mtngld / lsim

Learn Stereo, Infer Mono: Siamese Networks for Self-Supervised, Monocular, Depth Estimation

Learn Stereo, Infer Mono: Siamese Networks for Self-Supervised, Monocular, Depth Estimation

lsim_model.py is based on Monodepth, with modifications to allows siamese and mirroring:

@inproceedings{monodepth17,
  title     = {Unsupervised Monocular Depth Estimation with Left-Right Consistency},
  author    = {Cl{\'{e}}ment Godard and
               Oisin {Mac Aodha} and
               Gabriel J. Brostow},
  booktitle = {CVPR},
  year = {2017}
}

Requirements

• Tensorflow 1.13
• Pandas
• Numpy

You will also need our version of monodepth (minor changes to support Python 3), make sure to clone using

git clone --recursive https://github.com/mtngld/lsim.git

Data

Use monodepth excellent downloader in order to get the data for kitti. For cityscapes see https://www.cityscapes-dataset.com/

Acknowledgement

We thank Godard et al. for their excellent paper and code.

@inproceedings{monodepth17,
  title     = {Unsupervised Monocular Depth Estimation with Left-Right Consistency},
  author    = {Cl{\'{e}}ment Godard and
               Oisin {Mac Aodha} and
               Gabriel J. Brostow},
  booktitle = {CVPR},
  year = {2017}
}

Reference

@inproceedings{goldman2019lsim,
  title={Learn Stereo, Infer Mono: Siamese Networks for Self-Supervised, Monocular, Depth Estimation},
  author={Goldman, Matan and Hassner, Tal and Avidan, Shai},
  booktitle={Computer Vision and Pattern Recognition Workshops (CVPRW)},
  year=2019
}