Why isn't Deepstack correctly identifying cars?

Corvus85

Getting the hang of it
Joined
Aug 18, 2021
Messages
495
Reaction score
79
Location
Australia
I've noticed that one of my cameras isn't detecting cars during the day. The motion detection correctly detects motion, it triggers the camera, but Deepstack just ignores the snapshot. It doesn't even detect anything in the frame at all (even at a low confidence).

What's going on here and how can I fix it?

1652405916751.png
 

sebastiantombs

Known around here
Joined
Dec 28, 2019
Messages
11,511
Reaction score
27,690
Location
New Jersey
Start by changing the analysis time, making it shorter between frames. I use 500ms intervals for people and 200ms for vehicles. Cutting the pre-trigger buffer may help as well. I generally use 1 second for vehicles and 3 seconds for people. As with everything YMMV.
 

wittaj

IPCT Contributor
Joined
Apr 28, 2019
Messages
24,438
Reaction score
47,563
Location
USA
Is the confidence percentage too high? Lower it until you start getting false. Shorten duration times as well.

And also, while that is obvious to us it is a vehicle, AI gets iffy on edge of field of view analysis. Especially with the privacy blocks. And especially if the color of the vehicle matches the background as that appears it might be doing.
 

Corvus85

Getting the hang of it
Joined
Aug 18, 2021
Messages
495
Reaction score
79
Location
Australia
Start by changing the analysis time, making it shorter between frames. I use 500ms intervals for people and 200ms for vehicles. Cutting the pre-trigger buffer may help as well. I generally use 1 second for vehicles and 3 seconds for people. As with everything YMMV.
Considering the snapshot in this case seems to have perfectly captured the car unambiguously, why hasn't it properly analysed this image (even recognizing there's anything there to start with), and how would increasing the amount of snapshots (which it will also ignore) help?


Is the confidence percentage too high? Lower it until you start getting false. Shorten duration times as well.
So again, I can see why that would help if Deepstack is already correctly identifying that an object is there - but it's not even seeing anything there at all. In this case, why would lowering the confidence help? It seems to not be even detecting anything at all.
 

Corvus85

Getting the hang of it
Joined
Aug 18, 2021
Messages
495
Reaction score
79
Location
Australia
AI gets iffy on edge of field of view analysis.
Why is this? Why would it matter? Isn't it just analysing an image that BI sends to it? Why would it matter if it's on the edge of an image or in the centre?
 

wittaj

IPCT Contributor
Joined
Apr 28, 2019
Messages
24,438
Reaction score
47,563
Location
USA
Because it doesn't have surroundings to recognize motion and the objects tend to be smaller around the edges than if in the center of the frame. The fact that it isn't picking up the parked white truck as occupied is another clue...

DeepStack is a great tool, but it isn't magic and needs minimum requirements to be successful. Garbage In = Garbage out.

It is no different than camera AI having a tough time with identifying motion on the edges of the field of view.

Also confirm you do not have any weird zones that is not including the view.
 

sebastiantombs

Known around here
Joined
Dec 28, 2019
Messages
11,511
Reaction score
27,690
Location
New Jersey
What you see in the snapshot capture and what DeepStack may see real time are two different things. Contrast, as wittaj mentioned, is another consideration. When the object is at the edge of the image contrast gets more important. DeepStack is not human and your computer isn't a computer brain. There is a limited number of examples in the DeepStack "objects" file. If it doesn't see something close to the capture it can't detect.

Another hint is to not use "high resolution" for analysis captures. Blue Iris downsizes the images to 1080 anyway so high res isn't really much of a help IMHO, just a little more CPU load needed to downsize a 2K or 4K image.

I'd also suggest to try using the "combined" model rather than the default "objects" model. It has far fewer objects, no giraffe, elephant, zebra and so on, so it works much more quickly.


One more comment. This is real world, not Hollywood, and it takes some tuning and tweaking to get things working well.
 

wittaj

IPCT Contributor
Joined
Apr 28, 2019
Messages
24,438
Reaction score
47,563
Location
USA
Motion? I'm confused. I thought all Deepstack does is analyze still pictures, not motion. Can you clarify?
Sorry, that comment is more geared for camera AI, but the point stands that objects on the edges of field of views can be problematic whether it is internal camera AI or snapshots to DeepStack, especially when it is analyzed against a model set with vehicles more in the center of frames.

If you are using a less than ideal Field of View and/or cannot fine tune it to get the results you want, then you need to take your own photos and create your own model specific to your field of view.
 
