TopoMLP: A Simple yet Strong Pipeline for Driving Topology Reasoning

Dongming Wu, Jiahao Chang, Fan Jia, Yingfei Liu, Tiancai Wang, Jianbing Shen

TL;DR

TopoMLP is the 1st solution for 1st OpenLane Topology in Autonomous Driving Challenge. It suggests a first-detect-then-reason philosophy for better topology prediction. It includes two well-designed high-performance detectors and two elegant MLP networks with position embedding for topology reasoning.

• For lane detection, we represent each centerline as a smooth Bezier curve.
• For traffic detection, we propose to optionally improve the query-based detectors by elegantly incorporating an extra object detector YOLOv8.
• For lane-lane and lane-traffic topology prediction, MLPs is enough for better performance.

News

• [2024.04.08] Other backbones are released for the incoming Mapless Driving Challenge. Welcome star and cite!
• [2024.01.16] TopoMLP is accepted by ICLR2024.
• [2023.10.11] Code is released. TopoMLP paper is released at arXiv.
• [2023.06.16] Tech report is released at arXiv.
• [2023.06.02] We achieve the 1st for 1st OpenLane Topology in Autonomous Driving Challenge.

Getting Started

• Environment and Dataset Setup
• Training and Evaluation

Main Results

OpenLane-V2 subset-A val set

OpenLane-V2 devkit and metrics v1.0.0:

Method	Backbone	Pretrain	DETl	TOPll	DETt	TOPlt	OLS	Config	Weight/Log
TopoMLP	ResNet-50	-	28.5	7.1	49.5	23.4	38.3	config	weight/log
TopoMLP	VOV	FCOS3D	31.6	9.4	51.1	26.6	41.2	config	log
TopoMLP	Swin-B	-	31.6	9.2	54.2	28.6	42.4	config	log

OpenLane-V2 devkit and metrics v2.1.0:

Method	Backbone	Pretrain	DETl	TOPll	DETt	TOPlt	OLS	Config	Weight/Log
TopoMLP	ResNet-50	-	28.5	21.7	49.5	26.9	44.1	config	weight/log

OpenLane-V2 subset-B val set

OpenLane-V2 devkit and metrics v1.0.0:

Method	Backbone	Pretrain	DETl	TOPll	DETt	TOPlt	OLS	Config	Weight/Log
TopoMLP	ResNet-50	-	26.3	7.2	57.8	17.0	38.0	config	-

OpenLane-V2 devkit and metrics v2.1.0:

Method	Backbone	Pretrain	DETl	TOPll	DETt	TOPlt	OLS	Config	Weight/Log
TopoMLP	ResNet-50	-	26.3	21.2	57.8	19.4	43.6	config	-

Notes:

• Our code supports flash attention, which is not used in the above results. You can replace the PETRMultiheadAttention in the config file to PETRMultiheadFlashAttention to use it.
• ViT-Large can refer to StreamPETR. From our practice, it is not as good as Swin-B in this overall task, but it can perform well in sub-task centerline detection as reported in our technical report.

Citation

If you find TopoMLP is useful in your research or applications, please consider giving us a star 🌟 and citing it by the following BibTeX entry.

@article{wu2023topomlp,
  title={TopoMLP: An Simple yet Strong Pipeline for Driving Topology Reasoning},
  author={Wu, Dongming and Chang, Jiahao and Jia, Fan and Liu, Yingfei and Wang, Tiancai and Shen, Jianbing},
  journal={ICLR},
  year={2024}
}

@article{wu20231st,
  title={The 1st-place solution for cvpr 2023 openlane topology in autonomous driving challenge},
  author={Wu, Dongming and Jia, Fan and Chang, Jiahao and Li, Zhuoling and Sun, Jianjian and Han, Chunrui and Li, Shuailin and Liu, Yingfei and Ge, Zheng and Wang, Tiancai},
  journal={arXiv preprint arXiv:2306.09590},
  year={2023}
}

Acknowledgements

We thank the authors that open the following projects.

• MMDetection3d
• PETRv2
• MOTRv2
• OpenLane-v2
• TopoNet