This repository has the code needed to run the TRITON algorithm. This page is meant for reviewers, and so it is anonymous.
To train it, run the following commands:
cd translator
python3 train.py --config configs/alphabet_five_base.yaml
You will also need to install some python packages, including:
pip3 install torch rp einops numpy opencv-python OpenEXR
The code under 'translator' was originally based on https://gitlab.com/nct_tso_public/surgical-video-sim2real, from the following paper:
"Long-Term Temporally Consistent Unpaired Video Translation from Simulated Surgical 3D Data".
Dominik Rivoir, Micha Pfeiffer, Reuben Docea, Fiona Kolbinger, Carina Riediger, Jürgen Weitz, Stefanie Speidel.
International Conference on Computer Vision 2021.
Note that that repository was based on MUNIT from https://github.com/NVlabs/MUNIT:
@inproceedings{huang2018munit,
title={Multimodal Unsupervised Image-to-image Translation},
author={Huang, Xun and Liu, Ming-Yu and Belongie, Serge and Kautz, Jan},
booktitle={ECCV},
year={2018}
}
The code used in translator/pytorch_msssim.py is based on https://github.com/jorge-pessoa/pytorch-msssim
The code used in source/learnable_textures.py is based on https://github.com/ndahlquist/pytorch-fourier-feature-networks:
@article{DBLP:journals/corr/abs-2006-10739,
author = {Matthew Tancik and
Pratul P. Srinivasan and
Ben Mildenhall and
Sara Fridovich{-}Keil and
Nithin Raghavan and
Utkarsh Singhal and
Ravi Ramamoorthi and
Jonathan T. Barron and
Ren Ng},
title = {Fourier Features Let Networks Learn High Frequency Functions in Low
Dimensional Domains},
journal = {CoRR},
volume = {abs/2006.10739},
year = {2020},
url = {https://arxiv.org/abs/2006.10739},
eprinttype = {arXiv},
eprint = {2006.10739},
timestamp = {Thu, 14 Oct 2021 09:16:16 +0200},
biburl = {https://dblp.org/rec/journals/corr/abs-2006-10739.bib},
bibsource = {dblp computer science bibliography, https://dblp.org}
}