This is the official code for ICRA 2021 Paper: Multimodal Scale Consistency and Awareness for Monocular Self-Supervised Depth Estimation by Hemang Chawla, Arnav Varma, Elahe Arani and Bahram Zonooz.
G2S (GPS-to-Scale) Loss is a dynamically-weighted loss that can be added to the appearance-based losses to train any monocular self-supervised depth estimation architecture to get scale-consistant and scale-aware depth estimates at inference.
Here, we provide helper GPS dataloader and the G2S loss classes for using this loss with any model.
For details, please see the Paper and Presentation.
The GPS files containing geodesic gps information of raw kitti dataset in local coordinates for training with the g2s loss can be found in the assets
folder as kitti_gps_raw.zip
.
Unzip the file at /path/to/KITTI/raw_data/sync
to merge the GPS files in the expected directory tree structure.
You can use the G2S
class in lossG2S.py
within your project for scale-consistent and -aware predictions.
This requires using the copresent GPS modality along with images.
To load the GPS, please adopt the GPSDataloader
class within dataloaderGPS.py
into your images dataloader.
If you find the code useful in your research, please consider citing our paper:
@inproceedings{chawlavarma2021multimodal, author={H. {Chawla} and A. {Varma} and E. {Arani} and B. {Zonooz}}, booktitle={2021 IEEE International Conference on Robotics and Automation (ICRA)}, title={Multimodal Scale Consistency and Awareness for Monocular Self-Supervised Depth Estimation}, location={Xi’an, China}, publisher={IEEE (in press)}, year={2021} }
This project is licensed under the terms of the MIT license.