Added Dark Model... process times slower?

Read the rest of my last post. BI initializes DS at startup telling it to use custom models and where they are.
 
So you can specify which custom model for each situation and ignore the rest.
 
sorka said:
So are you saying every camera AI instance uses all custom models unless you specify which one you want and then it only uses what you specify for each camera?
Click to expand...

That is correct. I am using custom models and if I tell each camera to only use one model, my detection times are the same as if I didn't have the custom model.

So if I put in custom models box for the a particular camera dark,objects:0 then it only uses the dark model and the times are the same as prior to having the custom model.

This setup then allows one to only use dark at night and default DS during the day. In my field of view the custom model seems to work best, so I have objects:0 in most of my cameras.
 
Last edited:
  • Like
Reactions: sebastiantombs
wittaj said:
That is correct. I am using custom models and if I tell each camera to only use one model, my detection times are the same as if I didn't have the custom model.

So if I put in custom models box for the a particular camera dark;objects:0 then it only uses the dark model and the times are the same as prior to having the custom model.

This setup then allows one to only use dark at night and default DS during the day. In my field of view the custom model seems to work best, so I have objects:0 in most of my cameras.
Click to expand...

That doesn't work for me. When I add that string to the "custom models" text box, I get "nothing found" for each analyzed image. Also, the normal CPU usage of 5 to 7% for the entire machine is now sitting at 30% between python continually running and Windows Image Acquisition which is account for about 15% with python the other 10%. This is the case even when no motion events are happening and no images are being analyzed. Without the custom models folder checked there is no WIA process at all.
 
Or the model is corrupt.
 
So you know, without a doubt that it downloaded a perfect image and when it was moved into your model directory that was also a perfect operation. Confirming it's good simply by downloading again seems to be a fairly painless, quick double check to me. Especially given that you're having a problem not many others have.
 
sebastiantombs said:
So you know, without a doubt that it downloaded a perfect image and when it was moved into your model directory that was also a perfect operation. Confirming it's good simply by downloading again seems to be a fairly painless, quick double check to me. Especially given that you're having a problem not many others have.
Click to expand...

I did it multiple times before ever mentioning the issue here. I think the other thing to confirm is whether I really downloaded the right one or not. The one I downloaded is this:

github.com

Release DeepStack Dark Scene Objects Detection - Custom Model · OlafenwaMoses/DeepStack_ExDark

A DeepStack Custom Model for object detection API to detect objects in the dark/night images. It detects the following objects Bicycle Boat Bottle Bus Chair Car Cat Cup Dog Motorbike People Table ...
github.com github.com
 
That's the one. Next up is a corrupt installation as wittaj mentioned. If you haven't already, run a uninstall from control panel and check the program files directory to make sure the DS directory structure is gone. If not delete it, then try a reinstall.
 
The next step is to see if the :0 is working for you.

When you have motion and get the alert clip, is it showing findings from both models?

For example, all of the items under the objects default Deepstack will be lowercase.

The items under dark will start with a capital letter, except it is person for the default DeepStack and People for the dark model.

As you can see from this screenshot, for this camera I have in the custom models box dark,objects:0 which means only use the dark model and not the default model and you will see I have no all lower case here indicating that the custom model is the only one running for this camera and the ms timings are the same as if I eliminate the dark model.

1634437941826.png

Or are you getting both Car and car like this example and where my ms times are way higher:

1634438335192.png

So does yours show person and People or car and Car or is it only showing one of them?
 
  • Like
Reactions: sebastiantombs
wittaj said:
Yes because it is running two models.

I found that dark works better even during the day. So in the camera AI tab under Custom models, put in Objects:0 and it will skip the default detection. Or put in dark:0 during the day so it only uses default.

Only use one model and the detection will come down. The :0 tells it to ignore that model.
Click to expand...

The Objects:0 isn't working for me. I set "Use custom model folder ON with C:\DeepStack\MyModels in the Global AI tab and individual cameras (6) to Trigger-> Artificial Intelligence->Custom models--Objects:0 to disable the default detection. I noticed that I am getting both "Person" and "People" detection indicator. When I ran it through Testing and Tuning Analysis with DeepStack it shows me that both Object and Dark models are running. Log shows time to analyze still relatively low [514ms] for a particular clip, with times running from 348ms to 6637ms. I am going to modify the individual cameras Custom model to dark,objects:0 to see if that makes a difference. Any thoughts? Thanks.
 
Last edited:
  • Like
Reactions: sorka
Whenever I select "Use custom model folder" in the main AI settings tab, it causes a never ending loop of a Python task ramping to the top of the CPU usage list. The Python task coincides with a temp file that comes and goes quickly in the Deepstack temp folder every 2 or 3 seconds. This happens before I make any changes to the individual camera Trigger tab AI settings. I would really like to run dark only, day/night.
 
So, here are my results from attempting to further troubleshoot after setting up the dark custom model. Disclaimer: This last test is with some construction activity (a Bobcat grader spreading dirt) on three cameras in the back yard. Continuous + Alerts is set, and the alerts are disabled. :

I set dark,Objects:0 in AI for each camera. Both models are still running per DeepStack Testing & Tuning results. The total analysis times with both models running has also increased considerably (10k-14kms). Also when I get the DeepStack Error 100 message, it is always on the Objects model. Task Manager shows the CPU staying near 100% with two Python instances sometimes taking 27-40% CPU at peak. Turning off custom models (unchecking in global AI tab and removing from individual cameras AI tab) brings the CPU back down to 35-36% with one instance of Python running 4-9%, peaking in the 40s% range.

Clearly, with the dark and Objects models running together, performance takes a big hit and it is not clear to me how to disable the default model. Object:0 (restarting computer, Bluestacks, and Blue Iris service) hasn't worked for me so far.
 
Last edited:
you need:

objects:0

not

object:0

check that
 
I just noticed something interesting while testing. As soon as I turn on the use custom model setting, I mentioned that I start seeing temp files in the deepstack temp folder. These temp files are actually camera images from only motion. Why would these images be there if I have my camera is set to only do deepstack when the camera is triggered?
 
You must log in or register to reply here.
Top Bottom