Awesome

MACP: Efficient Model Adaptation for Cooperative Perception

The official repository for the WACV 2024 paper MACP: Efficient Model Adaptation for Cooperative Perception. This work proposes a novel method to adapt a single-agent pretrained model to a V2V cooperative perception setting. It achieves state-of-the-art performance on both the V2V4Real and the OPV2V datasets.

Setup

Our project is based on MMDetection3D v1.1.0. Please refer to the official documentation to set up the environment.

Data Preparation

Download the V2V4Real and OPV2V datasets.

Once the data is downloaded, it's necessary organize the data in the following structure:

├── $REPO_ROOT
│   ├── data
│   │   ├── v2v4real
│   │   │   ├── train
│   │   │   │   ├── testoutput_CAV_data_2022-03-15-09-54-40_0 # data folder
│   │   │   ├── test
|   |   ├── openv2v
│   │   │   ├── train
│   │   │   │   ├── 2021_08_16_22_26_54 # data folder
│   │   │   ├── test
|   |   |   ├── validate
|   |   |   ├── test_culver_city

Then, run the script files scripts/create_v2v4real.sh and scripts/create_openv2v.sh to prepare the cached data.

Notes

• The core code of our project is in the projects/Coperception folder.
• The voxelization OP in the original implementation of BEVFusion is different from the implementation in MMCV. Please refer here to compile the OP on CUDA.

MACP Weights

If you are interested in including any other pretrained weights or details, please open an issue or contact us.

Training

We train our model on one NVIDIA RTX 4090 GPU with 24GB memory. The training command is as follows:

cd /path/to/repo
export PYTHONPATH=$PWD:$PYTHONPATH
python tools/train.py path/to/config

Evaluation

The evaluation command is as follows:

cd /path/to/repo
export PYTHONPATH=$PWD:$PYTHONPATH
python tools/test.py path/to/config path/to/checkpoint

Citation

If you find our work useful in your research, please consider citing:

@inproceedings{ma2024macp,
  title={MACP: Efficient Model Adaptation for Cooperative Perception},
  author={Ma, Yunsheng and Lu, Juanwu and Cui, Can and Zhao, Sicheng and Cao, Xu and Ye, Wenqian and Wang, Ziran},
  booktitle={Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision},
  pages={3373--3382},
  year={2024}
}

Acknowledgement

This project is based on code from several open-source projects. We would like to thank the authors for their great work:

• MMDetection3D
• BEVFusion
• V2V4Real
• OpenCOOD