CoSDH (CVPR 2025)

CoSDH: Communication-Efficient Collaborative Perception via Supply-Demand Awareness and Intermediate-Late Hybridization

Paper

Installation

This project is based on the Coalign project. Please refer to the original project’s installation guide.

Please visit the feishu docs CoAlign Installation Guide for details!

Or you can refer to OpenCOOD data introduction and OpenCOOD installation guide to prepare data and install CoAlign. The installation is totally the same as OpenCOOD, except some dependent packages required by CoAlign.

Dataset

First, use mkdir to create a dataset folder in this project directory, then download or link the dataset using ln, organizing the dataset according to the format below:

dataset
├── my_dair_v2x 
│   └── v2x_c
├── OPV2V
│   ├── test
│   ├── train
│   └── validate
├── V2X-Sim-2.0
│   ├── v2.0
│   ├── maps
│   └── sweeps
└── v2xsim2_info
    ├── v2xsim_infos_test.pkl
    ├── v2xsim_infos_train.pkl
    └── v2xsim_infos_val.pkl

Note that

1. *.pkl file in v2xsim2_info can be found in Google Drive
2. use complemented annotation for DAIR-V2X in my_dair_v2x Google Drive, see Complemented Annotations for DAIR-V2X-C for more details.

Checkpoints

Download: Google Drive

Training and Inference

Train

torchrun --master_port 22334 --nproc_per_node=4 \
opencood/tools/train_ddp.py \
-y opencood/hypes_yaml/opv2v/lidar_only/pointpillar_cosdh.yaml

Inference

# intermediate–late hybrid fusion 
python opencood/tools/inference.py --model_dir opencood/logs/<ckpt-dir> --fusion_method intermediatelate
# intermediate fusion only
python opencood/tools/inference.py --model_dir opencood/logs/<ckpt-dir> --fusion_method intermediate

Citation

Please cite our work if you find it useful.

@InProceedings{Xu_2025_CVPR,
    author    = {Xu, Junhao and Zhang, Yanan and Cai, Zhi and Huang, Di},
    title     = {CoSDH: Communication-Efficient Collaborative Perception via Supply-Demand Awareness and Intermediate-Late Hybridization},
    booktitle = {Proceedings of the Computer Vision and Pattern Recognition Conference (CVPR)},
    month     = {June},
    year      = {2025},
    pages     = {6834-6843}
}

Acknowlege

This project based on the code of OpenCOOD and CoAlign, thanks to their great code framework.