15
Pa#ern Recogni-on and Applica-ons Lab University of Cagliari, Italy Department of Electrical and Electronic Engineering Sparse Support Faces Ba#sta Biggio, Marco Melis, Giorgio Fumera, Fabio Roli Dept. Of Electrical and Electronic Engineering University of Cagliari, Italy Phuket, Thailand, May 1922, 2015 ICB 2015

Sparse Support Faces - Battista Biggio - Int'l Conf. Biometrics, ICB 2015, Phuket, Thailand, May 19-22, 2015

Embed Size (px)

Citation preview

Pa#ern  Recogni-on    and  Applica-ons  Lab  

                               

 University  

of  Cagliari,  Italy  

 

Department  of  Electrical  and  Electronic  

Engineering  

Sparse Support Faces

Ba#sta  Biggio,  Marco  Melis,  Giorgio  Fumera,  Fabio  Roli        

Dept.  Of  Electrical  and  Electronic  Engineering  University  of  Cagliari,  Italy  

Phuket,  Thailand,  May  19-­‐22,  2015  ICB  2015  

 

http://pralab.diee.unica.it

Template-based Face Verification

2  

gc ≥ϑ c

genuine  

impostor  

true  

false  

s(x, tci ){ }i=1

p

Matcher    s(⋅, ⋅)

Fusion  rule  

gc (x)xFeature  extrac-on  

Verifica-on  is  based  on  how  similar  the  submi#ed  image  is  to  the  client’s  templates  

Client-­‐specific  one-­‐class  classifica:on  

mean gc (x) =1p

s(x, tci )

i=1

p

gc (x) = maxi=1,…,ps(x, tc

i )max

Claimed  Iden-ty   tc

1, …, tcp{ }

Claimed  iden-ty’s  templates  

 

http://pralab.diee.unica.it

Cohort-based Face Verification

3  

Verifica-on  is  based  on  how  similar  the  submi#ed  image  is  to  the  client’s  templates  and  on  how  different  it  is  from  the  cohorts’  templates  

Client-­‐specific  two-­‐class  classifica:on  (one-­‐vs-­‐all)  

gc ≥ϑ c

genuine  

impostor  

true  

false  

s(x, tci ){ }i=1

n

Matcher    s(⋅, ⋅)

Fusion  rule  

gc (x)xFeature  extrac-on  

tc1, …, tc

p{ }

Claimed  iden-ty’s  templates   Cohorts  

tcp+1, …, tc

n{ }Claimed  Iden-ty  

 

http://pralab.diee.unica.it

Cohort-based Fusion Rules

•  Cohort selection is heuristically driven –  e.g., selection of the closest cohorts to the client’s templates

•  Cohort-based fusion rules are also based on heuristics

–  Test-normalization [Auckenthaler et al., DSP 2000]

–  Aggarwal’s max rule [Aggarwal et al., CVPR-W 2006]

4  

gc (x) =1

σ c (x)1p

s(x, tci )

i=1

p

∑ −µc (x)#

$%

&

'(

gc (x) =maxi=1,…,p

s(x, tci )

maxj=p+1,…,n

s(x, tcj )

 

http://pralab.diee.unica.it

Open Issues

•  Fusion rules and cohort selection are based on heuristics –  No guarantees of optimality in terms of verification error

•  Our goal: to design a procedure to optimally select the reference templates and the fusion rule –  Optimal in the sense that it minimizes verification error (FRR and FAR)

•  Underlying idea: to consider face verification as a two-class classification problem in similarity space

5  

 

http://pralab.diee.unica.it

s(x, )

s(x, )

Face Verification in Similarity Space

•  The matching function maps faces onto a similarity space –  How to design an optimal decision function in this space?

6  

?  

