KITTI Scan Unfolding

Python implementation of KITTI scan unfolding.

We propose KITTI scan unfolding in our paper Scan-based Semantic Segmentation of LiDAR Point Clouds: An Experimental Study. If you use this code or the algorithm described in the paper, please make sure to cite this paper.

Fig. 1: Cylindrical point cloud projection: (a) Correcting for ego-motion leads to a projection that suffers from systematic point occlusions as some 3D points are projected into occupied pixels. Hidden points can not provide any information to the network and may not be accurately classified. (b) The scan unfolding method provides a dense projection without systematic discretization artifacts.

Description

A detailed explanation on the scan unfolding is given in our paper. Here we provide only a brief summary on how to use the code.

The unfolding method is not a general algorithm that can be applied to any LiDAR dataset. It is rather an engineering trick that can be applied to raw KITTI data and raw KITTI data only. You may find the raw KITTI dataset here. The data exhibits a unique pattern in which the files are created, easing the way to restore the original internal raw data structure of the sensor, i.e. the sensor view image.

Installation

Install package unfolding directly from GitHub with

pip install git+https://github.com/ltriess/kitti_scan_unfolding

It only requires numpy to run.

Usage

The package provides three functions. Two which return the indices for the rows and columns in the image-like projection and one that can be used to generate the projection directly. For a comprehensive guide take a look at this notebook.

We recommend using the projection function directly. It can unfold all additional channels accordingly to points.

import numpy as np
import unfolding

scan = np.fromfile("/path/to/kitti.bin", dtype=np.float32).reshape((-1, 4))
points = scan[:, :3]
remissions = scan[:, 3]

projection = unfolding.projection(points, remissions, image_size=(64, 2000))
proj_points = projection["points"]  # the projected points (64, 2000, 3)
proj_remissions = projection["channels"][0]  # the projected remissions (64, 2000)

Fig. 2: The unfolded scan: The top shows the color coded depth values, the bottom are the intensity values.

License

This project is licensed under the MIT License - see the LICENSE for details.

Citation

When you use this code or the algorithm from the paper, please cite

@article{triess2020,
    title={{Scan-based Semantic Segmentation of LiDAR Point Clouds: An Experimental Study}},
    author={Triess, Larissa T. and Peter, David and Rist, Christoph B. and Z\"ollner, J. Marius},
    journal={arxiv},
    year={2020}
}

References

[1] A. Milioto et al., “RangeNet++: Fast and Accurate LiDAR Semantic Segmentation,” in IEEE/RSJ Intl. Conf. on Intelligent Robots and Systems (IROS), 2019.

songw-zju / kitti_scan_unfolding