Lecture 1 4

54
Lecture 1 Lecture 1 4 4 Previous Works

description

Lecture 1 4. Previous Works. A target is assumed to be Scaled, Rotated Version Template With Edges Distorted. Problem: Target Image Search. Search on Images Database. Target. Template. Inspiration Jain et al [1] , “Object Matching Using Deformable Templates” Our Algorithm - PowerPoint PPT Presentation

Transcript of Lecture 1 4

Page 1: Lecture 1 4

Lecture 1Lecture 144

Previous Works

Page 2: Lecture 1 4

2

Problem: Target Image SearchProblem: Target Image Search

A target is assumed to be Scaled, Rotated Version Template With Edges Distorted

Search on Images Database

Template Target

Page 3: Lecture 1 4

3

MethodologyMethodology: An Overview: An Overview

InspirationJain et al [1] , “Object Matchin

g Using Deformable Templates”

Our Algorithm Finding Hypotheses : MGH

T Peak Clustering :

Watershed Method Contour Matching : Smo

oth Membrane Fitting

Template Target Contours Hypotheses 3 Peaks in 3D Hough Space

Rejected Hypothesis

Accepted Hypothesis

Rejected Hypothesis

Page 4: Lecture 1 4

4

Finding Candidate Target:Finding Candidate Target:Modified Generalized HoughModified Generalized Hough

2

r

L,

xc,yc

= 0

•[Nimkerdpol and Madarasmi, 2001]•A line at the contour edge is extended in the direction until it meets the other end of the contour •In MGHT, the relation between r,l) is stored as a linked list in R-Table, not as r) like in GHT

r1, 1, 1, l1 r2, 2, 2, l2 r3, 3, 3, l3 r4, 4, 4, l4

0...19 15,180,195,99 9,179,219,101 8,177,216,102 9,176,198,100

20...39 17,160,23,5 14,159,38,7 18,161,175,62 15,162,195,95

30…49 19,165,31,53 20,170,8,52 22,167,15,52 18,159,158,12

… … … … …

340...359

23,105,346,11 24,103,165,11 21,102,346,18 22,104,195,24

Page 5: Lecture 1 4

5

MGHT: Rotation/ScaleMGHT: Rotation/Scale

2

L,

= 0 2

L,= 0

= 30-200 = -170 = 190 = 300-110 = 190

Page 6: Lecture 1 4

6

MGHTMGHT

θθβ c

L

L S c

Rotation Factor:

Scaling Factor :

New ref. Point :

xc = x + S r cos (

yc = y + S r sin (

Page 7: Lecture 1 4

7

Watershed for Peak ClusteringWatershed for Peak Clustering

1. Shed, by labeling, at the first level, calculate peaks of each label2. Increase to higher level, shed again

2.1 Meet an area of previous level, shed to that area2.2 Not meet any area of previous level, make a new area ,

calculate a new peak

Page 8: Lecture 1 4

8

Deformation : Contour MatchingDeformation : Contour Matching

Parameter : xyor (u,v) range -7, -6, …,0,… 6, 7

Page 9: Lecture 1 4

9

Coarse and Fine MatchingCoarse and Fine Matching

Page 10: Lecture 1 4

10

Matching AlgorithmMatching Algorithm

Energy FunctionUpdate (u,v) : Gibbs Sampler with simulated annealing

Template

Target Edge

Page 11: Lecture 1 4

11

Experiment on Image SearchExperiment on Image Search

Template Target Edge Map ResultHough Space

Page 12: Lecture 1 4

12

Experiment on Image SearchExperiment on Image Search

Target Edge Hypotheses 1st Match

2nd Match 3rd Match 4th Match

Page 13: Lecture 1 4

13

Experiment on Image SearchExperiment on Image Search

Template Target Edge Map The Best Match Hough Space

Page 14: Lecture 1 4

14

Threshold Selection: Guitar 1Threshold Selection: Guitar 1

3.986274 0.929011 2.705226Template Target

EdgeHypotheses

Threshold : 1.0 - 2.6

Page 15: Lecture 1 4

15

Threshold Selection: Guitar 1Threshold Selection: Guitar 1

2.165488 0.965049 1.755835

Threshold : 1.0-1.6

Template Target Edge Hypotheses

Page 16: Lecture 1 4

16

Threshold Selection: Vase 1Threshold Selection: Vase 1

1.799267 1.114566 5.074061

Threshold : 1.2-1.6

Template Edge Map Hypotheses

Page 17: Lecture 1 4

17

Threshold Selection: Vase 2Threshold Selection: Vase 2

Template

TargetEdge

Hypotheses 0.868600 0.879799 3.799124

Threshold : 0.9-3.6

Page 18: Lecture 1 4

18

Threshold Selection: Vase 3Threshold Selection: Vase 3

Template Target Edge Hypotheses

1.293034 1.452130 3.364521 4.4185782

Threshold : 1.5-3.2

Page 19: Lecture 1 4

19

Energy ThresholdEnergy Threshold

0

0.5

1

1.5

2

2.5

3

3.5

4

accepted rejected

guitar1

guitar 2

vase

bottle

lamp

Page 20: Lecture 1 4

20

Experiment on Image QueryingExperiment on Image Querying

Database Search for Circle shape

Page 21: Lecture 1 4

21

Experiment on Image QueryingExperiment on Image Querying

Database

Search for bulb shape

Page 22: Lecture 1 4

22

ConclusionConclusion

A deformation model Contour Matching A method for image search Future work: large image database, efficient

method for minimizing energy, coarse-and-fine approach to computer vison modules

Page 23: Lecture 1 4

23

Accurate 3-D surface map using stereo vision

This proposal research addresses 2 issues:

• Find an accurate 3-D surface map using stereo vision

• Combine surface from different views to a single 3-D object for CAD applications.

Page 24: Lecture 1 4

24

Combine surface from multiple views to a single 3-D object.

To combine multiple view, we need to find the rotation and transformation matrices from each camera combined to the world or reference co-ordination system.

==================Rotation : 0.7122 0.7019 0.0130 Rotation : 0.0386 -0.0207 -0.9990 Rotation : -0.7010 0.7120 -0.0418

Translation : 16.6342 Translation : 32.6633 Translation : 181.5649 ==================

Page 25: Lecture 1 4

25

Expect Result After Combine Multiple View

Page 26: Lecture 1 4

26

A Relaxation Method for Shape A Relaxation Method for Shape from Contours from Contours

Input Contour Images:1. Geodesics Contours Only2. Developable Surface (No Folds/Twists)3. Non-Accidental View

1. Place Grid Points in X and Y Direction to have Smooth2. Draw Horizontal and Vertical Lines through Grid to

Form a Regular Square Texture3. Use Shape from Texture to Obtained Surface Normals

Page 27: Lecture 1 4

27

Experiment 2Experiment 2

Step 1

Step 2 Step 3

Original

Page 28: Lecture 1 4

28

Shape from ContourShape from Contour

  

Figure 1. Example of various surfaces with geodesic contours.

(A) LINE DRAWING IMAGE (B) GRAY SCALE IMAGE

FIGURE 2. COMPARING THE PERCEIVED SHAPE OF OBJECTS IN LINE-DRAWING VS. SHADED, GRAY-SCALE IMAGES.

Page 29: Lecture 1 4

29

Shape from ContourShape from Contour

FIGURE 3. EXAMPLES OF OBJECTS THAT CONSIST OF PARTS OF

QUADRILATERAL DEVELOPABLE CURVATURE SURFACES.

Figure 5. Examples of images that consist of only line-drawings. Even without shading and texture, a human viewer can determine the shape of the object.

Page 30: Lecture 1 4

30

Shape from ContourShape from Contour

(a)Occluding contour

(b)Texture contour

(c)Shadow contour

(d) Geodesic Contour (e) surface fold

contour Figure 4. Types of Contours (a) Occluding Contour (b) Texture Contour (c) Shadow Contour (d)Geodesic Contour (e) surface Fold Contour

Page 31: Lecture 1 4

31

Shape from ContourShape from Contour

(A) (B)

(C) (D)

FIGURE 7. (A) ORIGINAL CONTOUR (B) INTERSECTION POINTS DERIVED

FROM RELAXATION METHOD (C) CREATE LINES BY USING SPLINE

TECHNIQUE (D) SHAPE FROM TEXTURE DETERMINES THE NORMAL VECTOR

FOR EACH INTERSECTION POINT.

Page 32: Lecture 1 4

32

Shape from ContourShape from Contour

(A) (B)

FIGURE 9. EXAMPLE OF SOME SURFACES CAN BE PERCEIVED FROM

TEXTURE. (A) UNDULATE SINGLE TEXTURE (B) UNDULATE TEXTURE NET

LIKELIHOOD

(a) (b) (c)

FIGURE 10. COMPUTING SHAPE FROM A TEXTURE PATCH. (A) ARBITARY

PATCH WITH INTERSECTING OUTLINES. THE ANGLE SHOWN IN (B) IS

MEASURED IN THE IMAGE PLANE. (C) VECTORS U AND V ARE 3-D VECTORS

ON THE SURFACE, AND N IS THE SURFACE NORMAL AT THEIR INTERSECTION.

V

N

Page 33: Lecture 1 4

33

Shape from ContourShape from Contour

2222

1111)H

~H~

()H~

H~

()H~

H~

()H~

H~

(E)c,r()c,r()c,r()c,r()c,r()c,r()c,r()c,r( PPPPPPPPh

2222

1111)V

~V~

()V~

V~

()V~

V~

()V~

V~

(E)c,r()c,r()c,r()c,r()c,r()c,r()c,r()c,r( PPPPPPPPv

vhcontinuity EEE

22 )V ~V ~()H~H~(E)c,r()c,r()c,r()c,r( PPPPsmoothness

smoothnesssmoothnesscontinuitycontinuity EEE

T h e c o n s i d e r e d v e c t o r s w i l l b e d e n o t e d a s s y m b o l s b e l o w :

F I G U R E 8 . S M O O T H N E S S A N D C O N T I N U I T Y D E F I N E D . P ( R , C ) I S T H E P O I N T

U N D E R C O N S I D E R A T I O N W I T H I T S N E I G H B O R S S H O W N .

1 . 1 P ( r

, c ) = P

P ( r - 1 , c )

P ( r , c - 1 ) P ( r , c + 1 )

P ( r + 1 , c )

)c,r()c,r(P PPH~)c,r( 1

)c,r()c,r(P PPH~)c,r( 1

)c,r()c,r(P PPV~

)c,r( 1

)c,r()c,r(P PPV~)c,r( 1

Page 34: Lecture 1 4

34

Shape from ContourShape from Contour

a n d

A N D T I L T A N D T H E S L A N T A R E :

a n d N o w i f t h e i n t e r s e c t i n g p h y s i c a l c u r v e s a r e l i n e s o f c u r v a t u r e , t h e y a r e p e r p e n d i c u l a r a t t h e i n t e r s e c t i o n . H e n c e t h e d o t p r o d u c t o f U a n d V i s z e r o . T h e n S o t h e N o r m a l v e c t o r N w i l l r e s t o n l y a u n k n o w n v a r i a b l e

a,,U 01 b,sin,cosV

sin,cosa,sinaN,N,NN zyx

x

y

N

Ntan 1 21222

1

/

zyx

z

NNN

Ncos

acos

b

sin,cos

aa

,sinaN12

Page 35: Lecture 1 4

35

Shape from ContourShape from Contour

(A) SURFACE WITH

GEODESIC CONTOUR (B) SURFACE WITH

OPTIMAL

INTERSECTION POINT

SET

(C) SURFACE WITH

TEXTURE FORMED BY

THE INTERSECTION

POINT SET

Figure 16. Preliminary Experimental Results.

(A) (B) (C) (D)

Page 36: Lecture 1 4

36

Contour SearchContour Search

2

r

L,

xc,yc

= 0

Page 37: Lecture 1 4

37

Contour SearchContour Search

a) No Deformation b) Deformed by order

1 (N=M=1) c) Deformed by order 2 (N=M=2)

d) Deformed by order 3 (N=M=3)

Figure 2.3. Various deformation resolutions (copied from [1]).

Figure 2.4. An example of motion correspondence on 2 images

Page 38: Lecture 1 4

38

Contour SearchContour Search

Page 39: Lecture 1 4

39

Contour MatchingContour Matching

E T o ta l = E C o a rs e G rid + 1E F in e G rid + 2E In te r G rid (1 4 )

E G r i d = E C o n t o u r M a t c h i n g + E M a t c h i n g S m o o t h n e s s ( 1 5 )

j)(i,E Grid = j)(i,Ij))v(i,jj),u(i,(iI Tangent TemplateTangentTarget + (1 6 )

j)Nbr(i, nm,

j)u(i,n)u(m, +

j)Nbr(i, nm,

j)v(i,n)v(m,

E I n t e r G r i d = E u _ i n t e r g r i d + E v _ i n t e r g r i d ( 1 7 )

E u _ in te rg rid (i, j)fin e = 0 , |u (i, j) c o a rs e s c a le – u (i, j)fin e | s c a le /2 (1 8 )

1 , o th e rw is e

Page 40: Lecture 1 4

40

Template

Target Image

Target edge

(f) Iteration 25, Energy=8.5315

(g) Iteration 26, Energy=7.0977

(h) Iteration 29, Energy=2.6243

(i) Result at Iteration 50, Energy=1.3657

Page 41: Lecture 1 4

41

(a) Template

(b) Transformed template

(c) Target Image

(d) Target edge Map

(g) Iteration 25, Energy = 8.269747

(h) Iteration 27, Energy = 4.643370

(i) Iteration 35, Energy = 1.257369

(j) Final result at iteration 40, Energy = 1.159310

Page 42: Lecture 1 4

42

Template E=1.081968 E=1.174364 E=1.332886 E=1.190744 (a) (b) (c) (d) (e)

(a) Template

(b) Target Image

(c) Target edge

(d) Hough space

(g) It 26, E=5.9126

(h) It 27, E=3.8052

(i) It 29, E=2.0100

(j) It 35, E=1.284

Page 43: Lecture 1 4

43

(a) Template

(b) Target Edge

(c) Hough Space

(d) 1st match, Energy=1.799267, rejected

(e)2nd match 1.114566, accepted

(f) 1st vote, Energy=5.074061, rejected

Page 44: Lecture 1 4

44

(a)

