Research map / Detection and segmentation / CNN features
Improving Pixel-based MIM by Reducing Wasted Modeling Capability
Yuan Liu, Songyang Zhang, Jiacheng Chen, Zhaohui Yu, Kai Chen, Dahua Lin
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 |
|---|---|---|---|
| fuse a shallow layer (layer 0) with output layer for MAE reconstruction (ablation) | MAE ViT-B 300 ep · ADE20K | mIoU: 43.9 -> 46.4 | Improved |
| fuse a deep layer instead of a shallow layer with output layer (ablation) | MAE ViT-B 300 ep · ADE20K | mIoU: 43.9 -> 44.5 | Improved |
| fuse 6 layers rather than 2 (ablation) | MFF_MAE · ADE20K | mIoU: 46.4 -> 47.7 | Improved |
| fuse all 12 layers rather than 6 (ablation) | MFF_MAE · ADE20K | mIoU: 47.7 -> 46.8 | Hurt |
| use linear projection rather than none or non-linear (ablation) | MFF_MAE · ADE20K | mIoU: 46.2 (nonlinear) -> 47.7 | Improved |
| weighted-average fusion rather than self-attention (ablation) | MFF_MAE · ImageNet-1K linear probing | top-1 accuracy: 63.0 -> 63.3 | No clear effect |
| add MFF to MAE (800 ep) | MAE ViT-B · ImageNet-1K 1% low-shot | top-1 accuracy: 45.4 -> 48.0 | Improved |
| add MFF to MAE (300 ep) | MAE ViT-B · ImageNet-1K fine-tuning | top-1 accuracy: 82.8 -> 83.3 | Improved |
| add MFF to MAE | MAE ViT-B · ImageNet-C | mean corruption error: 51.7 -> 49.0 | Improved |
| add MFF to feature-based MIM (EVA, MILAN) | EVA, MILAN · ImageNet linear probing | top-1 accuracy: 69.0 -> 68.9 (EVA) | No clear effect |
| detach shallow layers to exclude deep supervision effect (ablation) | MFF_MAE · ImageNet-1K linear probing | top-1 accuracy: 67.0 -> 66.8 | No clear effect |
| add MFF to MAE | MAE ViT-B 800 ep · COCO detection | AP box: 47.3 -> 48.0 | Improved |
Closest papers by meaning
- Towards Latent Masked Image Modeling for Self-Supervised Visual Representation Learning (ECCV 2024 · similarity 0.91)
- Masked Image Modeling With Local Multi-Scale Reconstruction (CVPR 2023 · similarity 0.90)
- Revealing the Dark Secrets of Masked Image Modeling (CVPR 2023 · similarity 0.90)
- Wavelet-Driven Masked Image Modeling: A Path to Efficient Visual Representation (AAAI 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 0.88)
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.
