Livox Detection-simu V1.0: Trained on Simulated Data, Tested in the Real World [Livox Simu-dataset]


Livox Detection-simu is a robust and real-time detection package trained on Livox Simu-dataset. It only uses 14k frames of simulated data for training, and performs effective detection in the real world. The inference time is about 50ms on 2080Ti for 200m*100m range detection.
We hope this project can help you make better use of Livox Simu-dataset. In order to improve the performance of the detector, data augmentation such as object insertion and background mix-up is necessary.



  • python3.6+
  • tensorflow1.13+ (tested on 1.13.0)
  • pybind11
  • ros


  1. Clone this repository.
  2. Clone pybind11 from pybind11.

$ cd utils/lib_cpp
$ git clone
  1. Compile C++ module in utils/lib_cpp by running the following command.

$ mkdir build && cd build
$ cmake -DCMAKE_BUILD_TYPE=Release ..
$ make
  1. copy the to root directory:

$ cp ../../../
  1. Download the pre_trained model and unzip it to the root directory.


For sequence frame detection

Download the provided rosbags : rosbag and then

$ roscore

$ rviz -d ./config/show.rviz

$ python

$ rosbag play *.bag -r 0.1

The network inference time is around 25ms, but the point cloud data preprocessing module takes a lot of time based on python. If you want to get a faster real time detection demo, you can modify the point cloud data preprocessing module with c++.

To play with your own rosbag, please change your rosbag topic to /livox/lidar.


You can get support from Livox with the following methods :

  • Send email to [email protected] with a clear description of your problem and your setup
  • Report issue on github


GitHub - Livox-SDK/livox_detection_simu at
Trained on Simulated Data, Tested in the Real World - GitHub - Livox-SDK/livox_detection_simu at