Source-linked AI summary
ArcFace: Additive Angular Margin Loss for Deep Face Recognition
Jiankang Deng, Jia Guo, Jing Yang, Niannan Xue, Irene Kotsia, Stefanos Zafeiriou
TL;DR
Face recognition needs embeddings that are more discriminative for open-set recognition and methods that remain effective with massive noisy labels. The paper introduces ArcFace, extends it with sub-centers for noise isolation and web-face cleaning, and uses pretrained-model gradients with BN priors for face synthesis. The methods achieve state-of-the-art results across ten benchmarks, including 97.27% TPR@FPR=1e-4 on IJB-C, while producing identity-preserved and visually plausible face images.
Problem
Softmax embeddings are not sufficiently discriminative for open-set recognition, while large-scale clean annotations are difficult to obtain and noisy-label methods have practical limitations.
Method
ArcFace adds an angular margin to normalized target angles, while sub-center ArcFace uses K positive sub-centers to isolate clean, hard, and noisy faces during training.
Results
ArcFace achieves state-of-the-art performance on ten face-recognition benchmarks, reaching 97.27% TPR@FPR=1e-4 on IJB-C, and generates identity-preserved, visually plausible faces without an additional generator or discriminator.
Takeaways & Limitations
The proposed framework improves discriminative face embeddings, supports automatic large-scale web-face cleaning, and strengthens generative face synthesis.
Takeaways & Limitations
Existing noisy-data methods can remain dependent on initial model performance, and co-mining requires twin networks, reducing practicality for large models and datasets.
Abstract
from arXiv · showhide
Recently, a popular line of research in face recognition is adopting margins in the well-established softmax loss function to maximize class separability. In this paper, we first introduce an Additive Angular Margin Loss (ArcFace), which not only has a clear geometric interpretation but also significantly enhances the discriminative power. Since ArcFace is susceptible to the massive label noise, we further propose sub-center ArcFace, in which each class contains $K$ sub-centers and training samples only need to be close to any of the $K$ positive sub-centers. Sub-center ArcFace encourages one dominant sub-class that contains the majority of clean faces and non-dominant sub-classes that include hard or noisy faces. Based on this self-propelled isolation, we boost the performance through automatically purifying raw web faces under massive real-world noise. Besides discriminative feature embedding, we also explore the inverse problem, mapping feature vectors to face images. Without training any additional generator or discriminator, the pre-trained ArcFace model can generate identity-preserved face images for both subjects inside and outside the training data only by using the network gradient and Batch Normalization (BN) priors. Extensive experiments demonstrate that ArcFace can enhance the discriminative feature embedding as well as strengthen the generative face synthesis.
1 INTRODUCTION
The paper introduces ArcFace to improve open-set face-recognition embeddings through an additive angular margin, then extends it with sub-centers for massive label noise and explores gradient- and BN-prior-based face synthesis. Across these directions, the proposed methods target discriminative, robust, practical, and generative improvements.
- Motivation: Softmax features can separate closed-set identities but remain insufficiently discriminative for open-set face recognition, while the classifier matrix grows linearly with identities.Triplet loss also faces combinatorial triplet growth and difficult semi-hard sample mining on large-scale datasets.
- ArcFace: ArcFace adds an angular margin to the target angle after feature and center normalization, directly optimizing a geodesic distance margin on the normalized hypersphere.The modified target logit is rescaled by a fixed feature norm before continuing with the softmax procedure.
- Noisy labels: Margin-based softmax methods require clean annotations, while existing noisy-data approaches can depend on initial model quality, time-varying weighting, piece-wise losses, or twin-network training.A manually cleaned dataset cited in the paper required 50 annotators working continuously for one month.
- Sub-center ArcFace: Sub-center ArcFace assigns K sub-centers to each class, allowing samples to approach any positive sub-center and isolating hard or noisy faces from the dominant clean sub-class.This relaxed intra-class constraint is designed to improve robustness under massive real-world noise and supports automatic web-face cleaning.
- Related approaches: Triplet and Tuplet compare samples locally with Euclidean margins, whereas ArcFace and sub-center ArcFace compare samples globally with classes or subclasses using angular margins.The comparison is summarized in Figure 1.
- Generative face synthesis: Without an additional generator or discriminator, pretrained ArcFace generates identity-preserved and visually plausible faces using network gradients and Batch Normalization statistic priors.The inverse mapping targets face images from low-dimensional feature or latent representations.
- Reported outcomes: ArcFace achieves state-of-the-art performance on ten face-recognition benchmarks and reaches 97.27% TPR@FPR=1e-4 on IJB-C.The paper also reports that the model strengthens generative face synthesis for subjects inside and outside the training data.
2 RELATED WORK
Related work contrasts metric-learning approaches with margin-based softmax methods, reviews strategies for noisy-label handling, and situates multicenter losses and model inversion in face recognition.
- Face Recognition with Margin Penalty: Triplet loss uses local sample-to-sample comparisons but faces combinatorial triplet growth and difficult semi-hard mining at scale.These challenges make large-scale training computationally demanding and operationally difficult.
- Face Recognition with Margin Penalty: Margin-based softmax methods replace sample-to-sample comparisons with global sample-to-class comparisons, trading greater memory use for improved efficiency and stability.They retain class centers in the model to support these comparisons.
