RCooper: A Real-world Large-scale Dataset for Roadside Cooperative Perception
This is the official implementation of CVPR2024 paper. “RCooper: A Real-world Large-scale Dataset for Roadside Cooperative Perception”.
Ruiyang Hao*, Siqi Fan*, Yingru Dai, Zhenlin Zhang, Chenxi Li, Yuntian Wang, Haibao Yu, Wenxian Yang, Jirui Yuan, Zaiqing Nie†
Overview
Data Download
Please check this page website to download the data. As shown in the figure bellow.
After downloading the data, please put the data in the following structure:
├── RCooper
│ ├── calib
| |── lidar2cam
| |── lidar2world
│ ├── data
| |── folders named specific scene index
│ ├── labels
| |── folders named specific scene index
│ ├── original_label
| |── folders named specific scene index
Data Conversion
To facilitate the research of cooperative perception methods on RCooper. We provide the format converter from RCooper to other popular public cooperative perception datasets. After the conversion, researchers can directly employ the methods using several opensourced frameworks.
We now support the following conversions:
RCooper to V2V4Real
Setup the dataset path in codes/dataset_convertor/converter_config.py, and complete the conversion.
cd codes/dataset_converter
python rcooper2vvreal.py
RCooper to OPV2V
Setup the dataset path in codes/dataset_convertor/converter_config.py, and complete the conversion.
cd codes/dataset_converter
python rcooper2opv2v.py
RCooper to DAIR-V2X
Setup the dataset path in codes/dataset_convertor/converter_config.py, and complete the conversion.
cd codes/dataset_converter
python rcooper2dair.py
Quick Start
For detection training & inference, you can find instructions in docs/corridor_scene or docs/intersection_scene in detail. (Notes: you may need to set PYTHONPATH to call modified codes other than the pip-installed ones.)
For Tracking, you can find instructions in docs/tracking.md in detail.
All the checkpoints are released in link in the tabels below, you can save them in codes/ckpts/.
Benchmark
Results of Cooperative 3D object detection for corridor scenes
Results of Cooperative 3D object detection for intersection scenes
Results of Cooperative tracking for corridor scenes
| Method |
AMOTA(↑) |
AMOTP(↑) |
sAMOTA(↑) |
MOTA(↑) |
MT(↑) |
ML(↓) |
| No Fusion |
8.28 |
22.74 |
34.05 |
23.89 |
17.34 |
42.71 |
| Late Fusion |
9.60 |
25.77 |
35.64 |
24.75 |
24.37 |
42.96 |
| Early Fusion |
23.78 |
38.18 |
59.16 |
44.30 |
53.02 |
12.81 |
| AttFuse |
21.75 |
35.31 |
57.43 |
44.50 |
45.73 |
22.86 |
| F-Cooper |
22.47 |
35.54 |
58.49 |
45.94 |
47.74 |
22.11 |
| Where2Comm |
22.55 |
36.21 |
59.60 |
46.11 |
50.00 |
19.60 |
| CoBEVT |
21.54 |
35.69 |
53.85 |
47.32 |
47.24 |
18.09 |
Results of Cooperative tracking for corridor scenes
| Method |
AMOTA(↑) |
AMOTP(↑) |
sAMOTA(↑) |
MOTA(↑) |
MT(↑) |
ML(↓) |
| No Fusion |
18.11 |
39.71 |
58.29 |
49.16 |
35.32 |
41.64 |
| Late Fusion |
21.57 |
43.40 |
63.02 |
50.58 |
42.75 |
34.20 |
| Early Fusion |
21.38 |
47.71 |
62.93 |
50.15 |
36.80 |
42.75 |
| AttFuse |
11.84 |
36.63 |
46.92 |
39.32 |
29.00 |
53.90 |
| F-Cooper |
-4.86 |
14.71 |
0.00 |
-45.66 |
11.52 |
50.56 |
| Where2Comm |
14.21 |
38.48 |
50.97 |
42.27 |
29.00 |
45.72 |
| CoBEVT |
14.82 |
38.71 |
49.04 |
44.67 |
33.83 |
35.69 |
Citation
If you find RCooper useful in your research or applications, please consider giving us a star 🌟 and citing it by the following BibTeX entry.
@inproceedings{hao2024rcooper,
title={RCooper: A Real-world Large-scale Dataset for Roadside Cooperative Perception},
author={Hao, Ruiyang and Fan, Siqi and Dai, Yingru and Zhang, Zhenlin and Li, Chenxi and Wang, Yuntian and Yu, Haibao and Yang, Wenxian and Jirui, Yuan and Nie, Zaiqing},
booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
year={2024},
pages={22347-22357}
}
Acknowledgment
Sincere appreciation for their great contributions.
评论