Research map / Label-efficient and robust learning / Recognition and long-tail
WebFace260M: A Benchmark for Million-Scale Deep Face Recognition
Zheng Dong Zhu, Guan Huang, Jiankang Deng, Yun Ye, Junjie Huang, Xinze Chen, Jiagang Zhu, Tian Yang, Dalong Du, Jiwen Lu, Jie Zhou
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| train on WebFace42M instead of MS1MV2 | ArcFace ResNet-100 · IJB-C | TAR@FAR=1e-4: 96.03 -> 97.70 | Improved |
| train on WebFace42M instead of MS1MV2 | ArcFace ResNet-100 · Our test set SFR All | FNMR@FMR=1e-5: 9.88 -> 2.98 | Improved |
| train on WebFace42M instead of MS1MV2 | CurricularFace ResNet-100 · Our test set MFR All-Masked | FNMR@FMR=1e-5: 69.56 -> 42.97 | Improved |
| increase WebFace portion from 10% to 30% to 100% (ablation) | ArcFace ResNet-100 · IJB-C | TAR@FAR=1e-4: 96.77 -> 97.47 -> 97.70 | Improved |
| increase WebFace portion from 10% to 100% (ablation) | ArcFace ResNet-14 · IJB-C | TAR@FAR=1e-4: 93.13 -> 93.92 -> 94.22 | Improved |
| add simulated mask augmentation to training data (ablation) | ArcFace ResNet-100 on WebFace42M · Our test set MFR All-Masked | FNMR@FMR=1e-5: 47.25 -> 33.87 | Improved |
| add simulated mask augmentation to training data (ablation) | ArcFace ResNet-100 on WebFace4M · Our test set MFR All-Masked | FNMR@FMR=1e-5: 70.97 -> 59.89 | Improved |
| add more CAST iterations (1 to 3) (ablation) | ArcFace ResNet-100 on WebFace · IJB-C | TAR@FAR=1e-4: 97.28 -> 97.51 -> 97.70 | Improved |
| add a 4th CAST iteration (ablation) | ArcFace ResNet-100 on WebFace · IJB-C | TAR@FAR=1e-4: 97.70 -> 97.66 | No clear effect |
| use ResNet-200 instead of ResNet-100 as CAST teacher/student (ablation) | CAST · IJB-C | TAR@FAR=1e-4: 97.70 -> 97.71 | No clear effect |
| use Glint360K instead of MS1MV2 for initial teacher (ablation) | CAST · IJB-C | TAR@FAR=1e-4: 97.70 -> 97.68 | No clear effect |
| use MobileFaceNet instead of ResNet-100 as CAST teacher/student (ablation) | CAST · IJB-C | TAR@FAR=1e-4: 97.70 -> 95.82 | Hurt |
Closest papers by meaning
- WebFace260M: A Benchmark Unveiling the Power of Million-Scale Deep Face Recognition (CVPR 2021 · similarity 0.98)
- The MegaFace Benchmark: 1 Million Faces for Recognition at Scale (CVPR 2016 · similarity 0.91)
- BroadFace: Looking at Tens of Thousands of People at Once for Face Recognition (ECCV 2020 · similarity 0.89)
- Sub-center ArcFace: Boosting Face Recognition by Large-scale Noisy Web Faces (ECCV 2020 · similarity 0.88)
- Vec2Face: Scaling Face Dataset Generation with Loosely Constrained Vectors (ICLR 2025 · similarity 0.87)
- HyperFace: Generating Synthetic Face Recognition Datasets by Exploring Face Embedding Hypersphere (ICLR 2025 · similarity 0.86)
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.
