VoxelNet Implementation codes from "https://github.com/qianguih/voxelnet"

• python3.5+
• TensorFlow (tested on 1.4)
• opencv
• shapely
• numba
• easydict
• ROS(melodic)
• jsk package(ros-jsk-recognition)
• cv_brdge(ros)
• nodelet(ros)
• pcl_ros(ros)
• rviz(ros)

Download the test data from the raw KITII dataset

unzip to data folder,Directory structure should be:

data
----npy_lidar_0048(lidar point cloud)
----result_road_0048(road detection result)
----rgb_image_0048(correponding rgb left camera image)

before run the code, you may need to install:

• clone this repository
• move voxelnet_ros folder to your catkin_ws
• catkin_make
• roscd voxelnet/script/
• you probably also need to complie the box_overlap.so using numba
• python3 voxelnet_ros.py & python3 pub_kitti_point_cloud.py
  • unfortunately, rosrun won't work. because it's using Python 3.x instead of 2.x
• roslaunch voxelnet_ros convex_hull.launch
• rosrun rviz rviz
  • use the voxelnet.rviz to config rviz

• Retrain the model
• Still too slow, need more efficient implementation for the VoxelNet
• code clean