E3CM-Epipolar-Constrained-Cascaded-Correspondence-Matching

[Paper]

Abstract

Accurate and robust correspondence matching is of utmost importance for various 3D computer vision tasks. However, traditional explicit programming-based methods often struggle to handle challenging scenarios, and deep learning-based methods require large well-labeled datasets for network training. In this article, we introduce Epipolar-Constrained Cascade Correspondence (E3CM), a novel approach that addresses these limitations. Unlike traditional methods, E3CM leverages pre-trained convolutional neural networks to match correspondence, without requiring annotated data for any network training or fine-tuning. Our method utilizes epipolar constraints to guide the matching process and incorporates a cascade structure for progressive refinement of matches. We extensively evaluate the performance of E3CM through comprehensive experiments and demonstrate its superiority over existing methods.

News

[2024-5-30] The code is released!

[2023-8-31] The paper is accepted by Neurocomputing!

Installation

To set up the environment, run:

cd E3CM_deep_networks

conda create -n e3cm

conda activate e3cm

pip install -r requirements.txt

Test

We provide two images from the scene St. Peter's Basilica in the dataset Megadepth. You can test our model by running:

python3 test.py

Of course, you can try your own image pairs by changing the file address.

Experiments

We also provide the evaluation code on HPatches and Megadepth.

For HPatches, you can run:

python3 eval_hpatches.py --data_root [your dataset root]

For Megadepth, you can run:

python3 eval_megadepth.py --data_root [your dataset root]

BibTeX citation

If you use any ideas from the paper or code from this repo, please consider citing:

@article{zhou2023e3cm,
  title={E3CM: Epipolar-constrained cascade correspondence matching},
  author={Zhou, Chenbo and Su, Shuai and Chen, Qijun and Fan, Rui},
  journal={Neurocomputing},
  volume={559},
  pages={126788},
  year={2023},
  publisher={Elsevier}
}