Match constraint about fast-vqa-and-fastervqa HOT 2 CLOSED

Hi Dr Han:

Thank you for your interest.

I think it can be understood in an inverted way: the Patch Merging of Swin comes first, and then we design the fragments to make sure the merging only happens inside our mini-patches, rather than across them. And in general, the Patch Merging can be seen as a pooling layer with stride and kernel both at (2,2), so we call it a pooling layer (a special one, though, that can be matched) in our paper.

To perform ablation study, we simple modify the hyper-parameters during the sampling to make them not aligned.

Hope this can help.

Best,
Haoning

from fast-vqa-and-fastervqa.

Hi @teowu ,

Thanks for solving my puzzle on the match constraint.

Best,
Hanwei

from fast-vqa-and-fastervqa.