CooperScene: Multi-Modal Cooperative Autonomy Benchmark with C-V2X Communication Characterization
Introduction
🚗 This repository hosts the model configs and pre-trained checkpoints for CooperScene — the first real-world, multi-agent, multi-modal cooperative autonomy dataset with C-V2X communication characterization (three connected vehicles + one roadside unit, across intersections, highway ramps, and parking areas).
🚀 All training and inference code is open-sourced. See the project page and the GitHub repo for details.
💬 We welcome feedback and look forward to your comments!
What's here
Each model has its config and matching checkpoint together under
configs/<model>/:
| Cooperative detectors | BEVFusion |
|---|---|
cobevt |
bevfusion_single_lidar |
cosdh |
bevfusion_single_lidarcam |
ermvp |
bevfusion_coop_lidar |
v2vam |
bevfusion_coop_lidarcam |
v2vnet |
|
v2xvit |
All models run on a unified mmengine pipeline (proj_first=True, same global-sort
BEV/3D polygon-IoU AP @ 0.3 / 0.5 / 0.7), so the numbers are directly comparable.
Download
pip install -U huggingface_hub
hf download cisl-hf/CooperScene --local-dir assets
# -> assets/configs/<model>/{<model>.py, <model>.pth}
Usage
Clone the code repo, then evaluate or train with a downloaded config + checkpoint:
# evaluate (test split by default)
python tools/test.py assets/configs/ermvp/ermvp.py assets/configs/ermvp/ermvp.pth
# train (warm-start from a checkpoint, optional)
python tools/train.py assets/configs/ermvp/ermvp.py
See the GitHub README for data preparation and the Docker workflow.
Related links
🌐 Website: https://cisl.ucr.edu/CooperScene
💻 GitHub: https://github.com/UCR-CISL/CooperScene
🤗 Hugging Face: https://huggingface.co/cisl-hf/CooperScene