Location matters: A study of the effects of environment on facial recognition for biometric security

The term “in the wild” has become wildly popular in face recognition research. The term refers generally to use of datasets that are somehow less controlled or more realistic. In this work, we consider how face recognition accuracy varies according to the composition of the dataset on which the decision threshold is learned and the dataset on which performance is then measured. We identify different acquisition locations in the FRVT 2006 dataset, examine face recognition accuracy for within-environment image matching and cross-environment image matching, and suggest a way to improve biometric systems that encounter images taken in multiple locations. We find that false non-matches are more likely to occur when the gallery and probe images are acquired in different locations, and that false matches are more likely when the gallery and probe images were acquired in the same location. These results show that measurements of face recognition accuracy are dependent on environment.