Research map / Label-efficient and robust learning / Knowledge distillation
Localization Distillation for Dense Object Detection
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 |
|---|---|---|---|
| Distill box probability distributions with LD instead of using the GFocal baseline | GFocal ResNet-50 · MS COCO val2017 | AP: 40.1 -> 42.1 | Helped |
| Use LD rather than pseudo bounding-box regression (ablation) | GFocal ResNet-50 · MS COCO val2017 | AP: 40.5 -> 41.1 | Helped |
| Add classification KD on the main distillation region (ablation) | GFocal ResNet-50 · MS COCO val2017 | AP: 40.1 -> 40.2 | Helped |
| Add localization distillation on the main region (ablation) | GFocal ResNet-50 · MS COCO val2017 | AP: 40.1 -> 41.1 | Helped |
| Add LD on the valuable localization region to main-region LD (ablation) | GFocal ResNet-50 · MS COCO val2017 | AP: 41.1 -> 41.8 | Helped |
| Also add classification KD on the VLR to the Main KD + Main LD + VLR LD scheme (ablation) | GFocal ResNet-50 · MS COCO val2017 | AP: 42.1 -> 42.0 | Hurt |
| Use a softened localization distribution with temperature tau 10 instead of no LD (ablation) | GFocal ResNet-50 · MS COCO val2017 | AP: 40.1 -> 41.1 | Helped |
| Set LD temperature to 1 rather than the default tau 10 (ablation) | GFocal ResNet-50 with LD · MS COCO val2017 | AP: 41.1 -> 40.3 | Hurt |
| Add LD to lightweight ResNet-18 | GFocal ResNet-18 · MS COCO val2017 | AP: 35.8 -> 37.5 | Helped |
| Add LD to RetinaNet | RetinaNet ResNet-50 · MS COCO val2017 | AP: 36.9 -> 39.0 | Helped |
| Add LD to FCOS | FCOS ResNet-50 · MS COCO val2017 | AP: 38.6 -> 40.6 | Helped |
| Add LD to ATSS | ATSS ResNet-50 · MS COCO val2017 | AP: 39.2 -> 41.6 | Helped |
Closest papers by meaning
- Distilling Object Detectors With Fine-Grained Feature Imitation (CVPR 2019 · similarity 0.92)
- Improve Object Detection with Feature-based Knowledge Distillation: Towards Accurate and Efficient Detectors (ICLR 2021 · similarity 0.91)
- DCSF-KD: Dynamic Channel-wise Spatial Feature Knowledge Distillation for Object Detection (AAAI 2025 · similarity 0.90)
- CrossKD: Cross-Head Knowledge Distillation for Object Detection (CVPR 2024 · similarity 0.89)
- DetKDS: Knowledge Distillation Search for Object Detectors (ICML 2024 · similarity 0.88)
- DFD: Distilling the Feature Disparity Differently for Detectors (ICML 2024 · similarity 0.88)
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.