Voginip lezing 2015: Classificeren zonder voorbeelden

Visual Classification without examplesClassificeren van beeld zonder voorbeelden

Thomas MensinkVOGIN-IP-LEZING 2015

• What is an axolotl?

• Some examples

Preview

VOGIN-IP 20152

VOGIN-IP 20153

We can classify based on labeled examples(supervised learning)

Preview

• What is an aye-aye?

• Textual description:

– Is nocturnal

– Lives in trees

– Has large eyes

– Has long middle fingers

VOGIN-IP 20154

VOGIN-IP 20155

We can classify based on a textual description(and some prior knowledge)

VOGIN-IP 20156

Can a computer do the same?(yes, that is what this talk is about)

Agenda

• Supervised Visual Classification

• Attribute-Based Classification

• Co-occurrence Based Classification

VOGIN-IP 20157

Computer Vision – in the news

VOGIN-IP 20158

Visual Recognition

9

Cityscap

eOutdoor

…

tree

Buildin

gLamp

People

John

Dam

Slide credit: Jan van Gemert

VOGIN-IP 2015

Supervised Classification

• Obtain annotated examples

• Find a representation

• Train a generic classifier

VOGIN-IP 201510

Remarks:- New class: retrain on new examples- How to obtain training examples?- How to represent images?

Visual Classification: Two trends

Datasets

• 2005: motorbikes, bicycles, people, cars

• Since 2010: ImageNet15K

Representations

• 2005: Manual derived features/encodings

• 2012: Trained end-to-end, Deep Neural Nets

VOGIN-IP 201511

VOGIN-IP 201512

1000 classes, 5 guesses per imageCurrent state-of-the-art: 6.7% error

Estimating human performance

VOGIN-IP 201513

Andrej Karpathy: “I realized that I needed to go through the painfully long training process myself” Test set 1500 images

GoogleNet: 6.8% errorKarpathy: 5.1% error

VOGIN-IP 201514

Visual Classification:near-human performance

when ample train data is available

Attribute-based classification

VOGIN-IP 201515

1. Define vocabulary2. Train visual classifiers3. Class to attribute mapping4. Infer class

What are good attributes?

• Good attributes

– are task and category dependent;

– class discriminative, but not class specific;

– interpretable by humans; and

– detectable by computers

VOGIN-IP 201516

Quiz: What are good attributes?

• is grey?

• is made of atoms?

• lives in Amsterdam?

• is sunny?

• eat fish?

• has a SIFT descriptor with empty bin 3?

• has 4 wheels?

VOGIN-IP 201517

How many attributes?

• In theory k binary attributes can represent

– 2k classes

• In practice for c classes we need

– Many attributes

VOGIN-IP 201518

Animals with Attributes

VOGIN-IP 201519

Animals with Attributes - Vocabular

VOGIN-IP 201520

Class to attribute mapping

VOGIN-IP 201521

Attribute Based Prediction

1. Learn attribute classifiers from related classes

2. Train and Test set are disjoint

3. Infer attributes from new test image

4. Use attribute-to-class mapping to predict class

VOGIN-IP 201522

Animals with Attributes (results)

VOGIN-IP 201523

Disadvantages

• Unnatural distinction between

– Attributes to be detected

– Classes of interest

• Inherently multi-class zero-shot prediction

VOGIN-IP 201524

Classification based on co-occurrences

I’m looking for a label, which I have not seenbefore. However, this picture contains also:

1. Indoor

2. Living room

3. Table

4. Chair

VOGIN-IP 201525

VOGIN-IP 201526

We can classify based on context

COSTA: Design

VOGIN-IP 201527

COSTA: Design

VOGIN-IP 201528

• Many visual concepts can be described as an open set of concept-to-concept relations

• Describe image semantics with co-occurrences

• Exploit natural bias in natural images

Exploit natural bias in natural images

VOGIN-IP 201529

• k

•

• ŵl

slk ak"

•

:"concept)to)concept"inter)dependencies"reveal"a"

significant"part"of"the"latent"image"seman6cs"

Co)occurrences"from"Flickr"capture"well"the"concept)

to)concept"rela6ons

Regressing"classifiers"with"posi6ve"&"nega6ve"co)occurrences"lead"to"

surprisingly"good"results

Sink"is"usually"in"the"same"visual"space"as"a"cupboard,"a"stove,"and"a"dishwasher.

