energy minimization - university of chicagottic.uchicago.edu/~rurtasun/courses/cv/lecture15.pdfp =...

Energy Minimization

Raquel Urtasun

TTI Chicago

Feb 28, 2013

Raquel Urtasun (TTI-C) Computer Vision Feb 28, 2013 1 / 48

The ST-mincut problem

Suppose we have a graph G = {V ,E ,C}, with vertices V , Edges E andcosts C .

[Source: P. Kohli]Raquel Urtasun (TTI-C) Computer Vision Feb 28, 2013 2 / 48

An st-cut (S,T) divides the nodes between source and sink.

The cost of a st-cut is the sum of cost of all edges going from S to T

The st-mincut is the st-cut with the minimum cost

Back to our energy minimization

Construct a graph such that

1 Any st-cut corresponds to an assignment of x

2 The cost of the cut is equal to the energy of x : E(x)

St-mincut and Energy Minimization

[Source: P. Kohli]

How are they equivalent?

= A + 0 0

C-A C-A

0 B+C-A-D

if x1=1 add C-A

if x2 = 1 add D-C

B+C-A-D ! 0 is true from the submodularity of !ij

A = !ij (0,0) B = !ij(0,1) C = !ij

(1,0) D = !ij (1,1)

!ij (xi,xj) = !ij(0,0) + (!ij(1,0)-!ij(0,0)) xi + (!ij(1,0)-!ij(0,0)) xj + (!ij(1,0) + !ij(0,1) - !ij(0,0) - !ij(1,1)) (1-xi) xj

Graph Construction

How to compute the St-mincut?

[Source: P. Kohli]

How does the code look like

[Source: P. Kohli]

Graph cuts for multi-label problems

Exact Transformation to QPBF [Roy and Cox 98] [Ishikawa 03] [Schlesingeret al. 06] [Ramalingam et al. 08]

Very high computational cost

Computing the Optimal Move

Move Making Algorithms

[Source: P. Kohli]

Energy Minimization

Consider pairwise MRFs

E (f ) =∑

{p,q}∈N

Vp,q(fp, fq) +∑p

Dp(fp)

with N defining the interactions between nodes, e.g., pixels

Dp non-negative, but arbitrary.

This is the graph-cuts notation.

Important to notice it’s the same thing.

Energy Minimization

E (f ) =∑

{p,q}∈N

Vp,q(fp, fq) +∑p

Dp(fp)

Energy Minimization

E (f ) =∑

{p,q}∈N

Vp,q(fp, fq) +∑p

Dp(fp)

Metric vs Semimetric

Two general classes of pairwise interactions

Metric if it satisfies for any set of labels α, β, γ

V (α, β) = 0 ↔ α = β

V (α, β) = V (β, α) ≥ 0

V (α, β) ≤ V (α, γ) + V (γ, β)

Semi-metric if it satisfies for any set of labels α, β, γ

V (α, β) = 0 ↔ α = β

V (α, β) = V (β, α) ≥ 0

Metric vs Semimetric

Two general classes of pairwise interactions

Metric if it satisfies for any set of labels α, β, γ

V (α, β) = 0 ↔ α = β

V (α, β) = V (β, α) ≥ 0

V (α, β) ≤ V (α, γ) + V (γ, β)

Semi-metric if it satisfies for any set of labels α, β, γ

V (α, β) = 0 ↔ α = β

V (α, β) = V (β, α) ≥ 0

Examples for 1D label set

Truncated quadratic is a semi-metric

V (α, β) = min(K , |α− β|2)

with K a constant.

Truncated absolute distance is a metric

V (α, β) = min(K , |α− β|)

with K a constant.

For multi-dimensional, replace | · | by any norm.

Potts model is a metric

V (α, β) = K · T (α 6= β)

with T (·) = 1 if the argument is true and 0 otherwise.

V (α, β) = min(K , |α− β|2)

with K a constant.

V (α, β) = min(K , |α− β|)

with K a constant.

