Highlights

• Early Collaborative Paradigm: Injects lightweight collaborative voxels into the ego's local modality-fusion step, yielding compact yet informative 3D collaborative priors.
• Heatmap-driven Instance Communication: Only Top-K instance vectors from low-confidence, high-discrepancy regions are queried from peers — drastically reducing redundancy.
• Instance Completion & Refinement: Cross-attention completion recovers occluded objects; self-attention refinement enhances instance features across agents.
• 87.98% Bandwidth Reduction: Instance-centric messaging achieves superior performance with minimal communication overhead.

Results

Multi-modal Collaborative Detection

Method	OPV2V AP50	OPV2V AP70	DAIR-V2X AP50	DAIR-V2X AP70	Bandwidth (log2)
No Coll	63.74	58.32	65.02	53.82	0.00
V2VNet	93.13	89.00	72.22	52.95	25.43
V2X-ViT	93.66	86.06	71.87	55.46	24.00
CoBEVT	93.03	84.64	71.70	55.85	24.00
BM2CP	93.04	88.94	72.37	56.18	23.18
EIMC	94.71	89.16	73.01	58.37	20.16

Installation

This project is built upon the HEAL framework.

Step 1: Environment Setup

conda create -n eimc python=3.8
conda activate eimc
conda install pytorch==1.12.0 torchvision==0.13.0 torchaudio==0.12.0 cudatoolkit=11.6 -c pytorch -c conda-forge
pip install -r requirements.txt
python setup.py develop

Step 2: Install Spconv

pip install spconv-cu116  # match your CUDA version

Step 3: Compile CUDA Extensions

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

Data Preparation

Download and organize datasets under dataset/:

• OPV2V: Download
• DAIR-V2X: Download (with complemented annotations)

EIMC/dataset
├── OPV2V
│   ├── train
│   ├── validate
│   └── test
└── my_dair_v2x
    ├── v2x_c
    ├── v2x_i
    └── v2x_v

Usage

Training

# OPV2V
python opencood/tools/train.py -y opencood/hypes_yaml/opv2v/MM/DSfusion_V2.yaml

# DAIR-V2X
python opencood/tools/train.py -y opencood/hypes_yaml/dairv2x/MM/DSfusion_V2.yaml

Evaluation

python opencood/tools/inference.py --model_dir ${CHECKPOINT_FOLDER} --fusion_method intermediate

Citation

If you find this work useful, please cite:

@inproceedings{eimc2026icra,
  title={EIMC: Efficient Instance-aware Multi-modal Collaborative Perception},
  author={Kang Yang, Peng Wang, Lantao Li, Tianci Bu, Chen Sun, Deying Li, Yongcai Wang*},
  booktitle={IEEE International Conference on Robotics and Automation (ICRA)},
  year={2026}
}

Acknowledgements

This project is built upon HEAL and OpenCOOD. We thank the authors for their excellent work.

License

This project is released under the Academic Software License.