H-SUN

Sett ing SUB L1O ZS75 ZS50

Nr. Train labels 107 106 81 54

Baselines

Supervised SVM 21.5 – – –

Attributes – 12.8 13.0 12.3

COSTA

Co-oc Dice – 14.5 14.5 12.9

P&N Dice – 13.7 13.8 10.8

Reg P&N Dice – 17.0 16.4 15.0

COSTA: Internet co-occurrences

Web hit counts – 9.9 9.8 9.8

Imagehit counts – 12.7 9.1 9.3

Flickr hit counts – 15.1 13.4 10.1

1Zero 2 4 8 16 32 64 128 All

25

30

35

40

45

Number of images

mA

P

iCLEF10

Data

Data+ Prior

Data+ WebPrior

Zero 2 4 8 16 32 64 128 All

10

15

20

Few Many

Number of images

mA

P

H-SUN

Zero 2 4 8 16 32 64 128 All

11

12

13

14

15

16

Number of images

mA

P

CUB-Att

Figure3. Few-shot classificationusingregressionbasedzero-shot prior fromground-truthor webhit-counts(except CUB-Att). Including

theprior significantly benefitsperformance, especially inthepresenceof very fewpositiveimages. Notethelog-scaleonthex-axis.

References

[1] Z.Akata, F.Perronnin, Z.Harchaoui, andC.Schmid. Label-

embeddingforattribute-basedclassification. InCVPR,2013.

2, 5

[2] E.Bart andS.Ullman. Cross-generalization: Learningnovel

classes from a single example by feature replacement. In

CVPR, 2005. 2

[3] G. Chen, Y. Ding, J. Xiao, andT. Han. Detectionevolution

withmulti-order contextual co-occurrence. InCVPR, 2013.

3

[4] M. Choi, J. Lim, A. Torralba, and A. Willsky. Exploiting

hierarchical context onalargedatabaseof object categories.

InCVPR, 2010. 3, 5

[5] M. Everingham, L. Van Gool, C. Williams, J. Winn, and

A. Zisserman. ThePascal visual object classes(VOC) chal-

lenge. IJCV, 2010. 2, 5

[6] R.-E. Fan, K.-W. Chang, C.-J. Hsieh, X.-R. Wang, and C.-

J. Lin. Liblinear: A library for large linear classification.

JMLR, 2008. 4

[7] H. Grabner, J. Gall, and L. V. Gool. What makesachair a

chair? InCVPR, 2011. 1

[8] B. Hariharan, S. Vishwanathan, and M. Varma. Efficient

max-margin multi-label classification with applications to

zero-shot learning. MLJ, 2012. 3

[9] H. Jegou and O. Chum. Negative evidences and co-

occurrences in image retrieval: the benefit of PCA and

whitening. InECCV, 2012. 3

[10] H. Jegou, M. Douze, and C. Schmid. On theburstinessof

visual elements. InCVPR, 2009. 3

[11] L. Ladicky, C. R. P. Kohli, and P. Torr. Graph cut based

inferencewithco-occurrencestatistics. InECCV, 2010. 3

[12] C. Lampert, H. Nickisch, and S. Harmeling. Attribute-

based classification for zero-shot learning of object cate-

gories. IEEETrans. PAMI, 2013. 1, 2, 3, 6, 7

[13] F.-F. Li, R. Fergus, andP. Perona. One-shot learningof ob-

ject categories. IEEETrans. PAMI, 2006. 2

[14] T. MalisiewiczandA. Efros. Beyondcategories: thevisual

memex model for reasoning about object relationships. In

NIPS, 2009. 1, 3

[15] T.Mensink, J.Verbeek, andG.Csurka. Tree-structuredCRF

models for interactive image labeling. IEEE Trans. PAMI,

2012. 3, 5

[16] T. Mensink, J. Verbeek, F. Perronnin, and G. Csurka.

Distance-based image classification: Generalizing to new

classesat near-zerocost. IEEETrans. PAMI, 2013. 1, 2

[17] S.NowakandM.Huiskes. Newstrategiesfor imageannota-

tion: Overview of thephotoannotation task at ImageCLEF

2010. InWorkingNotesof CLEF, 2010. 5

[18] F. Orabona, C. Castellini, B. Caputo, A. E. Fiorilla, and

G. Sandini. Model adaptation with least-squares svm for

adaptivehandprosthetics. InIEEEICRA, 2009. 4

[19] M. Rastegari, A. Farhadi, andD. Forsyth. Attributediscov-

ery viapredictablediscriminativebinary codes. In ECCV,

2012. 2

[20] M. Rohrbach, M. Stark, andB. Schiele. Evaluatingknowl-

edgetransfer andzero-shot learning inalarge-scalesetting.

