Modeling the Factors Affecting Variability in the Human Face

As efforts applying advanced machine learning methods to facial shape start to include both genetic and non-genetic data the opportunity for biological insights that might help guide improved facial biometrics can and should be embraced. Human facial variation is as fascinating and manifold as it is useful for biometric validation and identification. The causes for population-specific biases in facial biometrics remain to be identified and might be assisted by biologically supervised investigations. Likewise, as the prospects for the combined application of multiple biometric systems, facial, iris, voice, and gait, for example, genetic information from DNA may prove especially useful in developing unbiased efficient algorithms. Our group has been working on describing the effects of the major genetic and non-genetic factors affecting the human face, among other traits, such as hair microstructure, skin, hair, and eye pigmentation, hand and foot photos, and voice characteristics, using large panels (N>12,000) of diverse participants. I will discuss our findings in the context of the broader practical applications of facial variation.

About the Speaker:

Professor of Anthropology Dr. Mark D. Shriver (Penn State University) heads projects emphasizing the practical applications of population genomic research. These projects are primarily focused on admixture mapping, signatures of natural selection, and the elucidation of the evolutionary-genetic architecture underlying phenotypic variability in common trait variation. A major goal of his work is to apply these methods and understanding of genomic variation to studies of typical-range variation, with a focus on superficial traits, such as pigmentation, hair form, and facial features, where he is interested in developing useful predictive biometric tests for use in a forensics context. Dr. Shriver has been funded by the NIH, NSF, NIJ, DTRA, DOD, and AHA and has published over 147 research articles, which have been cited over 24,500 times.

Demographic accuracy differentials in contemporary face recognition algorithms

The talk will describe demographic accuracy differentials in more than 180 mostly commercial algorithms submitted to NIST’s Face Recognition Vendor Test. These were used with 18 million images of 8 million people selected from four operational databases for which sex, age, and race or country of birth information is available. The talk will report false positive and false negative rate differentials from both verification and identification algorithms, and then discuss the importance and consequences of each type of error. The presentation will close by discussing next steps and advocating for specific methods for improved reporting of demographic effects in biometrics.

About the Speaker:

Patrick Grother is a scientist at the National Institute of Standards in Technology responsible for biometric standards and algorithm evaluation. He leads the Face Recognition Vendor Tests which include the world’s largest independent public assessments of face recognition performance. His research interests are biometric failure analysis, image quality assessment, demographic effects, fusion and scalability to large populations. Patrick assists several US Government agencies in biometrics performance assessment and standardization and, since 2018, has served as the chairman of the ISO/IEC/JTC 1 Subcommittee 37 on Biometrics.