MamV2XCalib

August 2, 2025 · View on GitHub

Overview

This repository provides a PyTorch implementation and checkpoint for MamV2XCalib (ICCV 2025). It leverages vehicle-to-infrastructure (V2X) collaboration methods to calibrate roadside cameras without targets.
For details, see the paper: MamV2XCalib: V2X-based Target-less Infrastructure Camera Calibration with State Space Model.

Installation

cd MamV2XCalib

conda create -n mamcalib python=3.10
conda activate mamcalib

pip install torch==2.1.1 torchvision==0.16.1 torchaudio==2.1.1 --index-url https://download.pytorch.org/whl/cu118
pip install -r requirements.txt   # Some packages may require source installation

cd VideoMamba
pip install -e causal-conv1d
pip install -e mamba

Checkpoint

You can download pretrained checkpoints from the following Google Drive folder:
👉 Google Drive - MamV2XCalib Checkpoints

Note: to be released

Quick Start

Dataset Download

You may refer to the following datasets for download and preparation:

  • DAIR-V2X: A large-scale dataset for vehicle-infrastructure cooperation perception tasks.
  • V2X-Seq: Sequence-level V2X dataset for spatiotemporal modeling.
  • TUMTraf Dataset: Real-world traffic scene dataset.

Evaluation

  • Modify the config and dataset path in evaluate_calib_mam.py according to your dataset.
  • Run the evaluation script:
python evaluate_calib_mam.py

Training

  • Modify the config, hyperparameters, and dataset paths in train_half.py according to your dataset.
  • Run the first stage of training:
python train_half.py
  • After obtaining the checkpoint, add it to train_mam.py.
  • Run the second stage of training:
python train_mam.py

Citation

If you find this work useful, please cite:

@inproceedings{MamV2XCalib,
title={MamV2XCalib: V2X-based Target-less Infrastructure Camera Calibration with State Space Model},
author={Yaoye Zhu, Zhe Wang, Yan Wang},
year={2025},
eprint={2507.23595},
archivePrefix={arXiv},
primaryClass={cs.CV},
url={https://arxiv.org/abs/2507.23595},
}

Acknowledgement

  • We appreciate help from: Public codes such as LCCNet, RAFT, VideoMamba, tum-traffic-dataset-dev-kit etc.
  • This work is supported by National Science and Technology Major Project (2022ZD0115502), and Wuxi Research Institute of Applied Technologies, Tsinghua University under Grant 20242001120.