CodeProject.AI Version 2.0

[spammenotinoz]

Well, on small I seem to be having accuracy issues. It's very easy for me to test because I'm on a high traffic street. I get a lot of false "bus" and "cat." Not looking too good here ;(

Make sure you are using ObjectDetection (Coral) not TF-Lite. Both work, but Coral has been optimised.
Possibly related, but with GPU\CPU something broke in the recent update in regards to custom models. They just are not working reliably, using the default and accuracy is back.
Could just be my test rig, as my other setups are still running older versions.

 

[tofu]

Make sure you are using ObjectDetection (Coral) not TF-Lite. Both work, but Coral has been optimised.
Possibly related, but with GPU\CPU something broke in the recent update in regards to custom models. They just are not working reliably, using the default and accuracy is back.
Could just be my test rig, as my other setups are still running older versions.

Thanks, but Chris just confirmed the latency we are seeing is expected for medium models

 

[Village Guy]

I have been monitoring the activities associated with the Coral accelerator for the last couple of months but for all the hype, IMHO it's performance is significantly less than my Nvidia 1060 card. Presently, Coral performance seems to be limited to a tiny or small model size which in turn substantially reduces reliable identification.

Coral is also said to use very low power in the range of 2-4 Watts. My 1060 card idles between 4-5 Watts and runs large models and very rarely needs to turn on the cooling fan.

So, my question is, am I missing something? I can imagine many less informed people jumping on the band wagon placing orders for Coral processors and then being very disappointed with the outcome.

Presently from everything I have read, the Coral accelerator is not suitable for serious security applications. Unless the accelerator has hidden untapped potential that I am unaware of, its use will be best utilised in other less demanding applications.

 

[David L]

I have been monitoring the activities associated with the Coral accelerator for the last couple of months but for all the hype, IMHO it's performance is significantly less than my Nvidia 1060 card. Presently, Coral performance seems to be limited to a tiny or small model size which in turn substantially reduces reliable identification.

Coral is also said to use very low power in the range of 2-4 Watts. My 1060 card idles between 4-5 Watts and runs large models and very rarely needs to turn on the cooling fan.

So, my question is, am I missing something? I can imagine many less informed people jumping on the wagon placing orders for Coral processors and then being very disappointed with the outcome.

Presently from everything I have read, the Coral accelerator is not suitable for serious security applications. Unless the accelerator has hidden untapped potential that I am unaware of, it use will be best utilised in other less demanding applications.

I would assume the price difference is a factor. $60 compared to several hundred for a video card. I just happened to have a 970 4Gig card I kept from my old gaming days that worked out for me.

 

[tofu]

I would assume the price difference is a factor. $60 compared to several hundred for a video card. I just happened to have a 970 4Gig card I kept from my old gaming days that worked out for me.

A 1060 is about $70, not several hundred

 

[Village Guy]

I agree price has to be taken into account but so does value for money. What is the point of spending $50 - $75 on an accelerator that may never be suitable for security cameras!

 

[David L]

A 1060 is about $70, not several hundred

Ok a couple of hundred now...prices have come down some, true...

Amazon.com : gtx 1060

 

[David L]

I agree price has to be taken into account but so does value for money. What is the point of spending $50 - $75 on an accelerator that may never be suitable for security cameras!

Agree, I was looking at the Coral awhile back for a RPi project I was thinking of doing. It has it's purpose but I am with you, go with a Video Card.

 

[Village Guy]

Ok a couple of hundred now...prices have come down some, true...

Amazon.com : gtx 1060

Price aside, my question remains. Will the Coral accelerator ever be able to support large or even medium size models?

 

[105437]

I have a raccoon problem all of the sudden. Overnight this past week they are damaging my lawn and beds and I'd like to know when they're outside the house. Have any of you had any success using the ipcam-animal model to successfully trigger on a raccoon? If so, I would appreciate any settings you used. Which CPAI detection model, which size model, etc.

I'll be using a clone of a few cameras to send critical alerts and wake me up so I can introduce them to my Gamo Swarm. Thanks

 

[Pentagano]

I have a raccoon problem all of the sudden. Overnight this past week they are damaging my lawn and beds and I'd like to know when they're outside the house. Have any of you had any success using the ipcam-animal model to successfully trigger on a raccoon? If so, I would appreciate any settings you used. Which CPAI detection model, which size model, etc.

