Why MIC work on DeepLab with ResNet backbone?

[daifeng2016]

Hi, this work is amzing. In MIC, image after masking is directly fed into student-teacher network. However, the masked areas will influence the convolution process and leads to much noise. I am wondering if MIC only work on transformer backbone (like MAE) and why it works on normal resnet backbone as presented in experimental part (DeepLab+ResNet101).

[Wenlve-Zhou]

Hi, it may not be entirely appropriate to regard MIC purely as an MIM technique. You might try understanding it from a perspective similar to FixMatch, where Masked Image can be seen as a form of strong augmentation, providing a semi-supervised regularization through a weak-to-strong approach.