Research map / Label-efficient and robust learning / Knowledge distillation
Discriminability Distillation in Group Representation Learning
Manyuan Zhang, Guanglu Song, Hang Zhou, Yu Liu
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 DDL aggregation instead of average pooling | ArcFace ResNet-101 · YouTube Face | verification accuracy (%): 97.97 -> 98.18 | Improved |
| keep only the top-1 discriminability frame | DDL · YouTube Face | verification accuracy (%): 98.18 -> 97.08 | Mixed |
| add DDL aggregation instead of average pooling | ResNet-101 · IQIYI-VID-FACE | TPR@FPR=1e-4 (%): 65.84 -> 69.05 | Improved |
| add re-scaling of discriminability scores (ablation) | DDL · IQIYI-VID-FACE | TPR@FPR=1e-4 (%): 67.38 -> 69.05 | Improved |
| add DDL aggregation instead of average pooling | ResNet-101 · IJB-C | TAR@FAR=1e-6 (%): 86.69 -> 92.39 | Improved |
| add DDL aggregation instead of average pooling | ResNet-101 · IJB-A | TAR@FAR=1e-3 (%): 97.71 -> 98.44 | Improved |
| add DDNet discriminability weighting (ablation) | MobileFaceNet · IJB-C | TAR@FAR=1e-6 (%): 79.88 -> 87.32 | Improved |
| add DDNet discriminability weighting (ablation) | ResNet-101 trained on IMDB-Face · IJB-C | TAR@FAR=1e-6 (%): 73.12 -> 88.35 | Improved |
| use wider or deeper DDNet (ResNet-34-CD, ResNet-18) (ablation) | DDL with ResNet-18-CD · IJB-C | TAR@FAR=1e-6 (%): 91.14 -> 91.13 | No clear effect |
| add DDL to CosFace-trained base (ablation) | CosFace · YouTube Face | verification accuracy (%): 97.68 -> 97.91 | Improved |
| add DDL to SphereFace-trained base (ablation) | SphereFace · YouTube Face | verification accuracy (%): 96.83 -> 97.12 | Improved |
| add DDL aggregation instead of average pooling | ResNet-50 · Mars | mAP (%): 74.1 -> 77.7 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 2 on most shared results.
- ArcFace — ArcFace: Additive Angular Margin Loss for Deep Face Recognition (CVPR 2019) · better on 4 of 4 shared results
- CosFace — CosFace: Large Margin Cosine Loss for Deep Face Recognition (CVPR 2018) · better on 1 of 1 shared result
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 0.87)
- Perceptual Group Tokenizer: Building Perception with Iterative Grouping (ICLR 2024 · similarity 0.86)
- Learning Task-Agnostic Representations through Multi-Teacher Distillation (NeurIPS 2025 · similarity 0.85)
- Group Distributionally Robust Dataset Distillation with Risk Minimization (ICLR 2025 · similarity 0.85)
- Label-Efficient Group Robustness via Out-of-Distribution Concept Curation (CVPR 2024 · similarity 0.84)
- Calibrating Multi-modal Representations: A Pursuit of Group Robustness without Annotations (CVPR 2024 · similarity 0.84)
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.
