Date of Graduation


Document Type

Problem/Project Report

Degree Type



Statler College of Engineering and Mineral Resources


Lane Department of Computer Science and Electrical Engineering

Committee Chair

Daryl Reynolds

Committee Member

Natalia Schmid

Committee Member

Xin Li


The performance of face recognition system components is traditionally reported using metrics such as the Receiver Operating Characteristic (ROC), Cumulative Match Characteristic (CMC), and Identification Error Tradeoff (IET). Recently, new metrics have been published to take advantage of annotation-dense datasets such as IARPA Janus Benchmark-Surveillance and IARPA Janus Benchmark-Multi Domain Face to describe end-to-end face recognition system performance. Unlike traditional (component-level) analysis, end-to-end analysis of a system produces a metric proportional to the experience of a user of a face recognition system. The End-to-End Cumulative Match Characteristic (E2ECMC) summarizes detection, identity consolidation, and identity retrieval performance. The End-to-End Subject Cumulative Match Characteristic (E2ESCMC) describes the lowest rank that subjects are retrieved in identification experiments. The End-to-End Identification Error Tradeoff (E2EIET) is a measure of the false alarm and miss performance of a system.

Until now, an evaluation utility capable of reporting the performance of individual components of a system and describing the user experience of a face recognition system was unavailable to the biometric community. Along with this problem report, a software package/framework capable of evaluating the performance of the following components will be released:

  • Face Detector
  • Face Verification
  • Face Identification
  • Face Clustering
  • End-to-End Face Recognition System Performance

In addition to providing a utility to researchers and system integrators, the evaluation framework is a C++17 library which may be incorporated into evolving/fine-tunable face recognition system pipelines as a means of providing performance snapshots over time.