Masked Image Modeling

Masked image modeling (MIM) is a self-supervised approach that hides patches or regions and trains a model to recover missing pixels, features, or semantic targets, reducing reliance on labeled images. Methods range from masked autoencoders and latent prediction to teacher-guided distillation, with random, saliency-guided, structure-aware, and domain-specific masking. Pretraining is used for transfer to classification, segmentation, detection, and scientific or medical imaging, often tested on benchmarks such as ImageNet, Cityscapes, and domain-specific datasets with limited labels. A central open tension is that masks and reconstruction objectives that improve pretraining loss do not always improve downstream performance, so research continues to test which information should be hidden and predicted.