How's the yolov9 compared to yolonas?

[lkaino]

I just changed to yolov9s from yolonas Frigate+ model. I will share my experiences later.

How have others liked it? Is it better or worse compared to the yolonas?

[NickM-27]

In my personal experience it is very similar except it is better an small animals like rabbit and squirrel. False positives are incredibly rare for me.

[lkaino]

Now that I have been running it for a while, I agree that the performance is very similar. Would have to run them side by side to notice the differences.

[lawul]

I switched to yolov9s to try it out when it became available (coming from mobiledet on coral, not yolonas). So far I can confirm that it's better at detecting small objects (birds etc). No noticeable difference on bigger objects like cats/dogs/humans.

I get a lot more false positives though. Feels a bit like the "early days" of the frigate+ mobiledet model.
I collected many samples and did another fine tuning which seems to have a bigger impact on false positive reduction compared to the mobiledet model. This is just my subjective perception, however. We've had some very windy days lately which tends to result in more false positives in general with motion everywhere.

[xconverge]

@lawul I have had the exact same experience, but I only have 1 fine tuning left, so I have a feeling I have some $5 tunings coming up 😆

[damianssj661]

Does anybody know how big difference is between yolo v9-s and v9-e?
I saw comparison table but I don't understand what is real life difference. I use frigate plus adjusted so i can't just download yolo v9-e.

I have RTX A4000 so GPU power is not a problem for bigger flops.

[NickM-27]

it is just a matter of accuracy, which may or may not manifest in noticeable ways in day to day usage