V (α, β) = K · T (α 6= β)

V (α, β) = min(K , |α− β|2)

with K a constant.

V (α, β) = min(K , |α− β|)

with K a constant.

V (α, β) = K · T (α 6= β)

V (α, β) = min(K , |α− β|2)

with K a constant.

V (α, β) = min(K , |α− β|)

with K a constant.

V (α, β) = K · T (α 6= β)

Binary Moves

α− β moves works for semi-metrics

α expansion works for V being a metric

Figure: Figure from P. Kohli tutorial on graph-cuts

For certain x1 and x2, the move energy is sub-modular QPBF

Swap Move

[Source: P. Kohli]

Expansion Move

More formally

Any labeling can be uniquely represented by a partition of image pixelsP = {Pl |l ∈ L}, where Pl = {p ∈ P|fp = l} is a subset of pixels assignedlabel l .

There is a one to one correspondence between labelings f and partitions P.

Given a pair of labels α, β, a move from a partition P (labeling f ) to a newpartition P’ (labeling f ′) is called an α− β swap if Pl = P ′ for any labell 6= α, β.

The only difference between P and P ′ is that some pixels that were labeledin P are now labeled in P ′, and vice-versa.

Given a label l , a move from a partition P (labeling f ) to a new partition P ′

(labeling f ′) is called an α-expansion if Pα ⊂ P ′α and P ′

l ⊂ Pl .

An α-expansion move allows any set of image pixels to change their labelsto α.

More formally

l ⊂ Pl .

More formally

l ⊂ Pl .

More formally

l ⊂ Pl .

More formally

l ⊂ Pl .

More formally

l ⊂ Pl .

Example

Figure: (a) Current partition (b) local move (c) α− β-swap (d) α-expansion.

Algorithms

Finding optimal Swap move

Given an input labeling f (partition P) and a pair of labels α, β we want tofind a labeling f that minimizes E over all labelings within one α− β-swapof f .

This is going to be done by computing a labeling corresponding to aminimum cut on a graph Gαβ = (Vαβ , Eαβ).

The structure of this graph is dynamically determined by the currentpartition P and by the labels α, β.

Graph Construction

The set of vertices includes the two terminals α and β, as well as imagepixels p in the sets Pα and Pβ (i.e., fp ∈ {α, β}).

Each pixel p ∈ Pαβ is connected to the terminals α and β, called t-links.

Each set of pixels p, q ∈ Pαβ which are neighbors is connected by an edgeep,q

Computing the Cut

Any cut must have a single t-link not cut.

This defines a labeling

There is a one-to-one correspondences between a cut and a labeling.

The energy of the cut is the energy of the labeling.

See Boykov et al, ”fast approximate energy minimization via graph cuts”PAMI 2001.

Properties

For any cut, then

Finding the optimal α expansion

Given an input labeling f (partition P) and a label α we want to find alabeling f that minimizes E over all labelings within one α-expansion of f .

This is going to be done by computing a labeling corresponding to aminimum cut on a graph Gα = (Vα, Eα).

The structure of this graph is dynamically determined by the currentpartition P and by the label α.

Different graph than the α− β swap.

Graph Construction

The set of vertices includes the two terminals α and α, as well as all imagepixels p ∈ P.

Additionally, for each pair of neighboring pixels p, q such that fp 6= fq wecreate an auxiliary node ap,q.

Each pixel p is connected to the terminals α and α, called t-links.

Each set of pixels p, q which are neighbors and fp = fq, we connect with andn-link.

For each pair of neighboring pixels such that fp 6= fq, we create a triplet{ep,a, ea,q, tαa }.

The set of edges is then

Graph Construction

Properties

There is a one-to-one correspondences between a cut and a labeling.

The energy of the cut is the energy of the labeling.

See Boykov et al, ”fast approximate energy minimization via graph cuts”PAMI 2001.

Global Minimization Techniques

Ways to get an approximate solution typically

Dynamic programming approximations

Sampling

Simulated annealing

Graph-cuts: imposes restrictions on the type of pairwise cost functions

Message passing: iterative algorithms that pass messages between nodes inthe graph.

Now we can solve for the MAP (approximately) in general energies. We can solve

for other problems than stereo

Let’s look at data/bechmarks

Benchmarks

Two benchmarks with very different characteristics

(Middlebury) (KITTI)

Middlebury Dataset

Middlebury Stereo Evaluation – Version 2

Laboratory

Lambertian

Rich in texture

Medium-size label set

Largely fronto-parallel

Middlebury Dataset

Laboratory

Lambertian

Rich in texture

Middlebury Dataset

Laboratory

Lambertian

Rich in texture

Middlebury Dataset

Laboratory

Lambertian

Rich in texture

Middlebury Dataset

Laboratory

Lambertian

Rich in texture

Benchmarks for Stereo and metrics

Best methods < 3% errors (for all non-occluded regions)

http://vision.middlebury.edu/stereo/data/

Benchmarks: KITTI Data Collection

Two stereo rigs (1392× 512 px, 54 cm base, 90◦ opening)

Velodyne laser scanner, GPS+IMU localization

6 hours at 10 frames per second!

The KITTI Vision Benchmark Suite

Novel Challenges

Fast guided cost-volume filtering (Rhemann et al., CVPR 2011)

Middlebury, Errors: 2.7%

Error threshold: 1 px (Middlebury) / 3 px (KITTI)

Novel Challenges

Fast guided cost-volume filtering (Rhemann et al., CVPR 2011)

Middlebury, Errors: 2.7% KITTI, Errors: 46.3%

Error threshold: 1 px (Middlebury) / 3 px (KITTI)

Novel Challenges

So what is the difference?

Middlebury

Laboratory

Lambertian

Rich in texture

Moving vehicle

Specularities

Sensor saturation

Large label set

Strong slants

Novel Challenges

Middlebury

Laboratory

Lambertian

Rich in texture

Moving vehicle

Specularities

Sensor saturation

Large label set

Strong slants

Novel Challenges

Middlebury

Laboratory

Lambertian

Rich in texture

Moving vehicle

Specularities

Sensor saturation

Large label set

Strong slants

Novel Challenges

Middlebury

d = 50 px

d = 16 px

Laboratory

Lambertian

Rich in texture

d = 150 px

d = 0 px

Moving vehicle

Specularities

Sensor saturation

Large label set

Strong slants

Novel Challenges

Middlebury

Laboratory

Lambertian

Rich in texture

Moving vehicle

Specularities

Sensor saturation

Large label set

Strong slants

Stereo Evaluation

MRFs for stereo

Global methods: define a Markov random field over

Pixel-level

Fronto-parallel planes

Slanted planes

Plane MRFs

First segment an image into small regions, i.e., superpixels

Assume that the 3D world is compose of small frontal/slanted planes

Good representation if the superpixels are small and respect boundaries

E (x1, · · · , xn) =∑i

C (xi ) +∑i

∑j∈Nj

C (xi , xj)

with xi ∈ < for the fronto-parallel planes, and xi ∈ <3 for the slanted planes

This are continuous variables. Is this a problem?

What can I do to solve this? Discretize the problem

The unitary are usually agreegation of cost over the local matching on thepixels in that superpixel

Pairwise is typically smoothness

Plane MRFs

E (x1, · · · , xn) =∑i

C (xi ) +∑i

∑j∈Nj

C (xi , xj)

Plane MRFs

E (x1, · · · , xn) =∑i

C (xi ) +∑i

∑j∈Nj

C (xi , xj)

Plane MRFs

E (x1, · · · , xn) =∑i

C (xi ) +∑i

∑j∈Nj

C (xi , xj)

Plane MRFs

E (x1, · · · , xn) =∑i

C (xi ) +∑i

∑j∈Nj

C (xi , xj)

Plane MRFs

E (x1, · · · , xn) =∑i

C (xi ) +∑i

∑j∈Nj

C (xi , xj)

