This folder contains our implementation of the perceptual loss presented in [Thomas2022].
Import the Features
class from loss.py
to use the loss, after instantiating you can call spatial_loss
and temporal_loss
. See model.py
for an example of how NPPD uses these.
-
Download color and semantic maps from ml-hypersim using Thomas Germer's script
-
Run
process_hypersim.py
to transform the training data to Zarr format -
Now you can train the perceptual loss running
python loss.py
You can skip the first two steps by downloading our Zarr file.
Note that loss.py
doesn't import the other Python files for inference; you only need those for training.