LiDAR Point Cloud Semantic Segmentation

An implementation of point painting for real-time point cloud semantic segmentation painting (labeling each liDAR point with a class) based on semantic segmentation maps using DeepLabV3+.

Dataset

Download the rectified stereo camera images and Velodyne sensor data from the KITTI-360 dataset. Save data in 2 folders respectively. To download them, you can use the the following shell script.

1. Recticied RGB camera images: bash download_2d_perspective.sh.
2. Velodyne point cloud bash download_3d_velodyne.sh.
3. camera intrinsics and extrinsics from here.

Checkpoint

For the semantic segmentation of RBG images, please download the pre-trained DeepLabV3+ checkpoint from here which is trained on Cityscapes dataset.

Run

python3 /Code/Wrapper.py --DataPath {dir/to/kitti/dataset} --SavePath {dir/to/your/saving/pcd/folder} --CkptPath {dir/to/DeepLabV3+/checkpoint} --ParseData {Parse the raw point cloud data, Default:0}

Demo

References

• DeepLabv3 https://towardsdatascience.com/review-deeplabv3-atrous-convolution-semantic-segmentation-6d818bfd1d74
• Point Painting https://github.com/AmrElsersy/PointPainting/tree/master