How about this method on classification task?

[gi2wzh]

Hi, thanks for sharing this great work.
I noticed that the current experiments are mainly on segmantation tasks. I am wondering whether this method can be also bring improvements to classification task?
Have you conducted any related experiments or have any insights on this? Thanks!

[ChengShiest]

Hi, thanks for your interest!

Yes, we did observe some improvement on classification during the early training stage, usually in the form of faster accuracy gains / better intermediate checkpoints. However, after full convergence, the final classification accuracy is very close to the standard ViT baseline under the same training recipe.

Our intuition is that image-level classification is a relatively forgiving objective: even if the model relies partly on background or shortcut patches, it can still predict the correct class once sufficiently trained.

[gi2wzh]

Got it. Thank you for clarifying this.