Research map / Label-efficient and robust learning / Domain adaptation and generalization
UMFN: Unified Multi-Domain Face Normalization for Joint Cross-domain Prototype Learning and Heterogeneous Face Recognition
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes UMFN, an unsupervised multi-domain face normalization network (autoencoder + GAN with domain-adversarial feature decoupling) that reconstructs frontal, neutral-expression prototypes across domains and fuses its features with an HFR network trained with contrastive learning.
Problem
Face normalization methods typically handle only one type of facial variation and a single domain, while practical heterogeneous face recognition needs removing multiple variations across visible, NIR and sketch domains without paired/identity-labeled data.
Approach
A unified encoder-mapping-decoder network with a domain classifier trained adversarially to decouple domain-agnostic identity features, a control mask selecting target domain and normalization, multi-scale patch discriminators with WGAN-GP, and an identity-feature similarity loss; the HFR head fuses domain-agnostic features with identity features of generated prototypes and uses a contrastive loss
Key results
- On IJB-A, ResNet-18 with UMFN-normalized faces reaches 95.92% VR versus 90.57% for the plain ResNet-18 baseline.
- On LAMP-HQ, UMFN HFR achieves 99.1% rank-1 and 98.7% VR@FAR=0.001, the best among compared methods such as NLPF (98.8 / 98.3).
- On CASIA NIR-VIS 2.0, UMFN reaches 99.9% rank-1 and 99.7% VR@FAR=0.001, the best among compared methods.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Apply UMFN normalization before ResNet-18 verification | ResNet-18 · LFW | VR: 99.23 -> 99.57 | Helped |
| Apply UMFN normalization before ResNet-18 verification | ResNet-18 · IJB-A | VR: 93.36±0.6 -> 95.92±0.7 | Helped |
| Use UMFN with fused domain-agnostic and identity-discriminative features | HFR network · LAMP-HQ NIR-VIS | Rank-1: 98.8 -> 99.1 | Helped |
| Use UMFN with fused domain-agnostic and identity-discriminative features | HFR network · CASIA NIR-VIS 2.0 | Rank-1: 99.7 -> 99.9 | Helped |
| Add feature fusion (ablation) | HFR network · LAMP-HQ | Rank-1: 91.2 -> 94.7 | Helped |
| Add feature fusion (ablation) | HFR network · CASIA NIR-VIS 2.0 | Rank-1: 95.7 -> 97.5 | Helped |
| Add contrastive learning (ablation) | HFR network · LAMP-HQ | Rank-1: 91.2 -> 97.3 | Helped |
| Add contrastive learning (ablation) | HFR network · CASIA NIR-VIS 2.0 | Rank-1: 95.7 -> 98.7 | Helped |
| Add identity feature similarity loss (ablation) | UMFN · Face normalization | Identity preservation | Helped |
| Add reconstruction loss (ablation) | UMFN · Cross-domain face normalization | Specified-domain prototype generation | Helped |
| Add feature decoupling loss (ablation) | UMFN · Cross-domain face normalization | Domain information removal | Helped |
| Add global path discriminator (ablation) | UMFN · Face normalization | Visual quality | Helped |
Limitations
- Sketch-to-VIS/NIR generation yields slight local distortions due to limited facial detail in sketches.
- Evaluated mainly on LAMP-HQ, BUAA, CUFSF and CASIA datasets; sketch domain data was simulated from NIR photos via software conversion in one experiment.
Experimental setup
Datasets: LAMP-HQ BUAA NIR-VIS CUFSF CASIA NIR-VIS 2.0 VGGFace2 MS1M IJB-A LFW
Benchmarks: LFW IJB-A LAMP-HQ CASIA NIR-VIS 2.0
Metrics: VR Symmetry Index Rank-1 accuracy VR@FAR=0.001
Baselines: PIM TP-GAN DR-GAN Hassner FNM R&R CFR PACH ADCANs MMTN NLPF LPL
Builds on: CycleGAN WGAN-GP PatchGAN
Closest papers by meaning
- Unsupervised Face Normalization With Extreme Pose and Expression in the Wild (CVPR 2019 · similarity 0.86)
- Link-based Contrastive Learning for One-Shot Unsupervised Domain Adaptation (CVPR 2025 · similarity 0.85)
- Cross-Domain Similarity Learning for Face Recognition in Unseen Domains (CVPR 2021 · similarity 0.85)
- UIFace: Unleashing Inherent Model Capabilities to Enhance Intra-Class Diversity in Synthetic Face Recognition (ICLR 2025 · similarity 0.85)
- Adversarial Discriminative Heterogeneous Face Recognition (AAAI 2018 · similarity 0.85)
- Escaping Data Scarcity for High-Resolution Heterogeneous Face Hallucination (CVPR 2022 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.