Research map / Label-efficient and robust learning / Recognition and long-tail
Sub-center ArcFace: Boosting Face Recognition by Large-scale Noisy Web Faces
Jiankang Deng, Jia Guo, Tongliang Liu, Mingming Gong, Stefanos Zafeiriou
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 |
|---|---|---|---|
| replace the single class center with K=3 sub-centers (ablation) | ArcFace trained on MS1MV0 · IJB-C (MS1MV0, ResNet-50) | TAR@FAR=1e-4: 90.27 -> 93.72 | Improved |
| increase sub-centers from K=3 to K=10 (ablation) | sub-center ArcFace on MS1MV0 · IJB-C (MS1MV0, ResNet-50) | TAR@FAR=1e-4: 93.72 -> 67.94 | Hurt |
| use max pooling instead of softmax pooling over sub-centers (ablation) | sub-center ArcFace K=3 · IJB-C (MS1MV0, ResNet-50) | TAR@FAR=1e-4: 93.55 -> 93.72 | Improved |
| drop non-dominant sub-centers and noisy samples beyond 75 degrees (ablation) | sub-center ArcFace K=3 on MS1MV0 · IJB-C (MS1MV0, ResNet-50) | TAR@FAR=1e-4: 93.72 -> 95.92 | Improved |
| drop non-dominant sub-centers and noisy samples instead of adding a compactness regularizer (ablation) | sub-center ArcFace K=3 · IJB-C (MS1MV0, ResNet-50) | TAR@FAR=1e-4: 93.64 -> 95.92 | Improved |
| vary drop threshold from 70 to 80 degrees (ablation) | sub-center ArcFace K=3 drop 1 · IJB-C (MS1MV0, ResNet-50) | TAR@FAR=1e-4: 95.91 / 95.92 / 95.74 | No clear effect |
| use sub-center ArcFace with drop instead of Co-mining | MS1MV0 noisy training · IJB-C (MS1MV0, ResNet-50) | TAR@FAR=1e-4: 93.82 -> 95.92 | Improved |
| use sub-centers on clean MS1MV3 | ArcFace · IJB-C (MS1MV3, ResNet-50) | TAR@FAR=1e-4: 96.50 -> 96.43 | No clear effect |
| add sub-centers with drop under 25% synthetic open-set noise | ArcFace · IJB-C (MS1MV3 open-set noise, ResNet-50) | TAR@FAR=1e-4: 94.52 -> 95.89 | Improved |
| add sub-centers (K=3, no drop) under 25% synthetic close-set noise | ArcFace · IJB-C (MS1MV3 close-set noise, ResNet-50) | TAR@FAR=1e-4: 94.28 -> 93.57 | Hurt |
| add dropping to sub-centers under 50% close-set noise (ablation) | sub-center ArcFace K=3 · IJB-C (MS1MV3 close-set noise, ResNet-50) | TAR@FAR=1e-4: 72.04 -> 78.12 | Improved |
| train on larger Celeb500K with sub-center drop instead of MS1MV3 ArcFace | ResNet-100 · IJB-C (ResNet-100) | TAR@FAR=1e-4: 96.61 -> 96.96 | Improved |
Closest papers by meaning
- EDFace-Celeb-1 M: Benchmarking Face Hallucination With a Million-Scale Dataset (TPAMI 2022 · similarity 0.93)
- ArcFace: Additive Angular Margin Loss for Deep Face Recognition (CVPR 2019 · similarity 0.90)
- WebFace260M: A Benchmark Unveiling the Power of Million-Scale Deep Face Recognition (CVPR 2021 · similarity 0.89)
- Rethinking Robust Representation Learning under Fine-Grained Noisy Faces (ECCV 2022 · similarity 0.89)
- AdaptiveFace: Adaptive Margin and Sampling for Face Recognition (CVPR 2019 · similarity 0.89)
- RepFace: Refining Closed-Set Noise with Progressive Label Correction for Face Recognition (AAAI 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.
