search

sniklaus / 3d-ken-burns

an implementation of 3D Ken Burns Effect from a Single Image using PyTorch
Other
1.51k stars 224 forks source link

Ground truth resolution #46

Closed dfrumkin closed 4 years ago

dfrumkin commented 4 years ago

Hello Simon! How do you reach the resolution of 1024 for the largest dimension if the resolution of the synthetic dataset is only 512x512?

sniklaus commented 4 years ago

The refinement network was trained to upsample a 128 input to a 512 output. At test time, it is applied to a 256 input such that it gives a 1024 output. This seems to work reasonably well despite the different resolutions.

dfrumkin commented 4 years ago

Thank you very much for your response! You train the refinement network not only independently of the coarse network (using elastic deformations instead of the typical input you would get), but also on a different resolution. Very interesting.