InCVPR, 2011. 1, 2

[21] M. Rohrbach, M. Stark, G. Szarvas, I. Gurevych, and

B. Schiele. What helpswhere–and why? semantic relat-

ednessfor knowledgetransfer. InCVPR, 2010. 2, 4, 6

[22] M. Sadeghi and A. Farhadi. Recognition using visual

phrases. InCVPR, 2011. 3

[23] R. Salakhutdinov, A. Torralba, andJ. Tenenbaum. Learning

tosharevisual appearancefor multiclassobject detection. In

CVPR, 2011. 2

[24] J. Sanchez, F. Perronnin, T. Mensink, and J. Verbeek. Im-

ageclassificationwiththefishervector: Theoryandpractice.

IJCV, 2013. 3, 4

[25] V.Sharmanska, N.Quadrianto, andC.Lampert. Augmented

attributerepresentations. InECCV, 2012. 2

[26] T. Tommasi and B. Caputo. Themoreyou know, the less

you learn: fromknowledgetransfer toone-shot learning of

object categories. InBMVC, 2009. 2

[27] T. Tommasi, F. Orabona, andB. Caputo. Safety innumbers:

Learning categories from few examples with multi model

knowledgetransfer. InCVPR, 2010. 4

[28] C.Wah, S.Branson, P.Welinder,P.Perona, andS.Belongie.

Thecaltech-ucsd birds-200-2011 dataset. Technical report,

Computation& Neural Systems, 2011. 1, 2, 3, 5

[29] J. Xiao, J. Hays, K. A. Ehinger, A. Oliva, and A. Torralba.

SUNdatabase: Large-scalescenerecognitionfromabbey to

zoo. InCVPR, 2010. 2, 5

8

Exis6ng"classifiers" Co)occurrences" Zero)shot"predic6on"

Co-occurrencestatistics. Co-occurrenceshavebeen re-

peatedlyconsideredforcapturinghigherorder relationships

betweenclasses, conceptsandlabels. They havebeenused

to improve image segmentation [11], object detection [3,

22], to reasonabout what toexpect inascene[4, 14], and

for image and attribute classification [8, 15]. In general,

wenoticethat co-occurrencesof objects, labelsandtextures

havebeenrecognizedasastrongcluefor label andattribute

prediction. However, wearenot awareof any works that

uselabel co-occurrencestoenablezero-shot classification.

3. Zero-shot multi-label classification

In thissectionwedefineour zero-shot framework using

label co-occurrences. Wefirst introduceour co-occurrence

basedclassificationmethod. Second, inSection3.2wede-

scribearegressioninterpretationtoobtainzero-shot classi-

fier. Third, inSection3.3weillustratetheuseof thezero-

shot model asprior for few-shot classification.

3.1. Co-occurrencebasedclassification

Our goal is to estimate aclassification function for an

unseenlabel l,usingaset of existingclassifiers.Weassume

thatwehaveasetof linearclassifiers, trainedonacollection

of annotated imageswithk labels. Theseclassifierscould

beobtained frombinary SVMsor logistic regression, and

arerepresented by their weight vectorswk 2 Rd⇥1. We

assume, without lossof generality, that theweight vectors

areaugmentedsuchthat thebiasesareincluded.

Weproposetoestimatetheweight vector wl toclassify

theunseenlabel l, as

wl =X

k

wk slk, (1)

where slk represents a measure of similarity between the

known label k and the unseen label l. In this paper, we

base these similarities on the co-occurrence statistics be-

tweenthenewlabel andexistinglabels.

Co-occurrence similarities. We explore different simi-

larity measuresbased on theco-occurrenceof two labels.

Let ci j denotethetotal number of imagesfor whichlabel i

and label j arerelevant according toanauxiliary resource,

for exampletheground-truthlabelling,awebsearchengine

or auser providedinput inthecaseof activelearning. Also,

ci denotesthetotal number of imagesdepictinglabel i, and

m denotesthetotal number of labels. Thesimilaritieswe

exploreare:

• Normalizedco-occurrences,

sni j =

ci j

ci

, (2)

where the similarity is directly proportional to the

number of co-occurrences.

• Binarized co-occurrences, motivated by thebinarized

class-to-attributes mappings used in attribute-based

zeroshot classification[12, 28]:

sbi j = [[ci j ≥ t]], (3)

wheret isaglobal thresholdset ast = 1m2

Pi ,j ci j .

