verlab
[Feature Inquiry] ONNX export support for lighterglue model
Hi! Thanks for your great work! We are exploring ways to deploy the model in cross-platform environments.
We've attempted several approaches to export the model, but encountered persistent issues.
Would there be an official script/tutorial for exporting the lighterglue model to ONNX format?
Hey Teddy,
I've added LighterGlue ONNX export in my fork here: https://github.com/stschake/accelerated_features/commit/acecc1561a60fc30904f0a79e0def825fc7224d1
Hey Teddy,
I've added LighterGlue ONNX export in my fork here: stschake@acecc15
Wow, thank you so much for sharing this! This is incredibly helpful for me!
This has same implement, easy export and with onnx demo.
您好，我尝试了使用了您的项目，非常好。但是我注意到和原版的lightgule相差较大，我在想精度上是否会有损失，请问您是否在Megadepth1500上测试过精度。
Which original version of LightGlue did you mean? As we know, this repo uses a much lighter version called LighterGlue, which has L=6 and only one attention head.
Which original version of LightGlue did you mean? As we know, this repo uses a much lighter version called LighterGlue, which has L=6 and only one attention head.
我执行onnx_runnner.py文件在自己的数据集上验证，发现精度很差。您用于验证的图片可能过于简单了，最好还是放到整个数据集上评估转为onnx后的模型精度。
pls pull the last commit code in my repo. a bug has fixed and one more thing is onnx is not support dynamic control flow, which means onnx not support width_confidence early stop in paper of lightglue.
pls pull the last commit code in my repo. a bug has fixed and one more thing is onnx is not support dynamic control flow, which means onnx not support width_confidence early stop in paper of lightglue.
非常感谢您的修复，我重新尝试了，已经成功了。但是我在尝试在Megadepth-1500上评估作者的XFeat+Lightglue模型，发现精度很差，有时候几何验证甚至会报错（因为符合条件的点数太少）。您是否在公共数据集上评估过您的模型。期待您的回复。
No, I didn't evaluate my onnx model file on any public dataset, it's just a simple model export without any training, and the exported model prediction result is as same as this repo..
是的，我已经尝试你的仓库了，它运行成功，但是精度非常差
Well, this is strange, because our company's product also uses the same ONNX export method. It works very well and is about to be put into production. And in principle, this kind of export should not have any impact on accuracy.