G2S

This is a reference implementation for using G2S loss described in the ICRA 2021 paper

Multimodal Scale Consistency and Awareness for Monocular Self-Supervised Depth Estimation

by Hemang Chawla, Arnav Varma, Elahe Arani and Bahram Zonooz.

in the Monodepth2 repository for KITTI Eigen Zhou split. The corresponding checkpoint can be found here.

The official code is available here.

This code is for non-commercial use following the original license from Monodepth2; please see the license file for terms.

If you find our work useful in your research please consider citing our paper:

@inproceedings{chawla2021multimodal,
  title={Multimodal scale consistency and awareness for monocular self-supervised depth estimation},
  author={Chawla, Hemang and Varma, Arnav and Arani, Elahe and Zonooz, Bahram},
  booktitle={2021 IEEE International Conference on Robotics and Automation (ICRA)},
  pages={5140--5146},
  year={2021},
  organization={IEEE}
}

⏳ Training

Monocular training:

python train.py --model_name g2s --data_path /path/to/KITTI/raw_data/sync --log_dir /path/to/log/dir/ --g2s --png (if images are in png)

Ground truth generation (Needs to be run once before first evaluation):

python export_gt_depth.py --data_path /path/to/KITTI/raw_data/sync --split eigen

Monocular evaluation:

python train.py evaluate_depth.py --eval_mono --data_path /path/to/KITTI/raw_data/sync --eval_split eigen --load_weights_folder /path/to/ckpt/folder --png (if images are in png)

👩‍⚖️ License

Please see the license file for terms.