45
Parametric measures to estimate and predict performance of identification techniques Amos Y. Johnson & Aaron Bobick STATISTICAL METHODS FOR COMPUTATIONAL EXPERIMENTS IN VISUAL PROCESSING & COMPUTER VISION NIPS 2002

Parametric measures to estimate and predict performance of identification techniques Amos Y. Johnson & Aaron Bobick STATISTICAL METHODS FOR COMPUTATIONAL

Embed Size (px)

Citation preview

Page 1: Parametric measures to estimate and predict performance of identification techniques Amos Y. Johnson & Aaron Bobick STATISTICAL METHODS FOR COMPUTATIONAL

Parametric measures to estimate and predict performance of identification techniques

Amos Y. Johnson & Aaron Bobick

STATISTICAL METHODS FOR COMPUTATIONAL EXPERIMENTS

IN VISUAL PROCESSING & COMPUTER VISIONNIPS 2002

Page 2: Parametric measures to estimate and predict performance of identification techniques Amos Y. Johnson & Aaron Bobick STATISTICAL METHODS FOR COMPUTATIONAL

Setup – for example

Given a particular human identification technique

Page 3: Parametric measures to estimate and predict performance of identification techniques Amos Y. Johnson & Aaron Bobick STATISTICAL METHODS FOR COMPUTATIONAL

Setup – for example

Given a particular human identification technique This technique measures 1 feature (q) from n individuals

n321 q ... q q qx

- 1D Feature Space -

Page 4: Parametric measures to estimate and predict performance of identification techniques Amos Y. Johnson & Aaron Bobick STATISTICAL METHODS FOR COMPUTATIONAL

Setup – for example

Given a particular human identification technique This technique measures 1 feature (q) from n individuals Measure the feature again

- 1D Feature Space -

n321 q ... q q qx

''''

n321q ... q q q

Page 5: Parametric measures to estimate and predict performance of identification techniques Amos Y. Johnson & Aaron Bobick STATISTICAL METHODS FOR COMPUTATIONAL

Setup – for example

Given a particular human identification technique This technique measures 1 feature (q) from n individuals Measure the feature again

- 1D Feature Space -

n321 q ... q q qx

''''

n321q ... q q q

Gallery

Probe

Page 6: Parametric measures to estimate and predict performance of identification techniques Amos Y. Johnson & Aaron Bobick STATISTICAL METHODS FOR COMPUTATIONAL

Setup – for example

Given a particular human identification technique This technique measures 1 feature (q) from n individuals Measure the feature again

- 1D Feature Space -

n321 q ... q q qx

''''

n321q ... q q q

Gallery

Probe

For template

Target

Page 7: Parametric measures to estimate and predict performance of identification techniques Amos Y. Johnson & Aaron Bobick STATISTICAL METHODS FOR COMPUTATIONAL

Setup – for example

Given a particular human identification technique This technique measures 1 feature (q) from n individuals Measure the feature again

- 1D Feature Space -

n321 q ... q q qx

''''

n321q ... q q q

Gallery

Probe

For template

Target Imposters

Page 8: Parametric measures to estimate and predict performance of identification techniques Amos Y. Johnson & Aaron Bobick STATISTICAL METHODS FOR COMPUTATIONAL

Question

For a given human identification technique, how should identification performance be evaluated?

- 1D Feature Space -

n321 q ... q q qx

''''

n321q ... q q q

Gallery

Probe

For template

Target Imposters

Page 9: Parametric measures to estimate and predict performance of identification techniques Amos Y. Johnson & Aaron Bobick STATISTICAL METHODS FOR COMPUTATIONAL

Possible ways to evaluate performance

For a given classification threshold, compute False accept rate (FAR) of impostors Correct accept rate (HIT) of genuine targets

- 1D Feature Space -

n321 q ... q q qx

''''

n321q ... q q q

Gallery

Probe

For template

Target Imposters

Page 10: Parametric measures to estimate and predict performance of identification techniques Amos Y. Johnson & Aaron Bobick STATISTICAL METHODS FOR COMPUTATIONAL

Possible ways to evaluate performance

For various classification thresholds, plot Multiple FAR and HIT rates (ROC curve)

Page 11: Parametric measures to estimate and predict performance of identification techniques Amos Y. Johnson & Aaron Bobick STATISTICAL METHODS FOR COMPUTATIONAL

