. • Proof sketch: 0 Main property of TRW-S • Theorem: lower bound never decreases. • Proof sketch: 0](https://slidetodoc.com/presentation_image_h/fac31c50397b539d78a2c9afaa3d2f68/image-35.jpg)
![Main property of TRW-S • Theorem: lower bound never decreases. • Proof sketch: 2 Main property of TRW-S • Theorem: lower bound never decreases. • Proof sketch: 2](https://slidetodoc.com/presentation_image_h/fac31c50397b539d78a2c9afaa3d2f68/image-36.jpg)
![TRW-S algorithm • Particular order of averaging and BP operations • Lower bound guaranteed TRW-S algorithm • Particular order of averaging and BP operations • Lower bound guaranteed](https://slidetodoc.com/presentation_image_h/fac31c50397b539d78a2c9afaa3d2f68/image-37.jpg)
![Efficient implementation Pick node p Run BP on all trees containing p “Average” node Efficient implementation Pick node p Run BP on all trees containing p “Average” node](https://slidetodoc.com/presentation_image_h/fac31c50397b539d78a2c9afaa3d2f68/image-38.jpg)
![Efficient implementation • Key observation: Node averaging operation preserves messages oriented towards this node Efficient implementation • Key observation: Node averaging operation preserves messages oriented towards this node](https://slidetodoc.com/presentation_image_h/fac31c50397b539d78a2c9afaa3d2f68/image-39.jpg)
![Efficient implementation • Algorithm: – Forward pass: 1 2 3 4 5 6 7 Efficient implementation • Algorithm: – Forward pass: 1 2 3 4 5 6 7](https://slidetodoc.com/presentation_image_h/fac31c50397b539d78a2c9afaa3d2f68/image-40.jpg)
![Efficient implementation • Algorithm: – Forward pass: 1 2 3 4 5 6 7 Efficient implementation • Algorithm: – Forward pass: 1 2 3 4 5 6 7](https://slidetodoc.com/presentation_image_h/fac31c50397b539d78a2c9afaa3d2f68/image-41.jpg)
![Memory requirements • Standard message passing: 2 messages per edge • TRW-S: 1 message Memory requirements • Standard message passing: 2 messages per edge • TRW-S: 1 message](https://slidetodoc.com/presentation_image_h/fac31c50397b539d78a2c9afaa3d2f68/image-42.jpg)
![Experimental results: stereo TRW-E TRW-T left image BP ground truth TRW-S • Global minima Experimental results: stereo TRW-E TRW-T left image BP ground truth TRW-S • Global minima](https://slidetodoc.com/presentation_image_h/fac31c50397b539d78a2c9afaa3d2f68/image-43.jpg)
![Conclusions • BP – Exact on trees • Gives min-marginals (unlike dynamic programming) – Conclusions • BP – Exact on trees • Gives min-marginals (unlike dynamic programming) –](https://slidetodoc.com/presentation_image_h/fac31c50397b539d78a2c9afaa3d2f68/image-44.jpg)
- Slides: 44
![ICCV 2007 tutorial Part III Messagepassing algorithms for energy minimization Vladimir Kolmogorov University College ICCV 2007 tutorial Part III Message-passing algorithms for energy minimization Vladimir Kolmogorov University College](https://slidetodoc.com/presentation_image_h/fac31c50397b539d78a2c9afaa3d2f68/image-1.jpg)
ICCV 2007 tutorial Part III Message-passing algorithms for energy minimization Vladimir Kolmogorov University College London
![Message passing p q Iteratively pass messages between nodes Message Message passing p q • Iteratively pass messages between nodes. . . • Message](https://slidetodoc.com/presentation_image_h/fac31c50397b539d78a2c9afaa3d2f68/image-2.jpg)
Message passing p q • Iteratively pass messages between nodes. . . • Message update rule? – Belief propagation (BP) – Tree-reweighted belief propagation (TRW) – max-product (minimizing an energy function, or MAP estimation) – sum-product (computing marginal probabilities) • Schedule? – Parallel, sequential, . . .
![Outline Belief propagation BP on a tree Minmarginals BP in Outline • Belief propagation – BP on a tree • Min-marginals – BP in](https://slidetodoc.com/presentation_image_h/fac31c50397b539d78a2c9afaa3d2f68/image-3.jpg)
Outline • Belief propagation – BP on a tree • Min-marginals – BP in a general graph – Distance transforms • Reparameterization • Tree-reweighted message passing – Lower bound via combination of trees – Message passing – Sequential TRW
![Belief propagation BP Belief propagation (BP)](https://slidetodoc.com/presentation_image_h/fac31c50397b539d78a2c9afaa3d2f68/image-4.jpg)
Belief propagation (BP)
![BP on a tree Pearl 88 leaf p q leaf r root Dynamic BP on a tree [Pearl’ 88] leaf p q leaf r root • Dynamic](https://slidetodoc.com/presentation_image_h/fac31c50397b539d78a2c9afaa3d2f68/image-5.jpg)
BP on a tree [Pearl’ 88] leaf p q leaf r root • Dynamic programming: global minimum in linear time • BP: – Inward pass (dynamic programming) – Outward pass – Gives min-marginals
![Inward pass dynamic programming p q r Inward pass (dynamic programming) p q r](https://slidetodoc.com/presentation_image_h/fac31c50397b539d78a2c9afaa3d2f68/image-6.jpg)
Inward pass (dynamic programming) p q r
![Inward pass dynamic programming p q r Inward pass (dynamic programming) p q r](https://slidetodoc.com/presentation_image_h/fac31c50397b539d78a2c9afaa3d2f68/image-7.jpg)
Inward pass (dynamic programming) p q r
![Inward pass dynamic programming p q r Inward pass (dynamic programming) p q r](https://slidetodoc.com/presentation_image_h/fac31c50397b539d78a2c9afaa3d2f68/image-8.jpg)
Inward pass (dynamic programming) p q r
![Inward pass dynamic programming p q r Inward pass (dynamic programming) p q r](https://slidetodoc.com/presentation_image_h/fac31c50397b539d78a2c9afaa3d2f68/image-9.jpg)
Inward pass (dynamic programming) p q r
![Inward pass dynamic programming p q r Inward pass (dynamic programming) p q r](https://slidetodoc.com/presentation_image_h/fac31c50397b539d78a2c9afaa3d2f68/image-10.jpg)
Inward pass (dynamic programming) p q r
![Inward pass dynamic programming p q r Inward pass (dynamic programming) p q r](https://slidetodoc.com/presentation_image_h/fac31c50397b539d78a2c9afaa3d2f68/image-11.jpg)
Inward pass (dynamic programming) p q r
![Inward pass dynamic programming p q r Inward pass (dynamic programming) p q r](https://slidetodoc.com/presentation_image_h/fac31c50397b539d78a2c9afaa3d2f68/image-12.jpg)
Inward pass (dynamic programming) p q r
![Outward pass p q r Outward pass p q r](https://slidetodoc.com/presentation_image_h/fac31c50397b539d78a2c9afaa3d2f68/image-13.jpg)
Outward pass p q r
![BP on a tree minmarginals p Minmarginal for node q and label j q BP on a tree: min-marginals p Min-marginal for node q and label j: q](https://slidetodoc.com/presentation_image_h/fac31c50397b539d78a2c9afaa3d2f68/image-14.jpg)
BP on a tree: min-marginals p Min-marginal for node q and label j: q r
![BP in a general graph Pass messages using same rules Empirically often BP in a general graph • Pass messages using same rules – Empirically often](https://slidetodoc.com/presentation_image_h/fac31c50397b539d78a2c9afaa3d2f68/image-15.jpg)
BP in a general graph • Pass messages using same rules – Empirically often works quite well • May not converge • “Pseudo” min-marginals • Gives local minimum in the “tree neighborhood” [Weiss&Freeman’ 01], [Wainwright et al. ’ 04] – Assumptions: • BP has converged • no ties in pseudo min-marginals
![Distance transforms Felzenszwalb Huttenlocher 04 Naïve implementation OK 2 Often can Distance transforms [Felzenszwalb & Huttenlocher’ 04] • Naïve implementation: O(K 2) • Often can](https://slidetodoc.com/presentation_image_h/fac31c50397b539d78a2c9afaa3d2f68/image-16.jpg)
Distance transforms [Felzenszwalb & Huttenlocher’ 04] • Naïve implementation: O(K 2) • Often can be improved to O(K) – Potts interactions, truncated linear, truncated quadratic, . . .
