Research map / Label-efficient and robust learning / Knowledge distillation
CoupleFace: Relation Matters for Face Recognition Distillation
Jiaheng Liu, Haoyu Qin, Yichao Wu, Jinyang Guo, Ding Liang, Ke Xu
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 |
|---|---|---|---|
| add FCD embedding alignment instead of classical KD | MobileNetV2 student · IJB-C (TAR@FAR=1e-4) | TAR: 86.12 -> 90.34 (IJB-B 1e-4); 89.03 -> 92.68 (IJB-C 1e-4) | Improved |
| add mutual relation distillation (RAD) on top of FCD | MobileNetV2 student · IJB-C (TAR@FAR=1e-5) | TAR: 87.74 -> 89.57 | Improved |
| add mutual relation distillation (RAD) on top of FCD | MobileNetV2 student · MegaFace (10^6 distractors) | rank-1 accuracy: 93.65 -> 94.27 | Improved |
| add ArcFace loss finetuning stage (CoupleFace+) | CoupleFace · IJB-B (TAR@FAR=1e-5) | TAR: 84.63 -> 85.12 | Improved |
| add selection of only valid relations (teacher closer than student) (ablation) | CoupleFace-A · IJB-B (TAR@FAR=1e-5) | TAR: 83.68 -> 84.23 | Improved |
| add margin q to RAD loss (ablation) | CoupleFace-B · IJB-B (TAR@FAR=1e-5) | TAR: 84.23 -> 84.63 | Improved |
| add margin q to RAD loss (ablation) | CoupleFace-B · IJB-C (TAR@FAR=1e-5) | TAR: 88.99 -> 89.57 | Improved |
| add informative relation mining with memory bank instead of batch-only relations (ablation) | CoupleFace-C · IJB-B (TAR@FAR=1e-5) | TAR: 83.18 -> 84.63 | Improved |
| mine most similar identities instead of random K=100 identities (ablation) | CoupleFace-D · IJB-C (TAR@FAR=1e-5) | TAR: 89.04 -> 89.57 | Improved |
| add mining with feature bank instead of mini-batch relations (ablation) | CoupleFace-E · IJB-B (TAR@FAR=1e-5) | TAR: 83.32 -> 84.63 | Improved |
| use R-50 teacher instead of MBNet to generate prototype set (ablation) | CoupleFace-MBN · IJB-B (TAR@FAR=1e-5) | TAR: 83.91 -> 84.63 | Improved |
| use R-100 instead of R-50 to generate prototype set (ablation) | CoupleFace · IJB-C (TAR@FAR=1e-5) | TAR: 89.57 -> 89.58 | No clear effect |
Later papers that compared with this method
1 later accepted paper reports CoupleFace as a baseline in their results tables; 1 of them report a better number on most of the results they share with it.
- ICD-Face: Intra-class Compactness Distillation for Face Recognition (ICCV 2023 · better on 6 of 6 shared results)
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Grouped Knowledge Distillation for Deep Face Recognition (AAAI 2023 · similarity 0.89)
- ICD-Face: Intra-class Compactness Distillation for Face Recognition (ICCV 2023 · similarity 0.88)
- Evaluation-Oriented Knowledge Distillation for Deep Face Recognition (CVPR 2022 · similarity 0.88)
- Rethinking Feature-Based Knowledge Distillation for Face Recognition (CVPR 2023 · similarity 0.87)
- Knowledge Distillation Meets Open-Set Semi-supervised Learning (IJCV 2024 · similarity 0.87)
- Teaching Where to Look: Attention Similarity Knowledge Distillation for Low Resolution Face Recognition (ECCV 2022 · similarity 0.87)
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.
