Images Similarity by Relative Dynamic Programming

37
Images Similarity by Relative Dynamic Programming M. Sc. thesis by Ady Ecker Supervisor: prof. Shimon Ullman

description

Images Similarity by Relative Dynamic Programming. M. Sc. thesis by Ady Ecker Supervisor: prof. Shimon Ullman. Overview. The similarity problem The Method: Overlapping patches Combining evidences Relative dynamic programming Analysis. Part 1: The Problem. Old problem. Important. - PowerPoint PPT Presentation

Transcript of Images Similarity by Relative Dynamic Programming

Images Similarity by Relative Dynamic

Programming

M. Sc. thesis by Ady Ecker

Supervisor: prof. Shimon Ullman

Overview

The similarity problem

The Method:Overlapping patchesCombining evidencesRelative dynamic programming

Analysis

The similarity problem

Old problem.

Important.

Difficult.

Natural to humans.

What is similarity?Similarity has many aspects, e.g. shape, color, size, outline, texture.

Relative and context dependent.

Depends on experience with classes in the real world.

Similarity is not a metric

Triangle inequality

Symmetry

Minimality

Previous approaches to similarity computation

Distance functions (L2, correlation).

Features detection, image signatures.

Minimal energy aligning transformation.

Dynamic programming measures.

Overlapping fragments (Ullman and Sali).

Motivation

Point-wise methods fail on very simple examples.

Motivation

We propose: common substructures of multiple sizes at multiple scales.

Similarity of sub-patchesat the local level

A1

B1

C1

A2

C2

B2

A1 A2 B1 B2 C1 C2

Overlapping sub-patchesoverlapping fragments impose similar structure.

Hierarchical decomposition Allows the sub-patches to move a little.

Hierarchical alignment

Similarity indicators

Each indicator has its marginal distribution, but the joint distribution is too complex.

The indicators are dependent.

FX

Xnormalization by 1-D distribution

Generalization groups example

The similarity is related to the intersection of the groups:Pr( (A,B) length ratio ratio(A,B) )

45º length ratio 2

Graphical illustration of the similarity score

independent case general case estimation error

zF

0 10

1z

(x,y)area zprob. F

F

z

0 1

z

(x0,y0)

(x1,y1)

zz

1

0

(x3,y3)

F

0 10

1

(x0,y0)

(x1,y1)

FF

F

Empirical example

10,000 points from 2D normal distribution.

Normalization:x → Pr ( X x )y → Pr ( Y y )

Only 233 points between the two curves.

The similarity score

The score combines several similarity indicators.The score is relative.Since the score is normalized, it can be used in hierarchical construction:

)(),( iXF xFFBADiiX

x1

x2

x3

y1

y2

z

Dynamic programming in 2D?

Vertical movement

Horizontal and vertical movements

Relative Dynamic Programming (RDP)

Works in 2D like dynamic programming algorithms, but:Scores are relative.No explicit objective function.Start from many points simultaneously.

I1 I2

D(x,x2)

D(x,x3)

D(x,x2)D(x,x1)

D(x,x3)

D(x,x)

D(x,x1)

T(i,)

3

2

1

0

123

x x

The similarity table

The minima tableAllow flexible movements.

Costs keep movement coherence.

Used for efficiency.

Bi-directional (two tables).

projected index

delta

Combining evidences

18

118

1

tableminimum thefrom score

i

i

i

vP

v

Flowchart

Patch size = N?

Input images Set patch size = 2 Basic similarity measure

Store scores in similarity table

Yes

Output the similarity score

Build minima tables

Normalization

Double the patch size

Geometric mean of sub-patches' scores

No

Weighting patches

ii

w

i

wi

patchi j

vP

iIjiIpatchw

118

1

)2,2(

)2,2(

2)()(

Multi-scaleTwo dimensions, cross-resolution:

T(r,z,i1,i2,1,2)=min(P, Vhigh-low, Vlow-low)

P

Vlow-low

Vhigh-low

low resolution:

high resolution:

DistributionsDistributions are sampled empirically.

The sampling is done layer by layer.

distribution of the geometric mean of independent uniform variables

distribution of the weighted geometric mean of real images

Performance

The complexity is O(N22).The number of operations is governed by the operations on the original resolution.Basic-similarity operations:

3232171716 4,735,000.Minima operations:

3232171792 5,327,000.Total 10,000,000.Takes a second on 1.5GH p.c.

ResultsB1 B2 B3 B4 B5 B6 B7 B8 B9 B10

D(Ai,Bj)

A1

0.008 0.008 0.914 0.844 0.656 0.844 0.211 0.547 0.508 0.695

A2

0.008 0.008 0.336 0.055 0.102 0.094 0.203 0.219 0.172 0.227

A3

0.906 0.336 0.008 0.234 0.141 0.016 0.898 0.781 0.133 0.016

A4

0.477 0.117 0.445 0.008 0.156 0.141 0.164 0.117 0.438 0.148

A5 0.445 0.063 0.07 0.055 0.008 0.164 0.938 0.891 0.586 0.32

A6

0.805 0.664 0.047 0.281 0.344 0.008 0.641 0.25 0.023 0.117

A7

0.633 0.297 0.992 0.602 0.984 0.727 0.008 0.016 0.57 0.859

A8

0.961 0.125 0.523 0.219 0.789 0.133 0.055 0.016 0.25 0.172

A9

0.289 0.289 0.133 0.375 0.492 0.086 0.297 0.305 0.008 0.008

A10

0.805 0.438 0.055 0.281 0.359 0.344 0.586 0.281 0.023 0.008

B1 B2 B3 B4 B5 B6 B7 B8 B9 B10 L2(Ai,Bj)

