Research map / Label-efficient and robust learning / Self-supervised and contrastive
Towards Latent Masked Image Modeling for Self-Supervised Visual Representation Learning
Yibing Wei, Abhinav Gupta, Pedro Morgado
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 |
|---|---|---|---|
| use momentum target encoder instead of joint optimization to prevent collapse (ablation) | Latent MIM · IN100 | NN accuracy: 0.9 -> 18.8 | Improved |
| replace MSE reconstruction with InfoNCE patch discrimination (ablation) | Latent MIM · IN100 | NN accuracy: 18.8 -> 30.0 | Improved |
| L1 loss instead of MSE for latent reconstruction (ablation) | Latent MIM · IN100 | NN accuracy: 18.8 -> 15.2 | Hurt |
| raise mask ratio from 0.75 to 0.9 (ablation) | Latent MIM · IN100 | NN accuracy: 30.0 -> 53.3 | Improved |
| add non-contiguous stochastic masking with gap 4 (ablation) | Latent MIM · IN100 | NN accuracy: 53.3 -> 54.6 | Improved |
| add patch similarity constraint regularizer (ablation) | Latent MIM · IN100 | NN accuracy: 54.6 -> 56.8 | Improved |
| switch from self-attention to cross-attention decoder with visual cues (ablation) | Latent MIM · IN100 | NN accuracy: 59.7 -> 63.3 | Improved |
| add MLP projector before decoder (ablation) | Latent MIM self-attention decoder · IN100 | NN accuracy: 56.8 -> 59.7 | Improved |
| increase decoder depth from 3 to 8 layers (ablation) | Latent MIM · IN100 | NN accuracy: 63.3 -> 35.9 | Hurt |
| latent MIM instead of pixel MIM (MAE) on linear probing | ViT-B/16 ImageNet-1k · ImageNet-1k | LP accuracy: 67.8 -> 72.0 | Improved |
| latent MIM instead of pixel MIM on IN100 fine-tuning (before decoder fixes) (ablation) | Latent MIM variants · IN100 | fine-tuning accuracy: 79.7 (momentum) vs 88.3 (MAE); 89.9 final | Mixed |
Closest papers by meaning
- Beyond [cls]: Exploring the True Potential of Masked Image Modeling Representations (ICCV 2025 · similarity 0.93)
- Masked Image Modeling: A Survey (IJCV 2025 · similarity 0.93)
- Revealing the Dark Secrets of Masked Image Modeling (CVPR 2023 · similarity 0.93)
- From Prototypes to General Distributions: An Efficient Curriculum for Masked Image Modeling (CVPR 2025 · similarity 0.91)
- Wavelet-Driven Masked Image Modeling: A Path to Efficient Visual Representation (AAAI 2025 · similarity 0.91)
- On the Role of Discrete Tokenization in Visual Representation Learning (ICLR 2024 · 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.
