Research map / Label-efficient and robust learning / Knowledge distillation
Distilling Object Detectors with Global Knowledge
Sanli Tang, Zhongyu Zhang, Zhanzhan Cheng, Jing Lu, Yunlu Xu, Yi Niu, Fan He
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 prototypes to DeFeat (distill weights) (ablation) | DeFeat · VOC, 1x schedule ablation (Res101 teacher, Res50 student) | AP50: 82.0 -> 82.4 | Improved |
| add prototypes to RKD (ablation) | RKD · VOC, 1x schedule ablation (Res101 teacher, Res50 student) | AP50: 81.6 -> 82.0 | Improved |
| add prototypes to Vanilla-KD (ablation) | Vanilla-KD · VOC, 1x schedule ablation (Res101 teacher, Res50 student) | AP50: 81.8 -> 82.2 | Improved |
| add global knowledge to feature-based and response-based local knowledge (ablation) | Ours · VOC, 1x schedule ablation (Res101 teacher, Res50 student) | AP50: 82.6 -> 82.9 | Improved |
| use global knowledge only versus local feature only (ablation) | Ours · VOC, 1x schedule ablation (Res101 teacher, Res50 student) | AP50: 82.0 -> 82.4 | Improved |
| use PGM prototypes instead of K-Means on student features (ablation) | Ours · VOC, 1x schedule ablation (Res101 teacher, Res50 student) | AP50: 82.3 -> 82.9 | Improved |
| use PGM prototypes instead of DBSCAN on teacher features (ablation) | Ours · VOC, 1x schedule ablation (Res101 teacher, Res50 student) | AP50: 82.2 -> 82.9 | Improved |
| use representative prototypes instead of ambiguous instances (ablation) | Ours · VOC, 1x schedule ablation (Res101 teacher, Res50 student) | AP50: 81.8 -> 82.9 | Improved |
| increase prototype update period T from 1 to 3 (ablation) | Ours · VOC, 1x schedule ablation (Res101 teacher, Res50 student) | AP50: 82.9 -> 82.2 | Hurt |
| use K=10 prototypes instead of K=1 (ablation) | Ours · VOC, 1x schedule ablation (Res101 teacher, Res50 student) | AP50: 82.0 -> 82.9 | Improved |
| raise lambda from 10 to 100 (ablation) | Ours · VOC, 1x schedule ablation (Res101 teacher, Res50 student) | AP50: 82.9 -> 82.0 | Hurt |
| raise alpha3 from 5 to 20 (ablation) | Ours · VOC, 1x schedule ablation (Res101 teacher, Res50 student) | AP50: 82.9 -> 82.2 | Hurt |
Closest papers by meaning
- Improve Object Detection with Feature-based Knowledge Distillation: Towards Accurate and Efficient Detectors (ICLR 2021 · similarity 0.93)
- Structured Knowledge Distillation for Accurate and Efficient Object Detection (TPAMI 2023 · similarity 0.93)
- Distilling Object Detectors With Fine-Grained Feature Imitation (CVPR 2019 · similarity 0.92)
- Learning Lightweight Object Detectors via Multi-Teacher Progressive Distillation (ICML 2023 · similarity 0.92)
- DFD: Distilling the Feature Disparity Differently for Detectors (ICML 2024 · similarity 0.91)
- DCSF-KD: Dynamic Channel-wise Spatial Feature Knowledge Distillation for Object Detection (AAAI 2025 · similarity 0.91)
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.
