CVPR-2021 Masked Autoencoders Are Scalable Vision Learners
Self-supervised, generative pre-training with ViT in the spirit of BERT
- Mask a larger fraction of patches so that the visible patches are less redundant, which makes the pretext task harder and pushes the model toward more robust representations.
- The encoder processes only unmasked tokens, which speeds up training; a transformer decoder reconstructs the raw pixel values directly, keeping the overall pipeline simple.
- Techniques introduced in the ViT line of work further stabilize and strengthen training.
<HR align=left color=#987cb9 SIZE=1>