Deepstack only working on some cameras...

E

erkme73

BIT Beta Team
Nov 9, 2014
1,570
1,486
Admittedly, I am having a hard time understanding the terminology and technology behind this new AI - so I'm sure my issues are operator error. That said...

I have multiple cameras on the street-facing side of my property. On two of the three utility pole mounted ptz cameras, it seems to be working reasonably well. I get images under the "flagged" group which have accurate identifying frames burned onto the image. However, on one of the cameras, I get nothing flagged. This, despite methodically comparing each of the trigger/AI/record tabs making sure they're all identical.

Short of posting screenshots to show that they're identical, I am left wondering what is different about the configuration. Cameras are Dahua, Amcrest, and Sunba. Frame rates and resolution are similar.


Works:
1621571939237.png
Works:
1621571976409.png

No Workie:
1621572097131.png



These are the settings:

1621572189796.png
1621572203810.png

1621572230944.png1621572617548.png

Any ideas?

Also, how does one get to the full resolution alert image that has the burned label?
 
Ok, after running a recorded clip on a non-working camera through the "testing and tuning" feature on the console, I see this:


1621573088951.png

And yet, under flagged alerts for that camera it shows nothing despite it being set to "auto-flag":

1621573350734.png
So the question is, why is it showing a confidence level higher than the threshold during testing, but not flagging it as the other cameras are doing?
 
The truck is close in color to the gravel road, so in real time it is probably struggling.

The tuning portion in BI I am not sure yet what it means, other than you can freeze frame playback and see if it can ID it. I am guessing at some point it will allow us to have it learn a certain vehicle for example?

You may also want to add more photos than 5 and lower the confidence.

And for kicks, turn AI off on the other cameras and have just this one and see if it picks it up then. In my limited testing, I think if too many cameras are calling DeepStack at the same time it is missing things.

If it picks it up with AI off on the other, then you know you need to tighten the parameters.
 
Thanks for the reply @wittaj. I feel like you're stalking me with how quickly you always seem to come to my rescue.

So, I'm pretty sure what's happening is unrelated to the number of cameras utilizing the DS service. I have a total of five cameras. Only three are doing any analysis.

I stayed up way too late last night watching a support video, and if I understand it correctly, if the analysis occurs, it should categorize the clips as confirmed or cancelled - whether "flagging" is turned on or not. On the two stubborn cameras, there is nothing in any of those three categories.

I-frame is set to 1.0 on all but the Sunba (which is a 0.5 due to a limit in the camera). It's almost like there is some missing setting, or incompatibility.

On an aside, thanks to prodding, I've got an i7-6700 PC coming so, after doing a rebuild, maybe this will be resolved.
 
LOL - just always seem to be in front of the computer!

Yeah, the flagging doesn't need to checked unless you want all of them saved for eternity LOL...and at some point you will run out of memory as they will not be auto-deleted.

With the addition of DeepStack, any camera with a Key of less than 1.0 will suffer due to the way it sends. The 0.5 means it could take twice as long to send the picture and then it could be missed.

In the log, how long is it saying it takes DS to analyze the picture?
 
1621647515073.png

Interesting but scanning the entire log, only a244, a249, and a248 have entries. The other two have zero DS entries.
 
Interesting indeed. I have seen elsewhere people are getting around 50-100ms returns, so at 1655ms, it could be struggling to keep up. If I turn it on too many I get in the thousands as well and seem to miss more.
 
I suppose I could disable AI on those that are working, leaving only the two that don't... I have to say though, as far as flagging/confirming/segregating the clips with action, it's been spot-on with the working cameras.

eta: disabled AI on the working cameras. Manual trigger on the suspect cameras did not generate a "canceled alert" clip - so I'm not holding out much hope. Acid test will be tomorrow when there's some traffic.
 
Last edited:
erkme73 said:
I suppose I could disable AI on those that are working, leaving only the two that don't... I have to say though, as far as flagging/confirming/segregating the clips with action, it's been spot-on with the working cameras.

eta: disabled AI on the working cameras. Manual trigger on the suspect cameras did not generate a "canceled alert" clip - so I'm not holding out much hope. Acid test will be tomorrow when there's some traffic.
Click to expand...
Ensure that in the alert tab for these cameras fire is not set to never. When it's set to never the image is not sent for processing..
 
  • Like
Reactions: VorlonFrog
fenderman said:
Ensure that in the alert tab for these cameras fire is not set to never. When it's set to never the image is not sent for processing..
Click to expand...

God do I feel like an idiot for not catching that. I have never turned off alerts on any of my cameras, and yet on those three, they were set either to never, or motion detect was unchecked.

I'm now getting canceled alerts on all cameras when I trigger the manually.

Thanks a bunch!
 
  • Like
Reactions: VorlonFrog, fenderman and sebastiantombs
You must log in or register to reply here.
Top Bottom