Research map / Label-efficient and robust learning / Recognition and long-tail
The Devil of Face Recognition is in the Noise
Fei Wang, Liren Chen, Cheng Li, Shiyao Huang, Yanjie Chen, Chen Qian, Chen Change Loy
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 |
|---|---|---|---|
| manually clean the training set instead of using the noisy sample (ablation) | A-Softmax on MS-Celeb-1M subset · MegaFace rank-1 | rank-1 accuracy (%): 70.56 -> 73.53 | Improved |
| manually clean the training set instead of using the noisy sample (ablation) | Center Loss on MegaFace subset · MegaFace rank-1 | rank-1 accuracy (%): 62.55 -> 67.64 | Improved |
| manually clean the training set instead of using the noisy sample (ablation) | Softmax on MegaFace subset · MegaFace rank-1 | rank-1 accuracy (%): 59.68 -> 62.86 | Improved |
| use cleaned subsets rather than sampled noisy subsets (average over losses and datasets) (ablation) | Attention-56 face models · MegaFace rank-1 | rank-1 accuracy (%): 4.14% | Improved |
| train on clean subset instead of 4x larger raw data (ablation) | A-Softmax on MS-Celeb-1M · MegaFace rank-1 | rank-1 accuracy (%): 73.99 -> 73.53 | Mixed |
| train on IMDb-Face instead of MS-Celeb-1M(V1) | A-Softmax · MegaFace rank-1 | rank-1 accuracy (%): 73.99 -> 84.06 | Improved |
| train on IMDb-Face instead of CASIA-WebFace | A-Softmax · MegaFace rank-1 | rank-1 accuracy (%): 70.89 -> 84.06 | Improved |
| use A-Softmax instead of Softmax on IMDb-Face | Attention-56 · MegaFace rank-1 | rank-1 accuracy (%): 74.75 -> 84.06 | Improved |
| inject label flips instead of outliers at equal noise ratio (ablation) | Softmax, Center, A-Softmax on IMDb-Face · MegaFace rank-1 | rank-1 accuracy | Hurt |
| increase label-flip noise to high ratios (ablation) | A-Softmax on IMDb-Face · MegaFace rank-1 | rank-1 accuracy | Hurt |
| dilute fixed clean data with equal, double or five times label-flip noise (ablation) | Attention-56 on IMDb-Face subsets · MegaFace rank-1 | rank-1 accuracy | Hurt |
| annotate with Scheme I (draw box) rather than yes/no Scheme III (ablation) | annotation scheme · user study on 30 identities | F1 / true positive rate: 90% | Improved |
Closest papers by meaning
- Global-Local GCN: Large-Scale Label Noise Cleansing for Face Recognition (CVPR 2020 · similarity 0.90)
- RepFace: Refining Closed-Set Noise with Progressive Label Correction for Face Recognition (AAAI 2025 · similarity 0.90)
- Noise-Robust Face Recognition via Non-target Similarity Distribution Guided Sample Selection (ECCV 2026 · similarity 0.89)
- Co-Mining: Deep Face Recognition With Noisy Labels (ICCV 2019 · similarity 0.89)
- Noise-Tolerant Paradigm for Training Face Recognition CNNs (CVPR 2019 · similarity 0.87)
- Robust Visual Recognition with Class-Imbalanced Open-World Noisy Data (AAAI 2024 · similarity 0.85)
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.
