1 Discovering Collocation Patterns: from Visual Words to Visual Phrases Junsong Yuan, Ying Wu and Ming Yang CVPR’07.

1

Discovering Collocation Patterns:from Visual Words to

Visual Phrases

Junsong Yuan, Ying Wu and Ming YangCVPR’07

2

Discovering Visual Collocation

VisualCollocations

VisualPrimimitives

AB AC

UV

W XY

ZT

U

TT

T B

V A A BCB

W S

T C... C

M JJ

AB AC A A BC

B C... C

...

...

Images

“Bag ofwords”

FeatureExtraction

FeatureQuantization

PatternDiscovery

3

An exciting idea: detour

• Related Work: J. Sivic et al. CVPR04, B. C. Russell et al. CVPR06, G. Wang et al. CVPR06, T. Quack et al. CIVR06, S. C. Zhu et al. IJCV05, …

Text &transaction

data

Imagedata

PLSALDAFIM

Textpattern

Visualpattern

Visualword

lexicon?

Image Data Text Data

?

4

Confrontation

• Spatial characteristics of images– over-counting co-occurrence frequency

• Uncertainty in visual patterns– Continuous visual feature quantized word– Visual synonym and polysemy

EDABCHTABCCSABKZTABMKFABVKSABTKAABEILABO

TRABW...

G12:G13:G67:G78:G79:G112:G113:G198:G215:G216:

AB

ID Group

5

Our Approach

Visual Primitive Dataset

,...},{ BA

A B C ... U V

A,BA,B,

C A,C B,C U,V

W ZYX

...

Clusteringe.g. K-means

Pattern Discovery e.g. FIM

PatternSummarization

New Metric: ADxd

MetricLearninge.g. NCA

wheelsCar

bodiesMeaningful

Patterns

Visualphraselexicon

Visualword

lexicon

6

Selecting visual phrases

• Visual collocations may occur by chance• Selecting phrases by a likelihood ratio test:

– H0: occurrence of phrase P is randomly generated– H1: phrase P is generated by a hidden pattern

• Prior: • Likelihood:

• Check if words are co-located together by chance or statistically meaningful

7

Frequent Word-sets ( |P|>=2 )

AB

ABF ABE

CD

CDE

DECEAE

BE BFAF

Discovery of visual phrases

A B F P

C D E S

A B F T

C D E X

A B D K

……

Closed

FIM

pair-wise student t-test

ranked by L(P)

AB

ABF

BF

AF

likelihood ratio

15.7

14.3

12.2

10.9

9.7CD}}{},{

},{},{},{{

CDBF

ABFAFAB

8

Frequent Itemset Mining (FIM)

• If an itemset is frequent then all of its subsets must also be frequent

nul l

AB AC AD AE BC BD BE CD CE DE

A B C D E

ABC ABD ABE ACD ACE ADE BCD BCE BDE CDE

ABCD ABCE ABDE ACDE BCDE

ABCDE

Pruned itemsets if{AB} is NOT frequent

9

Phrase Summarization

• Measuring the similarity between visual phrases by KL-divergence Yan et al., SIGKDD 05

• Clustering visual phrases by Normalized-cut

A,B

A,B,F

A,F

B,F

C,D

Normalizedcut

)()(

)()(),(

ji

jiji PGPG

PGPGPPS

G({A,B})

……

KDBA

XEDC

TFBA

SEDC

PFBA

……

KDBA

XEDC

TFBA

SEDC

PFBA

G({B,F})

……

KDBA

XEDC

TFBA

SEDC

PFBA

……

KDBA

XEDC

TFBA

SEDC

PFBA

10

Pattern Summarization Results

H1: wheelsPrc.: 97.5% Rec.: 22.8%

H2: car bodies Prc.: 71.3% Rec.:N.A.

Face database: summarizing top-10 phrases into 6 semantic phrase patterns

Car database: summarizing top-10 phrases into 2 semantic phrase patterns

11

Partition of visual word lexicon

backgroundwords

foregroundwords

}{: iWcodebook },...,,{ 21 MPPP

w2

w20

w77

w2

w1

W168

wk

w3

w4w5

PM

P3

P2

P1

P4

Visual phrases

Metric learning method: • Neighborhood component analysis (NCA).

Goldberger, et al., NIPS05

– improve the leave-one-out performance of the nearest neighbor classifier

12

Evaluation

• K-NN spatial group: K=5• Two image category database: car (123

images) and face (435 images)• Precision of visual phrase lexicon

– the percentage of visual phrases Pi ∈ Ψ that are located in the foreground object

• Precision of background word lexicon– the percentage of background words Wi ∈ Ω−

that are located in the background

• Percentage of images that are retrieved:

13

Results: visual phrases from car category

Visual phrase pattern 1: wheels different colors represent different semantic meanings

Visual phrase pattern 2: car bodies

14

Results: visual phrases from face category

15

Comparison