/DAIR-V2X

Primary LanguagePythonApache License 2.0Apache-2.0

DAIR-V2X and OpenDAIRV2X: Towards General and Real-World Cooperative Autonomous Driving

Project Page | Dataset Download | arXiv | OpenDAIRV2X



teaser

Table of Contents:

  1. Highlights
  2. News
  3. Dataset Download
  4. Getting Started
  5. Major Features
  6. Benchmark
  7. Citation
  8. Contaction

Highlights

  • DAIR-V2X: The first real-world dataset for research on vehicle-to-everything autonomous driving. It comprises a total of 71,254 frames of image data and 71,254 frames of point cloud data.
  • V2X-Seq: The first large-scale, real-world, and sequential V2X dataset, which includes data frames, trajectories, vector maps, and traffic lights captured from natural scenery. V2X-Seq comprises two parts: V2X-Seq-SPD (Sequential Perception Dataset), which includes more than 15,000 frames captured from 95 scenarios; V2X-Seq-TFD (Trajectory Forecasting Dataset), which contains about 80,000 infrastructure-view scenarios, 80,000 vehicle-view scenarios, and 50,000 cooperative-view scenarios captured from 28 intersections' areas, covering 672 hours of data.
  • OpenDAIR-V2X: An open-sourced framework for supporting the research on vehicle-to-everything autonomous driving.

News

  • [2024.04] 🔥 Our UniV2X available on arXiv. UniV2X is the first end-to-end framework that unifies all vital modules as well as diverse driving views into a network for cooperative autonomous driving. Code will be here.
  • [2024.03] 🔥 Our new Dataset RCooper, a real-world large-scale dataset for roadside cooperative perception, has been accepted by CVPR2024! Please follow RCooper for the latest news.
  • [2024.01] 🔥 Our QUEST has been been accpeted by ICRA2024.
  • [2023.10] We have released the code for V2X-Seq-SPD and V2X-Seq-TFD.
  • [2023.09] Our FFNET has been accpeted by Neurips2023.
  • [2023.05] V2X-Seq dataset is availale here. It can be unlimitedly downloaded within mainland China. Example dataset can be downloaded directly.
  • [2023.03] Our new dataset "V2X-Seq: A Large-Scale Sequential Dataset for Vehicle-Infrastructure Cooperative Perception and Forecasting" has been accepted by CVPR2023. Congratulations! We will release the dataset sooner. Please follow DAIR-V2X-Seq for the latest news.
  • [2023.03] We have released training code for our FFNET, and our OpenDAIRV2X now supports evaluating FFNET.
  • [2022.11] We have held the first VIC3D Object Detection challenge.
  • [2022.07] We have released the OpenDAIRV2X codebase v1.0.0. The current version can faciliate the researchers to use the DAIR-V2X dataset and reproduce the benchmarks.
  • [2022.03] Our Paper "DAIR-V2X: A Large-Scale Dataset for Vehicle-Infrastructure Cooperative 3D Object Detection" has been accepted by CVPR2022. Arxiv version could be seen here.
  • [2022.02] DAIR-V2X dataset is availale here. It can be unlimitedly downloaded within mainland China.

Dataset Download

Getting Started

Please refer to getting_started.md for the usage and benchmarks reproduction of DAIR-V2X dataset.

Please refer to get_started_spd.md for the usage and benchmarks reproduction of V2X-Seq-SPD dataset.

Benchmark

You can find more benchmark in SV3D-Veh, SV3D-Inf, VIC3D and VIC3D-SPD.

Part of the VIC3D detection benchmarks based on DAIR-V2X-C dataset:

Modality Fusion Model Dataset AP-3D (IoU=0.5) AP-BEV (IoU=0.5) AB
Overall 0-30m 30-50m 50-100m Overall 0-30m 30-50m 50-100m
Image VehOnly ImvoxelNet VIC-Sync 9.13 19.06 5.23 0.41 10.96 21.93 7.28 0.78 0
Late-Fusion ImvoxelNet VIC-Sync 18.77 33.47 9.43 8.62 24.85 39.49 14.68 14.96 309.38
Pointcloud VehOnly PointPillars VIC-Sync 48.06 47.62 63.51 44.37 52.24 30.55 66.03 48.36 0
Early Fusion PointPillars VIC-Sync 62.61 64.82 68.68 56.57 68.91 68.92 73.64 65.66 1382275.75
Late-Fusion PointPillars VIC-Sync 56.06 55.69 68.44 53.60 62.06 61.52 72.53 60.57 478.61
Late-Fusion PointPillars VIC-Async-2 52.43 51.13 67.09 49.86 58.10 57.23 70.86 55.78 478.01
TCLF PointPillars VIC-Async-2 53.37 52.41 67.33 50.87 59.17 58.25 71.20 57.43 897.91

Part of the VIC3D detection and tracking benchmarks based on V2X-Seq-SPD:

Modality Fusion Model Dataset AP 3D (Iou=0.5) AP BEV (Iou=0.5) MOTA MOTP AMOTA AMOTP IDs AB(Byte)
Image Veh Only ImvoxelNet VIC-Sync-SPD 8.55 10.32 10.19 57.83 1.36 14.75 4
Image Late Fusion ImvoxelNet VIC-Sync-SPD 17.31 22.53 21.81 56.67 6.22 25.24 47 3300

TODO List

  • Dataset Release
  • Dataset API
  • Evaluation Code
  • All detection benchmarks based on DAIR-V2X dataset
  • Benchmarks for detection and tracking tasks with different fusion strategies for Image based on V2X-Seq-SPD dataset
  • All benchmarks for detection and tracking tasks based on V2X-Seq-SPD dataset

Citation

Please consider citing our paper if the project helps your research with the following BibTex:

@inproceedings{v2x-seq,
  title={V2X-Seq: A large-scale sequential dataset for vehicle-infrastructure cooperative perception and forecasting},
  author={Yu, Haibao and Yang, Wenxian and Ruan, Hongzhi and Yang, Zhenwei and Tang, Yingjuan and Gao, Xu and Hao, Xin and Shi, Yifeng and Pan, Yifeng and Sun, Ning and Song, Juan and Yuan, Jirui and Luo, Ping and Nie, Zaiqing},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  year={2023},
}
@inproceedings{dair-v2x,
  title={Dair-v2x: A large-scale dataset for vehicle-infrastructure cooperative 3d object detection},
  author={Yu, Haibao and Luo, Yizhen and Shu, Mao and Huo, Yiyi and Yang, Zebang and Shi, Yifeng and Guo, Zhenglong and Li, Hanyu and Hu, Xing and Yuan, Jirui and Nie, Zaiqing},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={21361--21370},
  year={2022}
}

Contaction

If any questions and suggenstations, please email to dair@air.tsinghua.edu.cn.

Related Resources

Awesome