![Reparameterization Reparameterization](https://slidetodoc.com/presentation_image_h/fac31c50397b539d78a2c9afaa3d2f68/image-17.jpg)
Reparameterization
![Energy function visualization label 0 0 0 5 4 label 1 2 node Energy function - visualization label 0 0 0 5 4 label 1 2 node](https://slidetodoc.com/presentation_image_h/fac31c50397b539d78a2c9afaa3d2f68/image-18.jpg)
Energy function - visualization label 0 0 0 5 4 label 1 2 node p 3 1 edge (p, q) vector of all parameters 0 node q
![Energy function visualization label 0 0 0 5 4 label 1 2 node Energy function - visualization label 0 0 0 5 4 label 1 2 node](https://slidetodoc.com/presentation_image_h/fac31c50397b539d78a2c9afaa3d2f68/image-19.jpg)
Energy function - visualization label 0 0 0 5 4 label 1 2 node p 3 1 edge (p, q) vector of all parameters 0 node q
![Reparameterization 0 0 5 4 2 3 1 0 Reparameterization 0 0 5 4 2 3 1 0](https://slidetodoc.com/presentation_image_h/fac31c50397b539d78a2c9afaa3d2f68/image-20.jpg)
Reparameterization 0 0 5 4 2 3 1 0
![Reparameterization 0 0 3 5 4 d 2 1 d 0 d Reparameterization 0 0 3 5 4 d 2 1 d 0 +d](https://slidetodoc.com/presentation_image_h/fac31c50397b539d78a2c9afaa3d2f68/image-21.jpg)
Reparameterization 0 0 3 5 4 d 2 1 d 0 +d
![Reparameterization 0 0 4 d 2 3 5 1 d d Definition is Reparameterization 0 0 4 -d 2 3 5 1 -d d • Definition. is](https://slidetodoc.com/presentation_image_h/fac31c50397b539d78a2c9afaa3d2f68/image-22.jpg)
Reparameterization 0 0 4 -d 2 3 5 1 -d d • Definition. is a reparameterization of if they define the same energy: • Maxflow, BP and TRW perform reparameterisations
![BP as reparameterization Wainwright et al 04 Messages define reparameterization Mpq minmarginals for BP as reparameterization [Wainwright et al. 04] • Messages define reparameterization: Mpq min-marginals (for](https://slidetodoc.com/presentation_image_h/fac31c50397b539d78a2c9afaa3d2f68/image-23.jpg)
BP as reparameterization [Wainwright et al. 04] • Messages define reparameterization: Mpq min-marginals (for trees) j d +d d = Mpq( j ) • BP on a tree: reparameterize energy so that unary potentials become min-marginals
![Treereweighted message passing TRW Tree-reweighted message passing (TRW)](https://slidetodoc.com/presentation_image_h/fac31c50397b539d78a2c9afaa3d2f68/image-24.jpg)
Tree-reweighted message passing (TRW)
![Linear Programming relaxation Energy minimization NPhard problem Relax integrality constraint xp 0 Linear Programming relaxation • Energy minimization: NP-hard problem • Relax integrality constraint: xp {0,](https://slidetodoc.com/presentation_image_h/fac31c50397b539d78a2c9afaa3d2f68/image-25.jpg)
Linear Programming relaxation • Energy minimization: NP-hard problem • Relax integrality constraint: xp {0, 1} xp [0, 1] – LP relaxation [Schlesinger’ 76, Koster et al. ’ 98, Chekuri et al. ’ 00, Wainwright et al. ’ 03] • Try to solve dual problem: – Formulate lower bound on the function – Maximize the bound E Energy function with discrete variables E E LP relaxation Lower bound on the energy function
![Convex combination of trees Wainwright Jaakkola Willsky 02 Goal compute minimum of Convex combination of trees [Wainwright, Jaakkola, Willsky ’ 02] • Goal: compute minimum of](https://slidetodoc.com/presentation_image_h/fac31c50397b539d78a2c9afaa3d2f68/image-26.jpg)
Convex combination of trees [Wainwright, Jaakkola, Willsky ’ 02] • Goal: compute minimum of the energy for q : • Obtaining lower bound: – Split q into several components: q = q 1 + q 2 +. . . – Compute minimum for each component: – Combine F(q 1), F(q 2), . . . to get a bound on F(q) • Use trees!
