[Config Support]: converting from edgetpu to Intel

[deliverer33]

Describe the problem you are having

My coral card died so I need to switch. I'm using custom models from frigate + and wanted to validate this is the correct config to utilize YOLOv9 with my intel accelerated video card? Is there anything else I should need in my confiig? How do I decide between s and t sizes? I'm currently using small.

Relevant Config file below

Version

0.16.1-e664cb2

Frigate config file

detectors:



ffmpeg:
## changed from preset-vaapi due to latest version
  hwaccel_args: preset-intel-qsv-h264

  ov:
   type: openvino
   device: GPU

model:
   path: plus://xxxxxxxxxxxxxx (hashed but points to Frigate+ Model)

do I need to play anything else in the model section or does it get what it needs from the frigte plus model to support Yolov9?

Relevant Frigate log output

N/A

Relevant go2rtc log output

N/A

Frigate stats

No response

Operating system

Home Assistant OS

Install method

Home Assistant Add-on

docker-compose file or Docker CLI command

N/A

Object Detector

OpenVino

Screenshots of the Frigate UI's System metrics pages

No response

Any other information that may be helpful

No response

[inkeep-ai-assistant[bot]]

Your config for switching from Coral (edgetpu) to Intel with OpenVINO and Frigate+ YOLOv9 is on the right track. For Frigate+ models, when using YOLOv9 with OpenVINO, you should only specify the model path (the plus://... URL) in the model section—other model parameters (like width, height, etc.) should not be set, as the Frigate+ integration handles those automatically(1)(2).

For example:

detectors:
  ov:
    type: openvino
    device: GPU

model:
  path: plus://xxxxxxxxxxxxxx

(1), (2)

You do not need to add additional parameters for width, height, or model_type when using a Frigate+ YOLOv9 model—just the path is required(1)(2).

To choose between s and t sizes:

• "s" (small) and "t" (tiny) refer to model sizes. The "s" size is typically most similar in speed and accuracy to the small Frigate+ YOLO-NAS model, and is a good starting point for Intel GPUs(2).
• If you want the highest speed and your hardware is limited, try "t" (tiny). If you want better accuracy and can handle a bit more load, use "s" (small)(2).

You do not need to change anything else in your config for Frigate+ YOLOv9 support with OpenVINO on Intel when using the plus:// path(1)(2).

📚 Sources:

1. Supported Hardware
2. Models

---

Have another question? Just tag @inkeep.

[deliverer33]

I seem to be getting quite a few more false positives on this new setup. Deer are being detected as cows as an example. Is there something I need to do other than continue training the model?

[hawkeye217]

Did you update your object scores per the docs when you switched to a Frigate+ model? https://docs.frigate.video/plus/first_model/#step-4-adjust-your-object-filters-for-higher-scores

[deliverer33]

I did not, reviewing now. Why would you set a min value lower than the threshold? I don't understand that logic in how it works. Example, if I set deer to min score .65, with threshold of .85 that means unless the mean of the captures are =>.85 it will only capture it then? Why not just set the min score to .85?

[NickM-27]

min score is the minimum score to allow the detection to be considered valid for object tracking. The threshold is the score needed to consider the object a true positive.

If your min_score is too high then you will lose track of the object in difficult circumstances like when it is partially occluded by a shrub or is further away. Once the object passes the threshold, only the min_score matters.

[hawkeye217]

See the official documentation on how scores work: https://docs.frigate.video/configuration/object_filters#object-scores