The key to real-world generability is scale
Successor Lane Graph Prediction, Full Lane Graph Prediction, and Planning
Easy-to-use benchmarking API and dataset instructions
The UrbanLaneGraph dataset is a first-of-its-kind dataset for large-scale lane graph estimation from aerial images. The dataset contains aerial images from the cities of Austin, Miami, Pittsburgh, Palo Alto, Detroit, and Washington DC along with per-city lane graph annotations, aggregated from the Argoverse2 dataset.
The annotated regions
feature a diverse range of environments including urban, suburban, and rural regions with complex lane topologies.
Accumulated over all cities, the overall length of all lanes spans
over 5.000 km.
All images were captured at a resolution of 15 cm per pixel, enabling a detailed analysis of lane graphs, resolving even the smallest details.
To facilitate easier and quantitatively fair comparison between different approaches, we provide an easy-to-use graph prediction benchmarking API.
If you find the UrbanLaneGraph dataset useful for your research, please cite the following paper:
@article{buchner2023learning,
title={Learning and Aggregating Lane Graphs for Urban Automated Driving},
author={B{\"u}chner, Martin and Z{\"u}rn, Jannik and Todoran, Ion-George and Valada, Abhinav and Burgard, Wolfram},
journal={arXiv preprint arXiv:2302.06175},
year={2023}
}
Please also consider citing the following papers which introduce the Argoverse2 dataset:
@INPROCEEDINGS { Argoverse2,
author = {Benjamin Wilson and William Qi and Tanmay Agarwal and John Lambert and Jagjeet Singh and
Siddhesh Khandelwal and Bowen Pan and Ratnesh Kumar and Andrew Hartnett and Jhony Kaesemodel
Pontes and Deva Ramanan and Peter Carr and James Hays},
title = {Argoverse 2: Next Generation Datasets for Self-driving Perception and Forecasting},
booktitle = {Proceedings of the Neural Information Processing Systems Track on Datasets and Benchmarks
(NeurIPS Datasets and Benchmarks 2021)},
year = {2021}
}
@INPROCEEDINGS { TrustButVerify,
author = {John Lambert and James Hays},
title = {Trust, but Verify: Cross-Modality Fusion for HD Map Change Detection},
booktitle = {Proceedings of the Neural Information Processing Systems Track on Datasets and Benchmarks
(NeurIPS Datasets and Benchmarks 2021)},
year = {2021}
}
In addition to the UrbanLaneGraph dataset, we propose two novel methods for lane graph prediction.
Firstly, LaneGNN, a novel lane graph neural network that allows inferring ego-respective successor lane graphs.
Different to other methods, we directly predict in graph space instead of the image space:
Secondly, we publish a powerful method-agnostic merging scheme for aggregating lane graphs. We found that this merging scheme drastically improves the overall quality of the produced lane graph in the Full-LGP task:
We offer the UrbanLaneGraph dataset free of charge under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International Public License (“CC BY-NC-SA 4.0”).
By downloading the data, you accept the license agreement which is available here.
The UrbanLaneGraph dataset is based on the Argoverse2 dataset (https://www.argoverse.org/av2.html).
We thank the Argoverse2 team for making the Argoverse2 datset publicly available and for allowing the re-distribution of their dataset in remixed form.
We thank Hexagon and Woven by Toyota for providing the aerial images entailed in the dataset.
