How to use YOLOv9 in Frigate 0.16?

[saschabrockel]

Hey there,

I was a simple person. I have my Intel GPU and have set up everything to the default:

detectors:
  ov: # OpenVINO Detector needed for Intel iGPU
    type: openvino
    device: AUTO
    model_path: /openvino-model/ssdlite_mobilenet_v2.xml
model: # OpenVINO Detector needed for Intel iGPU
  width: 300
  height: 300
  input_tensor: nhwc
  input_pixel_format: bgr
  labelmap_path: /openvino-model/coco_91cl_bkgr.txt

But now I got the mail from Frigate that YOLOv9 is supported. But I'm not sure how I can use it. Do I need the YOLO_MODELS environment variable on my Docker container to have YOLOv9?

What should I put in the config.yml?

Do I still need to create my own onnx file? I thought the mail was telling that I don't need to do this anymore, but the docs say something different: https://docs.frigate.video/configuration/object_detectors/#yolo-v3-v4-v7-v9

I'm a free Frigate+ subscriber. Thanks for all the help!

[hawkeye217]

If you do not subscribe to Frigate+ but are only on the free plan, the email does not apply to you. You must be a paid Frigate+ subscriber to use YOLOv9 Frigate+ models.

[NickM-27]

You can follow https://docs.frigate.video/configuration/object_detectors/#yolo-v3-v4-v7-v9 to generate a YOLOv9 model that will be usable

[saschabrockel]

Hmm I followed the tutorial but end up in this problem:

2025-09-08 17:06:53.978999598  [2025-09-08 17:06:53] frigate.detectors.plugins.openvino ERROR   : SSD model output doesn't match. Found [1,84,2100].
2025-09-08 17:06:53.989861808  Process detector:ov:
2025-09-08 17:06:53.989864023  Traceback (most recent call last):
2025-09-08 17:06:53.989864807    File "/usr/lib/python3.11/multiprocessing/process.py", line 314, in _bootstrap
2025-09-08 17:06:53.989865461      self.run()
2025-09-08 17:06:53.989866267    File "/opt/frigate/frigate/util/process.py", line 41, in run_wrapper
2025-09-08 17:06:53.989866861      return run(*args, **kwargs)
2025-09-08 17:06:53.989867465             ^^^^^^^^^^^^^^^^^^^^
2025-09-08 17:06:53.989868089    File "/usr/lib/python3.11/multiprocessing/process.py", line 108, in run
2025-09-08 17:06:53.989868731      self._target(*self._args, **self._kwargs)
2025-09-08 17:06:53.989869302    File "/opt/frigate/frigate/object_detection/base.py", line 136, in run_detector
2025-09-08 17:06:53.989869895      detections = object_detector.detect_raw(input_frame)
2025-09-08 17:06:53.989870504                   ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
2025-09-08 17:06:53.989871281    File "/opt/frigate/frigate/object_detection/base.py", line 86, in detect_raw
2025-09-08 17:06:53.989871994      return self.detect_api.detect_raw(tensor_input=tensor_input)
2025-09-08 17:06:53.989872735             ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
2025-09-08 17:06:53.989873531    File "/opt/frigate/frigate/detectors/plugins/openvino.py", line 150, in detect_raw
2025-09-08 17:06:53.989874133      infer_request.infer(input_tensor)
2025-09-08 17:06:53.989874935    File "/usr/local/lib/python3.11/dist-packages/openvino/runtime/ie_api.py", line 132, in infer
2025-09-08 17:06:53.989875580      return OVDict(super().infer(_data_dispatch(
2025-09-08 17:06:53.989876135                    ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
2025-09-08 17:06:53.989876776  RuntimeError: Exception from src/inference/src/cpp/infer_request.cpp:116:
2025-09-08 17:06:53.989877403  Exception from src/inference/src/cpp/infer_request.cpp:66:
2025-09-08 17:06:53.989878331  Check 'port.get_element_type() == tensor->get_element_type()' failed at src/plugins/intel_gpu/src/plugin/sync_infer_request.cpp:142:
2025-09-08 17:06:53.989878964  [GPU] Mismatch tensor and port type: f32 vs u8

According to the docs I set the config to this:

detectors:
  ov:
    type: openvino
    device: GPU

model:
  path: /config/model_cache/yolov9-s.onnx

I'm trying to use YOLOv9s.

[NickM-27]

You need to provide the entire model config, not just the path, when using the COCO model

[nilo85]

Am I blind? Or did the docs change? To me it does not look like instructions how to convert from .pt to OpenVino?

EDIT: I was blind! / drowned in docs, it seems section of interest is now found under:
https://docs.frigate.video/configuration/object_detectors/#downloading-yolo-models

[kiloptero]

Hi.. im trying to use the yolo9-t but send me this error in frigate logs

2025-12-25 23:39:58.595293766 [2025-12-25 20:39:58] frigate.detectors.plugins.openvino ERROR : SSD model output doesn't match. Found [1,84,8400].
2025-12-25 23:39:58.609861840 [INFO] Starting go2rtc healthcheck service...

any ideas?

detectors:
ov:
type: openvino
device: AUTO

model:
width: 300
height: 300
input_tensor: nhwc
input_pixel_format: bgr
path: /config/yolov9-t.xml

[NickM-27]

Your config is incomplete, copy the full docs example