Research map / Label-efficient and robust learning / Recognition and long-tail
AdaFace: Quality Adaptive Margin for 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 |
|---|---|---|---|
| Use a feature-norm-based adaptive margin rather than CurricularFace | Face recognition trained on MS1MV2 · IJB-B and IJB-C | TAR@FAR=0.01%: 94.80 -> 95.67 (IJB-B); 96.10 -> 96.89 (IJB-C) | Helped |
| Use a feature-norm-based adaptive margin rather than CurricularFace | Face recognition trained on MS1MV2 · IJB-S and TinyFace | Rank-1: 62.43 -> 65.26; 63.81 -> 66.27; 19.54 -> 23.74; 63.68 -> 68.21 | Helped |
| Use a feature-norm-based adaptive margin rather than CurricularFace | Face recognition trained on MS1MV2 · High Quality datasets average | average 1:1 verification accuracy: 97.16 -> 97.19 | No clear effect |
| Add on-the-fly cropping, rescaling, and photometric augmentation (ablation) | AdaFace, ResNet50 trained on MS1MV2 · LQ datasets average | average performance: 40.95 -> 41.82 | Helped |
| Add on-the-fly cropping, rescaling, and photometric augmentation (ablation) | CurricularFace, ResNet50 trained on MS1MV2 · LQ datasets average | average performance: 41.00 -> 40.84 | Hurt |
| Add on-the-fly cropping, rescaling, and photometric augmentation (ablation) | AdaFace, ResNet50 trained on MS1MV2 · HQ datasets average | average performance: 96.72 -> 96.88 | No clear effect |
| Set norm concentration h to 0.22 rather than 0.33 (ablation) | AdaFace · HQ and LQ datasets average | average performance: 93.74 -> 93.67; 35.40 -> 34.92 | Hurt |
| Set norm concentration h to 0.66 rather than 0.33 (ablation) | AdaFace · HQ and LQ datasets average | average performance: 93.74 -> 93.70; 35.40 -> 35.29 | Hurt |
| Set margin m to 0.50 rather than 0.40 (ablation) | AdaFace · HQ and LQ datasets average | average performance: 93.74 -> 93.56; 35.40 -> 35.23 | Mixed |
| Set margin m to 0.75 rather than 0.40 (ablation) | AdaFace · HQ and LQ datasets average | average performance: 93.74 -> 93.37; 35.40 -> 35.69 | Mixed |
| Replace the feature-norm quality proxy with 1-BRISQUE (ablation) | AdaFace · HQ and LQ datasets average | average performance: 93.74 -> 93.43; 35.40 -> 34.55 | Hurt |
| Replace the feature-norm quality proxy with P_y_i (ablation) | AdaFace · HQ and LQ datasets average | average performance: 93.74 -> 93.46; 35.40 -> 35.17 | Hurt |
Closest papers by meaning
- AdaptiveFace: Adaptive Margin and Sampling for Face Recognition (CVPR 2019 · similarity 0.92)
- UniTSFace: Unified Threshold Integrated Sample-to-Sample Loss for Face Recognition (NeurIPS 2023 · similarity 0.89)
- ArcFace: Additive Angular Margin Loss for Deep Face Recognition (CVPR 2019 · similarity 0.88)
- MagFace: A Universal Representation for Face Recognition and Quality Assessment (CVPR 2021 · similarity 0.88)
- AQUAFace: Age-Invariant Quality Adaptive Face Recognition for Unconstrained Selfie vs ID Verification (AAAI 2025 · similarity 0.88)
- RepFace: Refining Closed-Set Noise with Progressive Label Correction for Face Recognition (AAAI 2025 · 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.