Research map / Label-efficient and robust learning / Knowledge distillation
AdaDistill: Adaptive Knowledge Distillation for Deep Face Recognition
Fadi Boutros, Vitomir Struc, Naser Damer
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 |
|---|---|---|---|
| distill teacher class centers through margin softmax instead of vanilla MSE feature KD (ablation) | MobileFaceNet student · IJB-C TAR@FAR=1e-5 | TAR (%): 79.79 -> 84.57 | Improved |
| adapt class centers with EMA instead of fixed teacher centers (ablation) | ArcDistill · IJB-C TAR@FAR=1e-5 | TAR (%): 84.57 -> 88.23 | Improved |
| weight alpha by hard-sample importance (ablation) | AdaArcDistill · IJB-C TAR@FAR=1e-5 | TAR (%): 88.23 -> 89.13 | Improved |
| use ArcFace margin 0.45 instead of 0.50 (ablation) | AdaArcDistill · IJB-C TAR@FAR=1e-4 | TAR (%): 92.97 -> 93.27 | Improved |
| use CosFace instead of ArcFace margin penalty (ablation) | AdaDistill · IJB-C TAR@FAR=1e-4 | TAR (%): 93.27 vs 93.05 | No clear effect |
| use stronger teacher ResNet100 instead of ResNet50 (ablation) | AdaArcDistill · IJB-C TAR@FAR=1e-5 | TAR (%): 89.32 -> 87.48 | Hurt |
| use weaker teacher ResNet18 instead of ResNet50 (ablation) | AdaArcDistill · IJB-C TAR@FAR=1e-5 | TAR (%): 89.32 -> 88.10 | Hurt |
| use a different student training dataset than the teacher (Casia-WebFace) (ablation) | AdaArcDistill · IJB-C TAR@FAR=1e-5 | TAR (%): 57.75 -> 86.25 | Improved |
| use synthetic IDiff-Face as student training data (ablation) | AdaArcDistill · IJB-C TAR@FAR=1e-5 | TAR (%): 3.43 -> 24.83 | Improved |
| use AdaArcDistill instead of ReFO+ on MegaFace and small benchmarks | MobileFaceNet student · MegaFace | rank-1 identification (%): AdaArcDistill 92.12 vs ReFO+ 92.41 | Mixed |
| use adaptive distillation instead of fixed centers (ablation) | training cost · training time per batch | seconds: 0.134 -> 0.247 | Hurt |
Closest papers by meaning
- Evaluation-Oriented Knowledge Distillation for Deep Face Recognition (CVPR 2022 · similarity 0.92)
- Grouped Knowledge Distillation for Deep Face Recognition (AAAI 2023 · similarity 0.89)
- Knowledge Distillation and Student-Teacher Learning for Visual Intelligence: A Review and New Outlooks (TPAMI 2021 · similarity 0.87)
- Exclusivity-Consistency Regularized Knowledge Distillation for Face Recognition (ECCV 2020 · similarity 0.87)
- DeepKD: A Deeply Decoupled and Denoised Knowledge Distillation Trainer (NeurIPS 2025 · similarity 0.86)
- VkD: Improving Knowledge Distillation using Orthogonal Projections (CVPR 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.
