so far

21
So far ... act methods for submodular energies roximations for non-submodular energie ve-making ( N_Variables >> N_Labels)

Upload: trinh

Post on 23-Feb-2016

66 views

Category:

Documents


0 download

DESCRIPTION

So far . Exact methods for submodular energies. Approximations for non- submodular energies. Move-making ( N_Variables >> N_Labels ). Inference for Learning Linear Programming Relaxation. Linear Integer Programming. min x g 0 T x. Linear function. s.t. g i T x ≤ 0. - PowerPoint PPT Presentation

TRANSCRIPT

Page 1: So far

So far ...

Exact methods for submodular energies

Approximations for non-submodular energies

Move-making ( N_Variables >> N_Labels)

Page 2: So far
Page 3: So far

Inference for Learning

Linear Programming Relaxation

Page 4: So far

Linear Integer Programming

minx g0Tx

s.t. giTx ≤ 0

hiTx = 0

Linear function

Linear constraints

Linear constraints

x is a vector of integers

For example, x {0,1}N

Hard to solve !!

Page 5: So far

Linear Programming

minx g0Tx

s.t. giTx ≤ 0

hiTx = 0

Linear function

Linear constraints

Linear constraints

x is a vector of reals

Easy to solve!!

For example, x [0,1]N

Relaxation

Page 6: So far

Roadmap

Express MAP as an integer program

Relax to a linear program and solve

Round fractional solution to integers

Page 7: So far

2

5

4

2

0

1 3

0V1 V2

Label ‘0’

Label ‘1’Unary Cost

Integer Programming Formulation

Unary Cost Vector u = [ 5

Cost of V1 = 0

2

Cost of V1 = 1

; 2 4 ]

Page 8: So far

2

5

4

2

0

1 3

0V1 V2

Label ‘0’

Label ‘1’Unary Cost

Unary Cost Vector u = [ 5 2 ; 2 4 ]T

Label vector x = [ 0

V1 0

1

V1 = 1

; 1 0 ]T

Integer Programming Formulation

Page 9: So far

2

5

4

2

0

1 3

0V1 V2

Label ‘0’

Label ‘1’Unary Cost

Unary Cost Vector u = [ 5 2 ; 2 4 ]T

Label vector x = [ 0 1 ; 1 0 ]T

Sum of Unary Costs = ∑i ui xi

Integer Programming Formulation

Page 10: So far

2

5

4

2

0

1 3

0V1 V2

Label ‘0’

Label ‘1’Pairwise Cost

Integer Programming Formulation

0 Cost of V1 = 0 and V1 = 00

00

0Cost of V1 = 0 and V2 = 0

3

Cost of V1 = 0 and V2 = 11 0

000 0

103 0

Pairwise Cost Matrix P

Page 11: So far

2

5

4

2

0

1 3

0V1 V2

Label ‘0’

Label ‘1’Pairwise Cost

Integer Programming Formulation

Pairwise Cost Matrix P

0 0

00

0 3

1 000

0 010

3 0

Sum of Pairwise Costs∑i<j Pij xixj

= ∑i<j Pij Xij

X = xxT

Page 12: So far

Integer Programming Formulation

Constraints

• Uniqueness Constraint

∑ xi = 1i Va

• Integer Constraints

xi {0,1}

X = x xT

Page 13: So far

Integer Programming Formulation

x* = argmin ∑ ui xi + ∑ Pij Xij

xi {0,1}

X = x xT

∑ xi = 1i Va

Page 14: So far

Roadmap

Express MAP as an integer program

Relax to a linear program and solve

Round fractional solution to integers

Page 15: So far

Integer Programming Formulation

x* = argmin ∑ ui xi + ∑ Pij Xij

∑ xi = 1i Va

xi {0,1}

X = x xT

Convex

Non-Convex

Page 16: So far

Integer Programming Formulation

x* = argmin ∑ ui xi + ∑ Pij Xij

∑ xi = 1i Va

xi [0,1]

X = x xT

Convex

Non-Convex

Page 17: So far

Integer Programming Formulation

x* = argmin ∑ ui xi + ∑ Pij Xij

∑ xi = 1i Va

xi [0,1]

Xij [0,1]

Convex

∑ Xij = xij Vb

Page 18: So far

Linear Programming Formulation

x* = argmin ∑ ui xi + ∑ Pij Xij

∑ xi = 1i Va

xi [0,1]

Xij [0,1]

Convex

∑ Xij = xij Vb

Schlesinger, 76; Chekuri et al., 01; Wainwright et al. , 01

Page 19: So far

Roadmap

Express MAP as an integer program

Relax to a linear program and solve

Round fractional solution to integers

Page 20: So far

Properties

Dominate many convex relaxations

Best known multiplicative bounds

2 for Potts (uniform) energies

2 + √2 for Truncated linear energies

O(log n) for metric labelingMatched by move-making

Kumar and Torr, 2008; Kumar and Koller, UAI 2009

Kumar, Kolmogorov and Torr, 2007

Page 21: So far

Algorithms

Tree-reweighted message passing (TRW)

Max-product linear programming (MPLP)

Dual decomposition

Komodakis and Paragios, ICCV 2007