Understanding Masked Image Modeling via Learning Occlusion Invariant Feature

Augmentation strategies. Generally in MAE, the source image which is send into encoder and the target image which is the target to reconstruct are always the same. Here we try to add additional augmentations on the source image. After the random-resize-cropped and random horizontal-flip augmentation, the image is cropped to 224×224. Then we try the compositions of three additional augmentation strategies on the source image. The definitions of augmentation strategies are described as follow: