Question regarding YOLO NAS S and input resolution to match sub stream.

[bobloadmire]

I'm switching over to YOLO NAS S and my input sub streams are 640x360. Besider higher inference times on intel iGPU, any reason I shouldn't modify the notebook to out put to that res exactly so no resizing occurs? I know it's going to increase inference times, but I'm well under 200ms that the 5fps needs. i'm at 17ms right now with the standard res. Essentially I'm reducing CPU, increasing GPU, and gaining more accuracy?

[NickM-27]

Frigate is specifically designed to run on smaller regions and not the entire frame. Trying to force it to pass the entire frame in every time will only cause problems

[bobloadmire]

interesting, have you found it noticably increases detection performance doing that?

[NickM-27]

like I said it really depends on the objects you are trying to detect. It absolutely makes a difference when trying to detect small animals on an overview camera, or people that are on the other side of the street (in my case).

[bobloadmire]

makes sense, it's been having a hard time with my cat. unfortunately some of my cameras don't offer to do a 1080p stream, only 1440p. I'll have to mess around with it, thank you!

[NickM-27]

either way it can be downscaled, will just use more GPU

[bobloadmire]

wow thats not bad at all. I just switched 7 cams to detect off of their 1080p/1440p main stream and downsize the detection to 720p and even with the yolo-nas 320x320 model, i'm running at about 10% on the cpu, and 6% on the GPU, Intel Celeron 8505 on one of those cheap chinese firewall boxes. QSV is pretty awesome.