![Convex combination of trees contd graph maximize tree T lower bound on the energy Convex combination of trees (cont’d) graph maximize tree T’ lower bound on the energy](https://slidetodoc.com/presentation_image_h/fac31c50397b539d78a2c9afaa3d2f68/image-27.jpg)
Convex combination of trees (cont’d) graph maximize tree T’ lower bound on the energy
![Maximizing lower bound Subgradient methods SchlesingerGiginyak 07 Komodakis et al 07 Maximizing lower bound • Subgradient methods – [Schlesinger&Giginyak’ 07], [Komodakis et al. ’ 07]](https://slidetodoc.com/presentation_image_h/fac31c50397b539d78a2c9afaa3d2f68/image-28.jpg)
Maximizing lower bound • Subgradient methods – [Schlesinger&Giginyak’ 07], [Komodakis et al. ’ 07] • Tree-reweighted message passing (TRW) – [Wainwright et al. ’ 02], [Kolmogorov’ 05]
![TRW algorithms Two reparameterization operations Ordinary BP on trees Node averaging TRW algorithms • Two reparameterization operations: – Ordinary BP on trees – Node averaging](https://slidetodoc.com/presentation_image_h/fac31c50397b539d78a2c9afaa3d2f68/image-29.jpg)
TRW algorithms • Two reparameterization operations: – Ordinary BP on trees – Node averaging
![TRW algorithms Two reparameterization operations Ordinary BP on trees Node averaging TRW algorithms • Two reparameterization operations: – Ordinary BP on trees – Node averaging](https://slidetodoc.com/presentation_image_h/fac31c50397b539d78a2c9afaa3d2f68/image-30.jpg)
TRW algorithms • Two reparameterization operations: – Ordinary BP on trees – Node averaging 0 4 1 0
![TRW algorithms Two reparameterization operations Ordinary BP on trees Node averaging TRW algorithms • Two reparameterization operations: – Ordinary BP on trees – Node averaging](https://slidetodoc.com/presentation_image_h/fac31c50397b539d78a2c9afaa3d2f68/image-31.jpg)
TRW algorithms • Two reparameterization operations: – Ordinary BP on trees – Node averaging 2 0. 5
![TRW algorithms Order of operations Affects performance dramatically Algorithms Wainwright TRW algorithms • Order of operations? – Affects performance dramatically • Algorithms: – [Wainwright](https://slidetodoc.com/presentation_image_h/fac31c50397b539d78a2c9afaa3d2f68/image-32.jpg)
TRW algorithms • Order of operations? – Affects performance dramatically • Algorithms: – [Wainwright et al. ’ 02]: parallel schedule (TRW-E, TRW-T) • May not converge – [Kolmogorov’ 05]: specific sequential schedule (TRW-S) • Lower bound does not decrease, convergence guarantees • Needs half the memory
![TRW algorithm of Wainwright et al with treebased updates TRWT Run BP on all TRW algorithm of Wainwright et al. with tree-based updates (TRW-T) Run BP on all](https://slidetodoc.com/presentation_image_h/fac31c50397b539d78a2c9afaa3d2f68/image-33.jpg)
TRW algorithm of Wainwright et al. with tree-based updates (TRW-T) Run BP on all trees “Average” all nodes • If converges, gives (local) maximum of lower bound • Not guaranteed to converge. • Lower bound may go down.
![Sequential TRW algorithm TRWS Kolmogorov 05 Pick node p Run BP on all trees Sequential TRW algorithm (TRW-S) [Kolmogorov’ 05] Pick node p Run BP on all trees](https://slidetodoc.com/presentation_image_h/fac31c50397b539d78a2c9afaa3d2f68/image-34.jpg)
Sequential TRW algorithm (TRW-S) [Kolmogorov’ 05] Pick node p Run BP on all trees containing p “Average” node p
![Main property of TRWS Theorem lower bound never decreases Proof sketch 0 Main property of TRW-S • Theorem: lower bound never decreases. • Proof sketch: 0](https://slidetodoc.com/presentation_image_h/fac31c50397b539d78a2c9afaa3d2f68/image-35.jpg)
Main property of TRW-S • Theorem: lower bound never decreases. • Proof sketch: 0 4 1 0
![Main property of TRWS Theorem lower bound never decreases Proof sketch 2 Main property of TRW-S • Theorem: lower bound never decreases. • Proof sketch: 2](https://slidetodoc.com/presentation_image_h/fac31c50397b539d78a2c9afaa3d2f68/image-36.jpg)
Main property of TRW-S • Theorem: lower bound never decreases. • Proof sketch: 2 0. 5
![TRWS algorithm Particular order of averaging and BP operations Lower bound guaranteed TRW-S algorithm • Particular order of averaging and BP operations • Lower bound guaranteed](https://slidetodoc.com/presentation_image_h/fac31c50397b539d78a2c9afaa3d2f68/image-37.jpg)
TRW-S algorithm • Particular order of averaging and BP operations • Lower bound guaranteed not to decrease • There exists limit point that satisfies weak tree agreement condition • Efficiency?
![Efficient implementation Pick node p Run BP on all trees containing p Average node Efficient implementation Pick node p Run BP on all trees containing p “Average” node](https://slidetodoc.com/presentation_image_h/fac31c50397b539d78a2c9afaa3d2f68/image-38.jpg)
Efficient implementation Pick node p Run BP on all trees containing p “Average” node p inefficient?
