LightGlue-OnnxRunner

Introduction

LightGlue-OnnxRunner is a repository hosts the C++ inference code of LightGlue in ONNX format. LightGlue is a lightweight feature matcher with high accuracy and blazing fast inference. It takes as input a set of keypoints and descriptors for each image and returns the indices of corresponding points.

• Offical Paper : LightGlue: Local Feature Matching at Light Speed
• Official Repository ：cvg/LightGlue
• Open Neural Network Exchange (ONNX) Repository : fabio-sim/LightGlue-ONNX

superpoint_lightglue_end2end.onnx renderings

disk_lightglue_end2end.onnx renderings

Attention⚠️

Currently, the interface only supports CPU execution.The specific experimental data and equipment used are shown below. And the inferface is only supported on Windows and may encounter issues when running on Linux.

Updates📰

• [2023.09.08] : LightGlueOnnxRunner supporting end-to-end model inference of SuperPoint and DISK
• [2023.09.11] : LightGlueDecoupleOnnxRunner supporting decouple model inference of SuperPoint/DISK + LightGlue
• [2023.09.12] : Support calling GPU inference and complete README.md experimental data

Development Enviroments🖥️

• Windows 11 Professional
• CUDA v11.7
• cmake version 3.26.2

Quick Start

Installation

Install this repo in the following ways :

git clone https://github.com/OroChippw/LightGlue-OnnxRunner.git
cd LightGlue-OnnxRunner

Requirements⚒️

# onnxruntime-cpu 3rdparty
This repository use onnxruntime-win-x64-1.14.1
# onnxruntime-gpu 3rdparty
This repository use onnxruntime-win-x64-gpu-1.15.0 # for CUDA 11.7
# opencv 3rdparty
This repository use opencv4.8.0
# CXX_STANDARD 17

Build and Run

# Enter the source code directory where CMakeLists.txt is located, and create a new build folder
mkdir build
# Enter the build folder and run CMake to configure the project
cd build
cmake ..
# Use the build system to compile/link this project
cmake --build .
# If the specified compilation mode is debug or release, it is as follows
# cmake --build . --config Debug
# cmkae --build . --config Release

Model Checkpoints

Experiment Record

Environment Device : i5-13500H + NVIDIA GeForce RTX 4060 Laptop GPU（8GB）.
All models are available in repository fabio-sim/LightGlue-ONNX
The inference speed of onnxruntime-GPU will be slower when the first image is loaded, and will return to normal speed later.

Decouple

Extractor Type	Extractor Model Name	CPU speed(ms)	GPU speed(ms)	Matcher Model Name	CPU speed(ms)	GPU speed(ms)
SuperPoint	superpoint.onnx	Debug:123ms Release: 73ms	Debug:15ms Release:13ms	superpoint_lightglue.onnx	Debug:2384ms Release: 2112ms	Debug:155ms Release:230ms
DISK	disk.onnx	Debug:341ms Release: 336ms	Debug:28ms Release:25ms	disk_lightglue.onnx	Debug:3347ms Release: 3257ms	Debug: 230ms Release:245ms

End-to-End🌟🌟🌟

Extractor Type	Model Name	Model Size(MB/GB)	CPU speed(ms)	GPU speed(ms)
SuperPoint	superpoint_lightglue_end2end.onnx	50.1MB	Debug:2181ms Release: 1829ms	Debug: 170ms Release:166ms
DISK	disk_lightglue_end2end.onnx	48.9MB	Debug:3312ms Release: 3287ms	Debug: 285ms Release:285ms

CHANGELOG

20231120

• 适配v1.0.0: Fused LightGlue-ONNX接口改动

• 适配linux平台，完成CPU下测试，移除部分Linux下用不到的代码

• 修复可视化时，两张输入图片大小不一致时匹配特征计算错误导致的特征显示错误

• 非end-to-end模式增加sp和lg的独立接口

License

This project is licensed under the MIT License.