FG 2025
Unconstrained face recognition
Train and adapt face-recognition models for images that are distant, small or degraded, where clean close-up training alone is not enough.
Co-first author
The quality gap
How it works
Large-scale pretraining and adaptation
A transformer learns face representations from large aligned datasets, then adapts to low-quality imagery through linear probing and full-model fine-tuning. Enlarged aligned crops preserve context around the face.
Separate pretraining settings
The BRIAR reported comparison uses WebFace12M with different architectures: AdaFace ResNet101 and Swin-B. A separate Swin-B model pretrained on WebFace42M reaches 81.38% Rank-1. The paper also evaluates TinyFace, IJB-S, IJB-B and IJB-C.
Staged adaptation
Context-aware crops
Results
The system accepts more genuine face matches while holding the false-accept rate at 0.1%. Both compared models use WebFace12M pretraining, but their architectures differ.
| Method | Backbone | TAR at FAR 0.1% |
|---|---|---|
| AdaFace | ResNet101 | 39.41% |
| Proposed method | Swin-B | 70.84% |
FG 2025, Table II. BRIAR, WebFace12M pretraining.
Architectures differ. The separate Rank-1 result uses WebFace42M.
Architectures differ: AdaFace ResNet101 versus Swin-B. The separate 81.38% Rank-1 result uses WebFace42M.
Presentation
