We have a custom fork of PhotonVisiom with ML running on the RKNN accelerator (only for an OrangePi).
We used it this year to detect notes and it has worked amazing. We’ve been using autonomous collection almost always and it worked without any problems.
To get started, you need an OrangePi, and to put the fork image / pv image on it.
After that, find or create a good training dataset. There are a lot we found / created.
From there, that’s pretty much it. You can play around with confidence threshold / normal vision settings, and access the data through the normal PV library using camera.getLatestResult().getBestTarget()
This is our implementation of the camera in code. (note this is with advantage kit style code, so it may be a bit different from what you’re used to.)