Efficient Online Transfer Learning for 3D Object Classification in Autonomous Driving

We are actively updating this repository (especially removing hard code and adding comments) to make it easy to use. If you have any questions, please open an issue. Thanks!

This is a ROS-based efficient online learning framework for object classification in 3D LiDAR scans, taking advantage of robust multi-target tracking to avoid the need for data annotation by a human expert. The system is only tested in Ubuntu 18.04 and ROS Melodic (compilation fails on Ubuntu 20.04 and ROS Noetic).

Please watch the videos below for more details.

Install & Build

Please read the readme file of each sub-package first and install the corresponding dependencies.

Run

1. Prepare dataset

• (Optional) Download the raw data from KITTI.

• (Optional) Download the sample data for testing.

• (Optional) Prepare a customized dataset according to the format of the sample data.

2. Manual set specific path parameters

 # launch/efficient_online_learning
 # autoware_tracker/config/params.yaml

3. Run the project

$ cd catkin_ws
$ source devel/setup.bash
$ roslaunch src/efficient_online_learning/launch/efficient_online_learning.launch

Citation

If you are considering using this code, please reference the following:

@article{efficient_online_learning,
   author = {Rui Yang and Zhi Yan and Tao Yang and Yassine Ruichek},
   title = {Efficient Online Transfer Learning for 3D Object Classification in Autonomous Driving},
   journal = {CoRR},
   volume = {abs/2104.10037},
   year = {2021},
   url = {http://arxiv.org/abs/2104.10037},
   archivePrefix = {arXiv},
   eprint = {2104.10037}
}