Plane MRFs

E (x1, · · · , xn) =∑i

C (xi ) +∑i

∑j∈Nj

C (xi , xj)

Slanted-plane MRFs

A more sophisticated occlusion model

MRF on continuous variables (slanted planes) and discrete var. (boundary)

Combines depth ordering (segmentation) and stereo

!"#$%&'(

)*+,*$- !"#$"%&'()*+),-").&$-*%/01/2.&$*/"3/4*+,*$-

./0%1)*2'()*+),-"5*.&6"$4782/9*-:**$/4*+,*$-4/

;/4-&-*4/

<==.#48"$ >8$+* ?"2.&$&'

?"$6$#"#4/@&'8&9.*/

184='*-*/@&'8&9.*/)#2*'28A*.4/BC?D/EF'9*.&*GH/*-/&.I/JKLLM/

////////////////////////&$%/)NO?/EF=7&$-&H/*-/&.I/JKLKMP/

Takes as input disparities computed by any local algorithm

Energy of PCBP-Stereo

y the set of slanted 3D planes, o the set of discrete boundary variables

E (y, o) = Ecolor (o) + Ematch(y, o) + Ecompatibility (y, o) + Ejunction(o)

!"#"$%&'()$)&' *"+,$-'.)'/,'()0$%1%&'

*,2'"#%3,'

4)$)&'5"#"$%&".-'

!"#"$%&'

6"55"#"$%&'

!"#$$%$&'()*'+(#$,-.'(/0(*&1-'(2*,13#*'4(%31(

5/%1-'$2(64(3&4(%3'7+*&"(%$'+/2((((89/2*:$2(,$%*;"./63.(%3'7+*&"<(

=*,13#*'4(%31( >.3&'$2(1.3&$(

?#-&73'$2(@-32#3A7(0-&7A/&

B&(6/-&23#4(CB77.-,*/&D(E(F/#$"#/-&2(,$"%$&'(/)&,(6/-&23#4(

!"#$%&'()*&+,*-) .*&$/,%)0*&$,-)!"#$%&

1',2,',$3,)"2)+"#$%&'()*&+,*) 45"0*&$&')6)78$9,)6):33*#-8"$;)

<&/3=)

>:33*#-8"$?)

>78$9,?)

>5"0*&$&'?)

i j i j

on boundary

in both segments

@<0"-,)0,$&*/()

!""#$%&'()*'$(+,-.)-/,%'(&(0)

123'%%&*#/)",%/%)

!""#$%&'()4-'(5)

6,"7)8&(0/)9'3#,(,-)

:;,#&7)<=>?@)

A/(,#&B/)&23'%%&*#/)C$("D'(%)

energy minimization - university of chicagottic.uchicago.edu/~rurtasun/courses/cv/lecture15.pdfp =...

Documents

fp tres olivos | fp tres olivos

sodium and potassium.. lgis

lighting legend -...

quadro per un giglio - pasquale aiezza: home per un giglio...

4.2.10. filtration practice (fp) fp-1....

39& fp o fp l

fp-510ii fp-510 fp-510(pfbm) fp-510k fp-510(bt) -...

fp alternancia presentacion_cordoba florent fp rules

dansani mido - laattapiste.fi · fp fp fp fp fp fp fp fp...

fp-2500/fp-2600 series user manual - pro-face touch … ·...

sparse matrix factorization - university of...

direct to gateway casinos...

nmisa - nla.org.za€¦ · gold fp 1064,180 °c silver fp...

urea cycle.. lgis

fp-488 fp-547h fp-594 fp-647h(min x...

fp-konesauma€¦ · 13 fp-lukkosauma fp-konesauma...

pkf6..fp.., pkf64.fp.c, pkf65.fp.., pkf6..fp.g, pkn6..fp

amazon s3projector+and...led micro fp-364d 300 micro sd/tf*,...

smipack fp sorozat - ajánlatkérés 2 perc alatt ·...

peoplesoft report id: nms046hr organizational listing...