[Detector Support]: Just got OpenVino capable N100. Any reason to keep using Coral? OpenVino vs Coral?

[rhatguy]

I recently purchased an N100 mini PC from Amazon. The N100 has Intel UHD 730 GPU. I've set frigate up running inside docker on Ubuntu 22.04 along with using OpenVino for detection. I'm running the newest Frigate .14 beta and the detector CPU graph shows usage usually less than 12% (I'm guessing this is averaged and there are probably spikes higher than that), and total CPU utilization hovers around 25%-30% (with 6 * 4k cameras with 720p detect streams). Initial impressions seem like OV is working great.

I also have a Coral that I previously used with a Raspberry PI 4 that I have not plugged into the N100 yet.

My question is, is there any advantage of using the Coral as opposed to the N100 GPU for detection? I understand the Coral is very efficient and would take workload off of the CPU, but this machine will be solely used for Frigate so as long as it works, I don't really care how loaded the CPU is. Is the Coral any better/worse at detection? Any other reasons to use Coral over OpenVino that I'm missing?

Version

0.14.0-bccffe6

Operating system

Debian

Install method

Docker Compose

Coral version

CPU (no coral)

Any other information that may be helpful

No response

[NickM-27]

you might have a tad lower power usage with the coral. Quality of detection comes down to the model that is run not the hardware itself

[stefajo]

I see you are a community developer for Frigate and the Home Assistant integration. I will report back after several days with CPU detection inference speeds comparison. I don't see too many AMD posts, so I'd like to contribute what data i can.

[HonkStonk]

FYI - my N200+OpenVino (GPU) now does 11-12ms inference speed always on 2 tapo cameras (one 1080 and one 1440) - integrates great with my HA on RPI5 as soon as I found the cards themselves had a lot of config options.

The Asus PN42 motherboard seemed impossible to configure for E-key coral instead of WiFi so I gave that up. But with my 11ms with no additional hardware I think I’m good.

The only thing that is a performance related issue now is that the setup is not quick enough to detect persons leaving the house - only entering…

[NickM-27]

The only thing that is a performance related issue now is that the setup is not quick enough to detect persons leaving the house - only entering…

are you sure that's the issue, or can you elaborate? if inference time is 11ms then unless you have more than 10 cameras it shouldn't be a matter of speed

[HonkStonk]

Well - I’m not sure where the ”bottleneck” in the system is but for sure you have to stand in front of the camera for 4 consecutive seconds before it logs and reports a person as detected or it will discard it altogether. I leave the house quicker than that but enter the house longer than that.

I should note the camera points down from porch ceiling and sees only heads from above. Very good accuracy when entering at least.

Does it have to do with the models (standard mobilenetssd) or with N200/OpenVino or with the mqtt connection to my Pi HA? I would expect the Frigate app in my standalone Ubuntu pc would record anything with motion and post process anything in due time and then report to HA it detected a person but from all my testing it does not report anything unless I stand in front of it for 4-5 seconds. Is there any filter to tune?

[NickM-27]

The default models trained on the COCO dataset have a difficult time with top-down camera angle due to no images like that being in the dataset. The easiest way to know what is going on is to see a recording of the debug live view during this scenario