Research map / Label-efficient and robust learning / Self-supervised and contrastive
Contrastive Masked Autoencoders are Stronger Vision Learners
Zhicheng Huang, Xiaojie Jin, Cheng-Ze Lu, Qibin Hou, Ming‐Ming Cheng, Dongmei Fu, Xiaohui Shen, Jiashi Feng
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 |
|---|---|---|---|
| add vanilla contrastive learning (random crop) to MAE (ablation) | MAE · ImageNet-1k (300 ep, ablation) | top-1 accuracy: 82.9 -> 83.1 | Improved |
| add pixel shifting augmentation to contrastive MAE (ablation) | CMAE · ImageNet-1k (300 ep, ablation) | top-1 accuracy: 83.1 -> 83.6 | Improved |
| add feature decoder (ablation) | CMAE · ImageNet-1k (300 ep, ablation) | top-1 accuracy: 83.6 -> 83.8 | Improved |
| use pixel shifting instead of random resized crop (ablation) | CMAE · ImageNet-1k (300 ep, ablation) | top-1 accuracy: 83.0 -> 83.4 | Improved |
| add color jittering on the momentum branch (ablation) | CMAE · ImageNet-1k (300 ep, ablation) | top-1 accuracy: 83.4 -> 83.8 | Improved |
| use InfoNCE instead of BYOL-style contrastive loss (ablation) | CMAE · ImageNet-1k (300 ep, ablation) | top-1 accuracy: 83.4 -> 83.8 | Improved |
| add a two-block feature decoder instead of none (ablation) | CMAE · ImageNet-1k (300 ep, ablation) | top-1 accuracy: 83.6 -> 83.8 | Improved |
| share weights between feature decoder and pixel decoder (ablation) | CMAE · ImageNet-1k (300 ep, ablation) | top-1 accuracy: 83.8 -> 83.4 | Hurt |
| increase feature decoder depth from 2 to 4 (ablation) | CMAE · ImageNet-1k (300 ep, ablation) | top-1 accuracy: 83.8 -> 83.8 | No clear effect |
| mask 75% of patches in the momentum branch (ablation) | CMAE · ImageNet-1k (300 ep, ablation) | top-1 accuracy: 83.8 -> 83.0 | Hurt |
| raise contrastive loss weight from 1.0 to 2.0 (ablation) | CMAE · ImageNet-1k (300 ep, ablation) | top-1 accuracy: 83.8 -> 83.2 | Hurt |
| add CMAE contrastive objective to MAE on a conv hybrid ViT | ConvMAE · ImageNet-1k fine-tuning (hybrid conv ViT-B) | top-1 accuracy: 84.6 -> 85.3 | Improved |
Closest papers by meaning
- Contrastive Tuning: A Little Help to Make Masked Autoencoders Forget (AAAI 2024 · similarity 0.90)
- Hybrid Distillation: Connecting Masked Autoencoders with Contrastive Learners (ICLR 2024 · similarity 0.89)
- Masked Autoencoders Are Scalable Vision Learners (CVPR 2022 · similarity 0.89)
- Masked Image Modeling with Denoising Contrast (ICLR 2023 · similarity 0.88)
- Learning Mask Invariant Mutual Information for Masked Image Modeling (ICLR 2025 · similarity 0.87)
- The Dynamic Duo of Collaborative Masking and Target for Advanced Masked Autoencoder Learning (AAAI 2025 · 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.
