Assembling Heterogeneous Domain Adap- tation Methods for Image Classification Boris Chidlovskii, Gabriela Csurka and Shalini Gangwar Xerox Research Centre Europe, France ImageCLEF, 16th September 2014 1 B. Chidlovskii et al, Assembling Heterogeneous DA
25
Embed
Assembling Heterogeneous Domain Adaptation Methods for ...imageclef.org/system/files/pres_ClefImage.pdf · Domain adaptation methods Instance Transfer I Instance weighting in source
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
Assembling Heterogeneous Domain Adap-tation Methods for Image Classification
Boris Chidlovskii, Gabriela Csurka and Shalini Gangwar
Xerox Research Centre Europe, France
ImageCLEF, 16th September 2014
1B. Chidlovskii et al, Assembling Heterogeneous DA
Domain Adaptation At Xerox
Transportation, image-based solutionsI Adapt learning components under data distribution
change, without a costly re-annotationI Changes caused by scene illumination, view angle,
background• Daylight to night, from inside to outside
• From one parking to another, other cameras, etc.
2B. Chidlovskii et al, Assembling Heterogeneous DA
ImageCLEF’14 Domain Adaptation Challenge
Domain adaptation scenario:I Multiple source domainsI Same labels between the source and the target domainsI Limited number of annotated data in the target domain
I Sources:• Caltech (C)• ImageNet (I)• Pascal (P)• Bing (B)
I Target:• SUN (S)
3B. Chidlovskii et al, Assembling Heterogeneous DA
Challenge setup
I 12 common classes:• airplane, bike, bird, boat, bus, car, ...
I No access to imagesI BOV features provided only
• 600 labeled features from each source (C, I, P, B)• 60 labeled and 600 unlabeled features from target (S)
I Source and target domains are semantically relevant butdifferent
I Target feature distribution changed between phases 1/2
Build a recognition system for target domain by leveraging theknowledge from source domains
4B. Chidlovskii et al, Assembling Heterogeneous DA
Outline
1. Assembling Heterogeneous Methods
2. Domain Adaptation by Instance Transfer
3. Domain Adaptation by Space Transformation
4. Ensemble Methods
5. Evaluation Results
6. Conclusion
5B. Chidlovskii et al, Assembling Heterogeneous DA
Domain adaptation methodsInstance Transfer
I Instance weighting in source domain (Dai et al. 2007, Xu2012)
I Selecting landmarks in source domain (Gong 2013)Feature Space Transformation
I Unsupervised transformation of domains• based on PCA projections (Gopalan et al. ICCV11, Gong et
al. CVPR12, Fernando et al. ICCV13, Baktashmotlagh etal. ICCV13)
I Learning transformation by exploiting class labels• based on metric learning (Zha et al. IJCAI09, Saeko et al.
ECCV10, Kulis et al. CVPR11, Hoffman et al. ECCV12)• Some methods exploit unlabeled target instances (e.g.
Duan et al. CVPR09, Saha et al. ECML11, Tomassi andCaputo ICCV13)
.6
B. Chidlovskii et al, Assembling Heterogeneous DA
XRCE approach
I Individual methods• Brute force: SVM cross validation
with all combinations
• Instance Weighting: Instancetransfer from sources to targetdomain using Boosting trick
• Space transformation: metriclearning-based domain adaptationto push together the same-classinstances from different domains
.I Ensemble techniques to aggregate
individual predictions
7B. Chidlovskii et al, Assembling Heterogeneous DA
Brute ForceI NSC = 2N
S − 1 = 15 source combinations SCi ,I For each source combination SCj :
• concatenate the target train set Tl with sources SCj
• train SVM in a cross validationI Multi-class SVM
• one kernel and same parameters for all classesI Binarised one-against-all SVM
• The best classifier for each class cj
• A specific set of parameter values, kernels and sourcecombinations
• For an unseen sample xi , take the classifier with thehighest confidence
ybsvm = argmaxcj∈Y
f cjbsvm(xi).
8B. Chidlovskii et al, Assembling Heterogeneous DA
Outline
1. Assembling Heterogeneous Methods
2. Domain Adaptation by Instance Transfer
3. Domain Adaptation by Space Transformation
4. Ensemble Methods
5. Evaluation Results
6. Conclusion
9B. Chidlovskii et al, Assembling Heterogeneous DA
Instance Transfer with AdaBoost
I Transfer AdaBoost is an extension of Adaboost to Transferlearning
I boost the accuracy of a weak learner by carefully adjustingthe weights of training instances and to learn a classifier
I In TrAdaboost:• Target training instances are weighted as in AdaBoost
• Source training instances are weighted differently
• Wrongly predicted source instances are the most dissimilar
• Their weights decrease to weaken their impact
10B. Chidlovskii et al, Assembling Heterogeneous DA
Transfer Adaptive Boosting with one sourceRequire: Target training set Tt = (Xt ,Y ); source training set Ts = (Xs,Y );
Learner; the number of iterations M.Ensure: Target learner f : Xt → Y .1: Initial weights: w1
T = (w1t1 , . . . ,w
1tNt
), w1S = (w1
s1 , . . . ,w1sNs
),
2: Set w = (wT ,wS), β = 1/(1 + 2√
ln Nt/M) and T = (Tt ,Ts).3: for r = 1, . . . ,M do4: Normalize wr = wr/|wr |.5: Call Learner on the training set T with wr to find fr : X → Y6: Calculate error of hr on Tt :
εr = min
(12 ,
1∑Nti=1 w r
ti
∑ni=1 w r
ti · [[fr (xti ) 6= yi ]]
).
7: Set βr = 1/2 log((1− εr )/εr ); Γr = 2(1− εr ).8: Update the weight vectors:
w r+1sj
= Γr w rsj
exp(−β [[fr (xsj ) 6= yj ]]), xs
j ∈ Xs,
w r+1ti
= w rti exp(2βr [[fr (xt
i ) 6= yi ]]), xti ∈ Xt .
9: end for10: Output the aggregated estimate ftra(x) =
(∑Mr=1 β
r fr (x))
.11
B. Chidlovskii et al, Assembling Heterogeneous DA
Transfer Adaptive Boosting: Two moons
12B. Chidlovskii et al, Assembling Heterogeneous DA
Outline
1. Assembling Heterogeneous Methods
2. Domain Adaptation by Instance Transfer
3. Domain Adaptation by Space Transformation
4. Ensemble Methods
5. Evaluation Results
6. Conclusion
13B. Chidlovskii et al, Assembling Heterogeneous DA
The Nearest Class Mean (NCM) classifier1
The NCM assigns an image to the closest class mean:
µc =1
|{xi |yi = c}|∑
xi∈{xi |yi =c}xi
Can be seen as the posterior of a GMM with wc = 1Nc
and Σ = I :
p(c|xi ) =wcp(xi |c)∑Nc
c′=1 w ′cp(xi |c′)=
wcN (xi ,µc , I)∑Ncc′=1 w ′cN (xi ,µc′ , I)
1T. Mensink, J. Verbeek, F. Perronnin and G. Csurka, Distance-based image classification: Generalizing
to new classes at near zero cost. PAMI 35(11), 2013
14B. Chidlovskii et al, Assembling Heterogeneous DA
ML for NCM2
Learning a projection W that maximizes the NCM accuracy: