Bayesian Face Recognition

Motivation(s)

Existing face recognition systems rely on similarity metrics based on Euclidean distance or normalized correlation, which corresponds to template-matching i.e. nearest-neighbor. The drawback of these metrics is that they do not exploit knowledge of which types of variation are critical in expressing similarity.

Proposed Solution(s)

The authors propose a probabilistic similarity measure based on the image differences to characterize facial image variations into intrapersonal and extrapersonal variations. This Bayesian formulation casts the M-ary classification problem into a binary classification problem i.e. \(Pr(\Omega_I \mid \Updelta) > Pr(\Omega_E \mid \Updelta)\).

Evaluation(s)

The proposed metric was compared with the eigenface matching algorithm using the ARPA FERET dataset. The cumulative match score showed that the probabilistic metric bested the baseline by 10-20%. An additional benefit is that the database only needs to store a single image of an individual through the use of the whitening transformation (7).

Future Direction(s)

• How does this compare to multi-linear models?

Question(s)

• Do the image differences of different individuals live in close proximity in the proposed subspace?

Analysis

Modeling some observations with a Gaussian establishes a good baseline. This approach can be viewed as an application of factor analyzers. The Bayesian formulation is brilliant, but I’m not too convinced of the image difference metric. As the authors admitted, multiple models should be used for dealing with large pose variations to avoid diluting the density models.

References

MJP00

Baback Moghaddam, Tony Jebara, and Alex Pentland. Bayesian face recognition. Pattern Recognition, 33(11):1771–1782, 2000.