Possible ways to evaluate performance

For various classification thresholds, plot Multiple FAR and HIT rates (ROC curve) Compute area under a ROC curve (AUROC)

Probability of correctclassification

Page 12: Parametric measures to estimate and predict performance of identification techniques Amos Y. Johnson & Aaron Bobick STATISTICAL METHODS FOR COMPUTATIONAL

Possible ways to evaluate performance

For various classification thresholds, plot Multiple FAR and HIT rates (ROC curve) Compute 1 - area under a ROC curve (1 -AUROC)

Probability of incorrectclassification

Page 13: Parametric measures to estimate and predict performance of identification techniques Amos Y. Johnson & Aaron Bobick STATISTICAL METHODS FOR COMPUTATIONAL

Problem Database size

If the database is not of sufficient size, then results may not estimate or predict performance on a larger population of people.

1 - AUROC

Page 14: Parametric measures to estimate and predict performance of identification techniques Amos Y. Johnson & Aaron Bobick STATISTICAL METHODS FOR COMPUTATIONAL

Our Goal

To estimate and predict identification performance with a small number subjects

1 - AUROC

Page 15: Parametric measures to estimate and predict performance of identification techniques Amos Y. Johnson & Aaron Bobick STATISTICAL METHODS FOR COMPUTATIONAL

Our Solution

Derive two parametric measures Expected Confusion (EC) Transformed Expected-Confusion (EC*)

Page 16: Parametric measures to estimate and predict performance of identification techniques Amos Y. Johnson & Aaron Bobick STATISTICAL METHODS FOR COMPUTATIONAL

Our Solution

Derive two parametric measures Expected Confusion (EC) Transformed Expected-Confusion (EC*)

Probability that an imposter’s feature vector is withinthe measurement variation of a target’s template

Page 17: Parametric measures to estimate and predict performance of identification techniques Amos Y. Johnson & Aaron Bobick STATISTICAL METHODS FOR COMPUTATIONAL

Our Solution

Derive two parametric measures Expected Confusion (EC) Transformed Expected-Confusion (EC*)

Probability that an imposter’s feature vector is closer to a target’s template, than the target’s feature vector

Page 18: Parametric measures to estimate and predict performance of identification techniques Amos Y. Johnson & Aaron Bobick STATISTICAL METHODS FOR COMPUTATIONAL

Our Solution

Derive two parametric measures Expected Confusion (EC) Transformed Expected-Confusion (EC*)

EC* = 1 - AUROC

Page 19: Parametric measures to estimate and predict performance of identification techniques Amos Y. Johnson & Aaron Bobick STATISTICAL METHODS FOR COMPUTATIONAL

Expected Confusion

Probability that an imposter’s feature vector is within the measurement variation of a target’s template

- 1D Feature Space -

n321 q ... q q qx

''''

n321q ... q q q

Gallery

Probe

For template

Target Imposters

Page 20: Parametric measures to estimate and predict performance of identification techniques Amos Y. Johnson & Aaron Bobick STATISTICAL METHODS FOR COMPUTATIONAL

Expected Confusion - Uniform

The templates of the n individuals, are from an uniform density Pp(x) = 1/n

- 1D Feature Space -

n321 q ... q q qx

P(x)

1/nPp(x)

Page 21: Parametric measures to estimate and predict performance of identification techniques Amos Y. Johnson & Aaron Bobick STATISTICAL METHODS FOR COMPUTATIONAL

Expected Confusion - Uniform

The measurement variation of a template is also uniform Pi(x) = 1/m

- 1D Feature Space -

n321 q ... q q qx

P(x)

1/nPp(x)

1/mPi(x)

Page 22: Parametric measures to estimate and predict performance of identification techniques Amos Y. Johnson & Aaron Bobick STATISTICAL METHODS FOR COMPUTATIONAL

Expected Confusion - Uniform

The probability that an imposter’s feature vector is within the measurement variation of template q3 is the area of overlap

True if m << n

- 1D Feature Space -

n321 q ... q q qx

P(x)

1/nPp(x)

1/mPi(x)

n

mA )q( 3

Page 23: Parametric measures to estimate and predict performance of identification techniques Amos Y. Johnson & Aaron Bobick STATISTICAL METHODS FOR COMPUTATIONAL

Expected Confusion - Uniform

The probability that an imposter’s feature vector is within the measurement variation of any template q

True if m << n

n321 q ... q q qx

P(x)

1/nPp(x)

