Research map / Label-efficient and robust learning / Knowledge distillation
Multi-scale Cross Distillation for Object Detection in Aerial Images
Kun Wang, Zi Wang, Zhang Li, Xichao Teng, Yang Li
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 |
|---|---|---|---|
| apply MSCD to single-scale RoI Trans | RoI Trans · DOTA-v1.0 | mAP: 73.99 -> 80.49 | Improved |
| apply MSCD to single-scale RoI Trans | RoI Trans · DIOR-R | mAP: 63.10 -> 70.22 | Improved |
| apply MSCD to Oriented RCNN versus its multi-scale inference model | Oriented RCNN · DOTA-v1.0 | mAP: 80.87 -> 79.54 | Hurt |
| apply MSCD to Rotated FR versus its multi-scale inference model | Rotated FR · DOTA-v1.0 | mAP: 78.15 -> 77.82 | Hurt |
| second-stage ACSKD training with LR 0.0005 for 5 epochs (ablation) | RoI Trans · DIOR-R | mAP: 69.08 -> 70.22 | Improved |
| use large LR 0.005 versus 0.0005 in stage 2 (ablation) | RoI Trans MSCD · DIOR-R | mAP: 70.22 -> 69.75 | Hurt |
| parallel multi-branch architecture instead of naive multi-scale training (ablation) | RoI Trans · DIOR-R | mAP: 65.63 -> 68.95 | Improved |
| add ACSKD second stage on top of PMBA (ablation) | RoI Trans · DOTA-v1.0 | mAP: 79.77 -> 80.49 | Improved |
| second-stage training with single-scale cross distillation only (ablation) | RoI Trans · DOTA-v1.0 | mAP: 75.63 -> 76.64 | No clear effect |
| cross distillation across scales and orientations in stage 2 versus none (ablation) | RoI Trans · DOTA-v1.0 | mAP: 75.63 -> 78.18 | Improved |
| single-scale inference with MSCD versus multi-scale inference | two-stage detectors · DOTA-v1.0 | inference time: about 7x less inference time than multi-scale | Improved |
Closest papers by meaning
- ScaleKD: Distilling Scale-Aware Knowledge in Small Object Detector (CVPR 2023 · similarity 0.89)
- Cross Domain Object Detection by Target-Perceived Dual Branch Distillation (CVPR 2022 · similarity 0.88)
- When Object Detection Meets Knowledge Distillation: A Survey (TPAMI 2023 · similarity 0.88)
- Multi-Scale Aligned Distillation for Low-Resolution Detection (CVPR 2021 · similarity 0.88)
- CrossKD: Cross-Head Knowledge Distillation for Object Detection (CVPR 2024 · similarity 0.86)
- DCSF-KD: Dynamic Channel-wise Spatial Feature Knowledge Distillation for Object Detection (AAAI 2025 · similarity 0.86)
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.
