TransFace: Parametric Attention based Transformer for Face Recognition

Face recognition has made remarkable success thanks to the excellent representational learning of deep Convolutional Neural Networks (CNNs). Recently, Transformer-based methods have achieved extraordinary results in various computer vision tasks, which motivates us to explore efficient Transformer for face recognition. Based on the observation that face patches at the same positions from different identities commonly represent similar semantics, we propose TransFace that improves the self-attention mechanism into parametirc-attention for face recognition. Our model is trained on MSCeleb-1M datasets, and evaluated on several benchmarks, i.e., LFW, CALFW, CPLFW, IJB-B and IJB-C datasets. Extensive experiments demonstrate that the cancellation of Q and K does not cause degradation of performance, and the modified parametric-attention has good portability. It is demonstrated that TransFace can achieve better results than CNNs with the same computational complexity.