CodeProject.AI Version 2.0

@Pentagano Here are the settings I'm using trying to detect and alert on the raccoons on my property. In CPAI, I'm using the Large model and YOLOv5 6.2. Should I put something in the To Cancel field so each trigger has all 16 images analyzed to decide the best result? Thanks
 

Attachments

  • Screen Shot 2023-07-17 at 5.41.32 PM.png
    Screen Shot 2023-07-17 at 5.41.32 PM.png
    517.6 KB · Views: 64
  • Screen Shot 2023-07-17 at 5.41.50 PM.png
    Screen Shot 2023-07-17 at 5.41.50 PM.png
    591.3 KB · Views: 66
  • Screen Shot 2023-07-17 at 5.43.14 PM.png
    Screen Shot 2023-07-17 at 5.43.14 PM.png
    492.2 KB · Views: 62
Village Guy said:
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.
Click to expand...
Excellent question, but I would challenge your 1060 card idles between 4-5 watts (that may be what it reports, but even under clocked\volted will be 10+), secondly that is idle most 1060 will be running 55w plus, but agree they won't hit their peaks.
On-Paper the Coral has extreme performance at 1.5watts.
In the real-world, paired with Frigate or ProjectAI agree the performance is far from extreme.
It's also far too new for Production Use, but here is the thing, it holds extreme potential to improve the efficiency of NVR solutions and is closer targeted at Image Recognition than a GPU.
My primary NVR has a 3060ti (just happened to be the most affordable GPU I could find at the time), and I was getting 30ms response times.
With Coral (default clock), and CodeProhect-Ai I get around 110ms response times. So much slower, but actually more than sufficient for an NVR use case.

If I was doing a new build, would I consider a Coral over a GPU? Honestly unsure, but the module that fits in the WiFi slot for $29 with 1.5w power draw is very tempting for a low power consumption build.

Another use case may be poor\low ventilation, removing the GPU from my rig, drastically reduced the heat output.
 
Last edited:
  • Like
Reactions: jrbeddow and teejnut
spammenotinoz said:
Excellent question, but I would challenge your 1060 card idles between 4-5 watts (that may be what it reports, but even under clocked\volted will be 10+), secondly that is idle most 1060 will be running 55w plus, but agree they won't hit their peaks.
On-Paper the Coral has extreme performance at 1.5watts.
In the real-world, paired with Frigate or ProjectAI agree the performance is far from extreme.
It's also far too new for Production Use, but here is the thing, it holds extreme potential to improve the efficiency of NVR solutions and is closer targeted at Image Recognition than a GPU.
My primary NVR has a 3060ti (just happened to be the most affordable GPU I could find at the time), and I was getting 30ms response times.
With Coral (default clock), and CodeProhect-Ai I get around 110ms response times. So much slower, but actually more than sufficient for an NVR use case.

If I was doing a new build, would I consider a Coral over a GPU? Honestly unsure, but the module that fits in the WiFi slot for $29 with 1.5w power draw is very tempting for a low power consumption build.

Another use case may be poor\low ventilation, removing the GPU from my rig, drastically reduced the heat output.
Click to expand...
My present understanding is that the Coral accelerator is unable to process medium and large models theirby reducing recognition accuracy. 110ms is an excellent response time for the Coral. Can you clarify if the 30ms and 110ms response times were using the same model size.

It is indeed a fact that software is limited in its capability to accurately determine power consumption. I am curious my self how much power is actually being consumed by the 1060 GPU and have decided to measure my dedicated system with and without the card. I will post the results tomorrow if I find the time.
 
  • Like
Reactions: actran, Pentagano and jrbeddow
Village Guy said:
My present understanding is that the Coral accelerator is unable to process medium and large models theirby reducing recognition accuracy. 110ms is an excellent response time for the Coral. Can you clarify if the 30ms and 110ms response times were using the same model size.

It is indeed a fact that software is limited in its capability to accurately determine power consumption. I am curious my self how much power is actually being consumed by the 1060 GPU and have decided to measure my dedicated system with and without the card. I will post the results tomorrow if I find the time.
Click to expand...


To clarify, it's 20-30ms on the cp console, and 100ms average on the blue iris status window.

Small and tiny are about the same for me, so I'm just using small.
 

Attachments

  • Screenshot_20230718_130200_Microsoft Remote Desktop.jpg
    Screenshot_20230718_130200_Microsoft Remote Desktop.jpg
    122.8 KB · Views: 26
As an Amazon Associate IPCamTalk earns from qualifying purchases.
  • Like
  • Love
Reactions: fenderman, actran and teejnut
MikeLud1 said:
Soon a module for the Orange Pi 5 will be released that has all my custom models. The Orange Pi 5 has a AI accelerator built-in similar to Coral.
Also an ALPR module for the Orange Pi 5 will be released


View attachment 167967
View attachment 167968
Click to expand...
Can you share any info on inference times with a medium size model with the OrangePi units?
 
As an Amazon Associate IPCamTalk earns from qualifying purchases.
  • Like
Reactions: actran
@MikeLud1 Was your animal model trained with any night images for the animals, or was it all day time images? I’m still fine-tuning things from the settings I showed above, just want to use the most accurate model and CPAI settings.

I have declared war on the raccoons, and I will be victorious! :p
 
  • Like
Reactions: Pentagano
105437 said:
@MikeLud1 Was your animal model trained with any night images for the animals, or was it all day time images? I’m still fine-tuning things from the settings I showed above, just want to use the most accurate model and CPAI settings.

I have declared war on the raccoons, and I will be victorious! :p
Click to expand...
A couple of years ago saved a load of cat jpegs at night and ran them through the deepstack training module process. Wasn't an easy process and took 12 hours with the cpu at max.
At the end I had my own cat py model.
Did work better with deepstack.
But now CPAI works so well I don't need it for the feline creatures that deficate on my grass (not now with with the automatic sprinklers now)
 
  • Like
Reactions: 105437
105437 said:
I'm curious how you trigger your sprinklers. Can you tell me about that?
Click to expand...
Nodered.
The sprinkler solenoid is connected to a simple sonoff device.
When nodered receives the mqtt payload for example ai/rear/motion with cat then this fires the sonoff device to open the pipe solenoid. Delay node for x minutes and close.
Result cat is scared shitless and runs off.
They are persistent though so need it on every night.
Knows when we leave the house so is on daytime when out.

1689719652024.png

1689719720579.png
 
  • Like
Reactions: 105437
105437 said:
@MikeLud1 Was your animal model trained with any night images for the animals, or was it all day time images? I’m still fine-tuning things from the settings I showed above, just want to use the most accurate model and CPAI settings.

I have declared war on the raccoons, and I will be victorious! :p
Click to expand...
Most of the raccoon images I used is from the below link. Also Dennis for CP made a critters model that you can try, link below.


How to Train a Custom YOLOv5 Model to Detect Objects

In this article, I will discuss what I experienced while creating a custom model for the detection of backyard pests. You can download the resulting critters.pt file.
www.codeproject.com www.codeproject.com
 
  • Like
Reactions: actran
You must log in or register to reply here.
Top Bottom