1/mPi(x)

n

mA )q( 3

n

mdqPAEC p

)q()q(

Page 24: Parametric measures to estimate and predict performance of identification techniques Amos Y. Johnson & Aaron Bobick STATISTICAL METHODS FOR COMPUTATIONAL

Following the same analysis, for the multidimensional Gaussian case

Expected Confusion - Gaussian

),()q( ppp Np : Population density

),q()( ii Nxp : Measurement variation

Page 25: Parametric measures to estimate and predict performance of identification techniques Amos Y. Johnson & Aaron Bobick STATISTICAL METHODS FOR COMPUTATIONAL

Expected Confusion - Gaussian

Following the same analysis, for the multidimensional Gaussian case True if the measurement variation is significantly less then the population variation

2/1

2/1

||

||EC

p

i

Probability that an imposter’s feature vector is within the measurement variation of a target’s template

Page 26: Parametric measures to estimate and predict performance of identification techniques Amos Y. Johnson & Aaron Bobick STATISTICAL METHODS FOR COMPUTATIONAL

Expected Confusion - Gaussian

Relationship to other metrics Mutual Information

The negative natural log of the EC is the mutual information of two Gaussian densities

)|ln(|)|ln(|)||

||ln( 2/12/1

2/1

2/1

ipp

i

Page 27: Parametric measures to estimate and predict performance of identification techniques Amos Y. Johnson & Aaron Bobick STATISTICAL METHODS FOR COMPUTATIONAL

Transformed Expected-Confusion

Probability that an imposter’s feature vector is closer to a target’s template, than the target’s feature vector

- 1D Feature Space -

n321 q ... q q qx

''''

n321q ... q q q

Gallery

Probe

For template

Target Imposters

Page 28: Parametric measures to estimate and predict performance of identification techniques Amos Y. Johnson & Aaron Bobick STATISTICAL METHODS FOR COMPUTATIONAL

Transformed Expected-Confusion

First: We find the probability that a target’s feature vector is some distance k away from its template

n321 q ... q q qx

''''

n321q ... q q q

For template

Target Imposters

k

)( dkkpqt

Page 29: Parametric measures to estimate and predict performance of identification techniques Amos Y. Johnson & Aaron Bobick STATISTICAL METHODS FOR COMPUTATIONAL

n321 q ... q q qx

''''

n321q ... q q q

For template

Target Imposters

k

)( dkkpqt

Transformed Expected-Confusion

Second: We find the probability that an imposter’s feature vector is less than or equal to that distance k

k

qim dvvp

0

)(

Page 30: Parametric measures to estimate and predict performance of identification techniques Amos Y. Johnson & Aaron Bobick STATISTICAL METHODS FOR COMPUTATIONAL

n321 q ... q q qx

''''

n321q ... q q q

Target Imposters

k

Transformed Expected-Confusion

Therefore: The probability that an imposter’s feature is closer to the target’s template, than the target’s feature (for a distance k) is

dkdqdvvpkpqpk

qim

qtp

00

)()()(

Page 31: Parametric measures to estimate and predict performance of identification techniques Amos Y. Johnson & Aaron Bobick STATISTICAL METHODS FOR COMPUTATIONAL

n321 q ... q q qx

''''

n321q ... q q q

Target Imposters

k

Transformed Expected-Confusion

Therefore: The probability that an imposter’s feature is closer to the target’s template, than the target’s feature (for any distance k) is

dkdqdvvpkpqpk

qim

qtp

00

)()()(

Page 32: Parametric measures to estimate and predict performance of identification techniques Amos Y. Johnson & Aaron Bobick STATISTICAL METHODS FOR COMPUTATIONAL

x

''''

n321q ... q q q

Transformed Expected-Confusion

Therefore: The expected value of this probability over all target’s templates is

dkdqdvvpkpqpk

qim

qtp

00

)()()(

n321 q ... q q q

Page 33: Parametric measures to estimate and predict performance of identification techniques Amos Y. Johnson & Aaron Bobick STATISTICAL METHODS FOR COMPUTATIONAL

Transformed Expected-Confusion

Next: Replace the density of the distance between a target’s feature-vectors and its template q

dkdqdvvpkpqpECk

qim

qtp

00

)()()(*

)(kpt

Page 34: Parametric measures to estimate and predict performance of identification techniques Amos Y. Johnson & Aaron Bobick STATISTICAL METHODS FOR COMPUTATIONAL

