http:// computer graphics laboratory, hiroshima city university all images are compressed

26
http://www.hiroshima-cu.ac.jp Computer Graphics Laboratory, Hiroshima City Univer All images are compressed.

Upload: shanon-mccoy

Post on 30-Dec-2015

214 views

Category:

Documents


1 download

TRANSCRIPT

http://www.hiroshima-cu.ac.jp/

Computer Graphics Laboratory, Hiroshima City University

All images are compressed.

http://www.hiroshima-cu.ac.jp/

Computer Graphics Laboratory, Hiroshima City University

Interactive shadow removal from a single image using hierarchical graph cut

Daisuke Miyazaki

Yasuyuki Matsushita

Katsushi Ikeuchi

http://www.hiroshima-cu.ac.jp/

CG Lab, Hiroshima City University

Motivation

http://www.hiroshima-cu.ac.jp/

CG Lab, Hiroshima City University

Prior work

multiple

singleautomatic

interactive

Weiss 01; Matsushita et al. 04

Finlayson et al. 04; Fredembach & Finlayson 06;Tappen et al. 05; Baba et al. 04; Arbel & Hel-Or 07;Nielsen & Madsen 07

Wu et al. 07

http://www.hiroshima-cu.ac.jp/

CG Lab, Hiroshima City University

Contribution

PriorPriorSmoothnessSmoothness

Graph cut

)(nO )(lognOAlpha expansion

User satisfaction

Robust

Intuitive

Fast

http://www.hiroshima-cu.ac.jp/

CG Lab, Hiroshima City University

Algorithm flow

Input Over-segmentation Lazy snapping

Shadow removal

Initial Interaction Output

11 22 33

44 55 66

http://www.hiroshima-cu.ac.jp/

CG Lab, Hiroshima City University

Shadow matte

FI

Prior of shadowless image

Prior of shadow image

Smoothness of shadowless image

Smoothness of shadow image

Solve

http://www.hiroshima-cu.ac.jp/

CG Lab, Hiroshima City University

Prior of shadowless image

Likelihood P

appropriate largesmall

)(log pFP

http://www.hiroshima-cu.ac.jp/

CG Lab, Hiroshima City University

Prior of shadow image

small largeappropriate

7.07.0

0 1 pp

L0.7-NormL1-NormL2-Norm

http://www.hiroshima-cu.ac.jp/

CG Lab, Hiroshima City University

Smoothness of shadowless image

small large

2

pF

appropriate

http://www.hiroshima-cu.ac.jp/

CG Lab, Hiroshima City University

Smoothness of shadow image

small largeappropriate

qp

http://www.hiroshima-cu.ac.jp/

CG Lab, Hiroshima City University

Smoothnessshadow

0

1

Interactive parameter optimization

Smoothnessshadowless

Priorshadowless

Priorshadow

0

0.25

0

0.25

0.4

0.7

InitInit InitialInitial1st1st2nd2nd3rd3rd4th4th5th5th6th6th

http://www.hiroshima-cu.ac.jp/

CG Lab, Hiroshima City University

Approach based on alpha expansion

Alpha expansion Hierarchical graph cut

alpha=2

alpha=2,4,...

alpha=4

3

1 52 4

1

1

3

5

5

1

1

3

5

5

3

512

12

3

54

54

12 1

25

4

http://www.hiroshima-cu.ac.jp/

CG Lab, Hiroshima City University

Hierarchical structure

32

16 48

8 24 40 56

8 16 24 32 40 48 56

16 32 48

8 16 24 32 40 48 56

16 32 48

20

http://www.hiroshima-cu.ac.jp/

CG Lab, Hiroshima City University

Graph construction example

p a q

),(),(,0max),()()( qpqpqpqp VVVDD

),(),(,0max),()()( qpqpqpqp VVVDD

),()()( qpqp VDD

),()()( qpqp VDD

),(),( qpqp VV

qp

http://www.hiroshima-cu.ac.jp/

CG Lab, Hiroshima City University

Algorithm

Alpha expansion Hierarchical graph cut

1. A = {{0}, {32}, {16}, {48},{8, 40}, {24, 56},{4, 20, 36, 52}, {12, 28, 44, 60}, ...,{3, 7, 11, ..., 63}}

2. for i = 0 to 11 doa. for all nodes&edges do

I. II. add nodes&edges using &

b. solve max-flow/min-cut problem3. Iterate 2 until convergence

piAp ][minarg

p p

1. A = {0, 1, 2, ..., 63}

2. for i = 0 to 63 doa. for all nodes&edges do

I. II. add nodes&edges using &

b. solve max-flow/min-cut problem3. Iterate 2 until convergence

][iA p

http://www.hiroshima-cu.ac.jp/

CG Lab, Hiroshima City University

Output of Hierarchical graph cut

Stereomatching

[Art]

Stereomatching[Laundry]

Imagerestoration

Shadowremoval

Input Ground truth Ishikawa 2003 expansion Hierarchical cut

http://www.hiroshima-cu.ac.jp/

CG Lab, Hiroshima City University

Speed of hierarchical graph cut

Stereo matching[Art]

Stereo matching[Laundry]

Image restoration

Shadow removal

Speed-up 6.8 11.4 16.6 3.4

Error difference +4.6% +3.0% -0.4% +0.0%

Hierarchical graph cut

Alpha expansion

[sec] Iteration250 7

EnergyEnergy

0

1.1107 1.7106

1.5106

http://www.hiroshima-cu.ac.jp/

CG Lab, Hiroshima City University

Incremental improvement

Input Truth Finlayson 2004 Wu 2007 Initial Interactive

[Ours]

Error

Number of strokes for parameter optimization 50

Finlayson 2004

Wu 2007Ours

http://www.hiroshima-cu.ac.jp/

CG Lab, Hiroshima City University

0 strokes (param. opt.) 1.4 sec 1 strokes 2.1 sec 6 strokes

Natural images

1.2 sec 4 strokes

http://www.hiroshima-cu.ac.jp/

CG Lab, Hiroshima City University

2.0 sec 12 strokes 1.5 sec 2 strokes

Natural images

1.9 sec 15 strokes1.5 sec 8 strokes

http://www.hiroshima-cu.ac.jp/

CG Lab, Hiroshima City University

5.8 sec 27 strokes 2.8 sec 40 strokes

Natural images

6.3 sec 32 strokes2.0 sec 26 strokes

http://www.hiroshima-cu.ac.jp/

CG Lab, Hiroshima City University

Aerial images

http://www.hiroshima-cu.ac.jp/

CG Lab, Hiroshima City University

Conclusions

Hierarchical graph cut

Cost function User interaction

http://www.hiroshima-cu.ac.jp/

CG Lab, Hiroshima City University

Discussions

Cost function

User interaction

Hierarchical graph cut

http://www.hiroshima-cu.ac.jp/

Computer Graphics Laboratory, Hiroshima City University

(c) Daisuke Miyazaki 2009All rights reserved.

http://www.cg.info.hiroshima-cu.ac.jp/Daisuke Miyazaki, Yasuyuki Matsushita, Katsushi Ikeuchi, "Interactive shadow removal from a single

image using hierarchical graph cut," Lecture Notes in Computer Science (Proc. Asian Conference on

Computer Vision), vol. 5994, part 1, 2009.