I’m working on a visual localization pipeline for Raspberry Pi 5 + Hailo AI HAT.
My goal is to keep all application logic (tile search, homography estimation, navigation, etc.) on the CPU, while offloading only the neural network inference to the Hailo accelerator.
The models I would like to deploy are:
-
SuperPoint (feature detector + descriptor)
-
LightGlue (feature matcher)
From my research, I found ONNX versions of both models, and I’ve also seen community discussions about LightGlue compilation and attention blocks. However, I’m still unsure about the recommended deployment workflow on Hailo.
My questions are:
-
Has anyone successfully deployed SuperPoint on Hailo-8 or Hailo-8L?
-
Has anyone managed to compile LightGlue into a HEF (either as a whole model or by splitting it into subgraphs)?
-
What is the recommended workflow for custom models such as SuperPoint/LightGlue?
-
PyTorch → ONNX → HAR → HEF?
-
Or is there a better approach?
-
-
Are there any unsupported operators or known issues that should be addressed before attempting compilation?
-
Would you recommend using Hailo Apps as the application framework for this type of pipeline, with custom post-processing and the remaining localization logic running on the CPU?
Any example projects, repositories, or lessons learned would be greatly appreciated.
Thank you!