Research map / Label-efficient and robust learning / Recognition and long-tail
AQUAFace: Age-Invariant Quality Adaptive Face Recognition for Unconstrained Selfie vs ID Verification
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 AQUAFace, an adaptive-margin contrastive plus ArcFace loss that uses a GMM-based age- and quality-aware likelihood ratio to handle joint age and quality variation in selfie vs ID face verification, and releases a synthetic AgeDB+MORPH dataset (Syn AM).
Problem
Face recognition under simultaneous age-gap and image-quality variation, as in selfie vs ID document matching, where existing quality-aware and age-invariant margin losses address only one factor.
Approach
An adaptive-margin contrastive loss whose margin is scaled by an age- and quality-aware likelihood ratio (AQUALR) estimated with a GMM over comparison scores, combined with ArcFace identity loss and fine-tuning a pretrained ArcFace model.
Key results
- On VerifyMe (real) easy/medium/hard bins at 0.1% FAR, AQUAFace (Real+Syn) reaches 88.56 / 80.89 / 42.25 GAR, above EQFace's 84.87 / 77.14 / 23.12 in easy and hard bins.
- With ResNet18 trained on VerifyMe and fine-tuned on Syn AM, AQUAFace attains 99.79 LFW, 97.82 CALFW, 94.57 CPLFW, 99.03 CFP-FP, 98.99 AgeDB, 95.23 IJB-B, 96.97 IJB-C, and 70.12 TinyFace TAR@FAR=0.01%.
- Training only on synthetic Syn AM yields lower performance than real VerifyMe training (e.g., TinyFace 58.59 vs 69.35), and fine-tuning on synthetic data lowers hard-bin real-data GAR (42.25 vs 51.33 at 0.1% FAR).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Train on VerifyMe and fine-tune on Syn AM rather than train on VerifyMe alone | AQUAFace · CALFW | accuracy: 96.36 -> 97.82 | Helped |
| Train on VerifyMe and fine-tune on Syn AM rather than train on VerifyMe alone | AQUAFace · IJB-C | TAR@FAR=0.01%: 96.45 -> 96.97 | Helped |
| Train on VerifyMe and fine-tune on Syn AM rather than train on VerifyMe alone | AQUAFace · TinyFace | Rank-1: 69.35 -> 70.12 | Helped |
| Train on VerifyMe and fine-tune on Syn AM rather than train on VerifyMe alone | AQUAFace · VerifyMe hard pairs | GAR@0.1%FAR: 51.33 -> 42.25 | Hurt |
| Train on Syn AM alone rather than VerifyMe | AQUAFace · CALFW | accuracy: 96.36 -> 84.37 | Hurt |
| Set margin parameter m to 0.5 rather than 0.2 (ablation) | AQUAFace · VerifyMe easy pairs | performance (Table 5): 73.42 -> 74.49 | Helped |
| Set margin parameter m to 1.0 rather than 0.5 (ablation) | AQUAFace · VerifyMe easy pairs | performance (Table 5): 74.49 -> 71.41 | Hurt |
| Set margin parameter m to 0.5 rather than 0.2 (ablation) | AQUAFace · Syn AM easy pairs | performance (Table 5): 89.76 -> 91.46 | Helped |
| Set margin parameter m to 1.0 rather than 0.5 (ablation) | AQUAFace · Syn AM easy pairs | performance (Table 5): 91.46 -> 86.17 | Hurt |
| Use AQUAFace rather than the strongest listed baseline | ResNet18 face verification · CALFW, CPLFW, CFP-FP, AgeDB, and TinyFace | benchmark metrics | Helped |
| Use AQUAFace rather than the strongest listed baseline | ResNet18 face verification · LFW, IJB-B | accuracy / TAR@FAR=0.01% | Mixed |
Surprising findings
- Fine-tuning on synthetic data improves easy and medium real-data performance but reduces hard-bin real-data performance relative to real-only training.
- Vanilla ArcFace (ResNet18, CASIA-WebFace) scores very low on the selfie vs ID bins (e.g., 29.38 easy at 0.1% FAR), far below quality-aware baselines.
Limitations
- Synthetic data lacks hard pairs, so fine-tuning hurts hard real-data performance.
- Main training data (VerifyMe) is private, limiting reproducibility.
- Evaluated mainly with ResNet18 backbone; ResNet100 results deferred to supplementary material.
Experimental setup
Datasets: VerifyMe AgeDB MORPH CASIA-WebFace LFW CFP-FP CALFW CPLFW IJB-B IJB-C TinyFace
Benchmarks: LFW CALFW CPLFW CFP-FP AgeDB IJB-B IJB-C TinyFace VerifyMe Syn AM
Metrics: GAR@0.1% FAR GAR@1% FAR 1:1 verification accuracy TAR@FAR=0.01% Rank-1 Rank-5
Baselines: LightCNN ElasticFace MagFace AdaFace CurricularFace EQFace ArcFace CosFace QMagFace
Builds on: ArcFace AdaFace MagFace QMagFace contrastive loss
Closest papers by meaning
- AdaFace: Quality Adaptive Margin for Face Recognition (CVPR 2022 · similarity 0.88)
- When Age-Invariant Face Recognition Meets Face Age Synthesis: A Multi-Task Learning Framework (CVPR 2021 · similarity 0.87)
- From Cradle to Cane: A Two-Pass Framework for High-Fidelity Lifespan Face Aging (NeurIPS 2025 · similarity 0.86)
- Look across Elapse: Disentangled Representation Learning and Photorealistic Cross-Age Face Synthesis for Age-Invariant Face Recognition (AAAI 2019 · similarity 0.85)
- Vec2Face: Scaling Face Dataset Generation with Loosely Constrained Vectors (ICLR 2025 · similarity 0.83)
- CemiFace: Center-based Semi-hard Synthetic Face Generation for Face Recognition (NeurIPS 2024 · similarity 0.82)
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.