- Production from OpenDriveLab. Jointly with our fellow collaborators at Huawei Noah's Ark Lab, Fudan University, and HKU MMLab.
- Primary contact: Tianyu Li ( [email protected] ) or/and Li Chen.
This repository contains the source code of TopoNet, Graph-based Topology Reasoning for Driving Scenes.
TopoNet is the first end-to-end framework capable of abstracting traffic knowledge beyond conventional perception tasks, i.e., reasoning connections between centerlines and traffic elements from sensor inputs. It unifies heterogeneous feature learning and enhances feature interactions via the graph neural network architecture and the knowledge graph design.
Instead of recognizing lanes, we adhere that modeling the lane topology is appropriate to construct road components within the perception framework, to facilitate the ultimate driving comfort.
This is in accordance with the UniAD philosophy.
Pinned:The leaderboard for Lane Topology Challenge is open for regular submissions year around. This Challengewouldbe back in 2024's edition.[2023/11]π₯The code and model of OpenLane-V2 subset-B is released![2023/08]The code and model of TopoNet is released![2023/04]TopoNet paper is available on arXiv.[2023/01]Introducing Autonomous Driving Challenge for Lane Topology at CVPR 2023.
We provide results on Openlane-V2 subset-A val set.
| Method | Backbone | Epoch | DETl | TOPll | DETt | TOPlt | OLS |
|---|---|---|---|---|---|---|---|
| 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 | 8.3 | 0.2 | 43.5 | 5.8 | 20.0 |
| MapTR* | ResNet-50 | 24 | 17.7 | 1.1 | 43.5 | 10.4 | 26.0 |
| TopoNet | ResNet-50 | 24 | 28.6 | 4.1 | 48.6 | 20.3 | 35.6 |
π₯: Based on the updated v1.1 OpenLane-V2 devkit and metrics, we have reassessed the performance of TopoNet and other SOTA models. For more details please see issue #76 of OpenLane-V2.
| Method | Backbone | Epoch | DETl | TOPll | DETt | TOPlt | OLS |
|---|---|---|---|---|---|---|---|
| STSU | ResNet-50 | 24 | 12.7 | 2.9 | 43.0 | 19.8 | 29.3 |
| VectorMapNet | ResNet-50 | 24 | 11.1 | 2.7 | 41.7 | 9.2 | 24.9 |
| MapTR | ResNet-50 | 24 | 8.3 | 2.3 | 43.5 | 8.9 | 24.2 |
| MapTR* | ResNet-50 | 24 | 17.7 | 5.9 | 43.5 | 15.1 | 31.0 |
| TopoNet | ResNet-50 | 24 | 28.6 | 10.9 | 48.6 | 23.8 | 39.8 |
*: evaluation based on matching results on Chamfer distance.
The result of TopoNet is from this repo.
| Method | Backbone | Epoch | DETl | TOPll | DETt | TOPlt | OLS |
|---|---|---|---|---|---|---|---|
| TopoNet | ResNet-50 | 24 | 24.4 | 6.7 | 52.6 | 16.7 | 36.0 |
The result is based on the updated
v1.1OpenLane-V2 devkit and metrics.
The result of TopoNet is from this repo.
| Model | Dataset | Backbone | Epoch | OLS | Memory | Config | Download |
|---|---|---|---|---|---|---|---|
| TopoNet-R50 | subset-A | ResNet-50 | 24 | 39.8 | 12.3G | config | ckpt / log |
| TopoNet-R50 | subset-B | ResNet-50 | 24 | 36.0 | 8.2G | config | ckpt / log |
- Linux
- Python 3.8.x
- NVIDIA GPU + CUDA 11.1
- PyTorch 1.9.1
We recommend using conda to run the code.
conda create -n toponet python=3.8 -y
conda activate toponet
# (optional) If you have CUDA installed on your computer, skip this step.
conda install cudatoolkit=11.1.1 -c conda-forge
pip install torch==1.9.1+cu111 torchvision==0.10.1+cu111 -f https://download.pytorch.org/whl/torch_stable.htmlInstall mm-series packages.
pip install mmcv-full==1.5.2 -f https://download.openmmlab.com/mmcv/dist/cu111/torch1.9.0/index.html
pip install mmdet==2.26.0
pip install mmsegmentation==0.29.1
pip install mmdet3d==1.0.0rc6Install other required packages.
pip install -r requirements.txtFollowing OpenLane-V2 repo to download the data and run the preprocessing code.
cd TopoNet
mkdir data && cd data
ln -s {PATH to OpenLane-V2 repo}/data/OpenLane-V2After setup, the hierarchy of folder data is described below:
data/OpenLane-V2
βββ train
| βββ ...
βββ val
| βββ ...
βββ test
| βββ ...
βββ data_dict_subset_A_train.pkl
βββ data_dict_subset_A_val.pkl
βββ data_dict_subset_B_train.pkl
βββ data_dict_subset_B_val.pkl
βββ ...
We recommend using 8 GPUs for training. If a different number of GPUs is utilized, you can enhance performance by configuring the --autoscale-lr option. The training logs will be saved to work_dirs/toponet.
cd TopoNet
mkdir -p work_dirs/toponet
./tools/dist_train.sh 8 [--autoscale-lr]You can set --show to visualize the results.
./tools/dist_test.sh 8 [--show]All assets and code are under the Apache 2.0 license unless specified otherwise.
If this work is helpful for your research, please consider citing the following BibTeX entry.
@article{li2023toponet,
title={Graph-based Topology Reasoning for Driving Scenes},
author={Li, Tianyu and Chen, Li and Wang, Huijie and Li, Yang and Yang, Jiazhi and Geng, Xiangwei and Jiang, Shengyin and Wang, Yuting and Xu, Hang and Xu, Chunjing and Yan, Junchi and Luo, Ping and Li, Hongyang},
journal={arXiv preprint arXiv:2304.05277},
year={2023}
}
@inproceedings{wang2023openlanev2,
title={OpenLane-V2: A Topology Reasoning Benchmark for Unified 3D HD Mapping},
author={Wang, Huijie and Li, Tianyu and Li, Yang and Chen, Li and Sima, Chonghao and Liu, Zhenbo and Wang, Bangjun and Jia, Peijin and Wang, Yuting and Jiang, Shengyin and Wen, Feng and Xu, Hang and Luo, Ping and Yan, Junchi and Zhang, Wei and Li, Hongyang},
booktitle={NeurIPS},
year={2023}
}We acknowledge all the open-source contributors for the following projects to make this work possible:
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent