Use the DeepStack interface to use yolo-models

[hoeflechner]

not a request I already have a solution, see below

Describe what you are trying to accomplish and why in non technical terms
I want to be able to use yolo-models running on my gpu so that I can perform fast and good detection

Describe the solution you'd like
a DeepStack-like service that is decoupled from frigate

Describe alternatives you've considered
converting yolo models to tensorrt oder onnx

Additional context
as mentioned here:

Originally posted by @alexyao2015 in #10717 (comment)

i hacked together a minimal deepstack-like service that uses yolo world. it can read frigates config and can detect objects mentioned there.

https://github.com/hoeflechner/yolostack

[NickM-27]

Frigate has built in support for deepstack / codeproject ai https://docs.frigate.video/configuration/object_detectors#deepstack--codeprojectai-server-detector

[hoeflechner]

thanks for your reply.

i know that deepstack support is build in. i used this deepstack support to call a service that implemets yoloworld (https://github.com/hoeflechner/yolostack)

so this is not a feature request but an anouncement for those who want to use yolo with frigate but i was not able to respond to #10717 (comment)

[jsapede]

Is openVINO hardware accélération possible with yolostack ?

[hoeflechner]

It is currently not supported. But I can look into it.

[jsapede]

And would it be available also for facial récognition ?

[hoeflechner]

It currently only uses yolo world. I have not looked into facial recognition yet (and no idea how to integrate that with frigate...

[jsapede]

Double-take makes this using frigate api deepstack api or compreface api and mqtt interaction but all thèse application only use NVIDIA hardware as accélération.

It works, but the inférence speed we gain with openVINO hardware for détection is destroyed by the CPU Time consummed by deepstack or compreface only running on CPU

Immich has machine learnig options with facial récognition accélérated by openVINO also and its quite efficient and available to api also maybe