This repository contains simple usage explanations of how the RangeNet++ inference works with the TensorRT and C++ interface.

Developed by Xieyuanli Chen, Andres Milioto and Jens Behley.

For more details about RangeNet++, one could find in LiDAR-Bonnetal.

First you need to install the nvidia driver and CUDA.

• CUDA Installation guide: Link

• Then you can do the other dependencies:

  $ sudo apt-get update 
  $ sudo apt-get install -yqq  build-essential python3-dev python3-pip apt-utils git cmake libboost-all-dev libyaml-cpp-dev libopencv-dev

• Then install the Python packages needed:

  $ sudo apt install python-empy
  $ sudo pip install catkin_tools trollius numpy

In order to infer with TensorRT during inference with the C++ libraries:

• Install TensorRT: Link.
• Our code and the pretrained model now only works with TensorRT version 5 (Note that you need at least version 5.1.0).
• To make the code also works for higher versions of TensorRT, one could have a look at here.

We use the catkin tool to build the library.

$ mkdir -p ~/catkin_ws/src
$ cd ~/catkin_ws/src
$ git clone https://github.com/ros/catkin.git 
$ git clone https://github.com/PRBonn/rangenet_lib.git
$ cd .. && catkin init
$ catkin build rangenet_lib

To run the demo, you need a pre-trained model, which can be downloaded here, model.

A single LiDAR scan for running the demo, you could find in the example folder example/000000.bin. For more LiDAR data, you could download from KITTI odometry dataset.

For more details about how to train and evaluate a model, please refer to LiDAR-Bonnetal.

To infer a single LiDAR scan and visualize the semantic point cloud:

# go to the root path of the catkin workspace
$ cd ~/catkin_ws
# use --verbose or -v to get verbose mode
$ ./devel/lib/rangenet_lib/infer -h # help
$ ./devel/lib/rangenet_lib/infer -p /path/to/the/pretrained/model -s /path/to/the/scan.bin --verbose

Notice: for the first time running, it will take several minutes to generate a .trt model for C++ interface.

Using rangenet_lib, we built a LiDAR-based Semantic SLAM system, called SuMa++.

You could find more implementation details in SuMa++.

OverlapNet is a LiDAR-based loop closure detection method, which uses multiple cues generated from LiDAR scans.

More information about our OverlapNet could be found here.

One could use our rangenet_lib to generate probabilities over semantic classes for training OverlapNet.

More detailed steps and discussion could be found here.

If you use this library for any academic work, please cite the original paper.

@inproceedings{milioto2019iros,
  author    = {A. Milioto and I. Vizzo and J. Behley and C. Stachniss},
  title     = {{RangeNet++: Fast and Accurate LiDAR Semantic Segmentation}},
  booktitle = {IEEE/RSJ Intl.~Conf.~on Intelligent Robots and Systems (IROS)},
  year      = 2019,
  codeurl   = {https://github.com/PRBonn/lidar-bonnetal},
  videourl  = {https://youtu.be/wuokg7MFZyU},
}

If you use SuMa++, please cite the corresponding paper:

@inproceedings{chen2019iros, 
  author    = {X. Chen and A. Milioto and E. Palazzolo and P. Giguère and J. Behley and C. Stachniss},
  title     = {{SuMa++: Efficient LiDAR-based Semantic SLAM}},
  booktitle = {Proceedings of the IEEE/RSJ Int. Conf. on Intelligent Robots and Systems (IROS)},
  year      = {2019},
  codeurl   = {https://github.com/PRBonn/semantic_suma/},
  videourl  = {https://youtu.be/uo3ZuLuFAzk},
}

Copyright 2019, Xieyuanli Chen, Andres Milioto, Jens Behley, Cyrill Stachniss, University of Bonn.

This project is free software made available under the MIT License. For details see the LICENSE file.