szc19990412 / TransMIL

TransMIL: Transformer based Correlated Multiple Instance Learning for Whole Slide Image Classification
363 stars 73 forks source link

Effectiveness of PPEG #37

Open zhwl2117 opened 1 year ago

zhwl2117 commented 1 year ago

Dear authors,

Thank you for your impressing work. I am curious about the motivation of the PPEG. Since the WSI patches within a bag should be unordered. Why do we still need the positional encoding?

In the paper, it is mentioned that adding zero-padding can provide more information for convolution. Could you please illustrate more on this?

Thanks!