mhariat / zbuffdepth

Unsupervised depth learning with z-buffering.

ZbuffDepth

An self-supervised monocular depth learning method utilizing image reconstruction loss, with the point occlusion issue solved by the novel z-buffer.

This repo trains and tests the model on stereo pairs using the Eigen splits.

Author(s)

Ziwen Chen (github: arthurhero)

Zixuan Guo (github: Olament)

Usage

Preparing

1. Create a blank folder.

mkdir raw_kitti && cd raw_kitti

2. Download the raw kitti dataset.

wget https://raw.githubusercontent.com/mrharicot/monodepth/master/utils/kitti_archives_to_download.txt 
wget -i kitti_archives_to_download.txt

3. Unzip the data.

unzip "*.zip"

4. Download KITTI's annotated depth map data set (14G) at KITTI's website
5. Merge the annotated depth data with the raw kitti

unzip data_depth_annotated.zip
mv -R data_depth_annotated/* raw_kitti

Training

1. Create folder mkdir logs and mkdir data
2. For training on stereo pairs, edit the parameters in train_stereo.sh, and run ./train_stereo.sh

Evaluating On Eigen

1. Open evaluate_kitti.py and scroll to bottom.
2. Modify first argument to KITTI data path and second argument to checkpoint path from training
3. Run ./evaluate_kitti.py

Files

train_stereo.sh : the training wrapper script for stereo, with cmd args

run_stereo.py : the training code

loss.py : code for all the losses used in the method, e.g., ssim loss and reconstruction loss

ops.py : code for all the operators used in the method, e.g., depth-to-point projection, z-buffering, etc.

data_process.py : code for all the data preprocessing and viewing methods, e.g., calculating egomotion matrix from gps coord

dataloader.py : code for the PyTorch dataloaders

utils.py : some basic utilities for PyTorch

eigen_*_files.txt : the standard eigen splits

generate_valid.py : reproducibly generate a validation split by randomly sampling from eigen validation split

evaluate_kitti.py : main evaluation code on eigen test split, adapted from Monodepth

evaluation_utils.py : utils for the evaluation, adapted from Monodepth

bts_orig.py : the encoder-decoder architecture, adapted from BTS (Lee et. al.)

Citation

If you find this code useful, please cite our paper:

Chen, Z., Guo, Z., and Weinman, J. Improved Point Transformation Methods for Self-Supervised Depth Prediction. In Proceedings of the 18th Conference on Robots and Vision (CRV). 2021.

@inproceedings{chen21improved,
    author = {Chen Ziwen and Zixuan Guo and Jerod Weinman},
    title = {Improved Point Transformation Methods for Self-Supervised Depth Prediction},
    booktitle = {2021 18th Conference on Computer and Robot Vision ({CRV})},
    year = {2021}
}

Acknowledgments

Thank Professor Jerod Weinman for his great support (including server provision and paper revision).