Instructions on how to create a LIDAR dataset with semantic labels in "semantic KITTI" Format

create_semantic_kitti_dataset.md

Prerequisites

Tested on Ubuntu 22.04 with an Nvidia RTX 3090

1. Install ROS2 Humble: LINK

2. Set up a ROS2 workspace: LINK
   2a. prepare automatic sourcing of the ROS2 installation and workspace:

   echo "source /opt/ros/humble/setup.bash" >> ~/.bashrc
   echo "source /home/user/ros2_ws/install/setup.bash" >> ~/.bashrc  
   

   where the second line must be adapted to your ros2 workspace. Afterwards, source the bashrc file (source ~/.bashrc) or simply close your terminal and open an new one (thus bashrc is loaded).

3. Clone the pointcloud merge node: LINK

4. Set up a Second workspace for Point Labeler and SuMa -> See Instructions at: LINK

5. Install Point Labeler: LINK

6. Install Surfel Based Mapping (SuMa): LINK 5a. One dependency of SuMa is GTSAM. On Ubuntu 22.04 there is currently no pre-built package for GTSAM. After installing GTSAM's dependencies, build GTSAM with:

   cd ~
   git clone https://github.com/borglab/gtsam.git
   cd gtsam
   mkdir build
   cd build
   cmake ..
   sudo make install
   

7. Install pcd2bin: LINK

Usage

1. Create a folder structure which meets the KITTI standard LINK:

   mkdir -p ~/dataset/pointclouds
   cd ~/dataset
   mkdir rosbags
   mkdir -p sequences/00/velodyne
   cd sequences/00
   mkdir labels
   

2. Record ROSBAG of your LiDAR Data. In our case, this data is made out of 5 x Livox Horizon LiDARs:
   ros2 bag record <params>

3. Start the pointcloud merge node inside ~/dataset/pointclouds. (the pcd files are stored in the folder in which the node is started):
   https://github.com/nerovalerius/pointcloud_merge_and_kitti.

4. Start playback of the ROSBAG within 5 seconds. Otherwise the pointcloud merge node quits.
   ros2 bag play rosbag2_2022_06_21-12_56_43_0.db3 -s sqlite3 --rate 0.5.

5. Convert the .pcd files with the pcd2bin node and store the .bin files inside ~/dataset/sequences/00/velodyne:
   ros2 run pcd2bin_kitti pcd2bin.

6. create a dummy calibration file inside your created sequence: nano ~/dataset/sequences/00/calib.txt and fill it with:

   P0: 1 0 0 0 0 1 0 0 0 0 1 0
   P1: 1 0 0 0 0 1 0 0 0 0 1 0
   P2: 1 0 0 0 0 1 0 0 0 0 1 0
   P3: 1 0 0 0 0 1 0 0 0 0 1 0
   Tr: 1 0 0 0 0 1 0 0 0 0 1 0
   

7. Run SuMa with: ~/catkin_ws/src/SuMa/bin/visualizer and open the first binary pointcloud of the sequence, e.g.: ~/dataset/sequences/00/full_cloud_00001.bin

   Run SuMa to get a poses.txt file for your sequence by pressing the PLAY button and then save poses.txt to the ~/dataset/sequences/00/ folder. SuMa creates a posegraph for your pointcloud sequence.

   Each pointcloud then gets a absolute position in a world_frame, which is used to load a complete sequence of pointclouds into the Point Labeler. This means that a large number of pointclouds can be labelled at once.

8. Open ~/dataset/sequences/00/ with Point Labeler. Run Point Labeler with:
   ~/catkin_ws/src/point_labeler/bin/labeler
   Point Labeler generates then generates the labes in the correct semantic KITTI format.

Perfect. Let me know if you need any help.

Best,
Armin