Skip to content

Face Recognition Architectures

Definition

This article covers the specific deep learning architectures and loss functions used in modern face recognition systems โ€” the models that generate face embeddings for identity verification.

Loss Function Evolution

graph LR
    A["Softmax<br/>(2014)"] --> B["Center Loss<br/>(2016)"]
    B --> C["SphereFace<br/>(2017)"]
    C --> D["CosFace<br/>(2018)"]
    D --> E["ArcFace<br/>(2019)"]
    E --> F["AdaFace<br/>(2022)"]
    E --> G["ElasticFace<br/>(2022)"]

    style E fill:#4051B5,color:#fff
    style F fill:#2E7D32,color:#fff

ArcFace (Additive Angular Margin)

Aspect Details
Core idea Add angular margin penalty m to the angle between feature and class center
Formula L = -log(exp(sยทcos(ฮธ + m)) / (exp(sยทcos(ฮธ + m)) + ฮฃ exp(sยทcos(ฮธj))))
Default params s=64, m=0.5
Benefit Creates clear geometric boundary between classes on hypersphere
Weakness Fixed margin โ€” same difficulty for easy and hard samples
Usage Most widely used in production eKYC systems

AdaFace (Quality-Adaptive)

Aspect Details
Core idea Adapt margin based on image quality โ€” harder margin for high-quality, softer for low-quality
Quality proxy Feature norm as quality indicator
Benefit Handles quality mismatch (ID photo vs selfie) that ArcFace struggles with
eKYC relevance Directly addresses the core eKYC challenge of cross-quality matching

ElasticFace

Aspect Details
Core idea Random margin sampled from distribution โ€” elastic class boundaries
Benefit More flexible decision boundaries, better generalization
Training Margin drawn from Gaussian or uniform distribution each iteration

Backbone Architectures

Standard Server Backbones

Backbone Params GFLOPs LFW AgeDB-30 CFP-FP Best For
IResNet-100 65M 24.2 99.83 98.35 99.07 Production server
IResNet-50 44M 12.3 99.80 97.95 98.62 Balanced server
IResNet-34 33M 7.4 99.78 97.60 98.20 Efficient server

Mobile/Edge Backbones

Backbone Params GFLOPs LFW Speed (Mobile) Best For
MobileNetV3-Large 5.4M 0.45 99.50 15-30ms Standard mobile
MobileFaceNet 0.99M 0.22 99.55 5-15ms Ultra-lightweight
EfficientNet-B0 5.3M 0.39 99.55 20-40ms Balanced mobile
GhostNet 5.2M 0.15 99.45 5-10ms Fastest mobile
EdgeNeXt-S 5.6M 1.0 99.60 15-25ms Modern efficient

Vision Transformers

Backbone Params GFLOPs LFW Notes
ViT-Small 22M 4.6 99.80 Competitive with CNNs
ViT-Base 86M 17.6 99.83 Matches ResNet-100
DeiT-Small 22M 4.6 99.78 Distilled, efficient

Training Pipeline

graph TD
    A[Training Data<br/>MS1MV2 / Glint360K] --> B[Data Augmentation<br/>Flip, crop, color jitter]
    B --> C[Backbone<br/>IResNet-100]
    C --> D[BN + Dropout]
    D --> E[FC Layer โ†’ 512-d embedding]
    E --> F[ArcFace Head<br/>s=64, m=0.5]
    F --> G[Cross-Entropy Loss]

    G --> H[SGD / AdamW Optimizer]
    H --> I[Learning Rate Schedule<br/>Cosine annealing / step decay]

    style F fill:#4051B5,color:#fff

Training Hyperparameters (Typical)

Parameter Value
Batch size 512 (distributed across GPUs)
Optimizer SGD (momentum=0.9, weight_decay=5e-4)
Learning rate 0.1, decayed at epochs 20, 28, 32
Total epochs 34-40
Embedding dim 512
Input size 112 ร— 112
Augmentation Horizontal flip, random erasing

Benchmarks

Verification Benchmarks

Benchmark What It Tests Top Performance
LFW Labeled Faces in the Wild โ€” general FR 99.83% (near-saturated)
AgeDB-30 Cross-age (30-year gap) 98.35%
CFP-FP Cross-pose (frontal vs profile) 99.07%
CALFW Cross-age with LFW pairs 96.20%
CPLFW Cross-pose with LFW pairs 93.37%
IJB-C Unconstrained verification (NIST) TAR=97% @ FAR=1e-4

Key Takeaways

Summary

  • ArcFace remains the standard โ€” angular margin loss on hypersphere creates discriminative embeddings
  • AdaFace is specifically relevant for eKYC โ€” handles quality mismatch between ID photos and selfies
  • IResNet-100 is the standard server backbone; MobileFaceNet for mobile
  • Vision Transformers are competitive but CNNs still dominate in production due to speed
  • Training requires large-scale datasets (5M+ images) with clean labels
  • LFW is saturated (99.8%+) โ€” real evaluation requires cross-age, cross-pose, and cross-quality benchmarks