(b) Target edge

(c) Hough space

Page 45: Lecture 1 4

45

Paper InspectionPaper Inspection

Page 46: Lecture 1 4

46

Multi-Layer StereoMulti-Layer Stereo

Page 47: Lecture 1 4

47

Multi-Layer StereoMulti-Layer Stereo

Page 48: Lecture 1 4

48

Multi-View StereoMulti-View Stereo

Z

Y

X

inateFirstCoorddinateSecondCoor

t

t

t

Z

Y

X

eeeeeeeeeeee

eeeeeeeeeeee

eeeeeeeeeeee

Z

Y

X

23

22

21

2010322031

103223

22

21

203021

2031302123

22

21

20

)(2)(2

)(2)(2

)(2)(2

.123

22

21

20 eeee

Page 49: Lecture 1 4

49

Multi-View StereoMulti-View Stereo

Page 50: Lecture 1 4

50

Rubber Sheet InspectionRubber Sheet Inspection

Page 51: Lecture 1 4

51

Input Image ofBackIlluminatedRubber Sheet

Preprocessing1: NoiseRemoval viaGaussian Blur

Preprocessing2: BoundaryErosion

Partition imageinto rectangularregions

AutomaticThresholding fromHistogram PerRegion

Threshold

BinaryImagePerRegion

AssembledBinaryImage ofDefects

b.

c.

d.

e.

a.

b.

c.

d.

e.

a.

b.

Page 52: Lecture 1 4

52

Garment Layout using GeneticGarment Layout using Genetic

W

-กำ��หนดจำ��นวน Population และสุ่��มเพื่��อสุ่ร้��ง Population-คั�ดเล�อกำ Solution ที่��ด�ที่��สุ่�ดออกำม�จำ��นวนหน��ง-น�� Solution ด�งกำล��วม�จำ�ดเร้�ยงจำ�กำปร้ะสุ่ ที่ธิ ภ�พื่ด�สุ่�ดไปแย�สุ่�ดRepeat

-สุ่��มคั��แบบ Uniformly Distributed Random ที่��ม�คั�� 0 ถึ�ง 1-กำ��หนดให� r คั�อคั��ที่��สุ่��มได� และ p คั�อคั��คัว�มน��จำะเป'นIf r < p (กำ��หนดให� p = 0.9) Then

- สุ่��มเล�อกำ Solution จำ�กำ Population ด�วยกำ�ร้ใช้� Linearly Biased Random Number Generator ออกำม�จำ��นวนหน��ง- สุ่��มเล�อกำ Strip จำ�กำ Solution ที่�)งหมดที่��เล�อกำม�ด�วยกำ�ร้ใช้� Linearly Biased Random Number Generator- สุ่ร้��ง Solution ม�ใหม�ด�วยกำ�ร้ Crossover โดยน�� Strip แต่�ละ Solution ที่��ได�สุ่��มเล�อกำม�น��ม� Crossover กำ�น และ Mutate ด�วยกำ�ร้น��ช้ )นง�นที่��เหล�อม�ไล�จำ�ดว�งใหม�- ที่��กำ�ร้ Adjust โดยเล�อกำ Strip ที่��ปร้ะสุ่ ที่ธิ ภ�พื่สุ่,งม�จำ�ดเร้�ยง และ น��ช้ )นง�นที่��เหล�อไล�จำ�ดว�งใหม�- น�� Solution ใหม�ที่��ได�ไปเป'น Population

Else (r p)- สุ่��มเล�อกำ Solution จำ�กำปร้ะช้�กำร้ แล�วสุ่��มเล�อกำ Strip โดยกำ�ร้สุ่��มแบบ Uniform Distributed Random เพื่��อน��ม�Crossover และ Mutate ด�วยกำ�ร้น��ม�ไล�จำ�ดว�งใหม�- น�� Solution ใหม�ที่��ได�ไปเป'น Population-จำ�ดเร้�ยง Population ด�วยคั�� Fitness และเกำ-บ Solution ที่��ต่ ดล��ด�บไว�

Until ปร้ะช้�กำร้ไม�เปล��ยน หร้�อ ปร้ะสุ่ ที่ธิ ภ�พื่ที่��ด�ที่��สุ่�ดคังที่�� หร้�อ คัร้บจำ��นวนคัร้�)ง ที่��กำ��หนด

Page 53: Lecture 1 4

53

Garment Layout using Garment Layout using GeneticGenetic

Page 54: Lecture 1 4

54

Food Inspection: TextureFood Inspection: Texture

ห� Co-occurrence matrixที่�� d[3,3] และ d[-3,3]

กำร้องเฉล��ย

LoG