Learnable Rotary Position Embeddings for Face

The evolution of face recognition models has witnessed remarkable advancements over the past decades, transforming from simple statistical approaches to sophisticated deep learning architectures. Face recognition systems have advanced to generate distinctive embeddings that effectively represent facial features. These embeddings encode the unique characteristics of an individual’s face and are mapped into a high-dimensional space. Within this space, embeddings corresponding to the same person are positioned closer together, facilitating accurate identity verification and comparison. In this paper, we introduce our technique to obtain face embeddings using Rotary Positional Embeddings with learnable parameters. In combination with facial landmarks, our model demonstrates impressive performance on face recognition and verification tasks.