Research map / Label-efficient and robust learning / Self-supervised and contrastive
Masked Frequency Modeling for Self-Supervised Visual Pre-Training
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 random low-pass or high-pass frequency masking rather than no mask (ablation) | MFM · ImageNet-1K | top-1 accuracy: 76.5 -> 83.1 | Helped |
| Sample both low-pass and high-pass filters rather than use only one (ablation) | MFM · ImageNet-1K | top-1 accuracy: 82.4 -> 83.1; 82.3 -> 83.1 | Helped |
| Use a fixed mask radius of 16 rather than the tested alternatives (ablation) | MFM · ImageNet-1K | top-1 accuracy: 82.8; 82.7; 82.6; 83.0 -> 83.1 | Mixed |
| Use a circle mask rather than square or rhombus masks (ablation) | MFM · ImageNet-1K | top-1 accuracy: 82.9 -> 83.1; 82.8 -> 83.1 | Helped |
| Sample low-pass and high-pass filters with equal probability (ablation) | MFM · ImageNet-1K | top-1 accuracy: 82.5; 82.7 -> 83.1 | Helped |
| Predict only masked frequencies rather than the full spectrum (ablation) | MFM · ImageNet-1K | top-1 accuracy: 82.4 -> 83.1 | Helped |
| Use frequency loss with gamma 1 rather than frequency gamma 2 or spatial losses (ablation) | MFM · ImageNet-1K | top-1 accuracy: 82.5; 82.3; 82.2 -> 83.1 | Helped |
| Use a linear decoder rather than a Transformer decoder (ablation) | MFM · ImageNet-1K | top-1 accuracy: 83.0; 83.1; 83.1; 83.1 vs. 83.1 | No clear effect |
| Use frequency masking rather than spatial patch masking (ablation) | MFM · ImageNet-1K | top-1 accuracy: 82.8 -> 83.1; 77.7 -> 78.5 | Helped |
| Combine SR with denoising rather than use either corruption individually (ablation) | ViT-B/16 pre-training · ImageNet-1K | top-1 accuracy: 82.4; 82.7 -> 82.2 | Hurt |
| Combine deblurring with denoising rather than use either corruption individually (ablation) | ViT-B/16 pre-training · ImageNet-1K | top-1 accuracy: 81.7; 82.7 -> 82.5 | Hurt |
| Use MFM rather than MAE | ViT-B/16 · ADE20K | mIoU: 48.1 -> 48.6 | Helped |
Closest papers by meaning
- Frequency-Guided Masking for Enhanced Vision Self-Supervised Learning (ICLR 2025 · similarity 0.91)
- From Prototypes to General Distributions: An Efficient Curriculum for Masked Image Modeling (CVPR 2025 · similarity 0.89)
- Wavelet-Driven Masked Image Modeling: A Path to Efficient Visual Representation (AAAI 2025 · similarity 0.88)
- Multiscale Low-Frequency Memory Network for Improved Feature Extraction in Convolutional Neural Networks (AAAI 2024 · similarity 0.87)
- Asymmetric Masked Distillation for Pre-Training Small Foundation Models (CVPR 2024 · similarity 0.87)
- Understanding and Enhancing Mask-Based Pretraining towards Universal Representations (NeurIPS 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.