Research map / Label-efficient and robust learning / Knowledge distillation
Look One and More: Distilling Hybrid Order Relational Knowledge for Cross-Resolution Image Recognition
Shiming Ge, Kangkai Zhang, Haolin Liu, Yingying Hua, Shengwei Zhao, Xin Jin, Hao Wen
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 |
|---|---|---|---|
| use a high-accuracy ResNet assistant (OUR-32-R) instead of lower-accuracy VGG assistant (ablation) | OUR-16 student · CIFAR100 test | accuracy (%): 69.87 -> 74.65 | Improved |
| HORKD 16x16 student vs KD baseline at 32x32 | low-res classification · CIFAR100 test | accuracy (%): 74.26 -> 74.65 | Improved |
| reduce input resolution 32x32 to 16x16 | HORKD student · CIFAR100 test | accuracy (%): 77.76 -> 74.65 | Hurt |
| HORKD vs SKD at 32x32 on LFW | low-res face verification · LFW | accuracy (%): 89.72 -> 93.83 | Improved |
| HORKD vs SKD at 16x16 on LFW | low-res face verification · LFW | accuracy (%): 85.87 -> 90.03 | Improved |
| HORKD vs SKD on UCCS 16x16 | low-res face identification · UCCS | top-1 error (%): 32.75 -> 22.19 | Improved |
| add cross-structure distillation stage (assistant) before cross-resolution (ablation) | cross-resolution metric learning · metric learning | accuracy: +3.03% | Improved |
| reduce student input resolution from 256 to 96 (ablation) | OUR-R student · Cars196 | score: 82.23 -> 73.41 | Hurt |
| HORKD student (ResNet18, 256) vs ResNet50 teacher | metric learning · Cars196 / CUB2011 / SOP | score: 75.91 -> 82.23 (Cars196) | Mixed |
| add highest-order terms (use all L1+L2+L3+LC) vs without L1 (ablation) | relational distillation loss · low-res face recognition (32x32) | accuracy | Hurt |
| increase order of relational loss terms (ablation) | relational distillation loss · low-res face recognition (32x32) | accuracy | Improved |
| distill directly cross-resolution without assistant stage (ablation) | metric learning student · metric learning | accuracy | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 0 of 1 on most shared results.
- ArcFace — ArcFace: Additive Angular Margin Loss for Deep Face Recognition (CVPR 2019) · better on 0 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
- Teaching Where to Look: Attention Similarity Knowledge Distillation for Low Resolution Face Recognition (ECCV 2022 · similarity 0.90)
- MTKD: Multi-Teacher Knowledge Distillation for Image Super-Resolution (ECCV 2024 · similarity 0.88)
- Multi-Scale Aligned Distillation for Low-Resolution Detection (CVPR 2021 · similarity 0.88)
- Knowledge Distillation and Student-Teacher Learning for Visual Intelligence: A Review and New Outlooks (TPAMI 2021 · similarity 0.87)
- NORM: Knowledge Distillation via N-to-One Representation Matching (ICLR 2023 · similarity 0.86)
- CoupleFace: Relation Matters for Face Recognition Distillation (ECCV 2022 · 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.
