Research map / Label-efficient and robust learning / Recognition and long-tail
Feature Transfer Learning for Face Recognition With Under-Represented Data
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 |
|---|---|---|---|
| Apply center-based feature transfer and alternating training to the softmax + m-L2 baseline | Face recognition model trained with under-represented classes · MS-Celeb-1M hold-out, 10K10K | IJB-A verification FAR@.01: 74.07 -> 80.25 | Helped |
| Apply center-based feature transfer and alternating training to the softmax + m-L2 baseline | Face recognition model trained with under-represented classes · MS-Celeb-1M hold-out, 10K10K | MS1M nearest-neighbor accuracy, UR: 84.10 -> 88.16 | Helped |
| Apply center-based feature transfer and alternating training to the softmax + m-L2 baseline | Face recognition model trained with under-represented classes · MS-Celeb-1M hold-out, 10K30K | IJB-A verification FAR@.001: 47.17 -> 61.04 | Helped |
| Apply center-based feature transfer and alternating training to the softmax + m-L2 baseline | Face recognition model trained with under-represented classes · MS-Celeb-1M hold-out, 10K50K | IJB-A verification FAR@.001: 53.44 -> 62.60 | Helped |
| Add m-L2 regularization to softmax (ablation) | Face recognition model · IJB-A | verification at FAR 0.001: 77.4 -> 80.2 | Helped |
| Add m-L2 regularization to softmax (ablation) | Face recognition model · LFW | accuracy: 98.60 -> 99.18 | Helped |
| Add m-L2 regularization to softmax (ablation) | LeNet++ · MNIST testing set | accuracy: 99.06% to 99.35% | Helped |
| Use the filtered identity feature f instead of the rich feature g (ablation) | FTL · MS-Celeb-1M hold-out, 10K10K | LFW accuracy: 96.72 -> 98.33 | Helped |
| Increase the number of under-represented classes from 10K to 50K (ablation) | FTL · MS-Celeb-1M hold-out | MS1M nearest-neighbor accuracy, UR: 88.16 -> 89.36 | Helped |
| Add media pooling and template adaptation (ablation) | FTL · IJB-A | verification at FAR 0.001: 82.9 -> 91.2 | Helped |
| Use FTL rather than the softmax + m-L2 baseline | Face recognition model · LFW | accuracy: 99.18 -> 99.55 | Helped |
| Use FTL rather than the softmax + m-L2 baseline | Face recognition model · IJB-A | identification Rank-1: 93.9 -> 94.8 | Helped |
Closest papers by meaning
- Meta Variance Transfer: Learning to Augment from the Others (ICML 2020 · similarity 0.85)
- Attributes Shape the Embedding Space of Face Recognition Models (ICML 2025 · similarity 0.83)
- Vec2Face: Scaling Face Dataset Generation with Loosely Constrained Vectors (ICLR 2025 · similarity 0.83)
- LAFS: Landmark-based Facial Self-supervised Learning for Face Recognition (CVPR 2024 · similarity 0.83)
- RepFace: Refining Closed-Set Noise with Progressive Label Correction for Face Recognition (AAAI 2025 · similarity 0.83)
- Masked Face Recognition with Generative-to-Discriminative Representations (ICML 2024 · similarity 0.82)
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.