On Optimizing Subspaces for Face Recognition

Subspace-based face recognition

Find subspaces that optimize some performance metric:
PCA, ICA: finding subspaces that preserves certain distributive properties of the data. FDA, Bayesian dual eigenspace, Bayesian Optimal LDA: finding subspaces that preserve certain discriminative properties of the data LPP, OLPP, MFA, NPE: finding subspaces that preserve some locality properties of the data

Two observations:
Different face recognition tasks (i.e., identification and verification) have different performance metrics. Direct optimization of the recognition performance scores may result in performance gain. Consistent appearance incoherences between the gallery and the probe data for the same subject could be used to guide the algorithm design, resulting in performance gain.

Notation

Optimal Subspace for Face Verification (OSFV) Face verification: verify whether two face images are from the same person or not. Performance metrics
false alarm rate (FAR):

false rejection rate (FRR)

verification error rate (PE)

Optimal Subspace for Face Verification (OSFV)

Objective function: Optimization method: Alternated updating of A and hT
Updating A: Gradient descent

Updating hT: globally optimal decision threshold [2]

Generate an ROC curve by sorting the distance scores Obtain the decision threshold that minimizes the verification error on the ROC curve Computational complexity: O(n log n), where n is the number of matchings.

[2] T. Fawcett. ROC graphs: Notes and practical considerations for researchers. Technical report, HP Laboratories, 2004.

OSFV: Updating A
PE is not differentiable:
PE PE f h A f h A
Sigmoid function
The step function f() is not differntiable

Approximate f with differentiable functions:

Exponential function

OSFV: Updating A
PE PE f h A f h A
Different distance measures may be used

Euclidean distance:

Exponential function

The colon operator : stands for the vectorization of a matrix.

OSFV: Updating hT

OSFV: Algorithm

OSFV Algorithm: Comments on Implementation and Parameterization

An initial guess for A0 can be obtained using one of the state-of-the-art subspace learning algorithms, such as LDA, LPP, MFA, etc. The parameter for f() is initialized with a large 1, and is gradually reduced to a smaller . Sigmoid Exponential

Parameter and

Optimal Subspace for Face Identification (OSFI)

Objective function:

OSFI Algorithm

Experimental Results
Data sets

Face verification

Face Identification

Face Verification for Airport Security Check-in

Conclusion
There exists different optimal face subspaces for different face recognition tasks. There could exists consistent appearance incoherences in the gallery and the probe set in real-world applications. Performance gains could be obtained by taking into consideration the differences of the performance score definitions of the different tasks and the intrinsic appearance incoherences between the gallery images and the probe images of the same subject.
OSFV OSFI

References
[1] Jilin Tu, Xiaoming Liu, Peter Tu, On Optimizing Subspaces for Face Recognition, ICCV 2009. [2] T. Fawcett. ROC graphs: Notes and practical considerations for researchers. Technical report, HP Laboratories, 2004.