DevNet: Self-supervised Monocular Depth Learning via Density Volume Construction - ECCV 2022.

Introduction

This is the PyTorch implementation of DevNet: Self-supervised Monocular Depth Learning via Density Volume Construction, a simple and efficient neural architecture for Self-supervised Monocular Depth Estimation.

Setup

DevNet provides support for multiple versions of Python and Torch, such as:

python==3.8 
pytorch==1.12.0
torchvision==0.13.0

Data

[KITTI]: To download this dataset, you can follow instruction of MonoDepth2

[KITTI Odometry]: To download this dataset, you can follow instruction of MonoDepth2

[NYU-V2]: To download this dataset, you can follow instruction of MonoDepth2

Running the code

Inference

You can access the trained models for partial experiments described in the paper by visiting this link. To store them, simply download the models and save them in the logs directory. Additionally, the inference script is named test_dev.py.

• DevNet on kitti with resnet18 backbone and 192 x 640 resolution:

sh start2test.sh

Training

Necessary training details could also be found in test_dev.py.

Citation

If you find this code useful for your research, please cite our paper

@inproceedings{zhou2022devnet,
  title={Devnet: Self-supervised monocular depth learning via density volume construction},
  author={Zhou, Kaichen and Hong, Lanqing and Chen, Changhao and Xu, Hang and Ye, Chaoqiang and Hu, Qingyong and Li, Zhenguo},
  booktitle={European Conference on Computer Vision},
  pages={125--142},
  year={2022},
  organization={Springer}
}

Acknowledgements

Our code partially builds on Monodepth2.