PRBonn / nuscenes2kitti

Converts panoptic nuScenes into SemanticKITTI-compatible format.

First you have to download the nuScene dataset, which include the Full dataset, the nuScenes-lidarseg dataset and the nuScenes-panoptic datset. The final data folder structure will look like:

Nuscene
├── lidarseg
├── maps
├── panoptic
├── samples
├── sweeps
├── v1.0-mini
├── v1.0-test
└── v1.0-trainval

warning: the v1.0-mini, v1.0-test, v1.0-trainval must have lidarseg.json and panotic.json. For example

Nuscene/v1.0-mini/
├── lidarseg.json
└── panoptic.json

Script usage:

python3 nuscenes2kitti.py --nuscenes_dir <nuscenes_directory> --output_dir <output_directory>

options:

--normalize_remission set remission values to range [0,1].

--mini compute only for the mini set of scenes.

--save_images save frontal image for each point cloud

This will create a directory for each scene in <output_directory> with the same structure as SemanticKITTI containing:

output_directory
└── scene_id
    ├── calib.txt           # dummy file containing just identity transforms
    ├── files_mapping.txt   # original filenames
    ├── labels              # directory containing the panoptic labels
    ├── lidar_tokens.txt    # token for each point cloud
    ├── poses.txt           # pose for each laserscan
    └── velodyne            # directory containing the point clouds

Config file

We provide the semantic-nuscenes.yaml file, which follows same structure as semantic-kitti.yaml and contains the labels, color_map, learning_map and splits.

License

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