Transformed Expected-Confusion

Answer: Probability that an imposter’s feature vector is closer to a target’s template, than the target’s feature vector

dkdqdvvpkpqpECk

qim

qtp

00

)()()(*

Page 35: Parametric measures to estimate and predict performance of identification techniques Amos Y. Johnson & Aaron Bobick STATISTICAL METHODS FOR COMPUTATIONAL

Transformed Expected-Confusion

This probability can be shown to be one minus the area under a ROC curve

Following the analysis of Green and Swets (1966)

dkdqdvvpkpqpECk

qim

qtp

00

)()()(*

Page 36: Parametric measures to estimate and predict performance of identification techniques Amos Y. Johnson & Aaron Bobick STATISTICAL METHODS FOR COMPUTATIONAL

Transformed Expected-Confusion

Integrate: With these assumptions

dkdqdvvpkpqpECk

qim

qtp

00

)()()(*

Page 37: Parametric measures to estimate and predict performance of identification techniques Amos Y. Johnson & Aaron Bobick STATISTICAL METHODS FOR COMPUTATIONAL

Transformed Expected-Confusion

Integrate: With these assumptions

dkdqdvvpkpqpECk

qim

qtp

00

)()()(*

),;( ppqN

Page 38: Parametric measures to estimate and predict performance of identification techniques Amos Y. Johnson & Aaron Bobick STATISTICAL METHODS FOR COMPUTATIONAL

Transformed Expected-Confusion

Integrate: With these assumptions

dkdqdvvpkpqpECk

qim

qtp

00

)()()(*

),;( ppqN

2

2

2)12/(

2/)2(

)2/(2i

k

ddi

d

ed

k

Page 39: Parametric measures to estimate and predict performance of identification techniques Amos Y. Johnson & Aaron Bobick STATISTICAL METHODS FOR COMPUTATIONAL

Transformed Expected-Confusion

Integrate: With these assumptions

),;( ppqN d

dp kVqp )(

2

2

2)12/(

2/)2(

)2/(2i

k

ddi

d

ed

k

dkdqdvvpkpqpECk

qim

qtp

00

)()()(*

Page 40: Parametric measures to estimate and predict performance of identification techniques Amos Y. Johnson & Aaron Bobick STATISTICAL METHODS FOR COMPUTATIONAL

Transformed Expected-Confusion

Integrate: Probability that an imposter’s feature vector is closer to a target’s template, than the target’s feature vector

dkdqdvvpkpqpECk

qim

qtp

00

)()()(* ECd

dVd

d

)2/()2(

)!1(2/

Page 41: Parametric measures to estimate and predict performance of identification techniques Amos Y. Johnson & Aaron Bobick STATISTICAL METHODS FOR COMPUTATIONAL

Transformed Expected-Confusion

Compare: EC* with 1 - AUROC

EC* = 1 - AUROC

Page 42: Parametric measures to estimate and predict performance of identification techniques Amos Y. Johnson & Aaron Bobick STATISTICAL METHODS FOR COMPUTATIONAL

Conclusion

Derive two parametric measures Expected Confusion

(EC) Transformed

Expected-Confusion (EC*)

Probability that an imposter’s feature vector is closer

to a target’s template, than the target’s feature vector

Page 43: Parametric measures to estimate and predict performance of identification techniques Amos Y. Johnson & Aaron Bobick STATISTICAL METHODS FOR COMPUTATIONAL

Conclusion

Derive two parametric measures Expected Confusion

(EC) Transformed

Expected-Confusion (EC*)

Probability that an imposter’s feature vector is within

the measurement variation of a target’s template

Probability that an imposter’s feature vector is closer

to a target’s template, than the target’s feature vector

Page 44: Parametric measures to estimate and predict performance of identification techniques Amos Y. Johnson & Aaron Bobick STATISTICAL METHODS FOR COMPUTATIONAL

Conclusion

Derive two parametric measures Expected Confusion

(EC) Transformed

Expected-Confusion (EC*)

Probability that an imposter’s feature vector is within

the measurement variation of a target’s template

Probability that an imposter’s feature vector is closer

to a target’s template, than the target’s feature vector

Page 45: Parametric measures to estimate and predict performance of identification techniques Amos Y. Johnson & Aaron Bobick STATISTICAL METHODS FOR COMPUTATIONAL

Future Work

Developing a mathematical model of the cumulative match characteristic (CMC) curve Benefit: To predict how the CMC curve

changes as more subjects are added