Research map / Label-efficient and robust learning / Noisy labels
RepFace: Refining Closed-Set Noise with Progressive Label Correction for Face Recognition
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes RepFace, a face recognition training framework that uses auxiliary synthetic noisy samples, a clean/ambiguous/noisy split by nearest-negative similarity, label fusion via a memory bank, and smoothed label correction to handle closed-set label noise.
Problem
Face recognition training is degraded by closed-set label noise (identity flips to another identity in the dataset), which early-stage training is not robust to and which existing correction methods handle poorly for low-confidence ambiguous samples.
Approach
Injects auxiliary samples with random labels to calibrate a noise-detection threshold early in training, splits samples into clean, ambiguous, and noisy groups by similarity to the positive and nearest negative centers, fuses ambiguous labels with a memory bank of accumulated predictions, and softly corrects noisy labels toward the nearest negative class.
Key results
- On CASIA-WebFace with 20% synthetic closed-set noise, RepFace reaches 95.76 average verification accuracy, matching BoundaryFace's 95.76 and above MV-Softmax's 94.73.
- On MS1Mv2 with 20% noise, RepFace reaches 90.91% TAR@FAR=1e-5 on IJB-C versus 90.79% for BoundaryFace.
- Ablation on CASIA-WebFace with 20% noise shows the average RFW accuracy rising from 88.00 (MV-Softmax) to 90.10 with ASC+SLC+LRF.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use RepFace instead of BoundaryFace | IR-50 trained on CASIA-WebFace with 10% closed-set noise · CASIA-WebFace 10% noise, verification average | accuracy: 95.96 -> 96.15 | Helped |
| Use RepFace instead of BoundaryFace | IR-50 trained on MS1Mv2 with 20% closed-set noise · MS1Mv2 20% closed-set noise, IJB-C FAR=0.0001 | TAR: 94.70 -> 94.77 | Helped |
| Add ASC to MV-Softmax (ablation) | MV-Softmax on CASIA-WebFace with 20% noise · LFW | accuracy: 99.12 -> 99.13 | Helped |
| Add SLC after ASC (ablation) | ASC on CASIA-WebFace with 20% noise · AgeDB | accuracy: 92.87 -> 93.83 | Helped |
| Add LRF to ASC+SLC (ablation) | ASC+SLC on CASIA-WebFace with 20% noise · RFW average | accuracy: 88.78 -> 90.10 | Helped |
| Combine label smoothing and cosine matrix smoothing in SLC (ablation) | SLC on CASIA-WebFace with 20% noise · AgeDB | accuracy: 93.90 -> 94.43 | Helped |
| Use label smoothing rather than direct negative-class assignment (ablation) | SLC on CASIA-WebFace with 20% noise · AgeDB | accuracy: 93.00 -> 93.60 | Helped |
| Set alpha to 0.05 rather than 0.03 (ablation) | RepFace on CASIA-WebFace with 20% noise · LFW | accuracy: 99.23 -> 99.40 | Helped |
| Set alpha to 0.05 rather than 0.07 (ablation) | RepFace on CASIA-WebFace with 20% noise · LFW | accuracy: 99.22 -> 99.40 | Helped |
| Set beta to 0.9 rather than 0.95 (ablation) | RepFace on CASIA-WebFace with 20% noise · LFW | accuracy: 99.33 -> 99.40 | Helped |
| Vary M among 24, 32, 40, and 48 (ablation) | RepFace on CASIA-WebFace with 20% noise · LFW | accuracy: 99.39, 99.40, 99.38, 99.39 | No clear effect |
| Use tau values 0.16, 0.18, 0.20, or 0.22 (ablation) | RepFace on datasets with 10% closed-set noise · LFW, AgeDB, CFP-FP, CALFW, SLLFW | verification accuracy | Mixed |
Surprising findings
- ASC and SLC alone give little improvement; the gains appear mainly once LRF is added.
- On MS1Mv2 with 20% noise, the reported average is 97.67 for RepFace versus 97.63 for BoundaryFace, a very small margin that is not clearly significant.
Limitations
- Recall of noise detection is modest because some closed-set noise is labeled ambiguous rather than noisy.
- Noise is synthetic (random label injection) and the real-world CASIA-WebFace result shows only a small margin.
- Hyperparameters such as tau, alpha, and beta were tuned on test benchmarks.
Experimental setup
Datasets: MS1MV2 CASIA-WebFace LFW SSLFW CALFW CPLFW AgeDB CFP-FP RFW IJB-B IJB-C
Benchmarks: LFW SSLFW CALFW CPLFW AgeDB-30 CFP-FP RFW IJB-B IJB-C
Metrics: 1:1 verification accuracy TAR@FAR=1e-5 TAR@FAR=1e-4
Baselines: ArcFace MV-Softmax CurricularFace RVFace AdaFace BoundaryFace
Builds on: BoundaryFace MV-Softmax ArcFace SELC
Closest papers by meaning
- Navigating Label Ambiguity for Facial Expression Recognition in the Wild (AAAI 2025 · similarity 0.86)
- Unlocking the Power of Open Set: A New Perspective for Open-Set Noisy Label Learning (AAAI 2024 · similarity 0.86)
- Enhanced Sample Selection with Confidence Tracking: Identifying Correctly Labeled Yet Hard-to-Learn Samples in Noisy Data (AAAI 2025 · similarity 0.86)
- Revisiting Interpolation for Noisy Label Correction (AAAI 2025 · similarity 0.85)
- Robust Visual Recognition with Class-Imbalanced Open-World Noisy Data (AAAI 2024 · similarity 0.85)
- NoiseGPT: Label Noise Detection and Rectification through Probability Curvature (NeurIPS 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.