A1

0.707 0.707 0.639 0.685 0.704 0.706 0.701 0.707 0.707 0.717

A2

0.707 1.000 0.623 0.685 0.713 0.699 0.704 0.707 0.707 0.697

A3

0.632 0.643 0.423 0.602 0.546 0.606 0.680 0.679 0.659 0.609

A4

0.707 0.707 0.593 0.661 0.667 0.692 0.690 0.723 0.707 0.697

A5 0.702 0.690 0.555 0.707 0.527 0.692 0.781 0.815 0.704 0.697

A6

0.707 0.707 0.621 0.707 0.718 0.494 0.706 0.673 0.729 0.708

A7

0.707 0.707 0.714 0.750 0.797 0.732 0.332 0.552 0.791 0.761

A8

0.707 0.707 0.702 0.750 0.843 0.703 0.626 0.306 0.750 0.724

A9

0.707 0.707 0.627 0.707 0.711 0.746 0.786 0.758 0.707 0.655

A10

0.707 0.707 0.658 0.707 0.692 0.719 0.708 0.740 0.685 0.708

B1 B2 B3 B4 B5 B6 B7 B8 B9 B10 correlation (Ai,Bj)

A1

0.500 0.500 0.502 0.468 0.495 0.498 0.490 0.500 0.500 0.515

A2

0.500 1.000 0.473 0.468 0.509 0.488 0.494 0.500 0.500 0.485

A3

0.504 0.524 0.375 0.496 0.480 0.494 0.481 0.471 0.555 0.477

A4

0.500 0.500 0.473 0.453 0.494 0.493 0.422 0.472 0.500 0.491

A5 0.491 0.472 0.513 0.560 0.387 0.523 0.482 0.543 0.495 0.502

A6

0.500 0.500 0.470 0.500 0.518 0.241 0.499 0.446 0.531 0.501

A7

0.500 0.500 0.505 0.518 0.503 0.497 0.144 0.406 0.644 0.574

A8

0.500 0.500 0.478 0.518 0.606 0.449 0.511 0.125 0.572 0.511

A9

0.500 0.500 0.481 0.500 0.507 0.557 0.633 0.586 0.500 0.429

A10

0.500 0.500 0.539 0.500 0.475 0.518 0.501 0.554 0.469 0.501

RDP

L2

correlation

ResultsB1 B2 B3 B4 B5 B6 B7 B8 B9 B10

D(Ai,Bj)

A1 0.055 0.008 0.461 0.102 0.484 0.07 0.148 0.422 0.063 0.148

A2 0.008 0.055 0.172 0.055 0.164 0.203 0.117 0.438 0.063 0.07

A3 0.359 0.117 0.055 0.008 0.125 0.125 0.359 0.57 0.391 0.219

A4 0.148 0.148 0.008 0.055 0.117 0.07 0.102 0.125 0.523 0.055

A5 0.359 0.102 0.148 0.172 0.055 0.008 0.391 0.414 0.625 0.172

A6 0.086 0.172 0.219 0.227 0.008 0.055 0.07 0.359 0.5 0.055

A7 0.219 0.359 0.117 0.328 0.461 0.273 0.398 0.016 0.625 0.102

A8 0.219 0.172 0.148 0.273 0.5 0.125 0.008 0.258 0.258 0.117

A9 0.055 0.125 0.5 0.125 0.461 0.063 0.102 0.203 0.133 0.023

A10 0.273 0.063 0.461 0.07 0.656 0.68 0.32 0.539 0.023 0.148

B1 B2 B3 B4 B5 B6 B7 B8 B9 B10 D(Ai,Bj)

A1 0.031 0.055 0.797 0.898 0.047 0.055 0.188 0.18 0.055 0.055

A2 0.055 0.031 0.773 0.922 0.055 0.055 0.141 0.086 0.055 0.055

A3 0.609 0.805 0.031 0.039 0.656 0.516 0.492 0.453 0.594 0.82

A4 0.719 0.898 0.055 0.047 0.523 0.523 0.383 0.125 0.75 0.555

A5 0.063 0.078 0.484 0.578 0.047 0.063 0.055 0.117 0.055 0.078

A6 0.055 0.055 0.758 0.781 0.055 0.039 0.188 0.125 0.055 0.094

A7 0.141 0.078 0.578 0.688 0.063 0.078 0.055 0.055 0.055 0.094

A8 0.164 0.258 0.242 0.484 0.359 0.281 0.234 0.039 0.203 0.359

A9 0.055 0.055 0.609 0.875 0.055 0.055 0.063 0.234 0.039 0.055

A10 0.055 0.055 0.852 0.93 0.055 0.078 0.141 0.234 0.055 0.039

1 2 3 4 5 6 7 8 9 10

A

D(Ai,B1) 0.031 0.031 0.031 0.039 0.047 0.055 0.055 0.055 0.055 0.055

B

D(A1,Bi) 0.031 0.031 0.047 0.055 0.055 0.055 0.094 0.180 0.180 0.234

Experimental conclusions

The algorithm is superior to point-wise distance functions (L2, correlation), even when the images are aligned by a global transformation.

The current implementation neglects important parts such as outlines and topology.

Qualitative properties

Similarity by parts.

Captures variety of transformations.

Weights the salient parts.

Adaptive to distributions in the domain.

Properties of the design

Robust.

Simple operations.

Can be implemented in parallel and neural-networks.

Intuitive.

Extendible to deal with many aspects of similarity.

SummaryWe presented a new scheme to the similarity problem in general.The scheme is based on the systematic evaluation of similarity on overlapping sub-patches. Scores are relative. Scores are combined hierarchically using empiric normalizations.The implementation captures aspects of perceptual similarity, but is still inferior to human’s judgment.