• Burstinesscorrectedco-occurrences. Burstinessisthe

phenomenathat thefrequencyof anobservationissig-

nificantly larger thanastatistically independent model

wouldpredict. Thesquare-root functionisusedinim-

ageretrieval andclassificationtocorrect for burstiness

of visual features[10,24]. Similarly,weaimtoreduce

theburstinessinlabellings, anduse:

ssi j =

pci j , (4)

• TheDice’scoefficient,

sdi j =

ci j

ci + cj

, (5)

isameasureusedinmany Natural LanguageProcess-

ingsystems. It aimstoestimatethesemantic related-

nessbetweentwoterms, basedonhit countsfromweb

searchengines.

Definingaconcept bywhat it isnot. Thesimilaritiesde-

finedabove,arebasedonlyonpositiveco-occurrences, i.e.,

how often two labelsarerelevant for an image. However,

knowingwhat isnot relatedtotheconcept might beavery

informativeclueabout thevisual scopeof aconcept, which

isalsoshowninanimageretrieval setting[9].

In addition to the positive co-occurrences, denoted by

c++i j , wealsousetheother possibleco-occurrencerelations:

thepresenceof label i with theabsenceof label j , theab-

senceof label i withthepresenceof label j ,andtheabsence

of bothlabels, denotedbyc+–i j ,c–+i j , andc––i j respectively. For

eachof thesedefinitionsof co-occurrenceweusethesimi-

laritymeasuresdefinedabove.

Using the positive and negative co-occurrences, the

weight vector wl of anunknownlabel canbeestimatedas:

wl =X

k

wk s++lk − wk s+–

lk − wk s–+lk + wk s––lk , (6)

Estimateco-occurrencefromwebdata. Sofar,wehave

not discussedhowtoobtaintheco-occurrencestatisticsre-

quired for our zero-shot recognition framework. To show

the potential of our framework, in most of the experi-

mentsweestimate label co-occurrences fromtheground-

truth labellingof our imagedatasets. Alternatively, theco-

occurrencestatisticscouldbeestimatedfromlargetext cor-

pora,e.g.,WordnetorWikipedia,or internetsearchengines,

suchasYahoo, GoogleandBing.

3

Following [21], we use the hit counts estimations of

theYahoowebsearch, Yahoo imagesearchandFlickr-tag

searchenginestoestimatetoco-occurrences. Weestimate

the similaritiesdefined above, using cHCi j denoting the hit

count for aquery consistingof label i and j . Thepositive

andnegativeco-occurrencesareestimatedby using thehit

countscHCi of theindividual labelsi, andanestimateof the

total number of imagescHCtotal =

Pi ci .

3.2. Regression toestimateclassifiers

Toimprovetheclassifier for theunseenlabel l, wepro-

posetolearnaweightedversionof Eq. (1), givenby:

wl =X

k

ak wk slk, (7)

whereak isaweight for classifier k, which isindependent

fromtheunseenlabel l.

Ideally,weset theclassifierweightsa 2 Rk⇥1 suchthat,

theestimatedweight vector wl equalstheideal weight vec-

tor wl , whichwouldhavebeenbeobtainedby learningon

thevisual datawith annotations available. This could be

seen asaregression problem, wherea isset to regress w

towardsw. However, sinceweaimfor zero-shot classifica-

tion, wedonot haveaccessto theunseen labels l, nor the

ideal weight vectorswl at traintime. Thereforeweusethe

knownlabelsk inaleave-one-out settingfor learning.

Weminimizethefollowingregressionsquared-loss:

Lreg =X

i

kwi −X

k

akwk si kk22, (8)

=X

i

X

d

wi d − a> vi d2, (9)

where index i and k both run over theknown labels, and

si i = 0. The vector vi d contains the k co-occurrence

weightedweight vectorsvi dk = si k wkd.

Notethat thelossisformulatedover trainclassesandnot

over train images. Moreover, Eq. (9) showsthat a can be

obtainedinclosed-formusingridge-regression. Inpractice

weobservethat regularization isnot needed for goodper-

formance, thedimensionality of a ismuchsmaller thanthe

number of traininginstances(k vs. dk).

3.3. Zero-shot prior for few-shot prediction

In a few-shot classification setting afew, e.g., up to 8,

positiveimagesareprovidedper label to learnaclassifier.

In such a setting a strong prior could benefit the perfor-

manceby guiding theSVM classifier. In this section we

consider asimplemodel adaptationmethodwherethezero-

shot model actsasaprior for thefew-shot classifier.

