Latent-Space Identity Feature Re-Registration Across Recognition Systems

A transformation model is introduced as an intermediary component that maps identity feature information from the old feature calculation device's representation domain to the new device's representation domain. This model learns the correspondence between the two domains and enables automatic conversion of registered information, allowing the system to benefit from the new device's performance while maintaining compatibility with existing user data.

The system performs self-service transformation of identity feature information automatically using the transformation model. Instead of requiring manual intervention for each user's re-registration, the model autonomously converts the feature information from the old domain to the new domain, significantly reducing the time and effort needed while ensuring all users are properly adapted to the new system.

The transformation model extracts and processes only the essential identity feature information from the existing database without requiring access to user images or videos. By working solely with the abstract feature vectors already stored in the database, the system achieves re-registration functionality while completely avoiding the handling of sensitive personal information, thus eliminating the associated security risks.

The transformation model is trained in advance using paired data from the old and new feature calculation devices before the actual re-registration process begins. This preliminary training phase allows the model to learn the transformation relationships and store them in its parameters. During deployment, the model can quickly apply these pre-learned transformations without requiring complex real-time computation, thus improving efficiency while keeping the operational system relatively simple.