InterFace:Adjustable Angular Margin Inter-class Loss for Deep Face Recognition

TL;DR

InterFace introduces an adjustable angular margin loss for face recognition that enhances class separability by applying dynamic margins between features and all class weights, outperforming fixed-margin methods on multiple benchmarks.

Contribution

The paper proposes a novel loss function, InterFace, which allows adjustable margins between features and all class weights, improving discriminative power in face recognition.

Findings

Outperforms fixed-margin methods on five benchmarks

Advances state-of-the-art in face recognition accuracy

Provides geometric explanation and extensive comparisons

Abstract

In the field of face recognition, it is always a hot research topic to improve the loss solution to make the face features extracted by the network have greater discriminative power. Research works in recent years has improved the discriminative power of the face model by normalizing softmax to the cosine space step by step and then adding a fixed penalty margin to reduce the intra-class distance to increase the inter-class distance. Although a great deal of previous work has been done to optimize the boundary penalty to improve the discriminative power of the model, adding a fixed margin penalty to the depth feature and the corresponding weight is not consistent with the pattern of data in the real scenario. To address this issue, in this paper, we propose a novel loss function, InterFace, releasing the constraint of adding a margin penalty only between the depth feature and the corresponding weight to push the separability of classes by adding corresponding margin penalties between the depth features and all weights. To illustrate the advantages of InterFace over a fixed penalty margin, we explained geometrically and comparisons on a set of mainstream benchmarks. From a wider perspective, our InterFace has advanced the state-of-the-art face recognition performance on five out of thirteen mainstream benchmarks. All training codes, pre-trained models, and training logs, are publicly released \footnote{$https://github.com/iamsangmeng/InterFace$}.