IP Cam Talk Custom Community DeepStack Model

[wittaj]

OK so this would be my first time running a custom model. In the BI manual it says "By default, all Custom model files are considered, although you may select a subset here by specifying one or more (comma delimited; omit the file extensions). Custom models are enabled on the global Settings AI page. You may also add faces:0 and/or objects:0 here to specifically skip face and/or default object detection."

If I just wanted to run just the custom model do I need to do anything like whats bold above? Sorry for hijacking the thread.

Yes, put in objects:0 so that is skips the default object detection. These custom models in this thread are to replace the default.

 

[sebastiantombs]

That's what works, "objects:0,(custom model name),(custom model name)".

 

[sebastiantombs]

^I am guessing people coming into your yard would be shadow backlit figures in color.

On the "west" camera they get front lit by the street light that's glaring in the "east" camera. The "east" camera will catch front features depending on the angles involved, direction of travel. The distances involved are another challenge. The "east" camera is actually about 100-150 feet from that streetlight. The "west" camera is only about 75 feet from that streetlight but faces away from it. I have privacy masks set in both cameras. "East" because of a highly reflective street sign, and "west" because DS keeps IDing both my and my across the street neighbors mailbox as a person both day and night.

 

[105437]

That's what works, "objects:0,(custom model name),(custom model name)".

So like this?
objects:0,dark,combined

 

[MikeLud1]

So like this?
objects:0,dark,combined

yes

 

[105437]

yes

I used that syntax and ran a clip through "Test with Deepstack" It showed all three being used.

 

[wittaj]

Test with DeepStack is going to use every model available and show every item that DS can identify and is used to see if you can identify all the items - it is not tied to what you set up. It isn't the same as the live running of the DS and identification.

If you want to see how it performs for your settings, you need to tell it to save the analysis .dat file and then open it up in BI Status and then DeepStack Tab

 

[105437]

Test with DeepStack is going to use every model available and show every item that DS can identify and is used to see if you can identify all the items - it is not tied to what you set up. It isn't the same as the live running of the DS and identification.

Gotcha, so with two models running I should see a percentage for both models. Like Car 93%, car 93%?

 

[wittaj]

Gotcha, so with two models running I should see a percentage for both models. Like Car 93%, car 93%?

That would be correct. So if you notice they are about the same, you can run either one. If you notice a consistent large difference, then go with the higher % one. But no real reason to run it through two models for the same object.

 

[105437]

That would be correct. So if you notice they are about the same, you can run either one. If you notice a consistent large difference, then go with the higher % one. But no real reason to run it through two models for the same object.

Yes, right now I'm testing the combined model to see how it's working.

 

[105437]

I just sent the dogs out and the log shows that the dark model recognized the dog, but not the combined model. When using two models that look for the same object, does it only use one of them?

3 12/13/2021 10:31:28.020 PM Left_Rear_Cam MOTION_A
3 12/13/2021 10:31:30.473 PM Right_Rear_Cam MOTION_A
0 12/13/2021 10:31:32.196 PM Right_Rear_Cam DeepStack: Dog:73% [978,68 1202,210] 675ms
0 12/13/2021 10:31:33.542 PM Left_Rear_Cam DeepStack: Alert cancelled [nothing found] 261ms
0 12/13/2021 10:31:43.683 PM New Delete: 3 items 51.5M [210.0G/210.0G, 19.1G free]
0 12/13/2021 10:31:43.707 PM Alerts Delete: 2 items 3.32M [3.00G/3.00G, 19.2G free]
3 12/13/2021 10:31:44.803 PM Right_Rear_Cam MOTION_A
3 12/13/2021 10:31:44.951 PM Left_Rear_Cam MOTION_A
0 12/13/2021 10:31:50.005 PM Right_Rear_Cam DeepStack: Dog:77% [980,70 1200,207] 264ms
0 12/13/2021 10:31:50.677 PM Left_Rear_Cam DeepStack: Alert cancelled [nothing found] 368ms
3 12/13/2021 10:32:56.480 PM Left_Rear_Cam MOTION_A
0 12/13/2021 10:32:58.809 PM Left_Rear_Cam DeepStack: Dog:77% [1418,326 1591,522] 568ms

 

[wittaj]

It will use both models if they are on. The dark model is designed for night, while the combined can work for the night, but it is more for the day, which was why the dark model was created.

 

[fresnoboy]

Hi. I run deepstack on a separate VM with lots of CPU power (BI runs on a dedicated system with no virtualization and intel GPU acceleration). How do I start standalone deepstack with the proper settings to use your custom model?

 

[MikeLud1]

Hi. I run deepstack on a separate VM with lots of CPU power (BI runs on a dedicated system with no virtualization and intel GPU acceleration). How do I start standalone deepstack with the proper settings to use your custom model?

I am not sure but the below link is the official DeepStack instruction on how to use a custom model

Deploying Your Model With DeepStack | DeepStack 2021.09.1 documentation

 

[MikeLud1]

I am starting to work on combined V2.0. This model will have all but two of the animales (coyote & possum). So if anyone has images with these animales please let me know and I can add them to V2.0

 

[MikeLud1]

FYI I am changing boar to pig because I found image labeled pig which has a mix of pigs and boars

 

[sebastiantombs]

If I ever see either one I'm racking a slug in the 12GA.

 

[MikeLud1]

If I ever see either one I'm racking a slug in the 12GA.

We can have pig roast

 

[CCTVCam]

I'm very interested to hear your comments on the comparison of the two. To my unprofessional eye the 5841 looks crisper but that's not to say that the 5441 isn't crisp either.

If you have that much light did you try running the camera in color at night. Below is a clip from my 5442T-ZE
View attachment 111930

Looks like DS needs Takeaway recognition and alert.

 

[MikeLud1]

I started training combined V2.0, is should be done about noon tomorrow. There are two of the animales (coyote & possum) that will not be in this model.