In this repos, we relize commom version of embedding layer for face recognize, meanwhile we add parameters m4 to improve the performace. In the cloth recognition, formula 1.2 improve the top 1 acc from 62.1% to 63.7% for resetnet50.
The embedding space for iras dataset show as follow. Our plan got a more compat embedding, which not only reduce the distance of inner cluster, but also gain the inter distance in "arc-face expand".
Reference:
- https://arxiv.org/pdf/1801.07698 arcface(insight face)
- https://arxiv.org/pdf/1801.09414 cosface
- https://arxiv.org/pdf/1503.03832 am-face
- https://arxiv.org/pdf/1704.08063 sphere-face
- https://arxiv.org/pdf/2002.10857 circle-loss