ArcFace

The core innovation of ArcFace is the direct addition of an angular margin penalty m to the target logit in the angular domain. Unlike cosine margin losses that operate on the cosine space, ArcFace adds the margin directly to the angle θ between the feature vector and the weight vector. This creates a geodesic distance margin on the hypersphere, which corresponds exactly to the arc length. The resulting decision boundary is linear in the angular space, providing a more rigorous geometric interpretation and leading to superior intra-class compactness and inter-class discrepancy.

ArcFace operates on L2-normalized embeddings and weight vectors projected onto a hypersphere of radius s. This normalization step is critical because it:

• Removes radial variation, forcing the model to learn only angularly discriminative features
• Makes the loss purely dependent on the angle between feature and weight vectors
• Stabilizes training by bounding feature magnitudes
• Enables the additive angular margin to function correctly in the geodesic space The feature scale parameter s controls the radius of the hypersphere, directly influencing the gradient dynamics and the temperature of the softmax distribution.

ArcFace explicitly optimizes two competing objectives simultaneously:

• Intra-class compactness: Features belonging to the same identity are pulled tightly together, minimizing angular variance within a class cluster
• Inter-class discrepancy: Features from different identities are pushed apart by the angular margin, creating clear separation boundaries This dual optimization results in highly discriminative embeddings where the maximum intra-class angular distance is constrained to be smaller than the minimum inter-class angular distance by at least the margin m. The margin acts as a buffer zone that prevents class overlap in the angular space.

ArcFace belongs to a family of margin-based softmax losses, each with distinct geometric properties:

• SphereFace (A-Softmax): Applies a multiplicative angular margin, which produces a non-linear decision boundary that varies with the angle, making optimization unstable at small angles
• CosFace (AM-Softmax): Adds a cosine margin directly to the cosine logit, operating in the cosine space rather than the angular space, which lacks the exact geodesic interpretation
• ArcFace: Adds the margin in the angular domain, achieving the most natural and stable decision boundary with exact correspondence to geodesic distance on the hypersphere Empirically, ArcFace consistently outperforms both predecessors on face verification benchmarks like LFW, CFP-FP, and AgeDB-30.

ArcFace supports flexible margin configurations to handle diverse training data quality:

• Hard margin: A fixed angular margin m applied uniformly to all samples, enforcing strict separation regardless of sample difficulty
• Easy margin: A dynamic margin that adapts based on the angular distance between the feature and the target weight, relaxing the penalty for already well-classified samples
• Combined margin: A hybrid approach that adds both an additive angular margin m1 and an additive cosine margin m2, providing fine-grained control over the decision boundary shape The easy margin variant is particularly useful when training on noisy datasets where some samples are inherently ambiguous or mislabeled, preventing the model from overfitting to outliers.

ArcFace exhibits superior training stability compared to multiplicative margin approaches due to its linear decision boundary in the angular space. Key convergence properties include:

• The gradient of the loss with respect to the angle is well-behaved across the entire angular range, avoiding the vanishing gradient problem at small angles that plagues SphereFace
• The additive nature of the margin ensures that the target logit decreases linearly with the margin, providing consistent optimization signals
• The feature scale parameter s can be tuned to control the concentration of the posterior probability distribution, with typical values ranging from 30 to 64
• ArcFace converges reliably without requiring complex learning rate schedules or margin annealing strategies, making it practical for large-scale deployment.

A foundational metric learning objective that operates on triplets of samples: an anchor, a positive (same class), and a negative (different class). The loss minimizes the distance between the anchor and positive while maximizing the distance to the negative by a fixed margin. Unlike ArcFace, which adds angular margin directly in the softmax logits, Triplet Loss requires careful hard negative mining to select informative triplets that violate the margin constraint, making training convergence highly dependent on sampling strategy.

The fundamental distance metric used in ArcFace's angular space. It measures the cosine of the angle between two L2-normalized vectors, ignoring magnitude and focusing purely on directional alignment. ArcFace transforms the standard softmax logit (dot product of weights and features) into the cosine of the angle between them. By applying the additive angular margin penalty directly to this cosine space, the decision boundary becomes a geodesic distance margin on the hypersphere, enforcing maximum intra-class compactness.

An extension of self-supervised contrastive methods that leverages explicit class labels to treat all samples from the same category as positive pairs. While ArcFace achieves a similar goal—tight intra-class clustering—through an angular margin in the classification layer, Supervised Contrastive Loss operates directly on the embedding representations. It pulls all same-class embeddings together in a batch while pushing apart all other classes, creating a more robust embedding space that generalizes beyond the closed-set training identities.

The strategic selection of negative samples that are deceptively similar to the anchor but belong to a different class. In face recognition, hard negatives are identities that look visually similar (e.g., doppelgangers). ArcFace's angular margin penalty implicitly handles hard negatives by creating a more stringent decision boundary—the additive margin forces the model to separate even the most confusable identities by a wider angular gap, reducing reliance on explicit mining strategies during training.

The standard classification objective that ArcFace fundamentally reformulates. Traditional softmax loss separates features by maximizing the posterior probability of the correct class, but lacks an explicit mechanism to minimize intra-class variance. ArcFace transforms this by: (1) normalizing weights and features, (2) computing the arc-cosine of the dot product to get the angle, and (3) adding a constant angular margin m to the target logit. This simple modification converts a separable feature space into a discriminative one with clear angular margins between identities.

The deployment architecture commonly used for ArcFace-based face recognition systems. A Bi-Encoder independently encodes the probe image and gallery images into normalized embedding vectors using the same ArcFace-trained model. Recognition is performed via nearest neighbor search using cosine similarity. This allows the gallery embeddings to be pre-computed offline, enabling real-time identification at scale. The angular margin learned during training ensures that embeddings of the same identity cluster tightly, while different identities are separated by at least the margin angle.