In thecasethat weemploy linear SVM classifierswith

squaredhinge-loss, theobjectivetominimizebecomes:

Lfew =C

2

X

i

⇥1− yi l w>

l x i

⇤2+

+1

2kwl − βwzk

22, (10)

wherewz is theprior obtained fromthezero-shot model,

andβ 2 (0,1) isascalingparameter tocontrol thedegree

to which thelabel classifier should besimilar to thezero-

shot model. It can beshown that theoptimal solution for

wl , for aspecificvalueof C, isgivenby [18, 27]:

wl = wg + βwz, (11)

wherewg istheweight vectorobtainedfromoptimizingthe

standardSVMformulation, i.e., usingEq. (10) withβ = 0.

Wewill usewg alsoasabaselinefew-shot classifier. Note

that theoptimal parameter C could differ when including

theprior. Inourexperiments,wefirstdeterminetheoptimal

valueof C forwg usingcross-validation, thenweobtainwl

byusingβ = 1.

4. Experiments

Inthissectionweexperimentallyvalidateourmodelsfor

zero-shot andfew-shot imagelabellingusingco-occurrence

statistics. Sinceweareinterested inazero-shot classifica-

tionsetting, for most experimentswesplit thelabelsof the

datasetsintotwodisjoint sets: theknownclassesandtheun-

seenclasses. Thetruezero-shot classifiersuselabelsfrom

theknownclassesonly. Sincewearenotawareof anymeth-

ods that do zero-shot recognition on multi-label datasets

nor usingmulti-label co-occurrencestatistics, wecompare

with thebinarized versionof co-occurences, theclosest to

what attributescouldbefor multi-labeleddatasets. Further-

more, for all datasetswereport resultsobtained in afully

supervisedsetting, wheretheSVMclassifiersusethetrain-

inglabelsfromall classes, beit knownor unseen.

4.1. Experimental set-upanddatasets

Imagefeatures. Forall ourexperimentsweusetheFisher

Vector (FV) image representation [24]. Per image a sin-

gleFV x isextracted, and wefollow acommon pipeline

and use: (i) SIFT descriptors, projected with PCA to 96-

dimensions; (ii) Mixture-of-Gaussian codebook with 16

components; (iii) power-normalization and `2 normaliza-

tion. Thefinal FV isonly3K dimensional, despitethecom-

pactnessitsperformanceisstill competitive.

Implementation. For all experimentswhereSVMclassi-

fiersareused, wetrain linear SVMs[6] and employ two-

fold cross-validation on the train data to set the value of

C. Performance is measured by mean AveragePrecision

(mAP). For the few-shot and zero-shot experiments, the

mAP is averaged only over test labels. For the few-shot

experiments, wefixβ = 1, seeEq. (11).

Wealso report thesupervised upper bound (SUB) per-

formance, obtained by training SVMson thefull ground-

truth annotations. The SUB serves as the ideal classifier

whichcouldbeobtainedfromthisdata.

4

Following [21], we use the hit counts estimations of

theYahoo websearch, Yahoo imagesearch andFlickr-tag

searchenginestoestimatetoco-occurrences. Weestimate

the similaritiesdefined above, using cHCi j denoting the hit

count for aquery consisting of label i and j . Thepositive

andnegativeco-occurrencesareestimatedby using thehit

countscHCi of theindividual labelsi, andanestimateof the

total number of imagescHCtotal =

Pi ci .

3.2. Regression toestimateclassifiers

Toimprovetheclassifier for theunseen label l, wepro-

posetolearnaweightedversionof Eq. (1), givenby:

wl =X

k

ak wk slk, (7)

whereak isaweight for classifier k, which isindependent

fromtheunseenlabel l.

Ideally,weset theclassifierweightsa 2 Rk⇥1 suchthat,

theestimatedweight vector wl equalstheideal weight vec-

tor wl , whichwouldhavebeenbeobtainedby learningon

the visual data with annotations available. This could be

seen asaregression problem, wherea isset to regress w

towardsw. However, sinceweaimfor zero-shot classifica-

tion, wedo not haveaccessto theunseen labels l, nor the

ideal weight vectorswl at traintime. Thereforeweusethe

knownlabelsk inaleave-one-out settingfor learning.

Weminimizethefollowingregressionsquared-loss:

Lreg =X

i

kwi −X

k

akwk si kk22, (8)

=X

i

X

d

wi d − a> vi d2, (9)

where index i and k both run over theknown labels, and

si i = 0. The vector vi d contains the k co-occurrence

weightedweight vectorsvi dk = si k wkd.

Notethat thelossisformulatedover trainclassesandnot

over train images. Moreover, Eq. (9) showsthat a can be

obtainedinclosed-formusingridge-regression. Inpractice

weobservethat regularization isnot needed for good per-

formance, thedimensionality of a ismuchsmaller thanthe

number of traininginstances(k vs. dk).

