Latent Identity Variables: Biometric Matching Without Explicit Identity Estimation

We present a new approach to biometrics that makes probabilistic inferences about matching without ever estimating an identity “template”. The biometric data is considered to have been created by a noisy generative process. This process consists of (i) a deterministic component, which depends entirely on an underlying representation of identity and (ii) a stochastic component which accounts for the fact that two biometric samples from the same person are not identical. In recognition, we make inferences about whether the underlying identity representation is the same without ever estimating it. Instead we treat identity as fundamentally uncertain and consider all possible values in our decision. We demonstrate these ideas with toy examples from face recognition. We compare our approach to the class-conditional viewpoint.