Research map / Label-efficient and robust learning / Noisy labels
Noise-Robust Face Recognition via Non-target Similarity Distribution Guided Sample Selection
Fanglong Wu, Youqiang Gui, Cheng Peng
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 progressive threshold schedule instead of a fixed upper threshold (ablation) | DDLN · Simulated #40-20-20 (CosFace) | precision / F1: 67.72 -> 99.39 precision, 80.70 -> 99.42 F1 | Improved |
| use rank k=10 instead of k=1 for the upper-tail estimate (ablation) | DDLN · Simulated #40-20-20 | F1: 98.93 -> 99.14 | Improved |
| use rank k=10 instead of k=50 (ablation) | DDLN · Simulated #40-20-20 | F1: 98.56 -> 99.14 | Improved |
| add slack alpha=0.02 (ablation) | DDLN · Simulated #40-20-20 | F1: 99.14 -> 99.42 | Improved |
| vary schedule length nmax among 8, 9, 10 (ablation) | DDLN · Simulated #40-20-20 | F1: 99.38, 99.42, 99.39 | No clear effect |
| vary EMA factor lambda over 0.005 to 0.1 (ablation) | DDLN · Simulated #40-20-20 | F1: 99.32, 99.42, 99.30, 99.29 | No clear effect |
| use k=1 instead of default k=10 under heavy noise (ablation) | DDLN · Simulated #80-40-40 | Avg: 78.11 -> 71.97 | Hurt |
| use k=50 instead of default k=10 under heavy noise (ablation) | DDLN · Simulated #80-40-40 | Avg: 78.11 -> 70.68 | Hurt |
| replace OTSU threshold (RVFace) with DDLN | CosFace · Simulated #80-40-40 | Avg: 60.06 -> 78.11 | Improved |
| replace OTSU threshold with DDLN: FNR at 60% noise | CosFace · Simulated 60% noise | FNR: 7.26 -> 1.24 | Improved |
| add DDLN to CosFace on WebFace2M-Noise | CosFace · WebFace2M-Noise | average over 8 benchmarks: 77.02 -> 84.43 | Improved |
| add DDLN to CosFace on CASIA-WebFace | CosFace · CASIA-WebFace IJB-C | TAR@FAR=1e-5: 67.97 -> 82.29 | 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.
- CosFace — CosFace: Large Margin Cosine Loss for Deep Face Recognition (CVPR 2018) · better on 2 of 2 shared results
- ArcFace — ArcFace: Additive Angular Margin Loss for Deep Face Recognition (CVPR 2019) · 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
- RepFace: Refining Closed-Set Noise with Progressive Label Correction for Face Recognition (AAAI 2025 · similarity 0.91)
- The Devil of Face Recognition is in the Noise (ECCV 2018 · similarity 0.89)
- Navigating Label Ambiguity for Facial Expression Recognition in the Wild (AAAI 2025 · similarity 0.87)
- Enhanced Sample Selection with Confidence Tracking: Identifying Correctly Labeled Yet Hard-to-Learn Samples in Noisy Data (AAAI 2025 · similarity 0.87)
- NoiseGPT: Label Noise Detection and Rectification through Probability Curvature (NeurIPS 2024 · similarity 0.86)
- Robust Visual Recognition with Class-Imbalanced Open-World Noisy Data (AAAI 2024 · 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.
