Awesome

Removert

What is removert?

• Static map construction in the wild.
• A dynamic points removing tool by constructing a static map
• The name is from the abbreviation of our title "Remove-then-revert" (IROS 2020): paper, video

What can we do using removert?

• We can easily construct and save a static map.
• We can easily parse dynamic points

Example

• Video 1: KITTI 09 with SuMa poses
• Video 2: MulRan dataset with poses from SC-LIO-SAM

Preparations

• Step 1: Get a set of LiDAR scans and corresponding poses by running any open source LiDAR odometry or SLAM algorithm (e.g., pose-and-scan saver of SC-LIO-SAM or pose-and-scan saver of SC-A-LOAM)
• Step 2: Make a pair of a scan's point cloud and a corresponding pose using associated timestamps. We note that you need to save a scan as a binary format as KITTI and the pose file as a single text file where SE(3) poses are written line-by-line (12 numbers for a single line), which is also the equivalent format as KITTI odometry's ground truth pose txt file.

Requirements

• Based on C++17
• ROS (and Eigen, PCL, OpenMP): the all examples in this readme are tested under Ubuntu 18.04 and ROS Melodic.
• FYI: We uses ROS's parameter parser for the convenience, despite no topic flows within our system (our repository currently runs at offline on the pre-prepared scans saved on a HDD or a SSD). But the speed is fast (over 10Hz for a single removing) and plan to extend to real-time slam integration in future.

How to use

• First, compile the source

$ mkdir -p ~/catkin/removert_ws/src
$ cd ~/catkin/removert_ws/src
$ git clone https://github.com/irapkaist/removert.git
$ cd ..
$ catkin_make
$ source devel/setup.bash

• Before to start the launch file, you need to replace data paths in the config/params.yaml file. More details about it, you can refer the above tutorial video (KITTI 09)

• Then, you can start the Removert

$ roslaunch removert run_kitti.launch # if you use KITTI dataset 
 or
$ roslaunch removert run_scliosam.launch # see this tutorial: https://youtu.be/UiYYrPMcIRU

• (Optional) we supports Matlab tools to visulaize comparasions of original/cleaned maps (see tools/matlab).

Further Improvements

• We propose combining recent deep learning-based dynamic removal (e.g., LiDAR-MOS) and our method for better map cleaning
  • Deep learning-based removal could run online and good for proactive removal of bunch of points.
  • Removert currently runs offline but good at finer cleaning for the remained 3D points after LiDAR-MOS ran.
• A tutorial video and an example result for the KITTI 01 sequence:

Contact

paulgkim@kaist.ac.kr

Cite Removert

@INPROCEEDINGS { gskim-2020-iros,
    AUTHOR = { Giseop Kim and Ayoung Kim },
    TITLE = { Remove, then Revert: Static Point cloud Map Construction using Multiresolution Range Images },
    BOOKTITLE = { Proceedings of the IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) },
    YEAR = { 2020 },
    MONTH = { Oct. },
    ADDRESS = { Las Vegas },
    NOTE = { Accepted. To appear. },
}

License

<a rel="license" href="http://creativecommons.org/licenses/by-nc-sa/4.0/"><img alt="Creative Commons License" style="border-width:0" src="https://i.creativecommons.org/l/by-nc-sa/4.0/88x31.png" /></a><br />This work is supported by Naver Labs Corporation and by the National Research Foundation of Korea (NRF). This work is also licensed under a <a rel="license" href="http://creativecommons.org/licenses/by-nc-sa/4.0/">Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License</a>.

TODO (in order)

Near future

• Full sequence cleaned-scan saver by automatically iterating batches (because using 50-100 scans for a single batch is recommended for computation speed)
• Adding revert steps (I think certainly removing dynamic points is generally more worthy for many applications, so reverting step is omitted currently)
• Automatically parse dynamic segments from the dynamic points in a scan (e.g., using DBSCAN on dynamic points in a scan)
• Exmaples from MulRan dataset (for showing removert's availability for various LiDAR configurations) — see this tutorial
• (scan, pose) pair saver using SC-LeGO-LOAM or SC-LIO-SAM, which includes a loop closing that can make a globally consistent map. — see this tutorial
• Examples from the arbitrary datasets using the above input data pair saver.
• Providing a SemanticKITTI (as a truth) evaluation tool (i.e., calculating the number of points of TP, FP, TN, and FN)
• (Not certain now) Changing all floats to double

Future

• Real-time LiDAR SLAM integration for better odometry robust to dynamic objects in urban sites (e.g., with LIO-SAM in the Riverside sequences of MulRan dataset)
• Multi-session (i.e., inter-session) change detection example
• Defining and measuring the quality of a static map
• Using the above measure, deciding when removing can be stopped with which resolution (generally 1-3 removings are empirically enough but for highly crowded environments such as urban roads)