← Back to Table of Contents

Chapter 22 β€” Face Recognition Architectures

β€œFace architectures must be fast enough for real-time, accurate enough for anti-spoofing, and discriminative enough to tell 7 billion faces apart.”

MTCNN: Cascaded Multi-Task Face Detection (2016)

Planned content: Three-stage cascade for face detection + 5-point landmark localization. P-Net (Proposal Net): stride-2 sliding window FCN β†’ fast face/non-face + coarse bbox; runs on full image. R-Net (Refine Net): 24Γ—24 input β†’ refine candidates, reduce FP. O-Net (Output Net): 48Γ—48 input β†’ precise localization + 5 landmarks. Multi-task loss: classification + regression + landmark.

πŸ“Š Planned diagram: MTCNN cascade β€” image pyramid β†’ P-Net filter β†’ R-Net refine β†’ O-Net final landmarks. FP reduction at each stage, with box count from 10K β†’ 100 β†’ 5.

RetinaFace: Multi-Task Face Detection (2019)

Planned content: FPN backbone (MobileNet or ResNet) + multi-task head: face classification, 4D bbox regression, 10D 5-point landmark regression, and 3D face reconstruction (optional). SSH context module for enhanced receptive field. State-of-art face detection backbone for pipelines.

πŸ“Š Planned diagram: RetinaFace head architecture β€” per anchor: 2 class logits + 4 box deltas + 10 landmark offsets + (optional) 3D shape coefficients.

SCRFD: Sample-Constrained Face Detection (2021)

Planned content: Efficient face detector. Sample constraint: limit positive/negative sample ratio during training. Competitive with RetinaFace at lower FLOPs. Good for mobile deployment.

MobileFaceNet: The Efficient FR Backbone (2018)

Planned content: Designed specifically for face recognition on mobile. Key differences from general MobileNetV2: bottleneck ratio reduced, global depthwise conv layer replacing GAP, Linear bottleneck to preserve face embedding quality. Input [B, 3, 112, 112] β†’ 20 layers β†’ [B, 512]. ~1M params, ~0.22 GFLOPs.

πŸ“Š Planned diagram: MobileFaceNet architecture β€” block diagram with each layer’s output shape [B, C, H, W]. Global DW conv layer highlighted.

πŸ“Š Planned table: MobileFaceNet vs. IResNet-50 vs. IResNet-100 β€” parameters, FLOPs, LFW acc, IJB-C TAR@FAR=1e-4.

IResNet: Improved ResNet for Face

Planned content: Residual networks adapted for face recognition. Remove max pooling, use stride-1 in early layers to preserve spatial detail. IR-SE (with SE blocks). IResNet-50/100. PartialFC trick: random sample a subset of classes per mini-batch for memory-efficient large-scale training.

FaceNet: Triplet Loss Approach (2015)

Planned content: Google’s approach. Inception backbone. L2-normalized 128D embedding. Triplet loss with online hard mining. 200M face training data. First to achieve near-human LFW accuracy.

ArcFace Network Design (2019)

Planned content: IResNet-50 backbone + ArcFace head. Input [B, 3, 112, 112] β†’ backbone β†’ [B, 512] β†’ L2 normalize β†’ ArcFace head [512, C_identities]. Training on MS1MV3 (93K identities). The complete training recipe: loss, optimizer, batch size, margin, scale.

πŸ“Š Planned diagram: Full ArcFace model β€” backbone producing embedding, then ArcFace head transforming cosine similarities into margined logits for CE loss.

AdaFace: Quality-Adaptive Margin (2022)

Planned content: Improvement over ArcFace for low-quality (blurry, occluded) faces. Embedding norm as image quality proxy. High norm β†’ large margin (similar to ArcFace). Low norm β†’ small margin. Achieves better in-the-wild recognition.

πŸ“Š Planned diagram: AdaFace margin function β€” margin vs. embedding norm curve, showing adaptive behavior.

Model Selection Guide

πŸ“Š Planned table: Face recognition model comparison β€” detector, backbone, landmark model, embedding size, LFW %, IJB-C TAR@FAR=1e-4, FPS on mobile.

Next: Chapter 23 β€” Tracking Architectures β†’


Last updated: May 2026