Spaces:
Running
on
T4
Running
on
T4
Very cool!
#2
by
torridgristle
- opened
This is very cool. I'm very interested in seeing what's done with this. Do you have any examples of projects using this?
It's also unexpected that ResNet50 has such blobby / blurry outputs, I would've expected more detailed features from the convolution model. Does a highpass filter expose more sharper details? Would training the upscaler on highpassed features make it react better to finer details? Or is this just the reality of ResNet50?
Thanks for the positive feedback! ResNet50 really agressively pools information so the blobbyness is definitely a result of that. See the main project page: https://aka.ms/featup for a link to the larger codebase and paper with additional experiments