Research map / Detection and segmentation / CNN features
ScaleKD: Distilling Scale-Aware Knowledge in Small Object Detector
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 ScaleKD to GFL ResNet50 | GFL ResNet50 · COCO val2017 | APS: 23.3 -> 25.9 | Helped |
| add ScaleKD to RetinaNet ResNet50 | RetinaNet ResNet50 · COCO val2017 | APS: 20.0 -> 24.8 | Helped |
| add ScaleKD to FCOS ResNet50 | FCOS ResNet50 · COCO val2017 | APS: 21.9 -> 27.8 | Helped |
| add ScaleKD | RetinaNet-MobileNetV2 · VisDrone | AP: 21.73 -> 26.08 | Helped |
| add ScaleKD | RetinaNet-ResNet50 · VisDrone | AP: 26.21 -> 29.45 | Helped |
| add the Cross-Scale Assistant (ablation) | RetinaNet ResNet50 ScaleKD · COCO | APS: 21.2 -> 23.5 | Helped |
| use multi-scale cross-attention instead of plain cross-attention (ablation) | RetinaNet ResNet50 ScaleKD · COCO | APS: 23.5 -> 24.8 | Helped |
| add scale-decoupled feature distillation (ablation) | RetinaNet ResNet50 ScaleKD · COCO | APS: 21.2 -> 24.8 | Helped |
| use weight sharing across parallel SDF branches instead of separate weights (ablation) | RetinaNet ResNet50 ScaleKD · COCO | AP: 41.6 -> 41.6 | No clear effect |
| add ScaleKD (ablation) | ResNet-18 detector · COCO | APS: 18.9 -> 20.1 | Helped |
| add ScaleKD (ablation) | ResNet-34 detector · COCO | APS: 21.5 -> 23.1 | Helped |
| perturb teacher bounding boxes by 6 pixels along the diagonal (ablation) | RetinaNet student distillation · COCO | APS: 21.9 -> 21.1 | Hurt |
Closest papers by meaning
- Knowledge Distillation for Object Detection via Rank Mimicking and Prediction-Guided Feature Imitation (AAAI 2022 · similarity 0.90)
- DCSF-KD: Dynamic Channel-wise Spatial Feature Knowledge Distillation for Object Detection (AAAI 2025 · similarity 0.89)
- ScaleKD: Strong Vision Transformers Could Be Excellent Teachers (NeurIPS 2024 · similarity 0.89)
- DetKDS: Knowledge Distillation Search for Object Detectors (ICML 2024 · similarity 0.89)
- DFD: Distilling the Feature Disparity Differently for Detectors (ICML 2024 · similarity 0.88)
- CrossKD: Cross-Head Knowledge Distillation for Object Detection (CVPR 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.