This is a Pytorch implementation of

Janai, J., Gney, F., Ranjan, A., Black, M. and Geiger, A., Unsupervised Learning of Multi-Frame Optical Flow with Occlusions. ECCV 2018.

[[Link to Paper](http://www.cvlibs.net/publications/Janai2018ECCV.pdf)] [[Project Page](https://avg.is.tuebingen.mpg.de/research_projects/back2future)] [[Original Torch Code](https://github.com/jjanai/back2future)]

Requirements

• Runs and tested on Pytorch 0.3.1, it should be compatible with higher versions with little/no modifications.

• Correlation package is taken from NVIDIA/flownet2-pytorch and it can be installed using

cd correlation_package
bash make.sh

If you are using Pytorch>0.3.1, you can use correlation layer from here.

Usage

To use the model, go to your favorite python environment

from back2future import Modelmodel = Model(pretrained='pretrained/path_to_your_favorite_model')

There are two pretrained models in pretrained/, that are fine tuned on Sintel and KITTI in an unsupervised way.

Refer to demo.py for more.

Testing

To test performance on KITTI, use

python3 test_back2future.py --pretrained-flow path/to/pretrained/model --kitti-dir path/to/kitti/2015/root

Training

Please use the [original torch code] for training new models.

License

This is a reimplementation. License for the original work can be found at JJanai/back2future.

While using this code, please cite

@inproceedings{Janai2018ECCV,
  title = {Unsupervised Learning of Multi-Frame Optical Flow with Occlusions },
  author = {Janai, Joel and G{"u}ney, Fatma and Ranjan, Anurag and Black, Michael J. and Geiger, Andreas},
  booktitle = {European Conference on Computer Vision (ECCV)},
  volume = {Lecture Notes in Computer Science, vol 11220},
  pages = {713--731},
  publisher = {Springer, Cham},
  month = sep,
  year = {2018},
  month_numeric = {9}
}