Research map / Label-efficient and robust learning / Recognition and long-tail
Data Uncertainty Learning in Face Recognition
Check your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| Learn feature means and uncertainty variances jointly rather than using deterministic embeddings | DUL_cls with AM-Softmax ResNet18 · IJB-C (TPR@FPR 0.001%) | TPR: 75.43 -> 88.25 | Helped |
| Learn feature means and uncertainty variances jointly rather than using deterministic embeddings | DUL_cls with AM-Softmax ResNet18 · CFP-FP | accuracy: 96.85 -> 97.28 | Helped |
| Learn feature means and uncertainty variances jointly rather than using deterministic embeddings | DUL_cls with ArcFace ResNet18 · IJB-C (TPR@FPR 0.001%) | TPR: 73.80 -> 87.22 | Helped |
| Add regression-based data uncertainty learning to the pretrained model | DUL_rgs with L2-Softmax ResNet18 · IJB-C (TPR@FPR 0.001%) | TPR: 77.60 -> 74.46 | Hurt |
| Learn mean and variance jointly instead of keeping the deterministic mean fixed | DUL_cls with AM-Softmax ResNet64 · IJB-C (TPR@FPR 0.001%) | TPR: 89.77 -> 88.18 | Hurt |
| Learn mean and variance jointly instead of keeping the deterministic mean fixed | DUL_cls with AM-Softmax ResNet64 · IJB-C (TPR@FPR 0.01%) | TPR: 94.14 -> 94.61 | Helped |
| Set KL regularization weight lambda to 0.0, removing the KL regularization (ablation) | DUL_cls · IJB-C (TPR@FPR 0.001%) | TPR: 88.25 -> 64.92 | Hurt |
| Increase KL regularization weight from lambda=0.01 to lambda=0.1 (ablation) | DUL_cls · YTF | accuracy: 96.46 -> 96.42 | No clear effect |
| Increase KL regularization weight from lambda=0.01 to lambda=0.5 (ablation) | DUL_cls · IJB-C (TPR@FPR 0.001%) | TPR: 88.25 -> 40.23 | Hurt |
| Increase KL regularization weight from lambda=0.01 to lambda=1.0 (ablation) | DUL_cls · IJB-C (TPR@FPR 0.001%) | TPR: 88.25 -> 1.770 | Hurt |
| Train with DUL_cls instead of the baseline under Gaussian-blurred training data | AM-Softmax ResNet18 · Noisy MS-Celeb-1M, 20% blurred, IJB-C (TPR@FPR 0.001%) | TPR: 43.52 -> 89.01 | Helped |
| Train with DUL_rgs instead of the baseline under Gaussian-blurred training data | AM-Softmax ResNet18 · Noisy MS-Celeb-1M, 40% blurred, YTF | accuracy: 95.51 -> 96.68 | Helped |
Closest papers by meaning
- Improving Training and Inference of Face Recognition Models via Random Temperature Scaling (AAAI 2023 · similarity 0.90)
- Striking the Right Balance With Uncertainty (CVPR 2019 · similarity 0.88)
- RepFace: Refining Closed-Set Noise with Progressive Label Correction for Face Recognition (AAAI 2025 · similarity 0.84)
- SURE: SUrvey REcipes for building reliable and robust deep networks (CVPR 2024 · similarity 0.83)
- From Risk to Uncertainty: Generating Predictive Uncertainty Measures via Bayesian Estimation (ICLR 2025 · similarity 0.83)
- A Data-Driven Measure of Relative Uncertainty for Misclassification Detection (ICLR 2024 · similarity 0.83)
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.