Awesome

EI-Nexus: Towards Unmediated and Flexible Inter-Modality Local Feature Extraction and Matching for Event-Image Data

<div align="center"> <a href="https://github.com/ZhonghuaYi" target="_blank">Zhonghua Yi</a> · <a href="https://github.com/MasterHow" target="_blank">Hao Shi</a> · <a href="https://github.com/zju-jiangqi" target="_blank">Qi Jiang</a> · <a href="https://scholar.google.com/citations?user=pKFqWhgAAAAJ" target="_blank">Kailun Yang</a> · <a href="https://github.com/flysoaryun" target="_blank">Ze Wang</a> · Diyang Gu · Yufan Zhang · <a href="https://scholar.google.com/citations?user=B6xWNvgAAAAJ" target="_blank">Kaiwei Wang</a>

<a href="https://arxiv.org/abs/2410.21743" target="_blank">Arxiv</a>

</div> :hammer_and_wrench: :construction_worker: :rocket:

News and ToDo

Release benchmark data.
Release code and checkpoints.

Requirements

The inference code was tested on

Ubuntu 22.04
Python 3.10
PyTorch 2.2.0
1 NVIDIA GPU (RTX A800) with CUDA version 12.1.

Setup an environment

conda create --name ei_nexus --file requirements.txt
conda activate ei_nexus

Note: Please follow the version of the dependencies in requirements.txt to avoid potential conflicts.

Data preparation

To prepare the MVSEC and EC data, download them from the following links:

MVSEC_RPE

EC_RPE

After downloading the datasets, create data folder and move extracted datasets in it. The struture of the data folder should be as following:

data
├── EC
│   ├── boxes_6dof
│   ├── boxes_rotation
│   ├── boxes_rotation_val.txt
│   ├── boxes_translation
│   ├── boxes_translation_val.txt
│   ├── calibration
│   ├── hdr_boxes
│   ├── new_boxes_rotation_val.txt
│   ├── new_boxes_translation_val.txt
│   ├── new_shapes_6dof_val.txt
│   ├── new_shapes_rotation_val.txt
│   ├── new_shapes_translation_val.txt
│   ├── poster_6dof
│   ├── poster_rotation
│   ├── poster_translation
│   ├── shapes_6dof
│   ├── shapes_6dof_val.txt
│   ├── shapes_rotation
│   ├── shapes_rotation_val.txt
│   ├── shapes_translation
│   └── shapes_translation_val.txt
└── MVSEC
    ├── indoor_flying
    ├── indoor_flying_calib
    ├── outdoor_day
    └── outdoor_day_calib

Reproduce Results

Checkpoints of MVSEC and EC data are in release page.

To reproduce keypoint extraction results, run the test_events-image_same_time.py:

python test_events-image_same_time.py

To reproduce matching results, run the test_events-image_different_time.py:

python test_events-image_different_time.py

Train

The training contains two stages, event extractor training stage and matcher training stage. To train event extractor, refer and modify the configs/train_EIM_stage1.yaml config file and simply run the train_extractor.py file:

CUDA_VISIBLE_DEVICES=0 python train_extractor.py

To train the matcher, set the matcher as LightGlue in configs/train_EIM_stage2.yaml and run the train_matcher.py file:

CUDA_VISIBLE_DEVICES=0 python train_matcher.py

The repository uses Wandb to manage expriments, comment related code if you don't want to use it.

Citation

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

@article{yi2024ei,
  title={EI-Nexus: Towards Unmediated and Flexible Inter-Modality Local Feature Extraction and Matching for Event-Image Data},
  author={Yi, Zhonghua and Shi, Hao and Jiang, Qi and Yang, Kailun and Wang, Ze and Gu, Diyang and Zhang, Yufan and Wang, Kaiwei},
  journal={arXiv preprint arXiv:2410.21743},
  year={2024}
}

Acknowledgement

EI-Nexus is built on top of LightGlue, SiLK and DeepEvTracker codebases. We appreciate the authors for their greate work and follow the License of EI-Nexus.