Research map / Label-efficient and robust learning / Knowledge distillation
Masked Autoencoders Are Stronger Knowledge Distillers
Shanshan Lao, Guanglu Song, Boxiao Liu, Yu Liu, Yujiu Yang
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 |
|---|---|---|---|
| add masked-image distillation (MKD) over MGD | RetinaNet-Res18 student · COCO val, 2x | mAP: 36.6 -> 37.3 | Improved |
| add MKD distillation to student | Faster R-CNN Res50 · COCO val, 2x | mAP: 38.4 -> 41.0 | Improved |
| add masked convolution in student backbone (ablation) | MKD · COCO ablation, 1x | mAP: 39.6 -> 39.9 | Improved |
| use random masking instead of grid masking (ablation) | MKD · COCO ablation, 1x | mAP: 37.8 -> 39.9 | Improved |
| use random masking instead of block masking (ablation) | MKD · COCO ablation, 1x | mAP: 39.7 -> 39.9 | Improved |
| align multi-scale features to H/32 resolution in SAM instead of H/16 (ablation) | MKD · COCO ablation, 1x | mAP: 38.9 -> 39.9 | Improved |
| increase decoder depth from 2 to 4 layers (ablation) | MKD · COCO ablation, 1x | mAP: 39.6 -> 39.9 | Improved |
| increase decoder depth from 4 to 8 layers (ablation) | MKD · COCO ablation, 1x | mAP: 39.9 -> 39.8 | No clear effect |
| use unshared mask tokens per FPN scale instead of shared (ablation) | MKD · COCO, RetinaNet-Res18 | mAP: 37.0 -> 37.3 | Improved |
| vary distillation loss weights lambda and gamma (ablation) | MKD · COCO ablation, 1x | mAP: 39.6 to 39.9 | No clear effect |
| add MKD decoder to MGD feature masking (ablation) | MGD · COCO, Retina Res101-Res18 | mAP: 36.6 -> 36.6 | No clear effect |
| apply MKD to semantic segmentation | PspNet-Res18 · Cityscapes | mIoU: 69.85 -> 73.99 | Improved |
Closest papers by meaning
- Masked Autoencoders Enable Efficient Knowledge Distillers (CVPR 2023 · similarity 0.93)
- ACAM-KD: Adaptive and Cooperative Attention Masking for Knowledge Distillation (ICCV 2025 · similarity 0.91)
- Exploring Target Representations for Masked Autoencoders (ICLR 2024 · similarity 0.90)
- DCSF-KD: Dynamic Channel-wise Spatial Feature Knowledge Distillation for Object Detection (AAAI 2025 · similarity 0.89)
- DFD: Distilling the Feature Disparity Differently for Detectors (ICML 2024 · similarity 0.88)
- Asymmetric Masked Distillation for Pre-Training Small Foundation Models (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.
