Research map / Label-efficient and robust learning / Self-supervised and contrastive
ColorMAE: Exploring data-independent masking strategies in Masked AutoEncoders
Carlos Hinojosa, Shuming Liu, Bernard Ghanem
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 |
|---|---|---|---|
| replace random masking with band-pass (Green) filtered noise masks (ablation) | MAE ViT-B · ADE20K, 800 epochs | mIoU: 46.46 -> 49.18 | Improved |
| replace random masking with Green masks (ablation) | MAE ViT-B · ImageNet-1K Top-1, 800 epochs | Top-1 accuracy: 83.17 -> 83.57 | Improved |
| replace random masking with Green masks (ablation) | MAE ViT-L · ADE20K, 800 epochs | mIoU: 50.29 -> 51.46 | Improved |
| replace random masking with Blue (high-pass) masks (ablation) | MAE ViT-B · ADE20K, 800 epochs | mIoU: 46.46 -> 44.81 | Hurt |
| replace random masking with Red (low-pass) masks (ablation) | MAE ViT-B · ADE20K, 800 epochs | mIoU: 46.46 -> 44.78 | Hurt |
| replace random masking with Purple (band-stop) masks (ablation) | MAE ViT-B · ADE20K, 800 epochs | mIoU: 46.46 -> 45.96 | Hurt |
| Green masking at short pre-training (100 epochs) (ablation) | MAE ViT-B · ImageNet-1K Top-1, 100 epochs | Top-1 accuracy: 81.69 -> 81.82 | No clear effect |
| Green masking at 300 epochs (ablation) | MAE ViT-B · COCO detection, 300 epochs | AP bbox: 48.50 -> 48.70 | Improved |
| use more complex void-and-cluster blue noise generation (ablation) | ColorMAE · downstream tasks | accuracy | No clear effect |
| use white noise instead of uniform noise for random masks (ablation) | MAE · pre-training and fine-tuning | accuracy | No clear effect |
| Blue masking lowers reconstruction loss (ablation) | MAE · pretraining loss vs downstream | downstream performance | No clear effect |
| Green masking on COCO detection vs MAE random at 1600 epochs (224px comparison) | MAE · COCO AP bbox | AP bbox: 50.6 -> 50.1 | Hurt |
Closest papers by meaning
- Understanding and Enhancing Mask-Based Pretraining towards Universal Representations (NeurIPS 2025 · similarity 0.90)
- R-MAE: Regions Meet Masked Autoencoders (ICLR 2024 · similarity 0.90)
- Multi-Modal Contrastive Masked Autoencoders: A Two-Stage Progressive Pre-training Approach for RGBD Datasets (CVPR 2025 · similarity 0.89)
- Learning Mask Invariant Mutual Information for Masked Image Modeling (ICLR 2025 · similarity 0.89)
- Self-Guided Masked Autoencoder (NeurIPS 2024 · similarity 0.89)
- The Dynamic Duo of Collaborative Masking and Target for Advanced Masked Autoencoder Learning (AAAI 2025 · similarity 0.89)
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.
