G-LSMEI: GPU-Accelerated Least-Square Matching Refiner for Epipolar Image Pairs

A project of the sub-pixel Least-Square Matching Refining Algorithm for Epipolar-Rectified Stereo Image Pairs in window-size areas. The classic algorithm is proposed by Ackermann in DIGITAL IMAGE CORRELATION: PERFORMANCE AND POTENTIAL APPLICATION IN PHOTOGRAMMETRY (1984), where the correspondence search reduces to a one-dimensional problem along epipolar lines, enabling both fast convergence and high accuracy.

We provide the source code in C++ and a Python API library built upon it. A PyQt Demo Application and the CUDA-based GPU-accelerated version is also provided.
Up to now we have successfully tested the C++ and Python built from source code on Windows 10 and Windows 11.

Project Page
Contributors: Haojun Tang, Jiahao Zhou
Acknowledgements: Thanks to the guidance of Yunsheng Zhang from Central South University.

About the Project

Environment

Application Structure

CUDA Executions

Some Results

Matching Results

Left Image	Right Image

Time Costs

Setups	CPU	GPU
Time (sec)	332.27	0.621

Improved by 500+ times in our GPU implementation!

Core Algorithm tested in C++ (only CPU-based)

Prerequisites
MSVC (Visual Studio 2022, with C++ CMake tools), CMake ≥ 3.20, and the source code:

git clone https://github.com/DonaldTrump-coder/G-LSMEI --recursive

Build for the Code
In the project directory, run the following commands:

mkdir build
cd build
cmake .. -G "Visual Studio 17 2022"
cmake --build . --config Release
.\Release\leastsquares_matching.exe

The output is from the main function in core\src\test.cpp
Use the C++ source code for testing the algorithm.

Image Matching Application Deployment (CPU and GPU support)

Prerequisites
You also need MSVC and CMake, conda (or any Python environment), as well as nvcc for CUDA (Already successfully tested on CUDA 12.4). Install them and start to build!

1. Install the Requirement Packages for Python
In the project directory, run

conda create -n matching python=3.11
conda activate matching
pip install -r requirements.txt

2. Build the C++ and CUDA Source for Python Application

cd python
python setup.py bdist_wheel
python -m pip install (Get-ChildItem dist\glsmei-*.whl).FullName --force-reinstall
cd ..

Then the installation of glsmei is done.

3. Use the Application in Python
run

python main.py

API Reference

Installation from pip (If your Python and CUDA environment is supported): pip install glsmei
Verify installation: python -c "import lsmatching; print(lsmatching.__version__)"

Matching in CPU

Matching(left_image_path: str, right_image_path: str)

APIs	Params	Returns	Description
set_params(windowsize, d_corr_threshold)	int=15, float=0.04	—	Set template window size and correlation change threshold
set_matching_params(windowsize, corr_threshold)	int=3, float=0.7	—	Set initial matching parameters
set_centers(x1, y1, x2, y2)	int × 4	—	Set initial conjugate points on left/right images
calculate()	—	—	Run single-point least-squares matching
get_left_window()	—	np.ndarray	Get the left image matching window
get_right_window()	—	np.ndarray	Get the right image matching window
get_matched_points(savepath)	str	—	Export matching results
get_matched_x()	—	float	Matched point x-coordinate
get_matched_y()	—	float	Matched point y-coordinate
get_delta0()	—	float	Standard error of unit weight
get_deltag()	—	float	Standard error of window
get_deltax()	—	float	Standard error of parameters
get_SNR()	—	float	SNR
get_h0() / get_h1()	—	float	Radiometric distortion parameters
get_a0() / get_a1() / get_a2()	—	float	Affine parameters (left → right, x-direction)
get_b0() / get_b1() / get_b2()	—	float	Affine parameters (left → right, y-direction)
gpu_device_count()	—	int	Number of available GPUs

Matching in GPU

batch_adjust_gpu(window_size=15, d_corr=0.04, max_iter=20, matching_wsize=3, corr_threshold=0.7, savepath=None)

GPU batch pipeline: Feature Extraction → Correlation Matching → Least-Squares Refinement

Parameter	Type	Default	Description
window_size	int	15	Least-squares refinement template window size
d_corr	float	0.04	Correlation change threshold (convergence criterion)
max_iter	int	20	Maximum iteration count
matching_wsize	int	3	Initial correlation matching template window size
corr_threshold	float	0.7	Initial matching correlation threshold
savepath	str	None	Output save path

License

This project is licensed under the Apache License 2.0. See LICENSE details.

Citation

If you use our work or our data in your research, please cite:

@misc{Tang2026GLSMEI,
  title        = {G-LSMEI: GPU-Accelerated Least-Square Matching Refiner for Epipolar Image Pairs},
  author       = {Haojun Tang and Jiahao Zhou},
  year         = {2026},
  howpublished = {\url{https://github.com/DonaldTrump-coder/G-LSMEI}},
  note         = {Version 1.0.2. Apache License 2.0}
}