Interesting, where is the removal documented? Is it only the models that are removed or also the integration code?

everything about yolo is removed so no more of it on 0.14 once it is released. more details in this commit: 14235c4

though, yolox is still there and they are adding yolo-nas as replacement here 4313fd9

Thank you for this information! I'll try the nas version out.

yes the reduced FP is the most obvious improvement of this model compared to v8.
btw im using the S model with size 736x736 with detect size 1280x736 and my camera is 1080p.

I am using the s version with size 256x256 for now, because of my limited hardware capability. I get an inference speed of 48ms out of a Celeron J4125 iGPU. I am considering purchasing a Core Ultra 7 with an Arc iGPU NUC to get more speed.

@philippedegeer what is your detect size? if i have my input size to be a lot smaller (like 320) and my camera is 1080p then it simply wont detect far objects despite the claim by frigate that it will cut required frame and send it to the detector. keeping the input and detect size at 736p fixed it for me.

Currently my detect size is 800x450 but I have played a lot with it, even up to 1920x1080. This clearly has an influence on far objects detection. What I have now seems ok for my use case. I am easily detecting persons 12 meters away from a wide angle (horizontal 129°) Axis camera.

i guess my camera quality is too bad on low res that i need to use the 1080p stream and resize it to 736p and use the model with the same size. 110ms speed is definitely a very bad number but since i only have 2 cameras and one of them i set to 2 fps detection it is not a big problem for me.

Small offtop: didn't try yolov10 yet, but I have strange bug with yolov8 on newest openvino. When I export it from onnx with ovc, detector starts see all the time every label in every frame :) And I have to export to openvino with 2022.3 mo to work it properly. Is there parameters I should add on train or export?

Small offtop: didn't try yolov10 yet, but I have strange bug with yolov8 on newest openvino. When I export it from onnx with ovc, detector starts see all the time every label in every frame :) And I have to export to openvino with 2022.3 mo to work it properly. Is there parameters I should add on train or export?

you could try the export function within newest ultralytics package. See if that changes anything. If it works, you can then look at their exporter.py code to see what they did different with ovc. With yolov8 and latest ultralytics package I can export to openvino and use that with frigate. Are you using compress_to_fp16 option with ovc? Or using the standard float32 model

No, standard export doesn't work for me. With ovc I tried both true and false for compress_to_fp16

I forgot to mention I do it on frigate 0.14 with ad-hoc patch that solves this error for me [GPU] Mismatch tensor and port type: f32 vs u8 . Maybe that's the problem :)

@canadarob thanks for sharing the exported models --I was able to get yolov10n-relu6_full_integer_quant_edgetpu320.tflite running on my Coral USB (with your patch) - working nicely & way less false positives!

Thanks. Tested the 320 relu6 model and got a Nice 20 -2ms inference speed with a USB edgetpu.

I can confirm - seeing similar numbers (~25ms) with Coral USB.

Does the ad-hoc patch work for frigate 0.14? When I use the patch on 0.14, I get a weird error about path and labelmap_path. Works on the default model but not on yolov10. My detector part of my config looks like this.

detectors:
coral:
type: edgetpu
device: usb
model:
path: /config/model_cache/yolov10/yolov10n-relu6_full_integer_quant_edgetpu320.tflite
width: 320
height: 320

The error is asking for path and labelpath (error stays the same when I add labelpath) for some reason:

frigate | 2024-08-24 20:16:44.861515807 *** Your config file is not valid! ***
frigate | 2024-08-24 20:16:44.861518399 *** Please check the docs at ***
frigate | 2024-08-24 20:16:44.861520807 *** https://docs.frigate.video/configuration/index ***
frigate | 2024-08-24 20:16:44.861662102 *************************************************************
frigate | 2024-08-24 20:16:44.861666342 *************************************************************
frigate | 2024-08-24 20:16:44.861668713 *** Config Validation Errors ***
frigate | 2024-08-24 20:16:44.861670972 *************************************************************
frigate | 2024-08-24 20:16:44.861686139 path: Field required
frigate | 2024-08-24 20:16:44.861702842 labelmap_path: Field required
frigate | 2024-08-24 20:16:44.861720027 *************************************************************
frigate | 2024-08-24 20:16:44.861723083 *** End Config Validation Errors ***