3.3. Zero-shot prior for few-shot prediction

In a few-shot classification setting a few, e.g., up to 8,

positiveimagesareprovided per label to learnaclassifier.

In such a setting a strong prior could benefit the perfor-

mance by guiding theSVM classifier. In this section we

consider asimplemodel adaptationmethodwherethezero-

shot model actsasaprior for thefew-shot classifier.

In thecasethat weemploy linear SVM classifierswith

squaredhinge-loss, theobjectivetominimizebecomes:

Lfew =C

2

X

i

⇥1− yi l w>

l x i

⇤2+

+1

2kwl − βwzk

22, (10)

wherewz is theprior obtained fromthezero-shot model,

andβ 2 (0,1) isascalingparameter tocontrol thedegree

to which thelabel classifier should besimilar to thezero-

shot model. It can beshown that theoptimal solution for

wl , for aspecificvalueof C, isgivenby [18, 27]:

wl = wg + βwz, (11)

wherewg istheweight vectorobtainedfromoptimizingthe

standardSVM formulation, i.e., usingEq. (10) withβ = 0.

Wewill usewg alsoasabaselinefew-shot classifier. Note

that theoptimal parameter C could differ when including

theprior. Inour experiments,wefirstdeterminetheoptimal

valueof C forwg usingcross-validation, thenweobtainwl

byusingβ = 1.

4. Experiments

Inthissectionweexperimentallyvalidateourmodelsfor

zero-shot andfew-shot imagelabellingusingco-occurrence

statistics. Sinceweareinterested inazero-shot classifica-

tionsetting, for most experimentswesplit thelabelsof the

datasetsintotwodisjoint sets: theknownclassesandtheun-

seenclasses. Thetruezero-shot classifiersuselabelsfrom

theknownclassesonly. Sincewearenotawareof anymeth-

ods that do zero-shot recognition on multi-label datasets

nor usingmulti-label co-occurrencestatistics, wecompare

with thebinarized version of co-occurences, theclosest to

what attributescouldbefor multi-labeleddatasets. Further-

more, for all datasetswereport resultsobtained in afully

supervisedsetting, wheretheSVMclassifiersusethetrain-

inglabelsfromall classes, beit knownor unseen.

4.1. Experimental set-upanddatasets

Imagefeatures. Forall ourexperimentsweusetheFisher

Vector (FV) image representation [24]. Per image a sin-

gleFV x isextracted, and wefollow acommon pipeline

and use: (i) SIFT descriptors, projected with PCA to 96-

dimensions; (ii) Mixture-of-Gaussian codebook with 16

components; (iii) power-normalization and `2 normaliza-

tion. Thefinal FV isonly3K dimensional, despitethecom-

pactnessitsperformanceisstill competitive.

Implementation. For all experimentswhereSVMclassi-

fiersareused, wetrain linear SVMs[6] and employ two-

fold cross-validation on the train data to set the value of

C. Performance is measured by mean Average Precision

(mAP). For the few-shot and zero-shot experiments, the

mAP is averaged only over test labels. For the few-shot

experiments, wefixβ = 1, seeEq. (11).

Wealso report thesupervised upper bound (SUB) per-

formance, obtained by training SVMson thefull ground-

truth annotations. The SUB serves as the ideal classifier

whichcouldbeobtainedfromthisdata.

4

Co-occurrencestatistics. Co-occurrences havebeen re-











Inthissectionwedefineour zero-shot framework using










of annotated imageswithk labels. Theseclassifierscould


are represented by their weight vectorswk 2 Rd⇥1. We





wl =X

k

wk slk, (1)








and label j arerelevant according toanauxiliary resource,





exploreare:


sni j =

ci j

ci

, (2)






sbi j = [[ci j ≥ t]], (3)


Pi ,j ci j .





ageretrieval andclassificationtocorrect for burstiness



ssi j =

pci j , (4)


sdi j =

ci j

ci + cj

, (5)




searchengines.


finedabove, arebasedonlyonpositiveco-occurrences, i.e.,









of bothlabels, denotedbyc+–i j ,c–+

i j , andc––i j respectively. For


larity measuresdefinedabove.



wl =X

k










pora,e.g.,WordnetorWikipedia,or internet searchengines,


3

Co-occurrencestatistics. Co-occurrences havebeen re-











In thissectionwedefineour zero-shot framework using










of annotated imageswith k labels. Theseclassifierscould


are represented by their weight vectorswk 2 Rd⇥1. We





wl =X

k

wk slk, (1)








andlabel j arerelevant according toanauxiliary resource,





exploreare:


sni j =

ci j

ci

, (2)






sbi j = [[ci j ≥ t]], (3)


