This is the official code for VISAPP 2022 paper Transformers in Self-Supervised Monocular Depth Estimation with Unknown Camera Intrinsics and its extended paper in Springer CCIS, Transformers in Unsupervised Structure-from-Motion.

Authors: Hemang Chawla, Arnav Varma, Elahe Arani and Bahram Zonooz.

We propose MT-SfMLearner v1 and v2(pdf) that show how transformers are more competitive and robust for monocular depth estimation.

Hardware details for original training of MT-SfMLearner (v2) can be found in respective papers.

git clone https://github.com/NeurAI-Lab/MT-SfMLearner.git
cd MT-SfMLearner
make docker-build

MT-SfMLearner (v2) is trained in a self-supervised manner from videos. For training, utilize a .yaml config file or a .ckpt model checkpoint file with scripts/train.py.

python scripts/train.py <config_file.yaml or model_checkpoint.ckpt>

Example config file to train MIMDepth can be found in configs folder.

A trained model can be evaluated by providing a .ckpt model checkpoint.

python scripts/eval.py --checkpoint <model_checkpoint.ckpt>

For running inference on a single image or folder,

python scripts/infer.py --checkpoint <checkpoint.ckpt> --input <image or folder> --output <image or folder> [--image_shape <input shape (h,w)>]

Pretrained Models for MT-SfMLearner and MIMDepth can be found Coming soon!

If you find the code useful in your research, please consider citing our papers:

@inproceedings{chawlavarma2022MTSfMLearnerv2,
  title={Transformers in Unsupervised Structure-from-Motion},
  author={Chawla, Hemang and Varma, Arnav and Arani, Elahe and Zonooz, Bahram},
  booktitle={International Joint Conference on Computer Vision, Imaging and Computer Graphics, Revised Selected Papers},
  pages={281--303},
  year={2022},
  doi={10.1007/978-3-031-45725-8_14},
  organization={Springer Nature}
}

@inproceedings{varmachawla2022MTSfMLearner,
	author={A. {Varma} and H. {Chawla} and E. {Arani} and B. {Zonooz}},
    booktitle={Proceedings of the 17th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications (VISIGRAPP 2022) - Volume 4: VISAPP},
    year={2022},
    pages={758-769},
    publisher={SciTePress},
    doi={10.5220/0010884000003124},
}

This project is licensed under the terms of the MIT license.