Research map / Label-efficient and robust learning / Knowledge distillation
Focal and Global Knowledge Distillation for Detectors
Check your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 FGD feature and global-relation distillation | RetinaNet-Res50 · COCO | mAP: 37.4 -> 39.6 | Helped |
| Add FGD feature and global-relation distillation | Faster RCNN-Res50 · COCO | mAP: 38.4 -> 40.4 | Helped |
| Add FGD feature and global-relation distillation | FCOS-Res50 · COCO | mAP: 38.5 -> 42.1 | Helped |
| Add teacher-to-student parameter inheritance | FGD RetinaNet-Res50 · COCO | mAP: 40.4 -> 40.7 | Helped |
| Combine focal and global losses rather than using either alone (ablation) | RetinaNet-Res50 with FGD · COCO | mAP: 40.2 -> 40.4 | Helped |
| Add focal distillation without global distillation (ablation) | RetinaNet-Res50 · COCO | mAP: 37.4 -> 40.2 | Helped |
| Add global distillation without focal distillation (ablation) | RetinaNet-Res50 · COCO | mAP: 37.4 -> 40.2 | Helped |
| Use spatial and channel attention together (ablation) | RetinaNet-Res50 focal distillation · COCO | mAP: 40.0 -> 40.2 | Helped |
| Use GcBlock instead of Non-Local for global relation distillation (ablation) | Faster RCNN-Res50 · COCO | mAP: 39.8 -> 41.5 | Helped |
| Set attention temperature to 0.5 rather than 1.0 (ablation) | RetinaNet-Res50 FGD · COCO | mAP: 40.2 -> 40.4 | Helped |
| Distill foreground and background separately with different weights (ablation) | RetinaNet-Res50 detector distillation · COCO | mAP: 38.9 -> 39.4 | Helped |
| Distill foreground and background together without splitting their weights (ablation) | RetinaNet-Res50 detector distillation · COCO | mAP: 39.3 -> 38.9 | Hurt |
Closest papers by meaning
- Improve Object Detection with Feature-based Knowledge Distillation: Towards Accurate and Efficient Detectors (ICLR 2021 · similarity 0.93)
- PKD: General Distillation Framework for Object Detectors via Pearson Correlation Coefficient (NeurIPS 2022 · 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 0.90)
- DetKDS: Knowledge Distillation Search for Object Detectors (ICML 2024 · similarity 0.88)
- CrossKD: Cross-Head Knowledge Distillation for Object Detection (CVPR 2024 · 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.