Gaussian Soft Margin Angular Loss for Face Recognition

Abstract

Advances in deep learning has lead to drastic improvements in face recognition. A key part of these deep models is their loss function. Consequently developing an efficient and suitable loss function has been an important topic in face recognition in the recent years. Angular-margin-based losses achieve an acceptable performance and inter-class separability. However they are held back by their enforcement of hard margins on all the samples of the training dataset, regardless of whether these samples actually differ from all the other classes enough to enforce a margin. It can be argued that in a large enough dataset with many different settings and age gaps, some faces will look similar to the faces of other classes. In an intuitive and expressive embedding, we expect some faces to be embedded near similar classes with a small margin. Thus we propose a loss function that while maximizing the inter-class distance and intra-class compactness, allows for the samples which naturally reside further from class center to have a smaller margin. We implement an extremely light and fast to train model using MobileNets and achieve accuracy comparable to state of the art method.