Research map / Label-efficient and robust learning / Recognition and long-tail
Hard Patches Mining for Masked Image Modeling
Haochen Wang, Kaiyou Song, Junsong Fan, Yuxi Wang, Jin Xie, Zhaoxiang Zhang
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 loss prediction and learned hard-patch masking (ablation) | MAE-style RGB MIM · ImageNet-1K | Top-1 accuracy: 82.23 -> 82.95 | Improved |
| Add loss prediction objective without using it to generate masks (ablation) | MAE-style RGB MIM · ImageNet-1K | Top-1 accuracy: 82.23 -> 82.49 | Improved |
| Use learned masking with α₀=0 and αT=0.5 rather than random masking (ablation) | HPM · ImageNet-1K | Top-1 accuracy: 82.49 -> 82.95 | Improved |
| Use fully loss-selected masking throughout instead of retaining randomness (ablation) | HPM · ImageNet-1K | Top-1 accuracy: 82.49 -> 81.40 | Hurt |
| Mask patches with the highest predicted loss rather than the lowest (ablation) | HPM · ImageNet-1K | Top-1 accuracy: 82.36 -> 82.95 | Improved |
| Use easy-to-hard rather than hard-to-easy masking (ablation) | HPM · ImageNet-1K | Top-1 accuracy: 81.71 -> 82.95 | Improved |
| Replace absolute MSE loss prediction with relative BCE (ablation) | HPM · ImageNet-1K | Top-1 accuracy: 82.77 -> 82.95 | Improved |
| Use DINO features rather than RGB pixels as the reconstruction target (ablation) | HPM · ImageNet-1K | Top-1 accuracy: 82.95 -> 84.13 | Improved |
| Add HPM to BEiT (ablation) | BEiT · ImageNet-1K | Top-1 accuracy: 80.9 -> 81.5 | Improved |
| Add HPM to iBOT (ablation) | iBOT · ImageNet-1K | Top-1 accuracy: 82.9 -> 83.4 | Improved |
| Add loss prediction and learned masking (ablation) | RGB-target pre-training · COCO object detection | AP_box: 40.45 -> 42.03 | Improved |
| Add loss prediction and learned masking (ablation) | RGB-target pre-training · ADE20k semantic segmentation | mIoU: 40.49 -> 42.09 | Improved |
Closest papers by meaning
- Bootstrap Masked Visual Modeling via Hard Patch Mining (TPAMI 2025 · similarity 0.95)
- Masked Image Modeling: A Survey (IJCV 2025 · similarity 0.90)
- From Prototypes to General Distributions: An Efficient Curriculum for Masked Image Modeling (CVPR 2025 · similarity 0.88)
- Morphing Tokens Draw Strong Masked Image Models (ICLR 2025 · similarity 0.87)
- Patch-Aware Sample Selection for Efficient Masked Image Modeling (AAAI 2024 · similarity 0.87)
- Wavelet-Driven Masked Image Modeling: A Path to Efficient Visual Representation (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.