Pi ,j ci j .





ageretrieval andclassification tocorrect for burstiness



ssi j =

pci j , (4)


sdi j =

ci j

ci + cj

, (5)




searchengines.


finedabove, arebasedonlyonpositiveco-occurrences, i.e.,









of bothlabels, denotedbyc+–i j ,c–+

i j , andc––i j respectively. For


larity measuresdefinedabove.



wl =X

k










pora,e.g.,WordnetorWikipedia,or internet searchengines,


3

Zero 2 4 8 16 32 64 128 All

25

30

35

40

45

Number of images

mA

P

iCLEF10

Data

Data+ Prior

Data+ WebPrior

Zero 2 4 8 16 32 64 128 All

10

15

20

Few Many

Number of images

mA

P

H-SUN

Zero 2 4 8 16 32 64 128 All

11

12

13

14

15

16

Number of images

mA

P

CUB-Att

Figure3. Few-shot classificationusingregressionbasedzero-shot prior fromground-truthor webhit-counts(except CUB-Att). Including

theprior significantly benefitsperformance, especially inthepresenceof very fewpositiveimages. Notethelog-scaleonthex-axis.

References

[1] Z.Akata,F.Perronnin, Z.Harchaoui, andC.Schmid. Label-

embeddingforattribute-basedclassification. InCVPR,2013.

2, 5

[2] E.Bart andS.Ullman. Cross-generalization: Learningnovel

classes from a single example by feature replacement. In

CVPR, 2005. 2

[3] G. Chen, Y. Ding, J. Xiao, andT. Han. Detectionevolution

withmulti-order contextual co-occurrence. InCVPR, 2013.

3

[4] M. Choi, J. Lim, A. Torralba, and A. Willsky. Exploiting

hierarchical context onalargedatabaseof object categories.

InCVPR, 2010. 3, 5

[5] M. Everingham, L. Van Gool, C. Williams, J. Winn, and

A. Zisserman. ThePascal visual object classes(VOC) chal-

lenge. IJCV, 2010. 2, 5

[6] R.-E. Fan, K.-W. Chang, C.-J. Hsieh, X.-R. Wang, and C.-

J. Lin. Liblinear: A library for large linear classification.

JMLR, 2008. 4

[7] H. Grabner, J. Gall, andL. V. Gool. What makesachair a

chair? InCVPR, 2011. 1

[8] B. Hariharan, S. Vishwanathan, and M. Varma. Efficient

max-margin multi-label classification with applications to

zero-shot learning. MLJ, 2012. 3

[9] H. Jegou and O. Chum. Negative evidences and co-

occurrences in image retrieval: the benefit of PCA and

whitening. InECCV, 2012. 3

[10] H. Jegou, M. Douze, and C. Schmid. On theburstinessof

visual elements. InCVPR, 2009. 3

[11] L. Ladicky, C. R. P. Kohli, and P. Torr. Graph cut based

inferencewithco-occurrencestatistics. InECCV, 2010. 3

[12] C. Lampert, H. Nickisch, and S. Harmeling. Attribute-

based classification for zero-shot learning of object cate-

gories. IEEETrans. PAMI, 2013. 1, 2, 3, 6, 7

[13] F.-F. Li, R. Fergus, andP. Perona. One-shot learningof ob-

ject categories. IEEETrans. PAMI, 2006. 2

[14] T. MalisiewiczandA. Efros. Beyondcategories: thevisual

memex model for reasoning about object relationships. In

NIPS, 2009. 1, 3

[15] T.Mensink, J.Verbeek,andG.Csurka. Tree-structuredCRF

models for interactive image labeling. IEEE Trans. PAMI,

2012. 3, 5

[16] T. Mensink, J. Verbeek, F. Perronnin, and G. Csurka.

Distance-based image classification: Generalizing to new

classesat near-zerocost. IEEETrans. PAMI, 2013. 1, 2

[17] S.NowakandM.Huiskes. Newstrategiesfor imageannota-

tion: Overview of thephotoannotation task at ImageCLEF

2010. InWorkingNotesof CLEF, 2010. 5

[18] F. Orabona, C. Castellini, B. Caputo, A. E. Fiorilla, and

G. Sandini. Model adaptation with least-squares svm for

adaptivehandprosthetics. InIEEEICRA, 2009. 4

[19] M. Rastegari, A. Farhadi, andD. Forsyth. Attributediscov-

ery viapredictablediscriminativebinary codes. In ECCV,

2012. 2

[20] M. Rohrbach, M. Stark, andB. Schiele. Evaluatingknowl-

edgetransfer andzero-shot learning inalarge-scalesetting.

InCVPR, 2011. 1, 2

[21] M. Rohrbach, M. Stark, G. Szarvas, I. Gurevych, and

B. Schiele. What helpswhere–and why? semantic relat-

ednessfor knowledgetransfer. InCVPR, 2010. 2, 4, 6

[22] M. Sadeghi and A. Farhadi. Recognition using visual

phrases. InCVPR, 2011. 3

[23] R. Salakhutdinov, A. Torralba, andJ. Tenenbaum. Learning

tosharevisual appearancefor multiclassobject detection. In

CVPR, 2011. 2

[24] J. Sanchez, F. Perronnin, T. Mensink, and J. Verbeek. Im-

ageclassificationwiththefishervector: Theoryandpractice.

IJCV, 2013. 3, 4

[25] V.Sharmanska, N.Quadrianto, andC.Lampert. Augmented

attributerepresentations. InECCV, 2012. 2

[26] T. Tommasi and B. Caputo. Themoreyou know, the less

you learn: fromknowledgetransfer toone-shot learning of

object categories. InBMVC, 2009. 2

[27] T. Tommasi, F. Orabona, andB. Caputo. Safety innumbers:

Learning categories from few examples with multi model

knowledgetransfer. InCVPR, 2010. 4

[28] C.Wah, S.Branson, P.Welinder,P.Perona, andS.Belongie.

Thecaltech-ucsd birds-200-2011 dataset. Technical report,

Computation& Neural Systems, 2011. 1, 2, 3, 5

[29] J. Xiao, J. Hays, K. A. Ehinger, A. Oliva, and A. Torralba.

SUNdatabase: Large-scalescenerecognitionfromabbey to

zoo. InCVPR, 2010. 2, 5

8

COSTA: Classifier

• Goal: Estimate classifier for unseen label

• Knowledge base:

– k trained classifiers

– Co-occurrences

• Zero-shot classifier:

VOGIN-IP 201530

Co-Occurrence Statistics

• Ground-truth data (proof-of-concept)

• Web search engines

• Flickr Tags

• Language resources

• Visual annotated data (eg Microsoft COCO)

VOGIN-IP 201531

Example: Beach Holiday

VOGIN-IP 201532

Concept Normalized Co-Oc Weight

Sea 0.1810Water 0.0992Summer 0.0548LandscapeNature 0.0435SunsetSunrise 0.0383

Sports 0.0367Travel 0.0347Ship 0.0346Sunny 0.0319Big Group 0.0282

Example: Beach Holidays

VOGIN-IP 201533

Results per concept

VOGIN-IP 201534

Co-occurrences from the Web

VOGIN-IP 201535

Conclusions

• Supervised visual classification performs well when ample train data is available

• Classification without examples:

– Define some set of base classifiers

– Transfer new class to space of these classifiers

– Two examples: attributes and co-occurrences

VOGIN-IP 201536

Thanks to:

• The organizers

• Christoph Lampert for slides and inspiration

• Authors of the cited papers

• Colleagues and supervisors (UvA: Amir, Cees, Jan, Spencer & Stratis, PhD: Cordelia, Florent, Gabriela, Jakob)

VOGIN-IP 201537

Literature

• Frome, Corrado, Shlens, Bengio, Dean, Ranzato,and Mikolov, “DeViSE: A Deep Visual-Semantic Embedding Model”, NIPS 2013

• Habibian, Mensink, and Snoek, “VideoStory: A New Multimedia Embedding for Few-Example Recognition and Translation of Events”, ACM MM 2014

• Lampert, Nickish, and Harmeling, “Attribute-Based Classification for Zero-Shot Learning of Object Categories”, TPAMI 2013

• Li, Gavves, Mensink, and Snoek, “Attributes Make Sense on Segmented Objects”, ECCV 2014

• Mensink, Gavves, and Snoek, “COSTA: Co-Occurrence Statistics for Zero-Shot Classification”, CVPR 2014

• Norouzi, Mikolov, Bengio, Singer, Shlens, Frome, Corrado, and Dean, “Zero-Shot Learning by Convex Combination of Semantic Embeddings”, ICLR 2014

VOGIN-IP 201538

Voginip lezing 2015: Classificeren zonder voorbeelden

Science

attributes voginip

prediction voginip

mapping voginip

computers voginip

design voginip

generic classifier voginip

van gemertvoginip

examples previewvoginip