AI works in debug but doesn't log it.

W

wpiman

Getting comfortable
Jul 16, 2018
416
303
massachusetts
View attachment Untitled.mov











So I am getting great shots of cars coming up the road. This is an Amazon truck with clear plates that it somehow missed.

Weirdly: I am playing back this video with AI debugging and in the right window it shows the plate, but on the left window I don't see it in the log on the left. It shows the Red window at least twice with 96 then 94% confidence.

My AI responds in about 134 ms, so I am trying 100ms captures and added more. Still no love.

Seperately, I am still missed some triggers. I went down to 15 FPS. Still issues. Maybe I will try 10 again.


Screenshot 2024-01-21 at 1.09.50 PM.png
 
wpiman said:
View attachment 183569











So I am getting great shots of cars coming up the road. This is an Amazon truck with clear plates that it somehow missed.

Weirdly: I am playing back this video with AI debugging and in the right window it shows the plate, but on the left window I don't see it in the log on the left. It shows the Red window at least twice with 96 then 94% confidence.

My AI responds in about 134 ms, so I am trying 100ms captures and added more. Still no love.

Seperately, I am still missed some triggers. I went down to 15 FPS. Still issues. Maybe I will try 10 again.


View attachment 183570
Click to expand...
Try setting your camera to 20 FPS and AI settings to + real-time images: to 20 and analyze one each: to 50ms
 
@wpiman If I view your video, the AI details on the left, for whatever reason, BI5 only sent 1 image to CP.AI, right?

And the image sent to CP.AI, did it have a license plate in the image? If no, then CP.AI won't have anything in image to pull plate #.

With respect to AI in debug mode, BI5 may be sending snapshots at different time points and if those snapshots have license plate in view, then CP.AI can process.

BTW, setting 110 images for each event seems excessive. You may be putting unexpected load on your CP.AI and thus your PC when enough events happens concurrently.
 
Last edited:
  • Like
Reactions: wpiman
You don't want to zoom out too far and then not be able to read plates.

It is a balance, and snow with salt-baked roads can be a problematic field of view based on color and dirtiness of car. While we can tell it is a car, the motion detection could miss it.

What are you motion settings page look like - probably need to increase motion sensitivity.
 
Gravity might have pulled it down a tad. Good catch. I’ll muck with that tomorrow.

Here are my motion settings. Zone A is basically the road minus those bushes in front. Zone b is left and zone c is ten pixels to the right or so. Tells me direction.


IMG_0463.png
IMG_0464.png
 
So the video I posed initially shows a cancelled alert. I NEVER get more than ONE of these even on a good capture that reports the plate..

ie. Mine looks nothing like yours...

Screenshot 2024-01-22 at 9.13.23 AM.png
 
I wonder if in some cases my trigger is too late...

Since I have my pre-trigger set up, I wonder if it would be possible to have Ken either start the AI process with frames that happened BEFORE the trigger by some configurable amount.. Or maybe search through them if it didn't get a hit. The data is there.
 
Looks like mine that get a "hit" go backwards too.

I guess I am getting a hit BEFORE the trigger. Mine are showing every 50 MS then going backwards.... to 200 ms, then looks like many are not printing.... Not quite sure where the ones in the 100ms range are?

I assume that is because it wasn't better than the one that happened at T = 0 until T - 858ms?
 
Mike..

Attached and also see post #9 for sub menus...

5 seconds after the trigger.... Maybe that is too long? Wouldn't that just mean more processing?

Thanks again guys....

Screenshot 2024-01-22 at 10.27.06 AM.png
 
You must log in or register to reply here.
Top Bottom