Face recognition performance using deep learning has seen dramatic improvements in recent years. This improvement has been fueled in part by the curation of large labeled training datasets with millions of images of hundreds of thousands of subjects.This results in effective generalization for matching over pose, illumination, expression and age variation, however these datasets have traditionally only focused on visible spectrum imageryCovert face recognition for the warfigher requires matching non-visible spectrum imagery against watch-lists derived from current visible facial databases.However, it would be impractical to collect non-visible training datasets at the scale necessary for end-to-end training.In this proposal, STR, and university partner Northeastern University, will investigate cross-modal transfer learning by applying domain transfer techniques for matching a visible face representation to low-light and night-time sensor data.The STR team will collect a cross-spectrum biometric dataset from LWIR/MWIR/NIR sensors in support of a feasibility study on this cross-modal transfer learning capability, leveraging the STR Janus face matcher for baseline performance evaluation.
