1 intro

• we propose an Additive Angular Margin Loss (ArcFace) to obtain highly discriminative features for face recognition.

existing loss:

softmax loss:

1. the size of linear transformation matrix increases linearly with the n;
2. learned features are separable for closed-set classification problem but not discriminative enough for open-set face recognition.

triplet loss:

1. leading to a significant increase in the number of iteration steps
2. semi-hard sample mining is a quite difficult problem for effective model training.

2 this paper

Propose a new ArcFace loss:

code:
https://github.com/deepinsight/insightface
https://github.com/deepinsight/insightface/tree/c2db41402c627cab8ea32d55da591940f2258276/recognition/arcface_torch
https://github.com/TreB1eN/InsightFace_Pytorch
https://github.com/ronghuaiyang/arcface-pytorch