Research map / Label-efficient and robust learning / Recognition and long-tail
Rethinking Robust Representation Learning under Fine-Grained Noisy Faces
Bingqi Ma, Guanglu Song, Boxiao Liu, Yu Liu
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 |
|---|---|---|---|
| add class-specific sub-centers to ArcFace (ablation) | ArcFace · IJB-C, Mixture of Noises | TAR@FAR=1e-4: 90.03 -> 93.76 | Improved |
| add sub-center merging (ablation) | ESL (sub-center) · IJB-C, Mixture of Noises | TAR@FAR=1e-4: 93.76 -> 94.51 | Improved |
| add sub-center producing (ablation) | ESL (sub-center + merging) · IJB-C, Mixture of Noises | TAR@FAR=1e-4: 94.51 -> 95.54 | Improved |
| add sub-center dropping (ablation) | ESL (without dropping) · IJB-C, Mixture of Noises | TAR@FAR=1e-4: 95.54 -> 96.22 | Improved |
| use lambda4 = 3 instead of 1 for merging threshold (ablation) | ESL · IJB-C, hyperparameter sweep | TAR@FAR=1e-4: 95.07 -> 96.22 | Improved |
| use lambda1 = 2 instead of 1 (ablation) | ESL · IJB-C, hyperparameter sweep | TAR@FAR=1e-4: 95.73 -> 96.22 | Improved |
| use initial sub-centers M=3 instead of 1 (ablation) | ESL · IJB-C, hyperparameter sweep | TAR@FAR=1e-4: 96.03 -> 96.22 | Improved |
| increase initial sub-centers from 3 to 4 or 5 (ablation) | ESL · IJB-C, hyperparameter sweep | TAR@FAR=1e-4: 96.22 -> 96.20 | No clear effect |
| replace Sub-center ArcFace with ESL | Sub-center ArcFace M=3 · IJB-C, MS1MV0 | TAR@FAR=1e-4: 93.72 -> 96.23 | Improved |
| replace SKH with ESL | SKH + ArcFace M=4 · IJB-C, Mixture of Noises | TAR@FAR=1e-4: 95.16 -> 96.22 | Improved |
| use Sub-center ArcFace on mixed noise instead of plain ArcFace | ArcFace · IJB-C, Mixture of Noises | TAR@FAR=1e-4: 90.03 -> 89.32 | Hurt |
| train ESL online instead of posterior cleaning with retraining | SKH + ArcFace M=3↓1 · IJB-C, MS1MV0 | GPU hours: 128 -> 80 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 2 on most shared results.
- ArcFace — ArcFace: Additive Angular Margin Loss for Deep Face Recognition (CVPR 2019) · better on 4 of 4 shared results
- CosFace — CosFace: Large Margin Cosine Loss for Deep Face Recognition (CVPR 2018) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Noise-Robust Face Recognition via Non-target Similarity Distribution Guided Sample Selection (ECCV 2026 · similarity 0.89)
- Sub-center ArcFace: Boosting Face Recognition by Large-scale Noisy Web Faces (ECCV 2020 · similarity 0.89)
- RepFace: Refining Closed-Set Noise with Progressive Label Correction for Face Recognition (AAAI 2025 · similarity 0.87)
- UIFace: Unleashing Inherent Model Capabilities to Enhance Intra-Class Diversity in Synthetic Face Recognition (ICLR 2025 · similarity 0.86)
- Robust Visual Recognition with Class-Imbalanced Open-World Noisy Data (AAAI 2024 · similarity 0.85)
- Ditch the Denoiser: Emergence of Noise Robustness in Self-Supervised Learning from Data Curriculum (NeurIPS 2025 · similarity 0.84)
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.
