Running LPR using Nomeroff-net on Hailo8L

Dear community,
I’m fairly new to Hailo as I bought the RPi and the HAilo8L Kit just couple of weeks ago and I’ve been trying to dive into it as deeply as possible.

I’m building a LPR solution for quite some time now and I got it running on some Intel NUC on CPU inference but I’m kind of unhappy with the performance. When I found out about the RPi + Hailo combination, I felt this was the way to go.

The project I’m using on the NUC is this one: GitHub - ria-com/nomeroff-net: Nomeroff Net. Automatic numberplate recognition system. and this one delivers by far the best results compared to any other LPR I have tried (for european license plates). I would like to have this project running using the Edge accelerator and I’m really stuck. This is the workflow of the project:

  1. license plate detection using custom yolo8 model (.pt file)
  2. license plate classification (detecting the region of the license plate) (.ckpt file)
  3. license plate reading (.ckpt file)

I previously managed to successfully convert a .pt file to .onnx and to .hef, but with these files I’m just stuck. I already failed at the first .pt file for various reasons - I got all kinds of different error messages, then tried re-exporting from .pt to .onnx with different settings and so on and so forth. To be honest, I lost myself in all this and at one point I just closed everything because I don’t really know what’s the best way to approach this. I think the .pt file has some special characteristics which won’t allow a “general” conversion.

I would really appreciate some help on this - I’m also not looking for somebody to solve the problem for me, I rather want to understand the backgrounds of it as it’s a really interesting topic altogether.

Here some relevant links:
Nomeroff project: GitHub - ria-com/nomeroff-net: Nomeroff Net. Automatic numberplate recognition system.
Weight files: Index of /config_model/nomeroff-net/

Thanks in advance!