# Additive Margin Softmax for Face Verification

**Authors:** Feng Wang, Weiyang Liu, Haijun Liu, Jian Cheng  
**arXiv:** [1801.05599](https://arxiv.org/abs/1801.05599)  
**Date:** 2018-01-17  
**Summary page:** https://www.summarizepaper.com/en/paper/1801.05599v1/

## Key points

- Authors propose a conceptually simple and geometrically interpretable objective function called additive margin Softmax (AM-Softmax) for deep face verification
- Goal is to learn large-margin face features with small intra-class variation and large inter-class difference for good performance
- Introduce a novel additive angular margin for the Softmax loss, more intuitively appealing than existing methods
- Emphasize the importance of feature normalization in their study
- Conduct experiments on LFW BLUFR and MegaFace datasets, consistently outperform state-of-the-art methods using same network architecture and training dataset
- Discuss two hyperparameters: scale (s) fixed at 30, margin (m) varied from 0.25 to 0.5; best performance achieved with margin between 0.35 and 0.4
- Presents a novel approach for deep face verification using AM-Softmax loss function, superior to existing methods
- Code implementation available for further research and development purposes

## Summary

In this paper, the authors propose a conceptually simple and geometrically interpretable objective function called additive margin Softmax (AM-Softmax) for deep face verification. The face verification task can be seen as a metric learning problem, where the goal is to learn large-margin face features with small intra-class variation and large inter-class difference in order to achieve good performance. Previous works have introduced Large-margin Softmax and Angular Softmax to incorporate the angular margin in a multiplicative manner. However, in this work, the authors introduce a novel additive angular margin for the Softmax loss which is more intuitively appealing and interpretable than existing methods. The authors also emphasize the importance of feature normalization in their study. They conduct experiments on LFW BLUFR and MegaFace datasets to evaluate their proposed AM-Softmax loss function. The results show that their method consistently outperforms state-of-the-art methods using the same network architecture and training dataset. The paper discusses two hyperparameters in their proposed loss function: scale (s) and margin (m). The scale parameter is fixed at 30 while they vary the margin parameter from 0.25 to 0.5. The experimental results demonstrate that setting the margin between 0.35 and 0.4 leads to the best performance. Overall, this paper presents a novel approach for deep face verification using an additive margin Softmax loss function which highlights its superiority over existing methods with code implementation made available for further research and development purposes.