Research map / Detection and segmentation / CNN features
Efficient Masked Autoencoders With Self-Consistency
Zhaowen Li, Yousong Zhu, Zhiyang Chen, Wei Li, Rui Zhao, Chaoyang Zhao, Ming Tang, Jinqiao Wang
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 |
|---|---|---|---|
| sample 4 random masks per iteration instead of 1 (ablation) | MAE ViT-B 200 ep · ImageNet-1k linear probing | top-1 accuracy: 58.8% -> 62.5% | Improved |
| use parallel disjoint masks instead of pure random masks (ablation) | MAE ViT-B 200 ep · ImageNet-1k linear probing | top-1 accuracy: 62.5% -> 63.4% | Improved |
| use complementary 25%/75% masks (ablation) | MAE ViT-B 200 ep · ImageNet-1k linear probing | top-1 accuracy: 58.8% -> 46.7% | Hurt |
| replace pixel reconstruction with pure self-consistency (ablation) | EMAE parallel mask · ImageNet-1k linear probing | top-1 accuracy: 63.4% -> 64.9% | Improved |
| add pixel reconstruction to self-consistency (ablation) | EMAE · ImageNet-1k linear probing | top-1 accuracy: 64.9% -> 65.0% | Improved |
| add parallel mask strategy (ablation) | MAE ViT-B 800 ep · ImageNet-1k linear probing | top-1 accuracy: 64.4% -> 68.4% | Improved |
| add self-consistency to parallel mask (ablation) | EMAE ViT-B 300 ep · ImageNet-1k linear probing | top-1 accuracy: 65.3% -> 68.4% | Improved |
| use K=14 (13/14 mask ratio) instead of K=4 (ablation) | EMAE ViT-B 800 ep · ImageNet-1k linear probing | top-1 accuracy: 70.4% -> 66.7% | Hurt |
| use EMAE K=7 versus MAE 6/7 mask ratio (ablation) | EMAE ViT-B 800 ep · ImageNet-1k linear probing | top-1 accuracy: 63.9% -> 68.1% | Improved |
| train EMAE 1600 epochs instead of 800 (ablation) | EMAE ViT-B · ImageNet-1k | linear probing top-1: 70.4% -> 71.2% | Improved |
| use EMAE instead of MAE as masked augmentor at 75% mask | ResNet50 · ImageNet-1k | top-1 accuracy: 75.9% -> 77.2% | Improved |
| replace MAE (1600 ep) with EMAE (800 ep) | ViT-B Mask R-CNN · COCO val2017 | box AP: 50.4% -> 51.4% | Improved |
Closest papers by meaning
- Context Autoencoder for Self-supervised Representation Learning (IJCV 2023 · similarity 0.91)
- Stare at What You See: Masked Image Modeling Without Reconstruction (CVPR 2023 · similarity 0.89)
- Masked Autoencoders Are Scalable Vision Learners (CVPR 2022 · similarity 0.89)
- Understanding and Enhancing Mask-Based Pretraining towards Universal Representations (NeurIPS 2025 · similarity 0.89)
- Self-Guided Masked Autoencoder (NeurIPS 2024 · similarity 0.89)
- Learning Mask Invariant Mutual Information for Masked Image Modeling (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.
