This PR is for part 1 of #1354. It focuses on adding a model selection
interface for models that exist in `photonvision_config/models/`. Upon
completion we can ship more than 1 model and users could upload their
own through `ssh` without deleting the shipped model. This PR also adds
the abstractions need to support more DNN backends (say OpenCV, or RPI
AI Kit)
Up next is adding a CRUD interface for managing models through the UI.
Our current code matches cameras in this order (which I think is objectively wrong and stupid)
- by-id (/dev/v4l/by-id/product-string)
- by path (/dev/videoN)
- product string/name, but ascii only
- asks cscore to reconnect to cameras using `path`, which on linux is actually /dev/videoN. This isn't guaranteed to stick to a camera if you replug them weirdly at runtime.
This is silly and does not consider the actual physical usb port. I propose instead, in this order:
- By physical usb port path and base name
- by physical usb port path and USB VID/PID
- By base name only (with a toggle switch to disable this, and create a new VisionModule instead)
- Give cscore /dev/video/by-path on Linux systems, pinning Photon USBCameras to a particular usb port once created.
This changes lots of things so stay paranoid!
Tested on Orange Pi 5 and Cool Pi 4B. Merge with parts of the OpenCV DNN PR.
Adds support for YOLOv5s models for Rockchip CPUs with a NPU. Right now hard coded to a note model from alex_idk. Very much still incubating and largely untested.
Allows logging software and live data view to see results. Also removes the requirement for AScope to keep up with the packet serde schema and instead just use the Protobuf descriptor.