Last edited:

MikeLud1

IPCT Contributor
Joined
Apr 5, 2017
Messages
2,141
Reaction score
4,118
Location
Brooklyn, NY
I've noticed that one of my cameras isn't detecting cars during the day. The motion detection correctly detects motion, it triggers the camera, but Deepstack just ignores the snapshot. It doesn't even detect anything in the frame at all (even at a low confidence).

What's going on here and how can I fix it?

View attachment 127927
What mode are you using? Try High.
 

Corvus85

Getting the hang of it
Joined
Aug 18, 2021
Messages
495
Reaction score
79
Location
Australia
I'd also suggest to try using the "combined" model rather than the default "objects" model. It has far fewer objects, no giraffe, elephant, zebra and so on, so it works much more quickly.
It's been a while since I've looked at that thread. The first post where the files are uploaded doesn't look like it's been updated since Dec 2021. Are those files still the latest ones? Does the combined model include the dark models now? Or do you still have to download the exdark models to use at night?
 

MikeLud1

IPCT Contributor
Joined
Apr 5, 2017
Messages
2,141
Reaction score
4,118
Location
Brooklyn, NY
It's been a while since I've looked at that thread. The first post where the files are uploaded doesn't look like it's been updated since Dec 2021. Are those files still the latest ones? Does the combined model include the dark models now? Or do you still have to download the exdark models to use at night?
I will be updating the custom models end of May early June. Currently the models do not have the dark images added, that is what I will work on.
 

Corvus85

Getting the hang of it
Joined
Aug 18, 2021
Messages
495
Reaction score
79
Location
Australia
I will be updating the custom models end of May early June. Currently the models do not have the dark images added, that is what I will work on.
I see. I'll probably just wait till whenever you get around to this to be honest. No point adding the custom models to my individual cameras and their individual day/night profiles when I'll have to do it again after everything's in the one model.
 

wittaj

IPCT Contributor
Joined
Apr 28, 2019
Messages
24,438
Reaction score
47,563
Location
USA
As I mentioned, the fact that it didn't pick up the white truck as occupied is another clue to either a bad field of view or you somehow have a Zone that is not delineated for that area.

I do not use the overview camera for identifying stuff on the edges, I use another camera optically zoomed in to the area.

The red car is parked and is blocked in blue meaning it is occupied (parked car). It appears yours is not picking up the parked vehicles either (or was the white truck also moving?)

1652408973727.png
 

sebastiantombs

Known around here
Joined
Dec 28, 2019
Messages
11,511
Reaction score
27,690
Location
New Jersey
You may want to reconsider using the current combined model. I, and many others, have found it identifies with a higher confidence level. For the amount of typing involved, especially if you use "copy and paste", it's not a whole lot of effort.
 

Corvus85

Getting the hang of it
Joined
Aug 18, 2021
Messages
495
Reaction score
79
Location
Australia
The red car is parked and is blocked in blue meaning it is occupied (parked car). It appears yours is not picking up the parked vehicles either (or was the white truck also moving?)
That's correct. It didn't pick up either of the cars. The white truck was parked and the black SUV was moving.

I, and many others, have found it identifies with a higher confidence level.
Even at night?
 

wittaj

IPCT Contributor
Joined
Apr 28, 2019
Messages
24,438
Reaction score
47,563
Location
USA
In that case confirm that in the settings you have it telling DeepStack to use the proper Zone and that the Zone is yellow up in that area.

If you are not using the whole field of view, then ideally, you should have Zone A be the whole field of view and then another Zone for just that area say Zone B and then tell DeepStack to analyze Zone B.

Provided all that is set up correctly, then it is a field of view issue and trying to do too much with one camera. Maybe tweaking some brightness and contrast may help, but the edge of images can be problematic.
 

Corvus85

Getting the hang of it
Joined
Aug 18, 2021
Messages
495
Reaction score
79
Location
Australia
In that case confirm that in the settings you have it telling DeepStack to use the proper Zone and that the Zone is yellow up in that area.
Yeah that seems to be correct. the entire section of the image from the road till the top edge of the image is covered under one zone, and Deepstack is set to detect on this zone, so I guess it's just the limitations of the model or DS itself.
 

Corvus85

Getting the hang of it
Joined
Aug 18, 2021
Messages
495
Reaction score
79
Location
Australia
Just got another of my cameras doing the same thing. The motion detection has clearly triggered (and motion is shown by the yellow trailing area) and this time, DS has identified a parked car in my driveway, however the old woman walking very slowly just behind it is ignored. She isn't even detected as a valid object.

1652411862188.png
 
Top