Introduction—About the Need of an Evaluation Framework in Biometrics

Abstract

How can scientific progress be measured? How do we know if a pattern recognition algorithm performs better on average than another one? How much data is needed to claim with confidence that one system performs better than another one? Is it possible to predict performance on a different data set? How will the performance rates achieved under laboratory conditions compare to those in larger populations? These are some of the questions that are central to this book. Biometrics is the application domain under concern. For applications related to verification, a person claims an identity. The system has in memory some training data for this identity claim (or a statistical model of it) and performs a comparison (or computes a likelihood) with the test data. The output is a score that is compared to a threshold to take a decision: accept or reject the identity claim. For applications related to identification, the system has in memory a list of identities and their training data. When a person presents biometric data, the system has to find out to whom the data belong. These two tasks, verification and identification, are grouped under the term of biometric recognition throughout this book.