FeatUp only comparable with 16x bilinear upsampling?

[andrehuang]

Hi,

I was testing FeatUp recently on semantic segmentation tasks on both COCOStuff (as in the paper) and Cityscapes. While I was able to reproduce the low-res baseline results and even got slightly higher performance for JBU than reported in the paper, I also found that simple 16x bilinear interpolation also achieves similar and sometimes even better results!

I want to ask the authors whether they also observed the similar results?