Skip to content

A non-learning geometry-based LiDAR depth completion model with components including outlier removal, normal calculation, distance transform, etc. IEEE Robotics and Automation Letters

Notifications You must be signed in to change notification settings

placeforyiming/RAL_Non-Learning_DepthCompletion

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

62 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

A Surface Geometry Model for LiDAR Depth Completion (IEEE RA-L)

Abstract:

We develop a non-learning depth completion pipeline with the assumption that the empty pixel shares the same surface with the nearest value. The pipeline has two major technical modules: an outlier removal and a surface model to explicitly calculate the depth with a mathematical equation. Please check our paper for more details:

The arxiv version: https://arxiv.org/pdf/2104.08466.pdf

The IEEE RAL version: https://ieeexplore.ieee.org/abstract/document/9387169?casa_token=m6hbd7riBioAAAAA:bsyIXoghUPIyvK_oASnZAATLqq1h1yt2gAEQbMPLSvdOl0zRsFadmPPuYTPnIEDwlrdW8bM

Method Pipeline:

pipeline

Outlier Removal Examples:

outlier_removal

Data Structure

Download the validation and test dataset from KITTI depth completion benchmark website, then put them under the fodler depth_selection with the stucture:

----- depth_selection
     |
     -------val_selection_cropped
     |
     -------test_depth_completion_anonymous

Environment Settings

Note: we have two different outlier removal methods. The function outlier_removal() in tools.py is a simple opencv filter-based algorithm. It can work on 64 line LiDAR but need mannually adjust two hyperparameters, thus can not directly extend on sparser 16 line or 32 line LiDAR. To design a more general outlier removal algoritem, we think from first principal rule and have an alternative one in non_learning_completion_multi_res.py. The function outlier_removal_mask() is parameter-free, that we believe more general. However, we only have one month to prepare the reply of our RA-L submission, so the current implementation utilize the extract_patch function in tensorflow. This is why the tensorflow 2.0 is temporary needed. we will implement a c++ version just like the convd operator in OpenCV in the short future.

Pull the docker image:

docker pull tensorflow/tensorflow:2.1.0-gpu-py3-jupyter

Activate the docker container with mounting the code in this repository.

Install packages with the certain version in the docker container:

bash install_dependency.sh

To subsample the 32 line and 16 line LiDAR, please run:

python subsample_Lidar_val.py

After installing all packages, you can open the jupyter note book by bash jupyter.sh, and use depth_completion.ipynb to explore each step of our method.

Run the Code

Run the completion with 64 line LiDAR on KITTI validation (simple outlier removal):

python non_learning_completion_CPU_64line.py

Run the completion with 64 line LiDAR on KITTI test (simple outlier removal):

python non_learning_completion_CPU_64line_test.py

Run the completion with 64 line LiDAR on KITTI validation (general outlier removal):

python non_learning_completion_multi_res.py

Run the completion with 64 line LiDAR on KITTI test (general outlier removal):

python non_learning_completion_multi_res_test.py

Publication

Please cite the paper if you use this code:

@article{zhao2021surface,
  title={A Surface Geometry Model for LiDAR Depth Completion},
  author={Zhao, Yiming and Bai, Lin and Zhang, Ziming and Huang, Xinming},
  journal={IEEE Robotics and Automation Letters},
  volume={6},
  number={3},
  pages={4457--4464},
  year={2021},
  publisher={IEEE}
}

About

A non-learning geometry-based LiDAR depth completion model with components including outlier removal, normal calculation, distance transform, etc. IEEE Robotics and Automation Letters

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published