![Efficient implementation Key observation Node averaging operation preserves messages oriented towards this node Efficient implementation • Key observation: Node averaging operation preserves messages oriented towards this node](https://slidetodoc.com/presentation_image_h/fac31c50397b539d78a2c9afaa3d2f68/image-39.jpg)
Efficient implementation • Key observation: Node averaging operation preserves messages oriented towards this node • Reuse previously passed messages! • Need a special choice of trees: – Pick an ordering of nodes – Trees: monotonic chains 1 2 3 4 5 6 7 8 9
![Efficient implementation Algorithm Forward pass 1 2 3 4 5 6 7 Efficient implementation • Algorithm: – Forward pass: 1 2 3 4 5 6 7](https://slidetodoc.com/presentation_image_h/fac31c50397b539d78a2c9afaa3d2f68/image-40.jpg)
Efficient implementation • Algorithm: – Forward pass: 1 2 3 4 5 6 7 8 9 • process nodes in the increasing order • pass messages from lower neighbours – Backward pass: • do the same in reverse order • Linear running time of one iteration
![Efficient implementation Algorithm Forward pass 1 2 3 4 5 6 7 Efficient implementation • Algorithm: – Forward pass: 1 2 3 4 5 6 7](https://slidetodoc.com/presentation_image_h/fac31c50397b539d78a2c9afaa3d2f68/image-41.jpg)
Efficient implementation • Algorithm: – Forward pass: 1 2 3 4 5 6 7 8 9 • process nodes in the increasing order • pass messages from lower neighbours – Backward pass: • do the same in reverse order • Linear running time of one iteration
![Memory requirements Standard message passing 2 messages per edge TRWS 1 message Memory requirements • Standard message passing: 2 messages per edge • TRW-S: 1 message](https://slidetodoc.com/presentation_image_h/fac31c50397b539d78a2c9afaa3d2f68/image-42.jpg)
Memory requirements • Standard message passing: 2 messages per edge • TRW-S: 1 message per edge – Similar observation for bipartite graphs and parallel schedule in [Felzenszwalb&Huttenlocher’ 04] standard message passing TRW-S
![Experimental results stereo TRWE TRWT left image BP ground truth TRWS Global minima Experimental results: stereo TRW-E TRW-T left image BP ground truth TRW-S • Global minima](https://slidetodoc.com/presentation_image_h/fac31c50397b539d78a2c9afaa3d2f68/image-43.jpg)
Experimental results: stereo TRW-E TRW-T left image BP ground truth TRW-S • Global minima for some instances with TRW [Meltzer, Yanover, Weiss’ 05] • See evaluation of MRF algorithms [Szeliski et al. ’ 07]
![Conclusions BP Exact on trees Gives minmarginals unlike dynamic programming Conclusions • BP – Exact on trees • Gives min-marginals (unlike dynamic programming) –](https://slidetodoc.com/presentation_image_h/fac31c50397b539d78a2c9afaa3d2f68/image-44.jpg)
Conclusions • BP – Exact on trees • Gives min-marginals (unlike dynamic programming) – If there are cycles, heuristic – Can be viewed as reparameterization • TRW – Tries to maximize a lower bound – TRW-S: • lower bound never decreases • limit point - weak tree agreement • efficient with monotonic chains – Not guaranteed to find an optimal bound! • See subgradient techniques [Schlesinger&Giginyak’ 07], [Komodakis et al. ’ 07]
Iccv area chair
Cvpr poster template
Iccv 2009
Iccv review
Cvpr buzz
Hamlet act iii scene ii
Dsp algorithms tutorial
Data structures and algorithms tutorial
Genetic algorithms tutorial
Ms project 2007 training
Ms project tutorial
Wavewatch iii tutorial
Hendrik tolman
Va handbook 5017 part iii appendix a
Iso 22301 utbildning
Novell typiska drag
Tack för att ni lyssnade bild
Vad står k.r.å.k.a.n för
Varför kallas perioden 1918-1939 för mellankrigstiden?
En lathund för arbete med kontinuitetshantering
Adressändring ideell förening
Tidbok yrkesförare
Sura för anatom
Densitet vatten
Datorkunskap för nybörjare
Stig kerman
Hur skriver man en tes
Magnetsjukhus
Nyckelkompetenser för livslångt lärande
Påbyggnader för flakfordon
Lufttryck formel
Publik sektor
Kyssande vind analys
Presentera för publik crossboss
Argument för teckenspråk som minoritetsspråk
Plats för toran ark
Klassificeringsstruktur för kommunala verksamheter
Luftstrupen för medicinare
Claes martinsson
Cks
Programskede byggprocessen
Mat för unga idrottare
Verktyg för automatisering av utbetalningar
Rutin för avvikelsehantering
Smärtskolan kunskap för livet