Research map / Label-efficient and robust learning / Knowledge distillation
Knowledge Distillation for Object Detection via Rank Mimicking and Prediction-Guided Feature Imitation
Read the paperCheck 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 RM to the student detector (ablation) | ResNet18 RetinaNet · COCO minival | AP: 31.9 -> 33.3 (+1.4) | Helped |
| Add PFI to the student detector (ablation) | ResNet18 RetinaNet · COCO minival | AP: 31.9 -> 34.2 (+2.3) | Helped |
| Combine RM and PFI (ablation) | ResNet18 RetinaNet · COCO minival | AP: 31.9 -> 34.8 (+2.9) | Helped |
| Replace traditional soft-label distillation with rank mimicking (ablation) | ResNet18 detectors · COCO | AP: 32.4 -> 33.3; 32.9 -> 33.4; 36.0 -> 36.4 | Helped |
| Add localization-quality rank mimicking to classification-score rank mimicking (ablation) | ResNet18 RetinaNet · COCO | AP: 32.9 -> 33.3 | Helped |
| Use PFI rather than unweighted whole-map feature imitation (ablation) | ResNet18 GFL · COCO | AP: 37.6 -> 38.1 | Helped |
| Restrict feature imitation to positive samples instead of whole maps (ablation) | ResNet18 GFL · COCO | AP: 37.6 -> 37.2 | Hurt |
| Restrict feature imitation to negative samples instead of whole maps (ablation) | ResNet18 GFL · COCO | AP: 37.6 -> 37.5 | Hurt |
| Restrict feature imitation to ground-truth boxes (ablation) | ResNet18 GFL · COCO | AP: 37.6 -> 37.5 | Hurt |
| Distill rank distribution and prediction-guided feature imitation | ResNet50 RetinaNet · COCO | AP: 36.9 -> 40.4 | Helped |
| Use RM + PFI rather than GID | ResNet50 RetinaNet with ResNet101 teacher · COCO | AP: 39.1 -> 39.6 | Helped |
| Use RM + PFI rather than DeFeat | ResNet50 RetinaNet with ResNet152 teacher · COCO | AP: 39.7 -> 40.4 | Helped |
Closest papers by meaning
- Distilling Object Detectors With Fine-Grained Feature Imitation (CVPR 2019 · similarity 0.93)
- PKD: General Distillation Framework for Object Detectors via Pearson Correlation Coefficient (NeurIPS 2022 · similarity 0.92)
- CrossKD: Cross-Head Knowledge Distillation for Object Detection (CVPR 2024 · similarity 0.92)
- DFD: Distilling the Feature Disparity Differently for Detectors (ICML 2024 · similarity 0.91)
- DetKDS: Knowledge Distillation Search for Object Detectors (ICML 2024 · similarity 0.90)
- DCSF-KD: Dynamic Channel-wise Spatial Feature Knowledge Distillation for Object Detection (AAAI 2025 · similarity 0.90)
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.