 

http://pralab.diee.unica.it

Support Face Machines (SFMs)

•  We learn a two-class SVM for each client –  using the matching score as the kernel function –  genuine client y=+1, impostors y=-1

•  SVM minimizes the classification error (optimal in that sense) –  FRR and FAR in our case

•  The fusion rule is a linear combination of matching scores •  The templates are automatically selected for each client

–  support vectors à support faces

7  

gc (x) = αis(x, tci )

i∑ − α js(x, tcj )

j∑ + b

 

http://pralab.diee.unica.it

Support Face Machines (SFMs)

8  

s(x, )

s(x, )

•  Maximum-margin classifiers

gc (x) = αis(x, tci )

i∑ − α js(x, tcj )

j∑ + b

 

http://pralab.diee.unica.it

Sparse Support Faces

•  Open issue: SFMs require too many support faces –  Number of support faces scales linearly with training set size

•  Our goal: to learn a much sparser combination of match scores

•  by jointly optimizing the weighting coefficients and support faces:

9  

hc (x) = βis(x, zck )+ b

k=1

m

∑ , m << n

minβ ,z

Ω β, z( ) = 1n

uk gc (xk )− hc (xk )( )2+λβTβ

i=1

n

 

http://pralab.diee.unica.it

z-­‐step

Sparse Support Faces

10  

SFM with 12 support faces

−5 0 5−5

0

5

−5

0

5SSFM with 4 virtual faces

−5 0 5−5

0

5

−5

0

5

β-­‐step  

Solu:on  algorithm  is  an  itera-ve  two-­‐step  procedure:  

If s(x,z) is not differentiable or analytically given, gradient can be approximated    

 

http://pralab.diee.unica.it

0.5 1 2 5 100

5

10

15

20 AT&T − RBF Kernel

FAR (%)

FR

R (

%)

mean (5)max (5)t−norm (10)aggarwal−max (10)SFM (37.5 ± 3.8)SFM−sel (10)SFM−red (2)SSFM (2)

Experiments

11  

Datasets: AT&T (40 clients, 10 images each) BioID (23 clients, 1,521 images) Matcher: PCA+RBF kernel (exact gradient) 5 repetitions, different clients in TR/TS splits TR: 5 images/client

0.5 1 2 5 100

10

20

30

40 BioID − RBF Kernel

FAR (%)

FR

R (

%)

mean (5)max (5)t−norm (10)aggarwal−max (10)SFM (23.9 ± 2.7)SFM−sel (10)SFM−red (2)SSFM (2)

 

http://pralab.diee.unica.it

Experiments

12  0.5 1 2 5 100

10

20

30

40 BioID − EBGM

FAR (%)

FR

R (

%)

mean (5)max (5)t−norm (10)aggarwal−max (10)SFM (15.0 ± 2.6)SFM−sel (5)SFM−red (5)SSFM (5)

0.5 1 2 5 100

5

10

15

20 AT&T − EBGM

FAR (%)

FR

R (

%)

mean (5)max (5)t−norm (10)aggarwal−max (10)SFM (19.5 ± 3.0)SFM−sel (5)SFM−red (5)SSFM (5)

Datasets: AT&T (40 clients, 10 images each) BioID (23 clients, 1,521 images) Matcher: EBGM (approx. gradient) 5 repetitions, different clients in TR/TS splits TR: 5 images/client

 

http://pralab.diee.unica.it

From Support Faces to Sparse Support Faces

•  A client’s gallery of 17 support faces (and weights) reduced to 5 virtual templates by our sparse support face machine –  Dataset: BioID –  Matching algorithm: EBGM

13  

4.040 2.854 −0.997 −3.525 −2.208

 

http://pralab.diee.unica.it

Conclusions and Future Research Directions

•  Sparse support face machines: –  reduce computational time and storing requirements during

verification without affecting verification accuracy –  by jointly learning an optimal combination of matching scores, and a

corresponding sparse set of virtual support faces

•  No explicit feature representation is required –  Matching algorithm exploited as kernel function –  Virtual templates created exploiting approximations of its gradient

•  Future work –  Fingerprint verification

–  Identification setting •  Joint reduction of virtual templates for each client-specific classifier

14  

 

http://pralab.diee.unica.it

?   Any questions Thanks  for  your  a#en-on!  

15  

Code available at: http://pralab.diee.unica.it/en/SSFCodeProject