TopoMLP: A Simple yet Strong Pipeline for Driving Topology Reasoning

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

[2024.01.16] 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.

Setup

For dataset preparation and environment requirements, please refer to OpenLane-V2.

Training and Evaluation

Training

If you want to train the model, please run the following command:

./tools/dist_train.sh ${CONFIG_FILE} ${GPU_NUM} [optional arguments]

For example, if you want to train TopoMLP on OpenLane-V2 subset-A train set, please run the following command:

./tools/dist_train.sh projects/configs/topomlp/topomlp_setA_r50_wo_yolov8.py 8 --work-dir=./work_dirs/topomlp_setA_r50_wo_yolov8

The training on 8 Nvidia A100 GPUs takes about 15 hours.

Evaluation

If you want to evaluate the model, please run the following command:

./tools/dist_test.sh ${CONFIG_FILE} ${CHECKPOINT_FILE} ${GPU_NUM} --eval=bbox

Main Results

OpenLane-V2 subset-A val set:

Method	Backbone	Epoch	DET_l	TOP_ll	DET_t	TOP_lt	OLS	Weight/Log
STSU	ResNet-50	24	12.7	0.5	43.0	15.1	25.4	-
VectorMapNet	ResNet-50	24	11.1	0.4	41.7	6.2	20.8	-
MapTR	ResNet-50	24	17.7	1.1	43.5	10.4	26.0	-
TopoNet	ResNet-50	24	28.5	4.1	48.1	20.8	35.6	-
TopoMLP	ResNet-50	24	28.3	7.2	50.0	22.8	38.2	weight/log
TopoMLP*	ResNet-50	24	28.8	7.8	53.3	30.1	41.2

$*$ means using YOLOv8 proposals.

OpenLane-V2 subset-B val set:

Method	Backbone	Epoch	DET_l	TOP_ll	DET_t	TOP_lt	OLS	Weight/Log
STSU	ResNet-50	24	8.2	0.0	43.9	9.4	21.2	-
VectorMapNet	ResNet-50	24	3.5	0.0	49.1	1.4	16.3	-
MapTR	ResNet-50	24	15.2	0.5	54.0	6.1	25.2	-
TopoNet	ResNet-50	24	24.3	2.5	55.0	14.2	33.2	-
TopoMLP	ResNet-50	24	26.6	7.6	58.3	17.8	38.7	weight/log

Citation

If you find our work useful in your research, please consider citing it.

@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={arXiv preprint},
  year={2023}
}

Acknowledgements

We thank the authors that open the following projects.

mengxingshifen1218/TopoMLP