Research map / Label-efficient and robust learning / Noisy labels
Adaptive Label Noise Cleaning With Meta-Supervision for Deep Face Recognition
Yaobin Zhang, Weihong Deng, Yaoyao Zhong, Jiani Hu, Xian Li, Dongyue Zhao, Dongchao Wen
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 adaptive threshold adapter on top of meta-learning (Simulation source) (ablation) | AMC cleaner · CASIA-WebFace cleaned, IJB-C TPR@FPR=1e-5 | TPR (%): 71.93 -> 73.59 | Improved |
| add adaptive threshold adapter on top of meta-learning (MS1M source) (ablation) | AMC cleaner · CASIA-WebFace cleaned, IJB-C TPR@FPR=1e-5 | TPR (%): 72.42 -> 73.78 | Improved |
| add adaptive threshold adapter on top of meta-learning (ablation) | AMC cleaner · MC-mini cleaned, IJB-C TPR@FPR=1e-5 | TPR (%): 92.14 -> 92.36 | Improved |
| add meta-learning instead of fine-tuning (MS1M source) (ablation) | cleaner · CASIA-WebFace cleaned, IJB-C TPR@FPR=1e-5 | TPR (%): 72.69 -> 72.42 for fine-tune vs meta; source 71.63 -> 72.42 | Mixed |
| add meta-learning instead of fine-tuning (Simulation source) (ablation) | cleaner · CASIA-WebFace cleaned, IJB-C TPR@FPR=1e-5 | TPR (%): 72.49 -> 71.93 | Hurt |
| add fine-tuning on pseudo-labeled target data to source-trained cleaner (ablation) | cleaner · MC-mini cleaned, IJB-C TPR@FPR=1e-5 | TPR (%): 92.33 -> 91.95 | Hurt |
| add fine-tuning on target data to source-trained cleaner (MS1M source) (ablation) | cleaner · CASIA-WebFace cleaned, IJB-B TPR@FPR=1e-5 | TPR (%): 62.84 -> 64.27 | Improved |
| use real noisy MS1M source instead of simulated source for AMC | AMC · CASIA-WebFace cleaned, IJB-C TPR@FPR=1e-5 | TPR (%): 73.59 -> 73.78 | Improved |
| use AMC cleaning instead of FaceGraph | ArcFace ResNet-34 · CASIA-WebFace cleaned, IJB-C TPR@FPR=1e-5 | TPR (%): 72.74 -> 73.78 | Improved |
| use AMC cleaning instead of FaceGraph | ArcFace ResNet-50 · MegaFace Challenge 1 | verification TPR (%): 97.76 -> 98.15 | Improved |
| use AMC cleaning instead of FaceGraph | ArcFace ResNet-50 · MC-mini cleaned, IJB-C TPR@FPR=1e-5 | TPR (%): 91.92 -> 92.36 | Improved |
| add manual cleaning to CASIA-WebFace | ArcFace ResNet-34 · CASIA-WebFace, CFP-FP | accuracy (%): 94.73 -> 94.11 | Hurt |
Closest papers by meaning
- Global-Local GCN: Large-Scale Label Noise Cleansing for Face Recognition (CVPR 2020 · similarity 0.87)
- Noise-Tolerant Paradigm for Training Face Recognition CNNs (CVPR 2019 · similarity 0.87)
- RepFace: Refining Closed-Set Noise with Progressive Label Correction for Face Recognition (AAAI 2025 · similarity 0.87)
- The Devil of Face Recognition is in the Noise (ECCV 2018 · similarity 0.87)
- Co-Mining: Deep Face Recognition With Noisy Labels (ICCV 2019 · similarity 0.86)
- Learning to Learn From Noisy Labeled Data (CVPR 2019 · 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.
