[論文レビュー] SphereFace: Deep Hypersphere Embedding for Face Recognition

Weiyang Liu, Yandong Wen+4 more|arXiv (Cornell University)|Apr 26, 2017

Face recognition and analysis参考文献 26被引用数 412

ひとこと要約

tldr: SphereFace introduces an angular margin loss (A-Softmax) to train CNNs for open-set face recognition by enforcing discriminative features on a hypersphere. It achieves state-of-the-art like results on LFW and YTF using WebFace data.

ABSTRACT

This paper addresses deep face recognition (FR) problem under open-set protocol, where ideal face features are expected to have smaller maximal intra-class distance than minimal inter-class distance under a suitably chosen metric space. However, few existing algorithms can effectively achieve this criterion. To this end, we propose the angular softmax (A-Softmax) loss that enables convolutional neural networks (CNNs) to learn angularly discriminative features. Geometrically, A-Softmax loss can be viewed as imposing discriminative constraints on a hypersphere manifold, which intrinsically matches the prior that faces also lie on a manifold. Moreover, the size of angular margin can be quantitatively adjusted by a parameter $m$. We further derive specific $m$ to approximate the ideal feature criterion. Extensive analysis and experiments on Labeled Face in the Wild (LFW), Youtube Faces (YTF) and MegaFace Challenge show the superiority of A-Softmax loss in FR tasks. The code has also been made publicly available.

研究の動機と目的

• Motivate open-set face recognition as learning discriminative features with controlled intra-class and inter-class angular distances.
• Introduce Angular Softmax (A-Softmax) loss to impose angular margins on a hypersphere manifold.
• Provide theoretical analysis and lower bounds for the angular margin parameter m to approximate ideal FR criteria.
• Demonstrate empirical effectiveness of SphereFace on standard benchmarks (LFW, YTF, MegaFace) using WebFace training data.

提案手法

• Derive angular softmax (A-Softmax) loss that replaces cosine similarities with an angular margin controlled by an integer m.
• Formulate a generalization of softmax where the decision boundary depends on cos(m theta) for the true class and cos(theta) for other classes, enabling an adjustable angular margin.
• Introduce a monotonic function psi(theta) to extend cos(m theta) beyond its original domain for end-to-end training.
• Embed features on a unit hypersphere (SphereFace) by forcing normalized weights and zero biases, making classification rely on angles.
• Derive a lower bound m_min for m in binary and multi-class settings, suggesting m=4 as a practical choice to approximate the open-set criterion.

実験結果

リサーチクエスチョン

• RQ1 Can angular margin learning on a hypersphere improve discriminability for open-set face recognition?
• RQ2 How should the angular margin parameter m be chosen to approximate the ideal intra-class vs inter-class angular distance criteria?
• RQ3 Do hypersphere-based embeddings learned with A-Softmax outperform traditional softmax and Euclidean-margin losses on standard FR benchmarks?
• RQ4 What are the empirical gains on benchmarks such as LFW, YTF, and MegaFace when using SphereFace with WebFace training data?

主な発見

• A-Softmax enforces a larger angular margin than softmax, yielding more discriminative features on a hypersphere.
• LFW accuracy with m=4 reaches 99.42% (from WebFace training with 64-layer CNN), and YTF accuracy reaches 95.0%.
• L-Softmax and other Euclidean-margin losses are outperformed by SphereFace across evaluated architectures on LFW/YTF.
• Deeper architectures gain more from A-Softmax than with standard softmax, with notable improvements on difficult benchmarks.
• SphereFace demonstrates competitive performance against models trained on private data, highlighting effectiveness with publicly available CASIA-WebFace data.