- Noisy Labels: Noisy-label methods include semantic bootstrapping, sample reweighting, hypothetical-label losses, and co-mining, but rely on iterative refinement, hyperparameters, initial-model quality, or twin networks.The reviewed approaches differ in how they detect or reduce the influence of noisy faces.
- Sub-class and Multicenter Methods: Softtriple uses multiple centers and class-wise regularization to capture complex data geometry and reduce intra-class variance, particularly benefiting fine-grained visual retrieval.Its multicenter formulation is presented as a concurrent approach related to the paper’s sub-center idea.
- Model Inversion: Earlier model inversion methods generate class images through input-gradient optimization but require heavy pixel-space regularization because image space is much larger than feature space.The cited regularizers include total variation and Gaussian blur.
3 PROPOSED APPROACH
ArcFace normalizes features and class centers, then adds an angular margin to improve geometric separation. Sub-center ArcFace extends this design to noisy data, while ArcFace inversion reconstructs identity-preserving faces using gradients and BN statistics.
- 3.1 ArcFace: Feature and weight normalization makes predictions depend only on angular relationships, placing embeddings on a hypersphere.The feature norm is fixed through normalization and rescaling, while each class weight is ℓ2-normalized.
- 3.1 ArcFace: ArcFace adds margin m to the target angle, enhancing intra-class compactness and inter-class discrepancy through a geodesic-distance margin.The target angle is computed with arc-cosine, modified by the additive angular margin, and mapped back with cosine.
- 3.1 ArcFace: ArcFace produces a more evident margin between nearest classes than Norm-Softmax in normalized 2D toy embeddings.Both methods use fixed-radius arc-space features, but Norm-Softmax leaves noticeable decision-boundary ambiguity.
- 3.1 ArcFace: ArcFace has a constant linear angular margin, whereas SphereFace and CosFace produce nonlinear angular margins in binary decision boundaries.The comparison is presented geometrically under the binary classification case.
- 3.2 Sub-center ArcFace: Sub-center ArcFace assigns K positive sub-centers per identity and uses the closest subclass similarity, separating hard and noisy faces from dominant clean subclasses.On CASIA with K = 10, hard and noisy samples are reported as separated from the majority of clean samples; on MS1MV0, dropping non-dominant centers affects intra-class compactness because hard samples can also be removed.
- 3.3 Inversion of ArcFace: ArcFace inversion reconstructs identity-preserving, visually plausible faces by jointly using ArcFace gradients and face-statistic priors stored in Batch Normalization layers.The approach optimizes identity and statistic losses; replacing classification loss with feature-pair ℓ2 loss enables open-set generation.
4 EXPERIMENTS
Experiments show that ArcFace improves face-recognition performance across standard benchmarks, while sub-center ArcFace increases robustness to noisy training data and supports automatic data cleaning.
- 4.2 Ablation Study on ArcFace: ArcFace achieves the highest verification accuracy across the three CASIA test sets, with a best angular margin of 0.5.The combined margin framework improves over individual SphereFace and CosFace margins but remains below ArcFace.
- 4.3 Ablation Study on Sub-center ArcFace: Sub-center ArcFace is more robust to massive noise, reducing the noisy-data drop from ArcFace’s 96.50% to 90.27% to 93.72%.The comparison uses clean MS1MV3 and noisy MS1MV0 training data.
- 4.3 Ablation Study on Sub-center ArcFace: Excessive sub-centers reduce accuracy from 93.72% to 67.94%, showing that noise tolerance must be balanced against intra-class compactness.The paper selects K=3 considering this trade-off and GPU memory consumption.
- 4.3 Ablation Study on Sub-center ArcFace: Sub-center ArcFace trained on noisy MS1MV0 reaches 95.92% versus 96.50% for ArcFace trained on manually cleaned MS1MV3.With enlarged training data, it reaches 96.91% versus 96.50%.
- 4.4 Benchmark Results: ArcFace models outperform strong baselines across LFW, YTF, CFP-FP, CPLFW, AgeDB, and CALFW, including pose- and age-variant benchmarks.Performance on LFW and YTF is near saturation, yet ArcFace remains among the most competitive methods.
- 4.4 Benchmark Results: On IJB-B and IJB-C, ArcFace improves over softmax by about 3–5%, reaches 96.02% and 97.27% TPR at FPR=1e-4, and reduces error versus CurricularFace by 22.57% and 29.09%.IBUG-500K training and ResNet100 provide the reported higher-performance configuration.
5 CONCLUSIONS
ArcFace enhances discriminative face embeddings, while sub-center ArcFace isolates clean faces from hard or noisy faces for web-face cleaning. The pretrained model also supports identity-preserving face synthesis, although generated pose and expression are not controllable.
- ArcFace enhances the discriminative power of deep feature embeddings for face recognition.
- Sub-center ArcFace relaxes intra-class constraints under massive real-world noise by isolating dominant clean-face sub-classes from non-dominant hard or noisy sub-classes.
- This automatic isolation supports cleaning large-scale web faces, and the method consistently outperforms the state of the art across comprehensive experiments.
- A pretrained ArcFace model can generate identity-preserved face images for subjects inside and outside the training data using network gradients and Batch Normalization priors.
- ArcFace inversion approximates the target identity feature, but generated facial poses and expressions are not controllable.