Official implementation for the paper: Match-Any-Events: Zero-Shot Motion-Robust Feature Matching Across Wide Baselines for Event Cameras
-
Realtime Demo -
Evaluation Script - Pretrained Weights (
Event matching| Cross-modality matching) - Data generation pipeline
- Training Script
- SETS Module Script
Code was tested with the following environment setup.
- System: Ubuntu 20.04 / 22.04
- Python: 3.10
- CUDA: 12.9
Install requirements as in the requirements.txt file.
pip install -r requirements.txtComing Soon!
Pretrained MatchAnyEvents model weights can be downloaded here.
Create a folder named pretrained and move the pretrained weights into this folder so that loading is supported directly in the demo and evaluation scripts.
Match-Any-Events/
├── data/
│ ├── ECM/
│ └── E-MegaDepth/
└── pretrained/
└── weights.pth
RealtimeDemo_tiny.mp4
- System: Ubuntu 22.04
- Cameras: 2 x Prophesee Gen3 VGA (640x480)
- SDK: Metavision SDK 4.6 https://docs.prophesee.ai/stable/installation/index.html
Run the demo script using the serial numbers of your cameras. If you are using Prophesee cameras, they are 8 digit numbers that can be found in Metavision Studio.
python realtime_demo.py --s1 YOUR_SERIAL_NUMBER1 --s2 YOUR_SERIAL_NUMBER2Coming Soon!
@article{zhang2026match,
title={Match-Any-Events: Zero-Shot Motion-Robust Feature Matching Across Wide Baselines for Event Cameras},
author={Zhang, Ruijun and Su, Hang and Daniilidis, Kostas and Wang, Ziyun},
journal={arXiv preprint arXiv:2604.18744},
year={2026}
}This project is built upon the foundational code provided by ELoFTR: https://github.com/zju3dv/efficientloftr. We thank the authors for releasing their code.
We would like to thank the authors for their open-source implementations, which inspired parts of this project: