[Question]: Any plans to update object detection to YOLOv9?

[Lensters]

What is your question:

Any plans to update object detection to YOLOv9?

[harakas]

At this moment it's too new. Also I can't find a pre-trained small or nano sized model for it. From one article I can see that it can have much worse inference speeds compared to yolov8: https://learnopencv.com/yolov9-advancing-the-yolo-legacy/#aioseo-yolov9-vs-yolov8 -- haven't tried myself though.

Let's hold off for a while to observe if it truly delivers on its promises as it matures.

[harakas]

For reference, the authors of the model say they will release the smaller models once their article gets accepted and published: WongKinYiu/yolov9#3 -- so lets wait for that.

But I'm not quite sure what to think of such a thing. Then again yolov8 has not been published at all, not even a preprint (ultralytics/ultralytics#1564).

After reading more I believe the slow inference with yolov9 mentioned before is likely due to the "auxiliary training branch" in the model being kept in -- it could be removed to speed the inference up.

[Lensters]

It is a little suspicious that they aren't releasing their models. That isn't normal for academics.

[blakeblackshear]

I am not sure if this applies to the yolov9 repository you lined, but see here: #10717

@harakas You should also remove your public fork with yolo8 support as you are likely in violation as well.

[harakas]

I have now done so: https://github.com/harakas/frigate?tab=License-1-ov-file

[blakeblackshear]

The reason I haven't taken that approach is because it isn't clear to me the degree to which you must now honor their interpretation that all outputs of your fork are now AGPL licensed too. I'm not going to relicense Frigate without serious consideration.

[harakas]

I believe it's a good thing that Frigate is under MIT. I will not say more.

[Lensters]

Without a copy of v9 to look at and no mention of its lineage in their press releases, it is impossible to say if v9 forked from v8. It is very possible that v9 is from v7 because they're both from the same authors.

[harakas]

Without a copy of v9 to look at and no mention of its lineage in their press releases, it is impossible to say if v9 forked from v8. It is very possible that v9 is from v7 because they're both from the same authors.

Larger models and training code for yolov9 are available on github (https://github.com/WongKinYiu/yolov9/). As is a preprint of their paper on arxiv.org. On a cursory glance I believe they build on their previous work with yolov7 (AF). So lets not delve into groundless speculation and hyperboles.

After some googling the current consensus seems to be that neural network model structure (node/layer structure) is not copyrightable. I'm sure you'll find a copyright lawyer who argues otherwise -- it's their job after all to try to copyright everything on this earth -- speaking from experience interacting with lawyers ;) But that's just wishful thinking on their part. It might be patentable (within a context of application, as far as I can tell). Also training your model with GPL or AGPL software does not make the resulting model/weights (A)GPL. It's a different thing if you use a pre-existing trained model (as we do with current yolov8 distribution in frigate -- pretrained weights from ultralytics AGPL release).

So as far as I can see one just needs to train brand new weights for yolov8 model from nothing, export an onnx file and that can then be licensed and distributed and used by the person who did it however they wish.