I'll be using a clone of a few cameras to send critical alerts and wake me up so I can introduce them to my Gamo Swarm. Thanks

I have issues with cats so use the ipcam combined model which works very well with the small creatures. Sprinklers are set off when a cat enters the grass area.
Using a large model and main stream for the triggers.

Regarding the coral I jumped on the bandwagon just before people here released their disappointing results. Bit unimpressed I spent money now.
Well I'll add it to my frigate container instead and make use of it there as a back up for BI

 

[105437]

I have issues with cats so use the ipcam combined model which works very well with the small creatures. Sprinklers are set off when a cat enters the grass area.
Using a large model and main stream for the triggers.

Thanks for the reply... On the 5 clones for raccoon alerts, are you suggesting that I only use ipcam-combined and then only alert if CPAI detects a raccoon? Is ipcam-combined more accurate at night (IR) than ipcam-animal for this use case in your experience?

 

[Pentagano]

Thanks for the reply... On the 5 clones for raccoon alerts, are you suggesting that I only use ipcam-combined and then only alert if CPAI detects a raccoon? Is ipcam-combined more accurate at night (IR) than ipcam-animal for this use case in your experience?

Honestly I believe they are the same but combined just has more objects
I tested both and gave the same results.
Needed person also so just use the combined now

 

[105437]

Honestly I believe they are the same but combined just has more objects
I tested both and gave the same results.
Needed person also so just use the combined now

You're confirming only on "cat"? What percentage have you found works for you. I just tested some clips and it seemed to confirm raccoons in my FOV between 40% - 50%. It also identified raccoons as fox and sometimes squirrels. I think I'll include those as well. Also, what motion sensor settings, do you use zone crossing or just trigger on motion for AI analysis.

 

[Pentagano]

You're confirming only on "cat"? What percentage have you found works for you. I just tested some clips and it seemed to confirm raccoons in my FOV between 40% - 50%. It also identified raccoons as fox and sometimes squirrels. I think I'll include those as well. Also, what motion sensor settings, do you use zone crossing or just trigger on motion for AI analysis.

cat and dog.
This is for the daytime trigger for example - some zones in there also due to the tree(shadow movement in summer). Night no zones (illuminated with some low level white light on full color mode)
I get very good results. Have 40% set for cats and dogs. Can get anywhere between 45 and 90% accuracy.
Have been tweaking it for a couple of years. Usually need to adjust it a bit between summer and winter.
sub stream for recording and main stream for triggers.

 

[105437]

If I'm not using zones, then the High Definition setting shouldn't matter, right? I do get a bunch of shadows during, but I'm only detecting for raccoons at night.

 

[Pentagano]

If I'm not using zones, then the High Definition setting shouldn't matter, right? I do get a bunch of shadows during, but I'm only detecting for raccoons at night.

I have the high def ticked and main stream ticked for analysis. Works much better in my case for that.
Also as a backup I have deepstack running with AITool just for the cats and dogs. BI sends out a jpeg every 1 sec to the AITool folder and deepstack on medium.
So basically both CPAI and DS. GPU works very well with both.

 

[Pentagano]

Well, on small I seem to be having accuracy issues. It's very easy for me to test because I'm on a high traffic street. I get a lot of false "bus" and "cat." Not looking too good here ;(

what size jpegs are you using for analysis? Frigate tends to use very small substream images as recommended at 5fps.
Have you tried with 1920x1080 and compared with substream res sizes?

 

[tofu]

what size jpegs are you using for analysis? Frigate tends to use very small substream images as recommended at 5fps.
Have you tried with 1920x1080 and compared with substream res sizes?

Hey, it's using my substream which is 640x360. I haven't increased the resolution to test if it tanks the performance

 

[Pentagano]

Sorry digressing a little from CPAI here. Last question about Frigate!

If I have the coral tpu installed will these detections in red squares currently using cpu, utilize the power of the coral tpu instead?

Update 16th July.
Omg frigate is a pita if you don't have a perfect view and scene.
Hours trying to get zones to work and tweaking different objects. Frustrating to say the least.
Blue iris is a godsend! Just hope the modules can be developed to get faster inference speeds with the coral.