Just want to confirm this implementation of PSPNet does not allow for the use of arbitrary/differing sized images during training and testing/inference? My understanding of the architecture is that it should be possible to do that. Is that only because of dilated convolutions in the original?
Just want to confirm this implementation of PSPNet does not allow for the use of arbitrary/differing sized images during training and testing/inference? My understanding of the architecture is that it should be possible to do that. Is that only because of dilated convolutions in the original?