Research map / Label-efficient and robust learning / Knowledge distillation
Distilling Object Detectors via Decoupled Features
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 decoupled object/background neck feature distillation instead of uniform neck distillation (ablation) | ResNet50-FPN student · COCO minival | mAP: 40.1 -> 40.4 | Helped |
| Add decoupled positive/negative proposal distillation rather than distilling all proposals equally (ablation) | ResNet50-FPN student with decoupled neck features · COCO minival | mAP: 40.5 -> 40.8 | Helped |
| Add backbone feature distillation (ablation) | ResNet50-FPN student with decoupled neck and proposal distillation · COCO minival | mAP: 40.8 -> 40.9 | Helped |
| Add background-region distillation to object-region distillation (ablation) | ResNet50-FPN student · COCO minival | mAP: 39.9 -> 40.4 | Helped |
| Replace the object/background selection mask with a random mask (ablation) | ResNet50-FPN student · COCO minival | mAP: 40.4 -> 40.0 | Hurt |
| Replace the FG-Mask with the ground-truth box mask (ablation) | ResNet50-FPN student · COCO minival | mAP: 40.4 -> 40.4 | No clear effect |
| Use only negative proposals rather than only positive proposals for distillation (ablation) | ResNet50-FPN student · COCO minival | mAP: 38.1 -> 38.6 | Helped |
| Use separately weighted positive and negative proposal distillation rather than shared loss weighting (ablation) | ResNet50-FPN student · COCO minival | mAP: 38.2 -> 38.9 | Helped |
| Subsample negative proposals to match the number of positive proposals (ablation) | ResNet50-FPN student · COCO minival | mAP: 38.4 -> 38.2 | Hurt |
| Use teacher-generated rather than student-generated proposals (ablation) | Decoupled classification distillation · COCO minival | mAP: 38.7 -> 38.9 | Helped |
| Apply DeFeat instead of FGFI | ResNet50-FPN student · COCO minival | mAP: 39.9 -> 40.9 | Helped |
| Apply DeFeat instead of TADF | ResNet50-FPN student · COCO minival | mAP: 40.1 -> 40.9 | Helped |
Closest papers by meaning
- Improve Object Detection with Feature-based Knowledge Distillation: Towards Accurate and Efficient Detectors (ICLR 2021 · similarity 0.92)
- Focal and Global Knowledge Distillation for Detectors (CVPR 2022 · 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 0.89)
- DCSF-KD: Dynamic Channel-wise Spatial Feature Knowledge Distillation for Object Detection (AAAI 2025 · similarity 0.89)
- 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.