V2XScenes

[ICCV 2025] V2XScenes: A Multiple Challenging Traffic Conditions Dataset for Large-Range Vehicle-Infrastructure Collaborative Perception

This repository contains a real-world multi-challenging-condition dataset covering large-range road sections for multi-modal V2X cooperative perception research.

Paper | Project Page | Access Request

📋 Overview

V2XScenes is a comprehensive dataset designed for vehicle-infrastructure cooperative perception research. It features:

• Multi-modal sensor data: LiDAR and camera data from both vehicle and roadside platforms
• Diverse traffic conditions: Multiple challenging scenarios including various weather and traffic states
• Large-range coverage: Data collected over extended road sections for realistic deployment scenarios
• High-quality annotations: Carefully labeled 3D bounding boxes and metadata

📥 Data Preparation

Step 1: Request Dataset Access

To obtain the V2XScenes dataset, please visit our Data Access Portal and submit your access request.

Step 2: Organize Dataset Structure

Create a dataset folder under the V2XScenes repository root and organize the downloaded data as follows:

dataset/v2xscenes/
├── calibration/                  # Sensor calibration parameters
├── data/
│   └── 20240712_111606_300_1720754373_to_1720754381_8_24/  # Scene data directory
│       ├── label_new/            # Annotation data
│       ├── road_camera/          # Roadside camera data
│       ├── road_lidar/           # Roadside LiDAR data
│       ├── veh_camera/           # Vehicle-mounted camera data
│       ├── veh_lidar/            # Vehicle-mounted LiDAR data
│       ├── visualization/        # Visualization results
│       ├── All_Path_Maps.txt     # Path mapping configuration
│       ├── gps.pkl               # GPS data (pickle format)
│       ├── imu.pkl               # IMU inertial measurement data (pickle format)
│       ├── map.pkl               # Map data (pickle format)
│       ├── odom.pkl              # Odometry data (pickle format)
│       ├── pose.pkl              # Pose data (pickle format)
│       ├── Timestamp.pkl         # Timestamp data (pickle format)
│       └── Timestamp.txt         # Timestamp data (text format)
│   └── ... (additional scenes)
└── tools/                        # Dataset preparation and visualization tools

Step 3: Create Symbolic Links (Optional)

To link data from a custom location, use the provided utility script:

python dataset/v2xscenes/tools/create_v2x_links.py --target /path/to/your/data

📦 Installation

Step 1: Environment Setup

# Create conda environment
conda env create -f environment.yml

# Activate environment
conda activate v2xscenes

# Install package in development mode
python setup.py develop

Step 2: Install Spconv

We use spconv (1.2.1 or 2.x) for voxel feature generation. Note: Checkpoints are stored in spconv 1.2.1 format and are not compatible with 2.x.

For spconv 2.x (recommended for easier installation): Check the official spconv table and run the appropriate command for your CUDA version. Example:

pip install spconv-cu116  # Replace cu116 with your CUDA version (cu111, cu118, etc.)

For spconv 1.2.1 (required if using provided checkpoints): Follow the spconv 1.2.1 installation guide. Alternatively, refer to the CoAlign Installation Documentation for detailed setup instructions.

Step 3: Compile CUDA Kernels

Compile the bounding box IoU CUDA kernels:

python opencood/utils/setup.py build_ext --inplace

🚀 Training and Evaluation

Training

Example training command using the Where2Comm configuration:

CUDA_VISIBLE_DEVICES=0 python opencood/tools/train_v2xscenes.py \
    -y ./opencood/hypes_yaml/v2xscenes/v2xsences_where2comm.yaml

Visualization and Debugging

To visualize training results and verify label-data alignment:

1. Open ./opencood/data_utils/datasets/intermediate_heter_fusion_dataset_v2xscenes.py
2. Change PLOT = False to PLOT = True

Example visualizations:

Fusion Result	No Fusion Result

Model Inference

CUDA_VISIBLE_DEVICES=0 python opencood/tools/train.py \
    --hypes_yaml ${CONFIG_FILE} \
    [--model_dir ${CHECKPOINT_FOLDER} \
    --half]

📚 Acknowledgements

This project builds upon the excellent work from:

• OpenCOOD - Collaborative perception framework
• HEAL - Heterogeneous agent learning

📖 Citation

If you use V2XScenes in your research, please cite our paper:

@inproceedings{wang2025v2xscenes,
  title={V2XScenes: A Multiple Challenging Traffic Conditions Dataset for Large-Range Vehicle-Infrastructure Collaborative Perception},
  author={Wang, Bowen and Wang, Yafei and Gong, Wei and Chen, Siheng and Liu, Genjia and Xiong, Minhao and Ng, Chin Long},
  booktitle={Proceedings of the IEEE/CVF International Conference on Computer Vision},
  pages={28385--28395},
  year={2025}
}

📄 License

Please refer to the LICENSE file for usage terms and conditions.

📞 Contact

For questions or issues regarding the dataset, please visit our Access Portal or open an issue on GitHub.