Evaluation Of 2D Still-Image Face Recognition Algorithms.pdf

Key findings include:

- Accuracy Variation: There's an order of magnitude difference in identification error rates among providers, highlighting the importance of core algorithmic capability.

- Application Modes: Face recognition algorithms can operate in investigational (human examiners review top candidates) or identification (high threshold for short candidate lists) modes, with different accuracies.

- Criminal Law Enforcement Applications: In scenarios like criminal recidivism, where a pool of examiners is available, highly accurate algorithms significantly reduce examiner workload.

- "Lights-Out" Identification: Algorithms can perform "lights-out" identification (no human review) by setting selectivity policies based on false candidate rates, with optimal thresholds varying by application needs.

- Image Quality Impact: Image quality, influenced by collection protocols and subject behavior, is a critical factor. Visa images, collected under controlled conditions, show higher accuracy than contemporary mugshot images due to variations in pose, illumination, and expression.

- Historical Comparison: While the most accurate algorithm demonstrated better non-frontal pose tolerance, verification accuracy on one historical database has declined over time, highlighting the need for continuous improvement and adaptation to changing image quality.

This evaluation is the first by NIST to measure face recognition algorithm accuracy, providing valuable insights into their performance across different scenarios and populations.