Mask-Based Panoptic LiDAR Segmentation for Autonomous Driving

This repository contains the implementation of our paper.

Installation

• Install this package by running in the root directory of this repo:

pip3 install -U -e .

• Install the packages in requirements.txt.

Data preparation

SemanticKITTI

Download the SemanticKITTI dataset inside the directory data/kitti/. The directory structure should look like this:

./
└── data/
    └── kitti
        └── sequences
            ├── 00/           
            │   ├── velodyne/	
            |   |	├── 000000.bin
            |   |	├── 000001.bin
            |   |	└── ...
            │   └── labels/ 
            |       ├── 000000.label
            |       ├── 000001.label
            |       └── ...
            ├── 08/ # for validation
            ├── 11/ # 11-21 for testing
            └── 21/
                └── ...

NuScenes

We use nuscenes2kitti to convert the nuScenes format into the SemanticKITTI format and store it in data/nuscenes/.

In the scripts, use the --nuscenes flag to train or evaluate using this dataset.

Pretrained models

• SemanticKITTI

• NuScenes

Reproducing results

python3 scripts/evaluate_model.py --w [path_to_model]

Training

python3 scripts/train_model.py

Citation

@article{marcuzzi2023ral,
  author = {R. Marcuzzi and L. Nunes and L. Wiesmann and J. Behley and C. Stachniss},
  title = {{Mask-Based Panoptic LiDAR Segmentation for Autonomous Driving}},
  journal = ral,
  volume = {8},
  number = {2},
  pages = {1141--1148},
  year = 2023,
  doi = {10.1109/LRA.2023.3236568},
  url = {https://www.ipb.uni-bonn.de/wp-content/papercite-data/pdf/marcuzzi2023ral.pdf},
}

Licence

Copyright 2023, Rodrigo Marcuzzi, Cyrill Stachniss, Photogrammetry and Robotics Lab, University of Bonn.

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