Research map / Label-efficient and robust learning / Knowledge distillation
Learning Lightweight Object Detectors via Multi-Teacher Progressive Distillation
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 |
|---|---|---|---|
| Progressively distill RetinaNet with HTC then DetectoRS instead of direct distillation from DetectoRS | RetinaNet R50 · MS COCO val2017 | Box AP: 39.5 -> 39.9 | Helped |
| Progressively distill Mask R-CNN with Teacher V then Teacher IV instead of direct distillation from Teacher IV | Mask R-CNN R50 · MS COCO val2017 | Box AP: 41.0 -> 41.4 | Helped |
| Progressively distill Mask R-CNN with Teacher V then Teacher IV instead of direct distillation from Teacher IV | Mask R-CNN R50 · MS COCO val2017 | Mask AP: 37.0 -> 37.3 | Helped |
| Use a progressive teacher sequence rather than a 3× training schedule | RetinaNet R50 · MS COCO val2017 | Box AP: 39.5 -> 39.9 | Helped |
| Use a progressive teacher sequence rather than a 3× training schedule | Mask R-CNN R50 · MS COCO val2017 | Box AP: 40.9 -> 41.4 | Helped |
| Use a progressive teacher sequence instead of simultaneous ensemble distillation (ablation) | Mask R-CNN R50 · MS COCO val2017 | Box AP: 39.8 -> 41.4 | Helped |
| Add Teacher III supervision rather than train the off-the-shelf student without distillation (ablation) | Mask R-CNN R50 · MS COCO val2017 | Box AP: 38.2 -> 40.2 | Helped |
| Add Teacher IV supervision rather than train the off-the-shelf student without distillation (ablation) | Mask R-CNN R50 · MS COCO val2017 | Box AP: 38.2 -> 40.8 | Helped |
| Add Teacher V supervision rather than train the off-the-shelf student without distillation (ablation) | Mask R-CNN R50 · MS COCO val2017 | Box AP: 38.2 -> 40.8 | Helped |
| Use Teacher V before Teacher IV rather than reverse the order (ablation) | Mask R-CNN R50 · MS COCO val2017 | Box AP: 41.0 -> 41.4 | Helped |
| Progressively distill a ResNet-18 Mask R-CNN with Teachers I→V→IV rather than use the off-the-shelf model | Mask R-CNN R18 · MS COCO val2017 | Box AP: 33.3 -> 37.0 | Helped |
| Progressively distill a low-resolution student with higher-resolution teachers | Mask R-CNN R50, 0.25× resolution · MS COCO val2017 | Box AP: 25.8 -> 31.5 | Helped |
Closest papers by meaning
- Distilling Image Classifiers in Object Detectors (NeurIPS 2021 · similarity 0.92)
- Learning Efficient Object Detection Models with Knowledge Distillation (NeurIPS 2017 · similarity 0.92)
- ScaleKD: Strong Vision Transformers Could Be Excellent Teachers (NeurIPS 2024 · similarity 0.89)
- DFD: Distilling the Feature Disparity Differently for Detectors (ICML 2024 · similarity 0.89)
- DCSF-KD: Dynamic Channel-wise Spatial Feature Knowledge Distillation for Object Detection (AAAI 2025 · similarity 0.89)
- DetKDS: Knowledge Distillation Search for Object 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.