Home

Awesome

<div align="center">

<img src="media/mysmallhelix.gif" width="50"> HelixNet Toolbox <img src="media/mysmallhelix.gif" width="50">

<h1>ECCV 2022 &nbsp;&nbsp;&nbsp;&nbsp; <img src="media/ECCV-logo3.png" width="120"></h1>

helixnet_sequences

</div>

This repository contains helper scripts to open, visualize, and process point clouds from the HelixNet dataset. It also contains scripts to process data from Semantic-Kitti [1, 2] in slices smaller than one full sensor rotation.

HelixNet

We introduce HelixNet, a new large-scale and open-access LiDAR dataset intended for the evaluation of real-time semantic segmentation algorithms. In contrast to other large-scale datasets, HelixNet includes fine-grained data about the sensor’s rotation and position, as well as the points’ release time.

:arrow_forward: 8.85B annotated points:arrow_forward: 20 sequences:arrow_forward: 78k LiDAR rotations
:arrow_forward: 9 classes:arrow_forward: 6 cities:arrow_forward: dense sensor information

Usage

The dataset can be downloaded from Zenodo

DOI

This repository contains:

📦HelixNet
 ┣ 📂configs       # hydra config files for both datasets
 ┣ 📂helixnet      # PytorchLightning datamodules for both datasets
 ┃ ┣ 📂laserscan   # .py files loading individual files
 ┣ 📂misc          # helpers for visualisation in CloudCompare
 ┣ 📂demos         # some illustrative notebooks

Each point is associated with the 9 following values: (1-3) Cartesian coordinates in a fixed frame of a reference, (4-6) cylindrical coordinate relative to the sensor at the time of acquisition, (7) intensity, (8) fiber index, and (9) packet output time.

See our notebooks in /demos for examples of data manipulation and several visualization functions for semantic segmentation.

Leaderboard

Please open an issue to submit new entries. Mention if the work has already been published and wether the code accessible for reproducibility. We require that at least a preprint is available to add an entry.

We use a 9-classes nomenclature: road (16.4% of all points), other surface (22.0%), building (31.3%), vegetation (8.5%), traffic signs (1.6%), static vehicle (4.9%), moving vehicle (2.1%), pedestrian (0.9%), and artifact (0.05%).

Methods that are real-time (Inf. time < Acq. time) are denoted with :heavy_check_mark:, methods that are not with :x:.


Semantic Segmentation 1/5 Frame (Online) : Acquisition Time = 21ms

Model name#Params (M)mIoUInf. (ms)Published
Helix4D1.078.7:heavy_check_mark: 19:heavy_check_mark: link
Cylinder3D*55.975.0:x: 54:heavy_check_mark: link
PolarNet*13.672.2:x: 36:heavy_check_mark: link
SPVNAS*10.869.9:x: 44:heavy_check_mark: link
SalsaNeXt*6.768.2:heavy_check_mark: 10:heavy_check_mark: link

Semantic Segmentation Frame-by-Frame : Acquisition Time = 104ms

Model name#Params (M)mIoUInf. (ms)Published
Helix4D1.079.4:heavy_check_mark: 45:heavy_check_mark: link
Cylinder3D*55.976.6:x: 108:heavy_check_mark: link
PolarNet*13.673.6:heavy_check_mark: 49:heavy_check_mark: link
SPVNAS*10.873.4:heavy_check_mark: 73:heavy_check_mark: link
SalsaNeXt*6.769.4:heavy_check_mark: 23:heavy_check_mark: link

Models that we re-trained ourselves are denoted with a star (*).


Citation

If you use this dataset and/or this API in your work, please cite our paper:

@article{loiseau22online,
  title={Online Segmentation of LiDAR Sequences: Dataset and Algorithm.},
  author={Romain Loiseau and Mathieu Aubry and Loic Landrieu},
  journal={ECCV},
  year={2022}
}

Credits


[1] J. Behley et al., SemanticKITTI: A Dataset for Semantic Scene Understanding of LiDAR Sequences, ICCV, 2016.
[2] A. Geiger et al., Are we ready for Autonomous Driving? The KITTI Vision Benchmark Suite, CVPR, 2012.