Padding instead of deconvolution in FCN

I read about deconvolution/upsampling and the only reason to use it (according to the info in the web) is to enlarge a picture.

Talking about Fully Convolutional Networks isn’t it better to just prevent "downsampling" by using zero-padding?

I understand that the results would differ but I haven’t found other reasons to choose deconvolution.
The only disadvantage of zero-padding I see is that more pixels is stored in hidden layers (though we don’t need a deconvolutional layer in this case). Plus we don’t loose the information from the borders.

So is it a good practice to use zero-padding in FCN instead of upsampling?