[Paper Review] Scalable Angular Discriminative Deep Metric Learning for Face Recognition

Bowen Wu, Wu, Huaming+1 more|arXiv (Cornell University)|Apr 29, 2018

Face recognition and analysis18 references3 citations

TL;DR

This paper proposes a scalable angular discriminative deep metric learning method for face recognition by enforcing larger intra-class cosine similarity than a margin during training, aligning training with testing. It introduces self-adaptive margin learning, feature normalization, and a progressive loss that considers both intra- and inter-class variations, achieving state-of-the-art performance on LFW, YTF, and IJB-A datasets with up to 1.5% accuracy gain.

ABSTRACT

With the development of deep learning, Deep Metric Learning (DML) has achieved great improvements in face recognition. Specifically, the widely used softmax loss in the training process often bring large intra-class variations, and feature normalization is only exploited in the testing process to compute the pair similarities. To bridge the gap, we impose the intra-class cosine similarity between the features and weight vectors in softmax loss larger than a margin in the training step, and extend it from four aspects. First, we explore the effect of a hard sample mining strategy. To alleviate the human labor of adjusting the margin hyper-parameter, a self-adaptive margin updating strategy is proposed. Then, a normalized version is given to take full advantage of the cosine similarity constraint. Furthermore, we enhance the former constraint to force the intra-class cosine similarity larger than the mean inter-class cosine similarity with a margin in the exponential feature projection space. Extensive experiments on Labeled Face in the Wild (LFW), Youtube Faces (YTF) and IARPA Janus Benchmark A (IJB-A) datasets demonstrate that the proposed methods outperform the mainstream DML methods and approach the state-of-the-art performance.

Motivation & Objective

• To address large intra-class variations in features learned via standard softmax loss in deep metric learning for face recognition.
• To align the training process with the testing protocol, which commonly uses cosine similarity for pair-wise similarity computation.
• To reduce reliance on manual hyperparameter tuning by introducing a self-adaptive margin strategy.
• To enhance feature discriminability by jointly modeling intra-class compactness and inter-class separability in a normalized feature space.
• To develop a scalable, differentiable loss function that maintains consistency across training and inference while improving generalization on unconstrained datasets.

Proposed method

• Proposes a novel loss function that enforces intra-class cosine similarity to exceed a fixed margin during training, improving consistency with cosine-based testing.
• Introduces a hard sample mining strategy to focus on misclassified samples, improving margin learning on challenging identities.
• Develops a self-adaptive margin updating mechanism that adjusts the margin based on local feature distribution, reducing manual tuning.
• Applies feature and weight vector normalization per mini-batch to prevent feature norm explosion and stabilize training.
• Proposes a more advanced loss (DLMC) that forces intra-class cosine similarity to exceed the mean of nearest inter-class similarities in exponential feature projection space.
• Uses a normalized exponential feature projection space to enhance discriminative power while maintaining numerical stability.

Experimental results

Research questions

• RQ1 Can enforcing a minimum intra-class cosine similarity margin during training improve face recognition performance?
• RQ2 Does hard sample mining improve the effectiveness of the proposed angular margin loss in deep metric learning?
• RQ3 Can a self-adaptive margin strategy reduce human intervention in hyperparameter tuning while maintaining or improving performance?
• RQ4 How does joint modeling of intra-class compactness and inter-class separability affect feature discriminability in unconstrained face recognition?
• RQ5 Can the proposed loss functions generalize well across diverse datasets with varying levels of pose, illumination, and expression variation?

Key findings

• The proposed method achieves a 0.6% to 0.8% accuracy improvement on the LFW dataset compared to standard softmax loss.
• On the YTF dataset, the method improves performance by 1% to 1.5% over baseline methods.
• The DLMC method achieves a Rank-1 identification rate of 97.34% on the IJB-A dataset, outperforming state-of-the-art methods like NormFace and L-Softmax.
• The ROC AUC at FAR=0.01 reaches 86.02% for the DLMC method on IJB-A, demonstrating strong verification performance.
• The self-adaptive margin strategy significantly reduces the need for manual margin tuning while maintaining or improving accuracy.
• The normalized loss (NLMC) and combined DLMC+MALMC variants show robustness and consistent gains across all three benchmark datasets.