Foundation Models for Biometric Recognition

Ubiquitous vision sensors not only facilitate the wide application of biometric recognition but also generate various heterogeneous sets of biometrics. Matching across different sensing modalities raises the problem of heterogeneous identity recognition or cross domain identity recognition. Due to significant difference in sensing processes, heterogeneous images of the same subject have large appearance variations, which has distinguished heterogeneous recognition from regular biometric recognition. During last several years, our research group have investigated a range such problems and developed several foundation models for biometric applications. This talk will focus on research and recent advances of heterogeneous face recognition, including fundamental models (lightCNN and light vision transformer), recognition method and recognition from synthesis.