Research map / Architectures and efficiency / Attention and transformers
Masked Image Modeling With Local Multi-Scale Reconstruction
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 |
|---|---|---|---|
| Use local multi-scale reconstruction instead of global reconstruction (ablation) | LocalMIM on ViT-B · ImageNet-1K | top-1 fine-tuning accuracy: 83.0 -> 83.3 | Helped |
| Use local multi-scale reconstruction instead of global reconstruction (ablation) | LocalMIM on Swin-B · ImageNet-1K | top-1 fine-tuning accuracy: 83.3 -> 83.8 | Helped |
| Use HOG rather than normalized pixels as reconstruction targets (ablation) | LocalMIM · ImageNet-1K, ViT-B | top-1 fine-tuning accuracy: 83.0 -> 83.3 | Helped |
| Use HOG rather than normalized pixels as reconstruction targets (ablation) | LocalMIM · ImageNet-1K, Swin-B | top-1 fine-tuning accuracy: 83.7 -> 83.8 | Helped |
| Increase the mask ratio from 75% to 90% (ablation) | LocalMIM · ImageNet-1K, ViT-B | top-1 fine-tuning accuracy: 83.3 -> 83.0 | Hurt |
| Reduce decoder size to 128D - 4H (ablation) | LocalMIM · ImageNet-1K, ViT-B | top-1 fine-tuning accuracy: 83.3 -> 83.2 | No clear effect |
| Add local reconstructions at earlier stages, reaching all four Swin stages (ablation) | LocalMIM · ImageNet-1K, Swin-B | top-1 fine-tuning accuracy: 83.3 -> 83.8 | Helped |
| Choose ViT-B local-loss locations [2,4,10,12] instead of [1,2,11,12] (ablation) | LocalMIM · ImageNet-1K, ViT-B | top-1 fine-tuning accuracy: 82.9 -> 83.3 | Helped |
| Use fine-to-coarse rather than coarse-to-fine supervision scales (ablation) | LocalMIM · ImageNet-1K, ViT-B | top-1 fine-tuning accuracy: 83.1 -> 83.3 | Helped |
| Use multi-scale rather than single-scale supervision for local reconstruction (ablation) | LocalMIM · ImageNet-1K, ViT-B | top-1 fine-tuning accuracy: 83.0 -> 83.3 | Helped |
| Replace local losses with feature fusion and global reconstruction (ablation) | LocalMIM · ImageNet-1K, Swin-B | top-1 fine-tuning accuracy: 83.8 -> 83.5 | Hurt |
| Isolate gradients after selected reconstruction layers (ablation) | LocalMIM · ImageNet-1K, ViT-B | top-1 fine-tuning accuracy: 83.3 -> 83.0 | No clear effect |
Closest papers by meaning
- Good Helper Is around You: Attention-Driven Masked Image Modeling (AAAI 2023 · similarity 0.90)
- Wavelet-Driven Masked Image Modeling: A Path to Efficient Visual Representation (AAAI 2025 · similarity 0.90)
- On Data Scaling in Masked Image Modeling (CVPR 2023 · similarity 0.90)
- From Prototypes to General Distributions: An Efficient Curriculum for Masked Image Modeling (CVPR 2025 · similarity 0.89)
- Learning Mask Invariant Mutual Information for Masked Image Modeling (ICLR 2025 · similarity 0.87)
- On the Role of Discrete Tokenization in Visual Representation Learning (ICLR 2024 · similarity 0.87)
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.